2 * Copyright (c) 1998-2002 Luigi Rizzo, Universita` di Pisa
3 * Portions Copyright (c) 2000 Akamba Corp.
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
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.
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
27 * $FreeBSD: src/sys/netinet/ip_dummynet.h,v 1.10.2.9 2003/05/13 09:31:06 maxim Exp $
28 * $DragonFly: src/sys/net/dummynet/ip_dummynet.h,v 1.4 2006/05/21 03:43:46 dillon Exp $
31 #ifndef _IP_DUMMYNET_H
32 #define _IP_DUMMYNET_H
35 * Definition of dummynet data structures. In the structures, I decided
36 * not to use the macros in <sys/queue.h> in the hope of making the code
37 * easier to port to other architectures. The type of lists and queue we
38 * use here is pretty simple anyways.
42 * We start with a heap, which is used in the scheduler to decide when
43 * to transmit packets etc.
45 * The key for the heap is used for two different values:
47 * 1. timer ticks- max 10K/second, so 32 bits are enough;
49 * 2. virtual times. These increase in steps of len/x, where len is the
50 * packet length, and x is either the weight of the flow, or the
52 * If we limit to max 1000 flows and a max weight of 100, then
53 * x needs 17 bits. The packet size is 16 bits, so we can easily
54 * overflow if we do not allow errors.
55 * So we use a key "dn_key" which is 64 bits. Some macros are used to
56 * compare key values and handle wraparounds.
57 * MAX64 returns the largest of two key values.
58 * MY_M is used as a shift count when doing fixed point arithmetic
59 * (a better name would be useful...).
61 typedef u_int64_t dn_key ; /* sorting key */
62 #define DN_KEY_LT(a,b) ((int64_t)((a)-(b)) < 0)
63 #define DN_KEY_LEQ(a,b) ((int64_t)((a)-(b)) <= 0)
64 #define DN_KEY_GT(a,b) ((int64_t)((a)-(b)) > 0)
65 #define DN_KEY_GEQ(a,b) ((int64_t)((a)-(b)) >= 0)
66 #define MAX64(x,y) (( (int64_t) ( (y)-(x) )) > 0 ) ? (y) : (x)
67 #define MY_M 16 /* number of left shift to obtain a larger precision */
70 * XXX With this scaling, max 1000 flows, max weight 100, 1Gbit/s, the
71 * virtual time wraps every 15 days.
75 * The OFFSET_OF macro is used to return the offset of a field within
76 * a structure. It is used by the heap management routines.
78 #define OFFSET_OF(type, field) ((int)&( ((type *)0)->field) )
81 * The maximum hash table size for queues. This value must be a power
84 #define DN_MAX_HASH_SIZE 65536
87 * A heap entry is made of a key and a pointer to the actual
88 * object stored in the heap.
89 * The heap is an array of dn_heap_entry entries, dynamically allocated.
90 * Current size is "size", with "elements" actually in use.
91 * The heap normally supports only ordered insert and extract from the top.
92 * If we want to extract an object from the middle of the heap, we
93 * have to know where the object itself is located in the heap (or we
94 * need to scan the whole array). To this purpose, an object has a
95 * field (int) which contains the index of the object itself into the
96 * heap. When the object is moved, the field must also be updated.
97 * The offset of the index in the object is stored in the 'offset'
98 * field in the heap descriptor. The assumption is that this offset
99 * is non-zero if we want to support extract from the middle.
101 struct dn_heap_entry {
102 dn_key key ; /* sorting key. Topmost element is smallest one */
103 void *object ; /* object pointer */
109 int offset ; /* XXX if > 0 this is the offset of direct ptr to obj */
110 struct dn_heap_entry *p ; /* really an array of "size" entries */
113 #if defined(_KERNEL) || defined(_KERNEL_STRUCTURES)
116 * struct dn_pkt identifies a packet in the dummynet queue, but
117 * is also used to tag packets passed back to the various destinations
118 * (ip_input(), ip_output(), bdg_forward() and so on).
119 * As such the first part of the structure must be a struct m_hdr,
120 * followed by dummynet-specific parameters. The m_hdr must be
123 * mh_flags = PACKET_TYPE_DUMMYNET;
124 * mh_next = <pointer to the actual mbuf>
126 * mh_nextpkt, mh_data are free for dummynet use (mh_nextpkt is used to
127 * build a linked list of packets in a dummynet queue).
131 #define DN_NEXT_NC(x) (x)->hdr.mh_nextpkt
132 #define DN_NEXT(x) (struct dn_pkt *)DN_NEXT_NC(x)
133 #define dn_m hdr.mh_next /* packet to be forwarded */
135 struct ip_fw *rule; /* matching rule */
136 int dn_dir; /* action when packet comes out. */
137 #define DN_TO_IP_OUT 1
138 #define DN_TO_IP_IN 2
139 #define DN_TO_BDG_FWD 3
140 #define DN_TO_ETH_DEMUX 4
141 #define DN_TO_ETH_OUT 5
143 dn_key output_time; /* when the pkt is due for delivery */
144 struct ifnet *ifp; /* interface, for ip_output */
145 struct sockaddr_in *dn_dst ;
146 struct route ro; /* route, for ip_output. MUST COPY */
147 int flags ; /* flags, for ip_output (IPv6 ?) */
153 * Overall structure of dummynet (with WF2Q+):
155 In dummynet, packets are selected with the firewall rules, and passed
156 to two different objects: PIPE or QUEUE.
158 A QUEUE is just a queue with configurable size and queue management
159 policy. It is also associated with a mask (to discriminate among
160 different flows), a weight (used to give different shares of the
161 bandwidth to different flows) and a "pipe", which essentially
162 supplies the transmit clock for all queues associated with that
165 A PIPE emulates a fixed-bandwidth link, whose bandwidth is
166 configurable. The "clock" for a pipe can come from either an
167 internal timer, or from the transmit interrupt of an interface.
168 A pipe is also associated with one (or more, if masks are used)
169 queue, where all packets for that pipe are stored.
171 The bandwidth available on the pipe is shared by the queues
172 associated with that pipe (only one in case the packet is sent
173 to a PIPE) according to the WF2Q+ scheduling algorithm and the
176 In general, incoming packets are stored in the appropriate queue,
177 which is then placed into one of a few heaps managed by a scheduler
178 to decide when the packet should be extracted.
179 The scheduler (a function called dummynet()) is run at every timer
180 tick, and grabs queues from the head of the heaps when they are
181 ready for processing.
183 There are three data structures definining a pipe and associated queues:
185 + dn_pipe, which contains the main configuration parameters related
186 to delay and bandwidth;
187 + dn_flow_set, which contains WF2Q+ configuration, flow
188 masks, plr and RED configuration;
189 + dn_flow_queue, which is the per-flow queue (containing the packets)
191 Multiple dn_flow_set can be linked to the same pipe, and multiple
192 dn_flow_queue can be linked to the same dn_flow_set.
193 All data structures are linked in a linear list which is used for
194 housekeeping purposes.
196 During configuration, we create and initialize the dn_flow_set
197 and dn_pipe structures (a dn_pipe also contains a dn_flow_set).
199 At runtime: packets are sent to the appropriate dn_flow_set (either
200 WFQ ones, or the one embedded in the dn_pipe for fixed-rate flows),
201 which in turn dispatches them to the appropriate dn_flow_queue
202 (created dynamically according to the masks).
204 The transmit clock for fixed rate flows (ready_event()) selects the
205 dn_flow_queue to be used to transmit the next packet. For WF2Q,
206 wfq_ready_event() extract a pipe which in turn selects the right
207 flow using a number of heaps defined into the pipe itself.
213 * per flow queue. This contains the flow identifier, the queue
214 * of packets, counters, and parameters used to support both RED and
217 * A dn_flow_queue is created and initialized whenever a packet for
218 * a new flow arrives.
220 struct dn_flow_queue {
221 struct dn_flow_queue *next ;
222 struct ipfw_flow_id id ;
224 struct dn_pkt *head, *tail ; /* queue of packets */
227 u_long numbytes ; /* credit for transmission (dynamic queues) */
229 u_int64_t tot_pkts ; /* statistics counters */
230 u_int64_t tot_bytes ;
233 int hash_slot ; /* debugging/diagnostic */
236 int avg ; /* average queue length est. (scaled) */
237 int count ; /* arrivals since last RED drop */
238 int random ; /* random value (scaled) */
239 u_int32_t q_time ; /* start of queue idle time */
242 struct dn_flow_set *fs ; /* parent flow set */
243 int heap_pos ; /* position (index) of struct in heap */
244 dn_key sched_time ; /* current time when queue enters ready_heap */
246 dn_key S,F ; /* start time, finish time */
248 * Setting F < S means the timestamp is invalid. We only need
249 * to test this when the queue is empty.
254 * flow_set descriptor. Contains the "template" parameters for the
255 * queue configuration, and pointers to the hash table of dn_flow_queue's.
257 * The hash table is an array of lists -- we identify the slot by
258 * hashing the flow-id, then scan the list looking for a match.
259 * The size of the hash table (buckets) is configurable on a per-queue
262 * A dn_flow_set is created whenever a new queue or pipe is created (in the
263 * latter case, the structure is located inside the struct dn_pipe).
266 struct dn_flow_set *next; /* next flow set in all_flow_sets list */
268 u_short fs_nr ; /* flow_set number */
270 #define DN_HAVE_FLOW_MASK 0x0001
271 #define DN_IS_RED 0x0002
272 #define DN_IS_GENTLE_RED 0x0004
273 #define DN_QSIZE_IS_BYTES 0x0008 /* queue size is measured in bytes */
274 #define DN_NOERROR 0x0010 /* do not report ENOBUFS on drops */
275 #define DN_IS_PIPE 0x4000
276 #define DN_IS_QUEUE 0x8000
278 struct dn_pipe *pipe ; /* pointer to parent pipe */
279 u_short parent_nr ; /* parent pipe#, 0 if local to a pipe */
281 int weight ; /* WFQ queue weight */
282 int qsize ; /* queue size in slots or bytes */
283 int plr ; /* pkt loss rate (2^31-1 means 100%) */
285 struct ipfw_flow_id flow_mask ;
287 /* hash table of queues onto this flow_set */
288 int rq_size ; /* number of slots */
289 int rq_elements ; /* active elements */
290 struct dn_flow_queue **rq; /* array of rq_size entries */
292 u_int32_t last_expired ; /* do not expire too frequently */
293 int backlogged ; /* #active queues for this flowset */
297 #define SCALE(x) ( (x) << SCALE_RED )
298 #define SCALE_VAL(x) ( (x) >> SCALE_RED )
299 #define SCALE_MUL(x,y) ( ( (x) * (y) ) >> SCALE_RED )
300 int w_q ; /* queue weight (scaled) */
301 int max_th ; /* maximum threshold for queue (scaled) */
302 int min_th ; /* minimum threshold for queue (scaled) */
303 int max_p ; /* maximum value for p_b (scaled) */
304 u_int c_1 ; /* max_p/(max_th-min_th) (scaled) */
305 u_int c_2 ; /* max_p*min_th/(max_th-min_th) (scaled) */
306 u_int c_3 ; /* for GRED, (1-max_p)/max_th (scaled) */
307 u_int c_4 ; /* for GRED, 1 - 2*max_p (scaled) */
308 u_int * w_q_lookup ; /* lookup table for computing (1-w_q)^t */
309 u_int lookup_depth ; /* depth of lookup table */
310 int lookup_step ; /* granularity inside the lookup table */
311 int lookup_weight ; /* equal to (1-w_q)^t / (1-w_q)^(t+1) */
312 int avg_pkt_size ; /* medium packet size */
313 int max_pkt_size ; /* max packet size */
317 * Pipe descriptor. Contains global parameters, delay-line queue,
318 * and the flow_set used for fixed-rate queues.
320 * For WF2Q+ support it also has 3 heaps holding dn_flow_queue:
321 * not_eligible_heap, for queues whose start time is higher
322 * than the virtual time. Sorted by start time.
323 * scheduler_heap, for queues eligible for scheduling. Sorted by
325 * idle_heap, all flows that are idle and can be removed. We
326 * do that on each tick so we do not slow down too much
327 * operations during forwarding.
330 struct dn_pipe { /* a pipe */
331 struct dn_pipe *next ;
333 int pipe_nr ; /* number */
334 int bandwidth; /* really, bytes/tick. */
335 int delay ; /* really, ticks */
337 struct dn_pkt *head, *tail ; /* packets in delay line */
340 struct dn_heap scheduler_heap ; /* top extract - key Finish time*/
341 struct dn_heap not_eligible_heap; /* top extract- key Start time */
342 struct dn_heap idle_heap ; /* random extract - key Start=Finish time */
344 dn_key V ; /* virtual time */
345 int sum; /* sum of weights of all active sessions */
346 int numbytes; /* bits I can transmit (more or less). */
348 dn_key sched_time ; /* time pipe was scheduled in ready_heap */
351 * When the tx clock come from an interface (if_name[0] != '\0'), its name
352 * is stored below, whereas the ifp is filled when the rule is configured.
354 char if_name[IFNAMSIZ];
356 int ready ; /* set if ifp != NULL and we got a signal from it */
358 struct dn_flow_set fs ; /* used with fixed-rate flows */
362 typedef int ip_dn_ctl_t(struct sockopt *); /* raw_ip.c */
363 typedef void ip_dn_ruledel_t(void *); /* ip_fw.c */
364 typedef int ip_dn_io_t(struct mbuf *m, int pipe_nr, int dir,
365 struct ip_fw_args *fwa);
366 extern ip_dn_ctl_t *ip_dn_ctl_ptr;
367 extern ip_dn_ruledel_t *ip_dn_ruledel_ptr;
368 extern ip_dn_io_t *ip_dn_io_ptr;
369 #define DUMMYNET_LOADED (ip_dn_io_ptr != NULL)
372 #endif /* _IP_DUMMYNET_H */