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 1.13 2007/11/05 15:16:46 sephe 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 to
43 * 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 sum
52 * If we limit to max 1000 flows and a max weight of 100, then x needs
53 * 17 bits. The packet size is 16 bits, so we can easily overflow if
54 * we do not allow errors.
56 * So we use a key "dn_key" which is 64 bits.
58 * MY_M is used as a shift count when doing fixed point arithmetic
59 * (a better name would be useful...).
61 typedef uint64_t dn_key; /* sorting key */
64 * Number of left shift to obtain a larger precision
66 * XXX With this scaling, max 1000 flows, max weight 100, 1Gbit/s, the
67 * virtual time wraps every 15 days.
74 * A heap entry is made of a key and a pointer to the actual object stored
77 * The heap is an array of dn_heap_entry entries, dynamically allocated.
78 * Current size is "size", with "elements" actually in use.
80 * The heap normally supports only ordered insert and extract from the top.
81 * If we want to extract an object from the middle of the heap, we have to
82 * know where the object itself is located in the heap (or we need to scan
83 * the whole array). To this purpose, an object has a field (int) which
84 * contains the index of the object itself into the heap. When the object
85 * is moved, the field must also be updated. The offset of the index in the
86 * object is stored in the 'offset' field in the heap descriptor. The
87 * assumption is that this offset is non-zero if we want to support extract
90 struct dn_heap_entry {
91 dn_key key; /* sorting key. Topmost element is smallest one */
92 void *object; /* object pointer */
98 int offset; /* XXX if > 0 this is the offset of direct ptr to obj */
99 struct dn_heap_entry *p; /* really an array of "size" entries */
103 * struct dn_pkt identifies a packet in the dummynet queue, but is also used
104 * to tag packets passed back to the various destinations (ip_input(),
105 * ip_output() and so on).
107 * It is a tag (PACKET_TAG_DUMMYNET) associated with the actual mbuf.
111 struct dn_pkt *dn_next;
113 struct ip_fw *rule; /* matching rule */
114 int dn_dir; /* action when packet comes out. */
115 #define DN_TO_IP_OUT 1
116 #define DN_TO_IP_IN 2
117 #define DN_TO_ETH_DEMUX 4
118 #define DN_TO_ETH_OUT 5
120 dn_key output_time; /* when the pkt is due for delivery */
121 struct ifnet *ifp; /* interface, for ip_output */
122 struct sockaddr_in *dn_dst;
123 struct route ro; /* route, for ip_output. MUST COPY */
124 int flags; /* flags, for ip_output (IPv6 ?) */
128 * Overall structure of dummynet (with WF2Q+):
130 * In dummynet, packets are selected with the firewall rules, and passed to
131 * two different objects: PIPE or QUEUE.
133 * A QUEUE is just a queue with configurable size and queue management policy.
134 * It is also associated with a mask (to discriminate among different flows),
135 * a weight (used to give different shares of the bandwidth to different flows)
136 * and a "pipe", which essentially supplies the transmit clock for all queues
137 * associated with that pipe.
139 * A PIPE emulates a fixed-bandwidth link, whose bandwidth is configurable.
140 * The "clock" for a pipe comes from an internal timer. A pipe is also
141 * associated with one (or more, if masks are used) queue, where all packets
142 * for that pipe are stored.
144 * The bandwidth available on the pipe is shared by the queues associated with
145 * that pipe (only one in case the packet is sent to a PIPE) according to the
146 * WF2Q+ scheduling algorithm and the configured weights.
148 * In general, incoming packets are stored in the appropriate queue, which is
149 * then placed into one of a few heaps managed by a scheduler to decide when
150 * the packet should be extracted. The scheduler (a function called dummynet())
151 * is run at every timer tick, and grabs queues from the head of the heaps when
152 * they are ready for processing.
154 * There are three data structures definining a pipe and associated queues:
156 * + dn_pipe, which contains the main configuration parameters related to
157 * delay and bandwidth;
158 * + dn_flow_set, which contains WF2Q+ configuration, flow masks, plr and
160 * + dn_flow_queue, which is the per-flow queue (containing the packets)
162 * Multiple dn_flow_set can be linked to the same pipe, and multiple
163 * dn_flow_queue can be linked to the same dn_flow_set.
164 * All data structures are linked in a linear list which is used for
165 * housekeeping purposes.
167 * During configuration, we create and initialize the dn_flow_set and dn_pipe
168 * structures (a dn_pipe also contains a dn_flow_set).
170 * At runtime: packets are sent to the appropriate dn_flow_set (either WFQ
171 * ones, or the one embedded in the dn_pipe for fixed-rate flows), which in
172 * turn dispatches them to the appropriate dn_flow_queue (created dynamically
173 * according to the masks).
175 * The transmit clock for fixed rate flows (ready_event()) selects the
176 * dn_flow_queue to be used to transmit the next packet. For WF2Q,
177 * wfq_ready_event() extract a pipe which in turn selects the right flow using
178 * a number of heaps defined into the pipe itself.
182 * Per flow queue. This contains the flow identifier, the queue of packets,
183 * counters, and parameters used to support both RED and WF2Q+.
185 * A dn_flow_queue is created and initialized whenever a packet for a new
188 struct dn_flow_queue {
189 struct dn_flow_queue *next;
190 struct ipfw_flow_id id;
192 struct dn_pkt *head, *tail; /* queue of packets */
195 u_long numbytes; /* credit for transmission (dynamic queues) */
197 uint64_t tot_pkts; /* statistics counters */
201 int hash_slot; /* debugging/diagnostic */
204 int avg; /* average queue length est. (scaled) */
205 int count; /* arrivals since last RED drop */
206 int random; /* random value (scaled) */
207 uint32_t q_time; /* start of queue idle time */
210 struct dn_flow_set *fs; /* parent flow set */
211 int heap_pos; /* position (index) of struct in heap */
212 dn_key sched_time; /* current time when queue enters ready_heap */
214 dn_key S, F; /* start time, finish time */
216 * Setting F < S means the timestamp is invalid. We only need
217 * to test this when the queue is empty.
222 * flow_set descriptor. Contains the "template" parameters for the queue
223 * configuration, and pointers to the hash table of dn_flow_queue's.
225 * The hash table is an array of lists -- we identify the slot by hashing
226 * the flow-id, then scan the list looking for a match.
227 * The size of the hash table (buckets) is configurable on a per-queue basis.
229 * A dn_flow_set is created whenever a new queue or pipe is created (in the
230 * latter case, the structure is located inside the struct dn_pipe).
233 struct dn_flow_set *next; /* next flow set in all_flow_sets list */
235 u_short fs_nr; /* flow_set number */
236 u_short flags_fs; /* see 'Flow set flags' */
238 struct dn_pipe *pipe; /* pointer to parent pipe */
239 u_short parent_nr; /* parent pipe#, 0 if local to a pipe */
241 int weight; /* WFQ queue weight */
242 int qsize; /* queue size in slots or bytes */
243 int plr; /* pkt loss rate (2^31-1 means 100%) */
245 struct ipfw_flow_id flow_mask;
247 /* hash table of queues onto this flow_set */
248 int rq_size; /* number of slots */
249 int rq_elements; /* active elements */
250 struct dn_flow_queue **rq; /* array of rq_size entries */
252 uint32_t last_expired; /* do not expire too frequently */
253 int backlogged; /* #active queues for this flowset */
256 int w_q; /* queue weight (scaled) */
257 int max_th; /* maximum threshold for queue (scaled) */
258 int min_th; /* minimum threshold for queue (scaled) */
259 int max_p; /* maximum value for p_b (scaled) */
260 u_int c_1; /* max_p/(max_th-min_th) (scaled) */
261 u_int c_2; /* max_p*min_th/(max_th-min_th) (scaled) */
262 u_int c_3; /* for GRED, (1-max_p)/max_th (scaled) */
263 u_int c_4; /* for GRED, 1 - 2*max_p (scaled) */
264 u_int *w_q_lookup; /* lookup table for computing (1-w_q)^t */
265 u_int lookup_depth; /* depth of lookup table */
266 int lookup_step; /* granularity inside the lookup table */
267 int lookup_weight; /* equal to (1-w_q)^t / (1-w_q)^(t+1) */
268 int avg_pkt_size; /* medium packet size */
269 int max_pkt_size; /* max packet size */
273 * Pipe descriptor. Contains global parameters, delay-line queue, and the
274 * flow_set used for fixed-rate queues.
276 * For WF2Q+ support it also has 3 heaps holding dn_flow_queue:
277 * + not_eligible_heap, for queues whose start time is higher than the
278 * virtual time. Sorted by start time.
279 * + scheduler_heap, for queues eligible for scheduling. Sorted by finish
281 * + idle_heap, all flows that are idle and can be removed. We do that on
282 * each tick so we do not slow down too much operations during forwarding.
284 struct dn_pipe { /* a pipe */
285 struct dn_pipe *next;
287 int pipe_nr; /* number */
288 int bandwidth; /* really, bytes/tick. */
289 int delay; /* really, ticks */
291 struct dn_pkt *head, *tail; /* packets in delay line */
294 struct dn_heap scheduler_heap; /* top extract - key Finish time*/
295 struct dn_heap not_eligible_heap; /* top extract- key Start time */
296 struct dn_heap idle_heap; /* random extract - key Start=Finish time */
298 dn_key V; /* virtual time */
299 int sum; /* sum of weights of all active sessions */
300 int numbytes; /* bits I can transmit (more or less). */
302 dn_key sched_time; /* time pipe was scheduled in ready_heap */
304 struct dn_flow_set fs; /* used with fixed-rate flows */
307 typedef int ip_dn_ctl_t(struct sockopt *); /* raw_ip.c */
308 typedef void ip_dn_ruledel_t(void *); /* ip_fw2.c */
309 typedef int ip_dn_io_t(struct mbuf *, int, int, struct ip_fw_args *);
311 extern ip_dn_ctl_t *ip_dn_ctl_ptr;
312 extern ip_dn_ruledel_t *ip_dn_ruledel_ptr;
313 extern ip_dn_io_t *ip_dn_io_ptr;
315 #define DUMMYNET_LOADED (ip_dn_io_ptr != NULL)
319 struct dn_ioc_flowid {
320 uint16_t type; /* ETHERTYPE_ */
335 struct dn_ioc_flowqueue {
343 int hash_slot; /* debugging/diagnostic */
344 dn_key S; /* virtual start time */
345 dn_key F; /* virtual finish time */
347 struct dn_ioc_flowid id;
348 uint8_t reserved[16];
351 struct dn_ioc_flowset {
352 u_short fs_type; /* DN_IS_{QUEUE,PIPE}, MUST be first */
354 u_short fs_nr; /* flow_set number */
355 u_short flags_fs; /* see 'Flow set flags' */
356 u_short parent_nr; /* parent pipe#, 0 if local to a pipe */
358 int weight; /* WFQ queue weight */
359 int qsize; /* queue size in slots or bytes */
360 int plr; /* pkt loss rate (2^31-1 means 100%) */
362 /* Hash table information */
363 int rq_size; /* number of slots */
364 int rq_elements; /* active elements */
367 int w_q; /* queue weight (scaled) */
368 int max_th; /* maximum threshold for queue (scaled) */
369 int min_th; /* minimum threshold for queue (scaled) */
370 int max_p; /* maximum value for p_b (scaled) */
371 int lookup_step; /* granularity inside the lookup table */
372 int lookup_weight; /* equal to (1-w_q)^t / (1-w_q)^(t+1) */
374 struct dn_ioc_flowid flow_mask;
375 uint8_t reserved[16];
379 struct dn_ioc_flowset fs; /* MUST be first */
381 int pipe_nr; /* pipe number */
382 int bandwidth; /* bit/second */
383 int delay; /* milliseconds */
385 dn_key V; /* virtual time */
387 uint8_t reserved[16];
393 #define DN_HAVE_FLOW_MASK 0x0001
394 #define DN_IS_RED 0x0002
395 #define DN_IS_GENTLE_RED 0x0004
396 #define DN_QSIZE_IS_BYTES 0x0008 /* queue size is measured in bytes */
397 #define DN_NOERROR 0x0010 /* do not report ENOBUFS on drops */
398 #define DN_IS_PIPE 0x4000
399 #define DN_IS_QUEUE 0x8000
405 #define SCALE(x) ((x) << SCALE_RED)
406 #define SCALE_VAL(x) ((x) >> SCALE_RED)
407 #define SCALE_MUL(x, y) (((x) * (y)) >> SCALE_RED)
409 #endif /* !_IP_DUMMYNET_H */