xref: /freebsd/sys/netinet/ip_dummynet.h (revision 3ff369fed2a08f32dda232c10470b949bef9489f)
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  * A heap entry is made of a key and a pointer to the actual
81  * object stored in the heap.
82  * The heap is an array of dn_heap_entry entries, dynamically allocated.
83  * Current size is "size", with "elements" actually in use.
84  * The heap normally supports only ordered insert and extract from the top.
85  * If we want to extract an object from the middle of the heap, we
86  * have to know where the object itself is located in the heap (or we
87  * need to scan the whole array). To this purpose, an object has a
88  * field (int) which contains the index of the object itself into the
89  * heap. When the object is moved, the field must also be updated.
90  * The offset of the index in the object is stored in the 'offset'
91  * field in the heap descriptor. The assumption is that this offset
92  * is non-zero if we want to support extract from the middle.
93  */
94 struct dn_heap_entry {
95     dn_key key ;	/* sorting key. Topmost element is smallest one */
96     void *object ;	/* object pointer */
97 } ;
98 
99 struct dn_heap {
100     int size ;
101     int elements ;
102     int offset ; /* XXX if > 0 this is the offset of direct ptr to obj */
103     struct dn_heap_entry *p ;	/* really an array of "size" entries */
104 } ;
105 
106 /*
107  * MT_DUMMYNET is a new (fake) mbuf type that is prepended to the
108  * packet when it comes out of a pipe. The definition
109  * ought to go in /sys/sys/mbuf.h but here it is less intrusive.
110  */
111 
112 #define MT_DUMMYNET MT_CONTROL
113 
114 /*
115  * struct dn_pkt identifies a packet in the dummynet queue. The
116  * first part is really an m_hdr for implementation purposes, and some
117  * fields are saved there. When passing the packet back to the ip_input/
118  * ip_output()/bdg_forward, the struct is prepended to the mbuf chain with type
119  * MT_DUMMYNET, and contains the pointer to the matching rule.
120  *
121  * Note: there is no real need to make this structure contain an m_hdr,
122  * in the future this should be changed to a normal data structure.
123  */
124 struct dn_pkt {
125     struct m_hdr hdr ;
126 #define dn_next	hdr.mh_nextpkt	/* next element in queue */
127 #define DN_NEXT(x)	(struct dn_pkt *)(x)->dn_next
128 #define dn_m	hdr.mh_next	/* packet to be forwarded */
129 #define dn_dir	hdr.mh_flags	/* action when pkt extracted from a queue */
130 #define DN_TO_IP_OUT	1
131 #define DN_TO_IP_IN	2
132 #define DN_TO_BDG_FWD	3
133 #define DN_TO_ETH_DEMUX	4
134 #define DN_TO_ETH_OUT	5
135 
136     dn_key  output_time;	/* when the pkt is due for delivery	*/
137     struct ifnet *ifp;		/* interface, for ip_output		*/
138     struct sockaddr_in *dn_dst ;
139     struct route ro;		/* route, for ip_output. MUST COPY	*/
140     int flags ;			/* flags, for ip_output (IPv6 ?)	*/
141 };
142 
143 /*
144  * Overall structure of dummynet (with WF2Q+):
145 
146 In dummynet, packets are selected with the firewall rules, and passed
147 to two different objects: PIPE or QUEUE.
148 
149 A QUEUE is just a queue with configurable size and queue management
150 policy. It is also associated with a mask (to discriminate among
151 different flows), a weight (used to give different shares of the
152 bandwidth to different flows) and a "pipe", which essentially
153 supplies the transmit clock for all queues associated with that
154 pipe.
155 
156 A PIPE emulates a fixed-bandwidth link, whose bandwidth is
157 configurable.  The "clock" for a pipe can come from either an
158 internal timer, or from the transmit interrupt of an interface.
159 A pipe is also associated with one (or more, if masks are used)
160 queue, where all packets for that pipe are stored.
161 
162 The bandwidth available on the pipe is shared by the queues
163 associated with that pipe (only one in case the packet is sent
164 to a PIPE) according to the WF2Q+ scheduling algorithm and the
165 configured weights.
166 
167 In general, incoming packets are stored in the appropriate queue,
168 which is then placed into one of a few heaps managed by a scheduler
169 to decide when the packet should be extracted.
170 The scheduler (a function called dummynet()) is run at every timer
171 tick, and grabs queues from the head of the heaps when they are
172 ready for processing.
173 
174 There are three data structures definining a pipe and associated queues:
175 
176  + dn_pipe, which contains the main configuration parameters related
177    to delay and bandwidth;
178  + dn_flow_set, which contains WF2Q+ configuration, flow
179    masks, plr and RED configuration;
180  + dn_flow_queue, which is the per-flow queue (containing the packets)
181 
182 Multiple dn_flow_set can be linked to the same pipe, and multiple
183 dn_flow_queue can be linked to the same dn_flow_set.
184 All data structures are linked in a linear list which is used for
185 housekeeping purposes.
186 
187 During configuration, we create and initialize the dn_flow_set
188 and dn_pipe structures (a dn_pipe also contains a dn_flow_set).
189 
190 At runtime: packets are sent to the appropriate dn_flow_set (either
191 WFQ ones, or the one embedded in the dn_pipe for fixed-rate flows),
192 which in turn dispatches them to the appropriate dn_flow_queue
193 (created dynamically according to the masks).
194 
195 The transmit clock for fixed rate flows (ready_event()) selects the
196 dn_flow_queue to be used to transmit the next packet. For WF2Q,
197 wfq_ready_event() extract a pipe which in turn selects the right
198 flow using a number of heaps defined into the pipe itself.
199 
200  *
201  */
202 
203 /*
204  * per flow queue. This contains the flow identifier, the queue
205  * of packets, counters, and parameters used to support both RED and
206  * WF2Q+.
207  *
208  * A dn_flow_queue is created and initialized whenever a packet for
209  * a new flow arrives.
210  */
211 struct dn_flow_queue {
212     struct dn_flow_queue *next ;
213     struct ipfw_flow_id id ;
214 
215     struct dn_pkt *head, *tail ;	/* queue of packets */
216     u_int len ;
217     u_int len_bytes ;
218     long numbytes ;		/* credit for transmission (dynamic queues) */
219 
220     u_int64_t tot_pkts ;	/* statistics counters	*/
221     u_int64_t tot_bytes ;
222     u_int32_t drops ;
223 
224     int hash_slot ;		/* debugging/diagnostic */
225 
226     /* RED parameters */
227     int avg ;                   /* average queue length est. (scaled) */
228     int count ;                 /* arrivals since last RED drop */
229     int random ;                /* random value (scaled) */
230     u_int32_t q_time ;          /* start of queue idle time */
231 
232     /* WF2Q+ support */
233     struct dn_flow_set *fs ;	/* parent flow set */
234     int heap_pos ;		/* position (index) of struct in heap */
235     dn_key sched_time ;		/* current time when queue enters ready_heap */
236 
237     dn_key S,F ;		/* start time, finish time */
238     /*
239      * Setting F < S means the timestamp is invalid. We only need
240      * to test this when the queue is empty.
241      */
242 } ;
243 
244 /*
245  * flow_set descriptor. Contains the "template" parameters for the
246  * queue configuration, and pointers to the hash table of dn_flow_queue's.
247  *
248  * The hash table is an array of lists -- we identify the slot by
249  * hashing the flow-id, then scan the list looking for a match.
250  * The size of the hash table (buckets) is configurable on a per-queue
251  * basis.
252  *
253  * A dn_flow_set is created whenever a new queue or pipe is created (in the
254  * latter case, the structure is located inside the struct dn_pipe).
255  */
256 struct dn_flow_set {
257     struct dn_flow_set *next; /* next flow set in all_flow_sets list */
258 
259     u_short fs_nr ;             /* flow_set number       */
260     u_short flags_fs;
261 #define DN_HAVE_FLOW_MASK	0x0001
262 #define DN_IS_RED		0x0002
263 #define DN_IS_GENTLE_RED	0x0004
264 #define DN_QSIZE_IS_BYTES	0x0008	/* queue size is measured in bytes */
265 #define DN_IS_PIPE		0x4000
266 #define DN_IS_QUEUE		0x8000
267 
268     struct dn_pipe *pipe ;	/* pointer to parent pipe */
269     u_short parent_nr ;		/* parent pipe#, 0 if local to a pipe */
270 
271     int weight ;		/* WFQ queue weight */
272     int qsize ;			/* queue size in slots or bytes */
273     int plr ;			/* pkt loss rate (2^31-1 means 100%) */
274 
275     struct ipfw_flow_id flow_mask ;
276 
277     /* hash table of queues onto this flow_set */
278     int rq_size ;		/* number of slots */
279     int rq_elements ;		/* active elements */
280     struct dn_flow_queue **rq;	/* array of rq_size entries */
281 
282     u_int32_t last_expired ;	/* do not expire too frequently */
283     int backlogged ;		/* #active queues for this flowset */
284 
285         /* RED parameters */
286 #define SCALE_RED               16
287 #define SCALE(x)                ( (x) << SCALE_RED )
288 #define SCALE_VAL(x)            ( (x) >> SCALE_RED )
289 #define SCALE_MUL(x,y)          ( ( (x) * (y) ) >> SCALE_RED )
290     int w_q ;			/* queue weight (scaled) */
291     int max_th ;		/* maximum threshold for queue (scaled) */
292     int min_th ;		/* minimum threshold for queue (scaled) */
293     int max_p ;			/* maximum value for p_b (scaled) */
294     u_int c_1 ;			/* max_p/(max_th-min_th) (scaled) */
295     u_int c_2 ;			/* max_p*min_th/(max_th-min_th) (scaled) */
296     u_int c_3 ;			/* for GRED, (1-max_p)/max_th (scaled) */
297     u_int c_4 ;			/* for GRED, 1 - 2*max_p (scaled) */
298     u_int * w_q_lookup ;	/* lookup table for computing (1-w_q)^t */
299     u_int lookup_depth ;	/* depth of lookup table */
300     int lookup_step ;		/* granularity inside the lookup table */
301     int lookup_weight ;		/* equal to (1-w_q)^t / (1-w_q)^(t+1) */
302     int avg_pkt_size ;		/* medium packet size */
303     int max_pkt_size ;		/* max packet size */
304 } ;
305 
306 /*
307  * Pipe descriptor. Contains global parameters, delay-line queue,
308  * and the flow_set used for fixed-rate queues.
309  *
310  * For WF2Q+ support it also has 3 heaps holding dn_flow_queue:
311  *   not_eligible_heap, for queues whose start time is higher
312  *	than the virtual time. Sorted by start time.
313  *   scheduler_heap, for queues eligible for scheduling. Sorted by
314  *	finish time.
315  *   idle_heap, all flows that are idle and can be removed. We
316  *	do that on each tick so we do not slow down too much
317  *	operations during forwarding.
318  *
319  */
320 struct dn_pipe {		/* a pipe */
321     struct dn_pipe *next ;
322 
323     int	pipe_nr ;		/* number	*/
324     int bandwidth;		/* really, bytes/tick.	*/
325     int	delay ;			/* really, ticks	*/
326 
327     struct	dn_pkt *head, *tail ;	/* packets in delay line */
328 
329     /* WF2Q+ */
330     struct dn_heap scheduler_heap ; /* top extract - key Finish time*/
331     struct dn_heap not_eligible_heap; /* top extract- key Start time */
332     struct dn_heap idle_heap ; /* random extract - key Start=Finish time */
333 
334     dn_key V ;			/* virtual time */
335     int sum;			/* sum of weights of all active sessions */
336     int numbytes;		/* bits I can transmit (more or less). */
337 
338     dn_key sched_time ;		/* time pipe was scheduled in ready_heap */
339 
340     /*
341      * When the tx clock come from an interface (if_name[0] != '\0'), its name
342      * is stored below, whereas the ifp is filled when the rule is configured.
343      */
344     char if_name[16];
345     struct ifnet *ifp ;
346     int ready ; /* set if ifp != NULL and we got a signal from it */
347 
348     struct dn_flow_set fs ; /* used with fixed-rate flows */
349 };
350 
351 #ifdef _KERNEL
352 typedef	int ip_dn_ctl_t(struct sockopt *); /* raw_ip.c */
353 typedef	void ip_dn_ruledel_t(void *); /* ip_fw.c */
354 typedef	int ip_dn_io_t(int pipe, int dir, struct mbuf *m,
355 	struct ifnet *ifp, struct route *ro, struct sockaddr_in * dst,
356 	struct ip_fw *rule, int flags); /* ip_{in,out}put.c, bridge.c */
357 extern	ip_dn_ctl_t *ip_dn_ctl_ptr;
358 extern	ip_dn_ruledel_t *ip_dn_ruledel_ptr;
359 extern	ip_dn_io_t *ip_dn_io_ptr;
360 #define	DUMMYNET_LOADED	(ip_dn_io_ptr != NULL)
361 #endif
362 
363 #endif /* _IP_DUMMYNET_H */
364