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 1.40.2.1 2008/04/25 10:26:30 oleg 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.
75 * The maximum hash table size for queues. This value must be a power
78 #define DN_MAX_HASH_SIZE 65536
81 * A heap entry is made of a key and a pointer to the actual
82 * object stored in the heap.
83 * The heap is an array of dn_heap_entry entries, dynamically allocated.
84 * Current size is "size", with "elements" actually in use.
85 * The heap normally supports only ordered insert and extract from the top.
86 * If we want to extract an object from the middle of the heap, we
87 * have to know where the object itself is located in the heap (or we
88 * need to scan the whole array). To this purpose, an object has a
89 * field (int) which contains the index of the object itself into the
90 * heap. When the object is moved, the field must also be updated.
91 * The offset of the index in the object is stored in the 'offset'
92 * field in the heap descriptor. The assumption is that this offset
93 * is non-zero if we want to support extract from the middle.
95 struct dn_heap_entry {
96 dn_key key ; /* sorting key. Topmost element is smallest one */
97 void *object ; /* object pointer */
103 int offset ; /* XXX if > 0 this is the offset of direct ptr to obj */
104 struct dn_heap_entry *p ; /* really an array of "size" entries */
109 * Packets processed by dummynet have an mbuf tag associated with
110 * them that carries their dummynet state. This is used within
111 * the dummynet code as well as outside when checking for special
112 * processing requirements.
113 * Note that the first part is the reinject info and is common to
114 * other forms of packet reinjection.
117 struct ipfw_rule_ref rule; /* matching rule */
119 /* second part, dummynet specific */
120 int dn_dir; /* action when packet comes out. */
121 /* see ip_fw_private.h */
123 dn_key output_time; /* when the pkt is due for delivery */
124 struct ifnet *ifp; /* interface, for ip_output */
125 struct _ip6dn_args ip6opt; /* XXX ipv6 options */
130 * Overall structure of dummynet (with WF2Q+):
132 In dummynet, packets are selected with the firewall rules, and passed
133 to two different objects: PIPE or QUEUE.
135 A QUEUE is just a queue with configurable size and queue management
136 policy. It is also associated with a mask (to discriminate among
137 different flows), a weight (used to give different shares of the
138 bandwidth to different flows) and a "pipe", which essentially
139 supplies the transmit clock for all queues associated with that
142 A PIPE emulates a fixed-bandwidth link, whose bandwidth is
143 configurable. The "clock" for a pipe can come from either an
144 internal timer, or from the transmit interrupt of an interface.
145 A pipe is also associated with one (or more, if masks are used)
146 queue, where all packets for that pipe are stored.
148 The bandwidth available on the pipe is shared by the queues
149 associated with that pipe (only one in case the packet is sent
150 to a PIPE) according to the WF2Q+ scheduling algorithm and the
153 In general, incoming packets are stored in the appropriate queue,
154 which is then placed into one of a few heaps managed by a scheduler
155 to decide when the packet should be extracted.
156 The scheduler (a function called dummynet()) is run at every timer
157 tick, and grabs queues from the head of the heaps when they are
158 ready for processing.
160 There are three data structures definining a pipe and associated queues:
162 + dn_pipe, which contains the main configuration parameters related
163 to delay and bandwidth;
164 + dn_flow_set, which contains WF2Q+ configuration, flow
165 masks, plr and RED configuration;
166 + dn_flow_queue, which is the per-flow queue (containing the packets)
168 Multiple dn_flow_set can be linked to the same pipe, and multiple
169 dn_flow_queue can be linked to the same dn_flow_set.
170 All data structures are linked in a linear list which is used for
171 housekeeping purposes.
173 During configuration, we create and initialize the dn_flow_set
174 and dn_pipe structures (a dn_pipe also contains a dn_flow_set).
176 At runtime: packets are sent to the appropriate dn_flow_set (either
177 WFQ ones, or the one embedded in the dn_pipe for fixed-rate flows),
178 which in turn dispatches them to the appropriate dn_flow_queue
179 (created dynamically according to the masks).
181 The transmit clock for fixed rate flows (ready_event()) selects the
182 dn_flow_queue to be used to transmit the next packet. For WF2Q,
183 wfq_ready_event() extract a pipe which in turn selects the right
184 flow using a number of heaps defined into the pipe itself.
190 * per flow queue. This contains the flow identifier, the queue
191 * of packets, counters, and parameters used to support both RED and
194 * A dn_flow_queue is created and initialized whenever a packet for
195 * a new flow arrives.
197 struct dn_flow_queue {
198 struct dn_flow_queue *next ;
199 struct ipfw_flow_id id ;
201 struct mbuf *head, *tail ; /* queue of packets */
206 * When we emulate MAC overheads, or channel unavailability due
207 * to other traffic on a shared medium, we augment the packet at
208 * the head of the queue with an 'extra_bits' field representsing
209 * the additional delay the packet will be subject to:
210 * extra_bits = bw*unavailable_time.
211 * With large bandwidth and large delays, extra_bits (and also numbytes)
212 * can become very large, so better play safe and use 64 bit
214 uint64_t numbytes ; /* credit for transmission (dynamic queues) */
215 int64_t extra_bits; /* extra bits simulating unavailable channel */
217 u_int64_t tot_pkts ; /* statistics counters */
218 u_int64_t tot_bytes ;
221 int hash_slot ; /* debugging/diagnostic */
224 int avg ; /* average queue length est. (scaled) */
225 int count ; /* arrivals since last RED drop */
226 int random ; /* random value (scaled) */
227 dn_key idle_time; /* start of queue idle time */
230 struct dn_flow_set *fs ; /* parent flow set */
231 int heap_pos ; /* position (index) of struct in heap */
232 dn_key sched_time ; /* current time when queue enters ready_heap */
234 dn_key S,F ; /* start time, finish time */
236 * Setting F < S means the timestamp is invalid. We only need
237 * to test this when the queue is empty.
242 * flow_set descriptor. Contains the "template" parameters for the
243 * queue configuration, and pointers to the hash table of dn_flow_queue's.
245 * The hash table is an array of lists -- we identify the slot by
246 * hashing the flow-id, then scan the list looking for a match.
247 * The size of the hash table (buckets) is configurable on a per-queue
250 * A dn_flow_set is created whenever a new queue or pipe is created (in the
251 * latter case, the structure is located inside the struct dn_pipe).
254 SLIST_ENTRY(dn_flow_set) next; /* linked list in a hash slot */
256 u_short fs_nr ; /* flow_set number */
258 #define DN_HAVE_FLOW_MASK 0x0001
259 #define DN_IS_RED 0x0002
260 #define DN_IS_GENTLE_RED 0x0004
261 #define DN_QSIZE_IS_BYTES 0x0008 /* queue size is measured in bytes */
262 #define DN_NOERROR 0x0010 /* do not report ENOBUFS on drops */
263 #define DN_HAS_PROFILE 0x0020 /* the pipe has a delay profile. */
264 #define DN_IS_PIPE 0x4000
265 #define DN_IS_QUEUE 0x8000
267 struct dn_pipe *pipe ; /* pointer to parent pipe */
268 u_short parent_nr ; /* parent pipe#, 0 if local to a pipe */
270 int weight ; /* WFQ queue weight */
271 int qsize ; /* queue size in slots or bytes */
272 int plr ; /* pkt loss rate (2^31-1 means 100%) */
274 struct ipfw_flow_id flow_mask ;
276 /* hash table of queues onto this flow_set */
277 int rq_size ; /* number of slots */
278 int rq_elements ; /* active elements */
279 struct dn_flow_queue **rq; /* array of rq_size entries */
281 u_int32_t last_expired ; /* do not expire too frequently */
282 int backlogged ; /* #active queues for this flowset */
286 #define SCALE(x) ( (x) << SCALE_RED )
287 #define SCALE_VAL(x) ( (x) >> SCALE_RED )
288 #define SCALE_MUL(x,y) ( ( (x) * (y) ) >> SCALE_RED )
289 int w_q ; /* queue weight (scaled) */
290 int max_th ; /* maximum threshold for queue (scaled) */
291 int min_th ; /* minimum threshold for queue (scaled) */
292 int max_p ; /* maximum value for p_b (scaled) */
293 u_int c_1 ; /* max_p/(max_th-min_th) (scaled) */
294 u_int c_2 ; /* max_p*min_th/(max_th-min_th) (scaled) */
295 u_int c_3 ; /* for GRED, (1-max_p)/max_th (scaled) */
296 u_int c_4 ; /* for GRED, 1 - 2*max_p (scaled) */
297 u_int * w_q_lookup ; /* lookup table for computing (1-w_q)^t */
298 u_int lookup_depth ; /* depth of lookup table */
299 int lookup_step ; /* granularity inside the lookup table */
300 int lookup_weight ; /* equal to (1-w_q)^t / (1-w_q)^(t+1) */
301 int avg_pkt_size ; /* medium packet size */
302 int max_pkt_size ; /* max packet size */
304 SLIST_HEAD(dn_flow_set_head, dn_flow_set);
307 * Pipe descriptor. Contains global parameters, delay-line queue,
308 * and the flow_set used for fixed-rate queues.
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
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.
320 struct dn_pipe { /* a pipe */
321 SLIST_ENTRY(dn_pipe) next; /* linked list in a hash slot */
323 int pipe_nr ; /* number */
324 int bandwidth; /* really, bytes/tick. */
325 int delay ; /* really, ticks */
327 struct mbuf *head, *tail ; /* packets in delay line */
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 */
334 dn_key V ; /* virtual time */
335 int sum; /* sum of weights of all active sessions */
337 /* Same as in dn_flow_queue, numbytes can become large */
338 int64_t numbytes; /* bits I can transmit (more or less). */
339 uint64_t burst; /* burst size, scaled: bits * hz */
341 dn_key sched_time ; /* time pipe was scheduled in ready_heap */
342 dn_key idle_time; /* start of pipe idle time */
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 */
354 /* fields to simulate a delay profile */
356 #define ED_MAX_NAME_LEN 32
357 char name[ED_MAX_NAME_LEN];
363 /* dn_pipe_max is used to pass pipe configuration from userland onto
364 * kernel space and back
366 #define ED_MAX_SAMPLES_NO 1024
369 int samples[ED_MAX_SAMPLES_NO];
372 SLIST_HEAD(dn_pipe_head, dn_pipe);
374 #endif /* _IP_DUMMYNET_H */