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.
115 struct ip_fw *rule; /* matching rule */
116 int dn_dir; /* action when packet comes out. */
117 #define DN_TO_IP_OUT 1
118 #define DN_TO_IP_IN 2
119 /* Obsolete: #define DN_TO_BDG_FWD 3 */
120 #define DN_TO_ETH_DEMUX 4
121 #define DN_TO_ETH_OUT 5
122 #define DN_TO_IP6_IN 6
123 #define DN_TO_IP6_OUT 7
124 #define DN_TO_IFB_FWD 8
126 dn_key output_time; /* when the pkt is due for delivery */
127 struct ifnet *ifp; /* interface, for ip_output */
128 struct _ip6dn_args ip6opt; /* XXX ipv6 options */
133 * Overall structure of dummynet (with WF2Q+):
135 In dummynet, packets are selected with the firewall rules, and passed
136 to two different objects: PIPE or QUEUE.
138 A QUEUE is just a queue with configurable size and queue management
139 policy. It is also associated with a mask (to discriminate among
140 different flows), a weight (used to give different shares of the
141 bandwidth to different flows) and a "pipe", which essentially
142 supplies the transmit clock for all queues associated with that
145 A PIPE emulates a fixed-bandwidth link, whose bandwidth is
146 configurable. The "clock" for a pipe can come from either an
147 internal timer, or from the transmit interrupt of an interface.
148 A pipe is also associated with one (or more, if masks are used)
149 queue, where all packets for that pipe are stored.
151 The bandwidth available on the pipe is shared by the queues
152 associated with that pipe (only one in case the packet is sent
153 to a PIPE) according to the WF2Q+ scheduling algorithm and the
156 In general, incoming packets are stored in the appropriate queue,
157 which is then placed into one of a few heaps managed by a scheduler
158 to decide when the packet should be extracted.
159 The scheduler (a function called dummynet()) is run at every timer
160 tick, and grabs queues from the head of the heaps when they are
161 ready for processing.
163 There are three data structures definining a pipe and associated queues:
165 + dn_pipe, which contains the main configuration parameters related
166 to delay and bandwidth;
167 + dn_flow_set, which contains WF2Q+ configuration, flow
168 masks, plr and RED configuration;
169 + dn_flow_queue, which is the per-flow queue (containing the packets)
171 Multiple dn_flow_set can be linked to the same pipe, and multiple
172 dn_flow_queue can be linked to the same dn_flow_set.
173 All data structures are linked in a linear list which is used for
174 housekeeping purposes.
176 During configuration, we create and initialize the dn_flow_set
177 and dn_pipe structures (a dn_pipe also contains a dn_flow_set).
179 At runtime: packets are sent to the appropriate dn_flow_set (either
180 WFQ ones, or the one embedded in the dn_pipe for fixed-rate flows),
181 which in turn dispatches them to the appropriate dn_flow_queue
182 (created dynamically according to the masks).
184 The transmit clock for fixed rate flows (ready_event()) selects the
185 dn_flow_queue to be used to transmit the next packet. For WF2Q,
186 wfq_ready_event() extract a pipe which in turn selects the right
187 flow using a number of heaps defined into the pipe itself.
193 * per flow queue. This contains the flow identifier, the queue
194 * of packets, counters, and parameters used to support both RED and
197 * A dn_flow_queue is created and initialized whenever a packet for
198 * a new flow arrives.
200 struct dn_flow_queue {
201 struct dn_flow_queue *next ;
202 struct ipfw_flow_id id ;
204 struct mbuf *head, *tail ; /* queue of packets */
209 * When we emulate MAC overheads, or channel unavailability due
210 * to other traffic on a shared medium, we augment the packet at
211 * the head of the queue with an 'extra_bits' field representsing
212 * the additional delay the packet will be subject to:
213 * extra_bits = bw*unavailable_time.
214 * With large bandwidth and large delays, extra_bits (and also numbytes)
215 * can become very large, so better play safe and use 64 bit
217 uint64_t numbytes ; /* credit for transmission (dynamic queues) */
218 int64_t extra_bits; /* extra bits simulating unavailable channel */
220 u_int64_t tot_pkts ; /* statistics counters */
221 u_int64_t tot_bytes ;
224 int hash_slot ; /* debugging/diagnostic */
227 int avg ; /* average queue length est. (scaled) */
228 int count ; /* arrivals since last RED drop */
229 int random ; /* random value (scaled) */
230 dn_key q_time; /* start of queue idle time */
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 */
237 dn_key S,F ; /* start time, finish time */
239 * Setting F < S means the timestamp is invalid. We only need
240 * to test this when the queue is empty.
245 * flow_set descriptor. Contains the "template" parameters for the
246 * queue configuration, and pointers to the hash table of dn_flow_queue's.
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
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).
257 SLIST_ENTRY(dn_flow_set) next; /* linked list in a hash slot */
259 u_short fs_nr ; /* flow_set number */
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_NOERROR 0x0010 /* do not report ENOBUFS on drops */
266 #define DN_HAS_PROFILE 0x0020 /* the pipe has a delay profile. */
267 #define DN_IS_PIPE 0x4000
268 #define DN_IS_QUEUE 0x8000
270 struct dn_pipe *pipe ; /* pointer to parent pipe */
271 u_short parent_nr ; /* parent pipe#, 0 if local to a pipe */
273 int weight ; /* WFQ queue weight */
274 int qsize ; /* queue size in slots or bytes */
275 int plr ; /* pkt loss rate (2^31-1 means 100%) */
277 struct ipfw_flow_id flow_mask ;
279 /* hash table of queues onto this flow_set */
280 int rq_size ; /* number of slots */
281 int rq_elements ; /* active elements */
282 struct dn_flow_queue **rq; /* array of rq_size entries */
284 u_int32_t last_expired ; /* do not expire too frequently */
285 int backlogged ; /* #active queues for this flowset */
289 #define SCALE(x) ( (x) << SCALE_RED )
290 #define SCALE_VAL(x) ( (x) >> SCALE_RED )
291 #define SCALE_MUL(x,y) ( ( (x) * (y) ) >> SCALE_RED )
292 int w_q ; /* queue weight (scaled) */
293 int max_th ; /* maximum threshold for queue (scaled) */
294 int min_th ; /* minimum threshold for queue (scaled) */
295 int max_p ; /* maximum value for p_b (scaled) */
296 u_int c_1 ; /* max_p/(max_th-min_th) (scaled) */
297 u_int c_2 ; /* max_p*min_th/(max_th-min_th) (scaled) */
298 u_int c_3 ; /* for GRED, (1-max_p)/max_th (scaled) */
299 u_int c_4 ; /* for GRED, 1 - 2*max_p (scaled) */
300 u_int * w_q_lookup ; /* lookup table for computing (1-w_q)^t */
301 u_int lookup_depth ; /* depth of lookup table */
302 int lookup_step ; /* granularity inside the lookup table */
303 int lookup_weight ; /* equal to (1-w_q)^t / (1-w_q)^(t+1) */
304 int avg_pkt_size ; /* medium packet size */
305 int max_pkt_size ; /* max packet size */
307 SLIST_HEAD(dn_flow_set_head, dn_flow_set);
310 * Pipe descriptor. Contains global parameters, delay-line queue,
311 * and the flow_set used for fixed-rate queues.
313 * For WF2Q+ support it also has 3 heaps holding dn_flow_queue:
314 * not_eligible_heap, for queues whose start time is higher
315 * than the virtual time. Sorted by start time.
316 * scheduler_heap, for queues eligible for scheduling. Sorted by
318 * idle_heap, all flows that are idle and can be removed. We
319 * do that on each tick so we do not slow down too much
320 * operations during forwarding.
323 struct dn_pipe { /* a pipe */
324 SLIST_ENTRY(dn_pipe) next; /* linked list in a hash slot */
326 int pipe_nr ; /* number */
327 int bandwidth; /* really, bytes/tick. */
328 int delay ; /* really, ticks */
330 struct mbuf *head, *tail ; /* packets in delay line */
333 struct dn_heap scheduler_heap ; /* top extract - key Finish time*/
334 struct dn_heap not_eligible_heap; /* top extract- key Start time */
335 struct dn_heap idle_heap ; /* random extract - key Start=Finish time */
337 dn_key V ; /* virtual time */
338 int sum; /* sum of weights of all active sessions */
340 /* Same as in dn_flow_queue, numbytes can become large */
341 int64_t numbytes; /* bits I can transmit (more or less). */
343 dn_key sched_time ; /* time pipe was scheduled in ready_heap */
346 * When the tx clock come from an interface (if_name[0] != '\0'), its name
347 * is stored below, whereas the ifp is filled when the rule is configured.
349 char if_name[IFNAMSIZ];
351 int ready ; /* set if ifp != NULL and we got a signal from it */
353 struct dn_flow_set fs ; /* used with fixed-rate flows */
355 /* fields to simulate a delay profile */
357 #define ED_MAX_NAME_LEN 32
358 char name[ED_MAX_NAME_LEN];
364 /* dn_pipe_max is used to pass pipe configuration from userland onto
365 * kernel space and back
367 #define ED_MAX_SAMPLES_NO 1024
370 int samples[ED_MAX_SAMPLES_NO];
373 SLIST_HEAD(dn_pipe_head, dn_pipe);
376 typedef int ip_dn_ctl_t(struct sockopt *); /* raw_ip.c */
377 typedef void ip_dn_ruledel_t(void *); /* ip_fw.c */
378 typedef int ip_dn_io_t(struct mbuf **m, int dir, struct ip_fw_args *fwa);
379 extern ip_dn_ctl_t *ip_dn_ctl_ptr;
380 extern ip_dn_ruledel_t *ip_dn_ruledel_ptr;
381 extern ip_dn_io_t *ip_dn_io_ptr;
382 #define DUMMYNET_LOADED (ip_dn_io_ptr != NULL)
385 * Return the IPFW rule associated with the dummynet tag; if any.
386 * Make sure that the dummynet tag is not reused by lower layers.
388 static __inline struct ip_fw *
389 ip_dn_claim_rule(struct mbuf *m)
391 struct m_tag *mtag = m_tag_find(m, PACKET_TAG_DUMMYNET, NULL);
393 mtag->m_tag_id = PACKET_TAG_NONE;
394 return (((struct dn_pkt_tag *)(mtag+1))->rule);
399 #endif /* _IP_DUMMYNET_H */