bsd/netinet/ip_dummynet.h

   1 /*
   2  * Copyright (c) 2000 Apple Computer, Inc. All rights reserved.
   3  *
   4  * @APPLE_LICENSE_HEADER_START@
   5  *
   6  * The contents of this file constitute Original Code as defined in and
   7  * are subject to the Apple Public Source License Version 1.1 (the
   8  * "License").  You may not use this file except in compliance with the
   9  * License.  Please obtain a copy of the License at
  10  * http://www.apple.com/publicsource and read it before using this file.
  11  *
  12  * This Original Code and all software distributed under the License are
  13  * distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY KIND, EITHER
  14  * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
  15  * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE OR NON-INFRINGEMENT.  Please see the
  17  * License for the specific language governing rights and limitations
  18  * under the License.
  19  *
  20  * @APPLE_LICENSE_HEADER_END@
  21  */
  22 /*
  23  * Copyright (c) 1998 Luigi Rizzo
  24  *
  25  * Redistribution and use in source forms, with and without modification,
  26  * are permitted provided that this entire comment appears intact.
  27  *
  28  * Redistribution in binary form may occur without any restrictions.
  29  * Obviously, it would be nice if you gave credit where credit is due
  30  * but requiring it would be too onerous.
  31  *
  32  * This software is provided ``AS IS'' without any warranties of any kind.
  33  *
  34  * $FreeBSD: src/sys/netinet/ip_dummynet.h,v 1.10.2.3 2001/02/01 20:25:09 luigi Exp $
  35  */
  36
  37 #ifndef _IP_DUMMYNET_H
  38 #define _IP_DUMMYNET_H
  39 #include <sys/appleapiopts.h>
  40 #ifdef __APPLE_API_PRIVATE
  41
  42 /*
  43  * Definition of dummynet data structures. In the structures, I decided
  44  * not to use the macros in <sys/queue.h> in the hope of making the code
  45  * easier to port to other architectures. The type of lists and queue we
  46  * use here is pretty simple anyways.
  47  */
  48
  49 /*
  50  * We start with a heap, which is used in the scheduler to decide when
  51  * to transmit packets etc.
  52  *
  53  * The key for the heap is used for two different values:
  54  *
  55  * 1. timer ticks- max 10K/second, so 32 bits are enough;
  56  *
  57  * 2. virtual times. These increase in steps of len/x, where len is the
  58  *    packet length, and x is either the weight of the flow, or the
  59  *    sum of all weights.
  60  *    If we limit to max 1000 flows and a max weight of 100, then
  61  *    x needs 17 bits. The packet size is 16 bits, so we can easily
  62  *    overflow if we do not allow errors.
  63  * So we use a key "dn_key" which is 64 bits. Some macros are used to
  64  * compare key values and handle wraparounds.
  65  * MAX64 returns the largest of two key values.
  66  * MY_M is used as a shift count when doing fixed point arithmetic
  67  * (a better name would be useful...).
  68  */
  69 typedef u_int64_t dn_key ;      /* sorting key */
  70 #define DN_KEY_LT(a,b)     ((int64_t)((a)-(b)) < 0)
  71 #define DN_KEY_LEQ(a,b)    ((int64_t)((a)-(b)) <= 0)
  72 #define DN_KEY_GT(a,b)     ((int64_t)((a)-(b)) > 0)
  73 #define DN_KEY_GEQ(a,b)    ((int64_t)((a)-(b)) >= 0)
  74 #define MAX64(x,y)  (( (int64_t) ( (y)-(x) )) > 0 ) ? (y) : (x)
  75 #define MY_M    16 /* number of left shift to obtain a larger precision */
  76
  77 /*
  78  * XXX With this scaling, max 1000 flows, max weight 100, 1Gbit/s, the
  79  * virtual time wraps every 15 days.
  80  */
  81
  82 /*
  83  * The OFFSET_OF macro is used to return the offset of a field within
  84  * a structure. It is used by the heap management routines.
  85  */
  86 #define OFFSET_OF(type, field) ((int)&( ((type *)0)->field) )
  87
  88 /*
  89  * A heap entry is made of a key and a pointer to the actual
  90  * object stored in the heap.
  91  * The heap is an array of dn_heap_entry entries, dynamically allocated.
  92  * Current size is "size", with "elements" actually in use.
  93  * The heap normally supports only ordered insert and extract from the top.
  94  * If we want to extract an object from the middle of the heap, we
  95  * have to know where the object itself is located in the heap (or we
  96  * need to scan the whole array). To this purpose, an object has a
  97  * field (int) which contains the index of the object itself into the
  98  * heap. When the object is moved, the field must also be updated.
  99  * The offset of the index in the object is stored in the 'offset'
 100  * field in the heap descriptor. The assumption is that this offset
 101  * is non-zero if we want to support extract from the middle.
 102  */
 103 struct dn_heap_entry {
 104     dn_key key ;        /* sorting key. Topmost element is smallest one */
 105     void *object ;      /* object pointer */
 106 } ;
 107
 108 struct dn_heap {
 109     int size ;
 110     int elements ;
 111     int offset ; /* XXX if > 0 this is the offset of direct ptr to obj */
 112     struct dn_heap_entry *p ;   /* really an array of "size" entries */
 113 } ;
 114
 115 /*
 116  * MT_DUMMYNET is a new (fake) mbuf type that is prepended to the
 117  * packet when it comes out of a pipe. The definition
 118  * ought to go in /sys/sys/mbuf.h but here it is less intrusive.
 119  */
 120
 121 #define MT_DUMMYNET MT_CONTROL
 122
 123 /*
 124  * struct dn_pkt identifies a packet in the dummynet queue. The
 125  * first part is really an m_hdr for implementation purposes, and some
 126  * fields are saved there. When passing the packet back to the ip_input/
 127  * ip_output()/bdg_forward, the struct is prepended to the mbuf chain with type
 128  * MT_DUMMYNET, and contains the pointer to the matching rule.
 129  *
 130  * Note: there is no real need to make this structure contain an m_hdr,
 131  * in the future this should be changed to a normal data structure.
 132  */
 133 struct dn_pkt {
 134         struct m_hdr hdr ;
 135 #define dn_next hdr.mh_nextpkt  /* next element in queue */
 136 #define DN_NEXT(x)      (struct dn_pkt *)(x)->dn_next
 137 #define dn_m    hdr.mh_next     /* packet to be forwarded */
 138 #define dn_dir  hdr.mh_flags    /* action when pkt extracted from a queue */
 139 #define DN_TO_IP_OUT    1
 140 #define DN_TO_IP_IN     2
 141 #define DN_TO_BDG_FWD   3
 142
 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 ?) */
 148 };
 149
 150 /*
 151  * Overall structure of dummynet (with WF2Q+):
 152
 153 In dummynet, packets are selected with the firewall rules, and passed
 154 to two different objects: PIPE or QUEUE.
 155
 156 A QUEUE is just a queue with configurable size and queue management
 157 policy. It is also associated with a mask (to discriminate among
 158 different flows), a weight (used to give different shares of the
 159 bandwidth to different flows) and a "pipe", which essentially
 160 supplies the transmit clock for all queues associated with that
 161 pipe.
 162
 163 A PIPE emulates a fixed-bandwidth link, whose bandwidth is
 164 configurable.  The "clock" for a pipe can come from either an
 165 internal timer, or from the transmit interrupt of an interface.
 166 A pipe is also associated with one (or more, if masks are used)
 167 queue, where all packets for that pipe are stored.
 168
 169 The bandwidth available on the pipe is shared by the queues
 170 associated with that pipe (only one in case the packet is sent
 171 to a PIPE) according to the WF2Q+ scheduling algorithm and the
 172 configured weights.
 173
 174 In general, incoming packets are stored in the appropriate queue,
 175 which is then placed into one of a few heaps managed by a scheduler
 176 to decide when the packet should be extracted.
 177 The scheduler (a function called dummynet()) is run at every timer
 178 tick, and grabs queues from the head of the heaps when they are
 179 ready for processing.
 180
 181 There are three data structures definining a pipe and associated queues:
 182
 183  + dn_pipe, which contains the main configuration parameters related
 184    to delay and bandwidth;
 185  + dn_flow_set, which contains WF2Q+ configuration, flow
 186    masks, plr and RED configuration;
 187  + dn_flow_queue, which is the per-flow queue (containing the packets)
 188
 189 Multiple dn_flow_set can be linked to the same pipe, and multiple
 190 dn_flow_queue can be linked to the same dn_flow_set.
 191 All data structures are linked in a linear list which is used for
 192 housekeeping purposes.
 193
 194 During configuration, we create and initialize the dn_flow_set
 195 and dn_pipe structures (a dn_pipe also contains a dn_flow_set).
 196
 197 At runtime: packets are sent to the appropriate dn_flow_set (either
 198 WFQ ones, or the one embedded in the dn_pipe for fixed-rate flows),
 199 which in turn dispatches them to the appropriate dn_flow_queue
 200 (created dynamically according to the masks).
 201
 202 The transmit clock for fixed rate flows (ready_event()) selects the
 203 dn_flow_queue to be used to transmit the next packet. For WF2Q,
 204 wfq_ready_event() extract a pipe which in turn selects the right
 205 flow using a number of heaps defined into the pipe itself.
 206
 207  *
 208  */
 209
 210 /*
 211  * per flow queue. This contains the flow identifier, the queue
 212  * of packets, counters, and parameters used to support both RED and
 213  * WF2Q+.
 214  */
 215 struct dn_flow_queue {
 216     struct dn_flow_queue *next ;
 217     struct ipfw_flow_id id ;
 218     struct dn_pkt *head, *tail ;        /* queue of packets */
 219     u_int len ;
 220     u_int len_bytes ;
 221     long numbytes ;             /* credit for transmission (dynamic queues) */
 222
 223     u_int64_t tot_pkts ;        /* statistics counters  */
 224     u_int64_t tot_bytes ;
 225     u_int32_t drops ;
 226     int hash_slot ;     /* debugging/diagnostic */
 227
 228     /* RED parameters */
 229     int avg ;                   /* average queue length est. (scaled) */
 230     int count ;                 /* arrivals since last RED drop */
 231     int random ;                /* random value (scaled) */
 232     u_int32_t q_time ;          /* start of queue idle time */
 233
 234     /* WF2Q+ support */
 235     struct dn_flow_set *fs ; /* parent flow set */
 236     int heap_pos ;      /* position (index) of struct in heap */
 237     dn_key sched_time ; /* current time when queue enters ready_heap */
 238
 239     dn_key S,F ; /* start-time, finishing time */
 240     /* setting F < S means the timestamp is invalid. We only need
 241      * to test this when the queue is empty.
 242      */
 243 } ;
 244
 245 /*
 246  * flow_set descriptor. Contains the "template" parameters for the
 247  * queue configuration, and pointers to the hash table of dn_flow_queue's.
 248  *
 249  * The hash table is an array of lists -- we identify the slot by
 250  * hashing the flow-id, then scan the list looking for a match.
 251  * The size of the hash table (buckets) is configurable on a per-queue
 252  * basis.
 253  */
 254 struct dn_flow_set {
 255     struct dn_flow_set *next; /* next flow set in all_flow_sets list */
 256
 257     u_short fs_nr ;             /* flow_set number       */
 258     u_short flags_fs;
 259 #define DN_HAVE_FLOW_MASK       0x0001
 260 #define DN_IS_PIPE              0x4000
 261 #define DN_IS_QUEUE             0x8000
 262 #define DN_IS_RED               0x0002
 263 #define DN_IS_GENTLE_RED        0x0004
 264 #define DN_QSIZE_IS_BYTES       0x0008  /* queue measured in bytes */
 265
 266     struct dn_pipe *pipe ;              /* pointer to parent pipe */
 267     u_short parent_nr ;         /* parent pipe#, 0 if local to a pipe */
 268
 269     int weight ; /* WFQ queue weight */
 270     int qsize ;         /* queue size in slots or bytes */
 271     int plr ;           /* pkt loss rate (2^31-1 means 100%) */
 272
 273     struct ipfw_flow_id flow_mask ;
 274     /* hash table of queues onto this flow_set */
 275     int rq_size ;               /* number of slots */
 276     int rq_elements ;           /* active elements */
 277     struct dn_flow_queue **rq;  /* array of rq_size entries */
 278     u_int32_t last_expired ;    /* do not expire too frequently */
 279         /* XXX some RED parameters as well ? */
 280     int backlogged ;            /* #active queues for this flowset */
 281
 282         /* RED parameters */
 283 #define SCALE_RED               16
 284 #define SCALE(x)                ( (x) << SCALE_RED )
 285 #define SCALE_VAL(x)            ( (x) >> SCALE_RED )
 286 #define SCALE_MUL(x,y)          ( ( (x) * (y) ) >> SCALE_RED )
 287     int w_q ;               /* queue weight (scaled) */
 288     int max_th ;            /* maximum threshold for queue (scaled) */
 289     int min_th ;            /* minimum threshold for queue (scaled) */
 290     int max_p ;             /* maximum value for p_b (scaled) */
 291     u_int c_1 ;             /* max_p/(max_th-min_th) (scaled) */
 292     u_int c_2 ;             /* max_p*min_th/(max_th-min_th) (scaled) */
 293     u_int c_3 ;             /* for GRED, (1-max_p)/max_th (scaled) */
 294     u_int c_4 ;             /* for GRED, 1 - 2*max_p (scaled) */
 295     u_int * w_q_lookup ;    /* lookup table for computing (1-w_q)^t */
 296     u_int lookup_depth ;    /* depth of lookup table */
 297     int lookup_step ;       /* granularity inside the lookup table */
 298     int lookup_weight ;     /* equal to (1-w_q)^t / (1-w_q)^(t+1) */
 299     int avg_pkt_size ;      /* medium packet size */
 300     int max_pkt_size ;      /* max packet size */
 301 } ;
 302
 303 /*
 304  * Pipe descriptor. Contains global parameters, delay-line queue,
 305  * and the flow_set used for fixed-rate queues.
 306  *
 307  * For WF2Q support it also has 4 heaps holding dn_flow_queue:
 308  *   not_eligible_heap, for queues whose start time is higher
 309  *      than the virtual time. Sorted by start time.
 310  *   scheduler_heap, for queues eligible for scheduling. Sorted by
 311  *      finish time.
 312  *   backlogged_heap, all flows in the two heaps above, sorted by
 313  *      start time. This is used to compute the virtual time.
 314  *   idle_heap, all flows that are idle and can be removed. We
 315  *      do that on each tick so we do not slow down too much
 316  *      operations during forwarding.
 317  *
 318  */
 319 struct dn_pipe {                        /* a pipe */
 320         struct dn_pipe *next ;
 321
 322     int pipe_nr ;               /* number       */
 323         int     bandwidth;              /* really, bytes/tick.  */
 324         int     delay ;                 /* really, ticks        */
 325
 326     struct      dn_pkt *head, *tail ;   /* packets in delay line */
 327
 328     /* WF2Q+ */
 329     struct dn_heap scheduler_heap ; /* top extract - key Finish time*/
 330     struct dn_heap not_eligible_heap; /* top extract- key Start time */
 331     struct dn_heap idle_heap ; /* random extract - key Start=Finish time */
 332
 333     dn_key V ; /* virtual time */
 334     int sum;    /* sum of weights of all active sessions */
 335     int numbytes;       /* bit i can transmit (more or less). */
 336
 337     dn_key sched_time ; /* first time pipe is scheduled in ready_heap */
 338
 339     /* the tx clock can come from an interface. In this case, the
 340      * name is below, and the pointer is filled when the rule is
 341      * configured. We identify this by setting the if_name to a
 342      * non-empty string.
 343      */
 344     char if_name[16];
 345     struct ifnet *ifp ;
 346     int ready ; /* set if ifp != NULL and we got a signal from it */
 347
 348     struct dn_flow_set fs ; /* used with fixed-rate flows */
 349 };
 350
 351 #ifdef KERNEL
 352
 353 MALLOC_DECLARE(M_IPFW);
 354
 355 typedef int ip_dn_ctl_t __P((struct sockopt *)) ;
 356 extern ip_dn_ctl_t *ip_dn_ctl_ptr;
 357
 358 void dn_rule_delete(void *r);           /* used in ip_fw.c */
 359 int dummynet_io(int pipe, int dir,
 360         struct mbuf *m, struct ifnet *ifp, struct route *ro,
 361         struct sockaddr_in * dst,
 362         struct ip_fw_chain *rule, int flags);
 363 #endif
 364
 365 #endif /* __APPLE_API_PRIVATE */
 366 #endif /* _IP_DUMMYNET_H */