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 $
30 #ifndef _IP_DUMMYNET_H
31 #define _IP_DUMMYNET_H
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.
41 * We start with a heap, which is used in the scheduler to decide when
42 * to transmit packets etc.
44 * The key for the heap is used for two different values:
46 * 1. timer ticks- max 10K/second, so 32 bits are enough;
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
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...).
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 */
69 * XXX With this scaling, max 1000 flows, max weight 100, 1Gbit/s, the
70 * virtual time wraps every 15 days.
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.
77 #define OFFSET_OF(type, field) ((int)&( ((type *)0)->field) )
80 * The maximum hash table size for queues. This value must be a power
83 #define DN_MAX_HASH_SIZE 65536
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.
100 struct dn_heap_entry {
101 dn_key key ; /* sorting key. Topmost element is smallest one */
102 void *object ; /* object pointer */
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 */
113 * struct dn_pkt identifies a packet in the dummynet queue, but
114 * is also used to tag packets passed back to the various destinations
115 * (ip_input(), ip_output(), bdg_forward() and so on).
116 * As such the first part of the structure must be a struct m_hdr,
117 * followed by dummynet-specific parameters. The m_hdr must be
120 * mh_flags = PACKET_TYPE_DUMMYNET;
121 * mh_next = <pointer to the actual mbuf>
123 * mh_nextpkt, mh_data are free for dummynet use (mh_nextpkt is used to
124 * build a linked list of packets in a dummynet queue).
128 #define DN_NEXT(x) (struct dn_pkt *)(x)->hdr.mh_nextpkt
129 #define dn_m hdr.mh_next /* packet to be forwarded */
131 struct ip_fw *rule; /* matching rule */
132 int dn_dir; /* action when packet comes out. */
133 #define DN_TO_IP_OUT 1
134 #define DN_TO_IP_IN 2
135 #define DN_TO_BDG_FWD 3
136 #define DN_TO_ETH_DEMUX 4
137 #define DN_TO_ETH_OUT 5
139 dn_key output_time; /* when the pkt is due for delivery */
140 struct ifnet *ifp; /* interface, for ip_output */
141 struct sockaddr_in *dn_dst ;
142 struct route ro; /* route, for ip_output. MUST COPY */
143 int flags ; /* flags, for ip_output (IPv6 ?) */
147 * Overall structure of dummynet (with WF2Q+):
149 In dummynet, packets are selected with the firewall rules, and passed
150 to two different objects: PIPE or QUEUE.
152 A QUEUE is just a queue with configurable size and queue management
153 policy. It is also associated with a mask (to discriminate among
154 different flows), a weight (used to give different shares of the
155 bandwidth to different flows) and a "pipe", which essentially
156 supplies the transmit clock for all queues associated with that
159 A PIPE emulates a fixed-bandwidth link, whose bandwidth is
160 configurable. The "clock" for a pipe can come from either an
161 internal timer, or from the transmit interrupt of an interface.
162 A pipe is also associated with one (or more, if masks are used)
163 queue, where all packets for that pipe are stored.
165 The bandwidth available on the pipe is shared by the queues
166 associated with that pipe (only one in case the packet is sent
167 to a PIPE) according to the WF2Q+ scheduling algorithm and the
170 In general, incoming packets are stored in the appropriate queue,
171 which is then placed into one of a few heaps managed by a scheduler
172 to decide when the packet should be extracted.
173 The scheduler (a function called dummynet()) is run at every timer
174 tick, and grabs queues from the head of the heaps when they are
175 ready for processing.
177 There are three data structures definining a pipe and associated queues:
179 + dn_pipe, which contains the main configuration parameters related
180 to delay and bandwidth;
181 + dn_flow_set, which contains WF2Q+ configuration, flow
182 masks, plr and RED configuration;
183 + dn_flow_queue, which is the per-flow queue (containing the packets)
185 Multiple dn_flow_set can be linked to the same pipe, and multiple
186 dn_flow_queue can be linked to the same dn_flow_set.
187 All data structures are linked in a linear list which is used for
188 housekeeping purposes.
190 During configuration, we create and initialize the dn_flow_set
191 and dn_pipe structures (a dn_pipe also contains a dn_flow_set).
193 At runtime: packets are sent to the appropriate dn_flow_set (either
194 WFQ ones, or the one embedded in the dn_pipe for fixed-rate flows),
195 which in turn dispatches them to the appropriate dn_flow_queue
196 (created dynamically according to the masks).
198 The transmit clock for fixed rate flows (ready_event()) selects the
199 dn_flow_queue to be used to transmit the next packet. For WF2Q,
200 wfq_ready_event() extract a pipe which in turn selects the right
201 flow using a number of heaps defined into the pipe itself.
207 * per flow queue. This contains the flow identifier, the queue
208 * of packets, counters, and parameters used to support both RED and
211 * A dn_flow_queue is created and initialized whenever a packet for
212 * a new flow arrives.
214 struct dn_flow_queue {
215 struct dn_flow_queue *next ;
216 struct ipfw_flow_id id ;
218 struct dn_pkt *head, *tail ; /* queue of packets */
221 u_long numbytes ; /* credit for transmission (dynamic queues) */
223 u_int64_t tot_pkts ; /* statistics counters */
224 u_int64_t tot_bytes ;
227 int hash_slot ; /* debugging/diagnostic */
230 int avg ; /* average queue length est. (scaled) */
231 int count ; /* arrivals since last RED drop */
232 int random ; /* random value (scaled) */
233 u_int32_t q_time ; /* start of queue idle time */
236 struct dn_flow_set *fs ; /* parent flow set */
237 int heap_pos ; /* position (index) of struct in heap */
238 dn_key sched_time ; /* current time when queue enters ready_heap */
240 dn_key S,F ; /* start time, finish time */
242 * Setting F < S means the timestamp is invalid. We only need
243 * to test this when the queue is empty.
248 * flow_set descriptor. Contains the "template" parameters for the
249 * queue configuration, and pointers to the hash table of dn_flow_queue's.
251 * The hash table is an array of lists -- we identify the slot by
252 * hashing the flow-id, then scan the list looking for a match.
253 * The size of the hash table (buckets) is configurable on a per-queue
256 * A dn_flow_set is created whenever a new queue or pipe is created (in the
257 * latter case, the structure is located inside the struct dn_pipe).
260 struct dn_flow_set *next; /* next flow set in all_flow_sets list */
262 u_short fs_nr ; /* flow_set number */
264 #define DN_HAVE_FLOW_MASK 0x0001
265 #define DN_IS_RED 0x0002
266 #define DN_IS_GENTLE_RED 0x0004
267 #define DN_QSIZE_IS_BYTES 0x0008 /* queue size is measured in bytes */
268 #define DN_NOERROR 0x0010 /* do not report ENOBUFS on drops */
269 #define DN_IS_PIPE 0x4000
270 #define DN_IS_QUEUE 0x8000
272 struct dn_pipe *pipe ; /* pointer to parent pipe */
273 u_short parent_nr ; /* parent pipe#, 0 if local to a pipe */
275 int weight ; /* WFQ queue weight */
276 int qsize ; /* queue size in slots or bytes */
277 int plr ; /* pkt loss rate (2^31-1 means 100%) */
279 struct ipfw_flow_id flow_mask ;
281 /* hash table of queues onto this flow_set */
282 int rq_size ; /* number of slots */
283 int rq_elements ; /* active elements */
284 struct dn_flow_queue **rq; /* array of rq_size entries */
286 u_int32_t last_expired ; /* do not expire too frequently */
287 int backlogged ; /* #active queues for this flowset */
291 #define SCALE(x) ( (x) << SCALE_RED )
292 #define SCALE_VAL(x) ( (x) >> SCALE_RED )
293 #define SCALE_MUL(x,y) ( ( (x) * (y) ) >> SCALE_RED )
294 int w_q ; /* queue weight (scaled) */
295 int max_th ; /* maximum threshold for queue (scaled) */
296 int min_th ; /* minimum threshold for queue (scaled) */
297 int max_p ; /* maximum value for p_b (scaled) */
298 u_int c_1 ; /* max_p/(max_th-min_th) (scaled) */
299 u_int c_2 ; /* max_p*min_th/(max_th-min_th) (scaled) */
300 u_int c_3 ; /* for GRED, (1-max_p)/max_th (scaled) */
301 u_int c_4 ; /* for GRED, 1 - 2*max_p (scaled) */
302 u_int * w_q_lookup ; /* lookup table for computing (1-w_q)^t */
303 u_int lookup_depth ; /* depth of lookup table */
304 int lookup_step ; /* granularity inside the lookup table */
305 int lookup_weight ; /* equal to (1-w_q)^t / (1-w_q)^(t+1) */
306 int avg_pkt_size ; /* medium packet size */
307 int max_pkt_size ; /* max packet size */
311 * Pipe descriptor. Contains global parameters, delay-line queue,
312 * and the flow_set used for fixed-rate queues.
314 * For WF2Q+ support it also has 3 heaps holding dn_flow_queue:
315 * not_eligible_heap, for queues whose start time is higher
316 * than the virtual time. Sorted by start time.
317 * scheduler_heap, for queues eligible for scheduling. Sorted by
319 * idle_heap, all flows that are idle and can be removed. We
320 * do that on each tick so we do not slow down too much
321 * operations during forwarding.
324 struct dn_pipe { /* a pipe */
325 struct dn_pipe *next ;
327 int pipe_nr ; /* number */
328 int bandwidth; /* really, bytes/tick. */
329 int delay ; /* really, ticks */
331 struct dn_pkt *head, *tail ; /* packets in delay line */
334 struct dn_heap scheduler_heap ; /* top extract - key Finish time*/
335 struct dn_heap not_eligible_heap; /* top extract- key Start time */
336 struct dn_heap idle_heap ; /* random extract - key Start=Finish time */
338 dn_key V ; /* virtual time */
339 int sum; /* sum of weights of all active sessions */
340 int numbytes; /* bits I can transmit (more or less). */
342 dn_key sched_time ; /* time pipe was scheduled in ready_heap */
345 * When the tx clock come from an interface (if_name[0] != '\0'), its name
346 * is stored below, whereas the ifp is filled when the rule is configured.
348 char if_name[IFNAMSIZ];
350 int ready ; /* set if ifp != NULL and we got a signal from it */
352 struct dn_flow_set fs ; /* used with fixed-rate flows */
356 typedef int ip_dn_ctl_t(struct sockopt *); /* raw_ip.c */
357 typedef void ip_dn_ruledel_t(void *); /* ip_fw.c */
358 typedef int ip_dn_io_t(struct mbuf *m, int pipe_nr, int dir,
359 struct ip_fw_args *fwa);
360 extern ip_dn_ctl_t *ip_dn_ctl_ptr;
361 extern ip_dn_ruledel_t *ip_dn_ruledel_ptr;
362 extern ip_dn_io_t *ip_dn_io_ptr;
363 #define DUMMYNET_LOADED (ip_dn_io_ptr != NULL)
366 #endif /* _IP_DUMMYNET_H */