28 */ 29 30#define DEB(x) 31#define DDB(x) x 32 33/* 34 * This module implements IP dummynet, a bandwidth limiter/delay emulator 35 * used in conjunction with the ipfw package. 36 * Description of the data structures used is in ip_dummynet.h 37 * Here you mainly find the following blocks of code: 38 * + variable declarations; 39 * + heap management functions; 40 * + scheduler and dummynet functions; 41 * + configuration and initialization. 42 * 43 * NOTA BENE: critical sections are protected by splimp()/splx() 44 * pairs. One would think that splnet() is enough as for most of 45 * the netinet code, but it is not so because when used with 46 * bridging, dummynet is invoked at splimp(). 47 * 48 * Most important Changes: 49 * 50 * 011004: KLDable 51 * 010124: Fixed WF2Q behaviour 52 * 010122: Fixed spl protection. 53 * 000601: WF2Q support 54 * 000106: large rewrite, use heaps to handle very many pipes. 55 * 980513: initial release 56 * 57 * include files marked with XXX are probably not needed 58 */ 59 60#include <sys/param.h> 61#include <sys/systm.h> 62#include <sys/malloc.h> 63#include <sys/mbuf.h> 64#include <sys/kernel.h> 65#include <sys/module.h> 66#include <sys/proc.h> 67#include <sys/socket.h> 68#include <sys/socketvar.h> 69#include <sys/time.h> 70#include <sys/sysctl.h> 71#include <net/if.h> 72#include <net/route.h> 73#include <netinet/in.h> 74#include <netinet/in_systm.h> 75#include <netinet/in_var.h> 76#include <netinet/ip.h> 77#include <netinet/ip_fw.h> 78#include <netinet/ip_dummynet.h> 79#include <netinet/ip_var.h> 80 81#include <netinet/if_ether.h> /* for struct arpcom */ 82#include <net/bridge.h> 83 84/* 85 * We keep a private variable for the simulation time, but we could 86 * probably use an existing one ("softticks" in sys/kern/kern_timer.c) 87 */ 88static dn_key curr_time = 0 ; /* current simulation time */ 89 90static int dn_hash_size = 64 ; /* default hash size */ 91 92/* statistics on number of queue searches and search steps */ 93static int searches, search_steps ; 94static int pipe_expire = 1 ; /* expire queue if empty */ 95static int dn_max_ratio = 16 ; /* max queues/buckets ratio */ 96 97static int red_lookup_depth = 256; /* RED - default lookup table depth */ 98static int red_avg_pkt_size = 512; /* RED - default medium packet size */ 99static int red_max_pkt_size = 1500; /* RED - default max packet size */ 100 101/* 102 * Three heaps contain queues and pipes that the scheduler handles: 103 * 104 * ready_heap contains all dn_flow_queue related to fixed-rate pipes. 105 * 106 * wfq_ready_heap contains the pipes associated with WF2Q flows 107 * 108 * extract_heap contains pipes associated with delay lines. 109 * 110 */ 111 112MALLOC_DEFINE(M_DUMMYNET, "dummynet", "dummynet heap"); 113 114static struct dn_heap ready_heap, extract_heap, wfq_ready_heap ; 115 116static int heap_init(struct dn_heap *h, int size) ; 117static int heap_insert (struct dn_heap *h, dn_key key1, void *p); 118static void heap_extract(struct dn_heap *h, void *obj); 119 120static void transmit_event(struct dn_pipe *pipe); 121static void ready_event(struct dn_flow_queue *q); 122 123static struct dn_pipe *all_pipes = NULL ; /* list of all pipes */ 124static struct dn_flow_set *all_flow_sets = NULL ;/* list of all flow_sets */ 125 126static struct callout_handle dn_timeout; 127 128#ifdef SYSCTL_NODE 129SYSCTL_NODE(_net_inet_ip, OID_AUTO, dummynet, 130 CTLFLAG_RW, 0, "Dummynet"); 131SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, hash_size, 132 CTLFLAG_RW, &dn_hash_size, 0, "Default hash table size"); 133SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, curr_time, 134 CTLFLAG_RD, &curr_time, 0, "Current tick"); 135SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, ready_heap, 136 CTLFLAG_RD, &ready_heap.size, 0, "Size of ready heap"); 137SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, extract_heap, 138 CTLFLAG_RD, &extract_heap.size, 0, "Size of extract heap"); 139SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, searches, 140 CTLFLAG_RD, &searches, 0, "Number of queue searches"); 141SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, search_steps, 142 CTLFLAG_RD, &search_steps, 0, "Number of queue search steps"); 143SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, expire, 144 CTLFLAG_RW, &pipe_expire, 0, "Expire queue if empty"); 145SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, max_chain_len, 146 CTLFLAG_RW, &dn_max_ratio, 0, 147 "Max ratio between dynamic queues and buckets"); 148SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_lookup_depth, 149 CTLFLAG_RD, &red_lookup_depth, 0, "Depth of RED lookup table"); 150SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_avg_pkt_size, 151 CTLFLAG_RD, &red_avg_pkt_size, 0, "RED Medium packet size"); 152SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_max_pkt_size, 153 CTLFLAG_RD, &red_max_pkt_size, 0, "RED Max packet size"); 154#endif 155 156static int config_pipe(struct dn_pipe *p); 157static int ip_dn_ctl(struct sockopt *sopt); 158 159static void rt_unref(struct rtentry *); 160static void dummynet(void *); 161static void dummynet_flush(void); 162void dummynet_drain(void); 163static ip_dn_io_t dummynet_io; 164static void dn_rule_delete(void *); 165 166int if_tx_rdy(struct ifnet *ifp); 167 168static void 169rt_unref(struct rtentry *rt) 170{ 171 if (rt == NULL) 172 return ; 173 if (rt->rt_refcnt <= 0) 174 printf("-- warning, refcnt now %ld, decreasing\n", rt->rt_refcnt); 175 RTFREE(rt); 176} 177 178/* 179 * Heap management functions. 180 * 181 * In the heap, first node is element 0. Children of i are 2i+1 and 2i+2. 182 * Some macros help finding parent/children so we can optimize them. 183 * 184 * heap_init() is called to expand the heap when needed. 185 * Increment size in blocks of 16 entries. 186 * XXX failure to allocate a new element is a pretty bad failure 187 * as we basically stall a whole queue forever!! 188 * Returns 1 on error, 0 on success 189 */ 190#define HEAP_FATHER(x) ( ( (x) - 1 ) / 2 ) 191#define HEAP_LEFT(x) ( 2*(x) + 1 ) 192#define HEAP_IS_LEFT(x) ( (x) & 1 ) 193#define HEAP_RIGHT(x) ( 2*(x) + 2 ) 194#define HEAP_SWAP(a, b, buffer) { buffer = a ; a = b ; b = buffer ; } 195#define HEAP_INCREMENT 15 196 197static int 198heap_init(struct dn_heap *h, int new_size) 199{ 200 struct dn_heap_entry *p; 201 202 if (h->size >= new_size ) { 203 printf("heap_init, Bogus call, have %d want %d\n", 204 h->size, new_size); 205 return 0 ; 206 } 207 new_size = (new_size + HEAP_INCREMENT ) & ~HEAP_INCREMENT ; 208 p = malloc(new_size * sizeof(*p), M_DUMMYNET, M_DONTWAIT ); 209 if (p == NULL) { 210 printf(" heap_init, resize %d failed\n", new_size ); 211 return 1 ; /* error */ 212 } 213 if (h->size > 0) { 214 bcopy(h->p, p, h->size * sizeof(*p) ); 215 free(h->p, M_DUMMYNET); 216 } 217 h->p = p ; 218 h->size = new_size ; 219 return 0 ; 220} 221 222/* 223 * Insert element in heap. Normally, p != NULL, we insert p in 224 * a new position and bubble up. If p == NULL, then the element is 225 * already in place, and key is the position where to start the 226 * bubble-up. 227 * Returns 1 on failure (cannot allocate new heap entry) 228 * 229 * If offset > 0 the position (index, int) of the element in the heap is 230 * also stored in the element itself at the given offset in bytes. 231 */ 232#define SET_OFFSET(heap, node) \ 233 if (heap->offset > 0) \ 234 *((int *)((char *)(heap->p[node].object) + heap->offset)) = node ; 235/* 236 * RESET_OFFSET is used for sanity checks. It sets offset to an invalid value. 237 */ 238#define RESET_OFFSET(heap, node) \ 239 if (heap->offset > 0) \ 240 *((int *)((char *)(heap->p[node].object) + heap->offset)) = -1 ; 241static int 242heap_insert(struct dn_heap *h, dn_key key1, void *p) 243{ 244 int son = h->elements ; 245 246 if (p == NULL) /* data already there, set starting point */ 247 son = key1 ; 248 else { /* insert new element at the end, possibly resize */ 249 son = h->elements ; 250 if (son == h->size) /* need resize... */ 251 if (heap_init(h, h->elements+1) ) 252 return 1 ; /* failure... */ 253 h->p[son].object = p ; 254 h->p[son].key = key1 ; 255 h->elements++ ; 256 } 257 while (son > 0) { /* bubble up */ 258 int father = HEAP_FATHER(son) ; 259 struct dn_heap_entry tmp ; 260 261 if (DN_KEY_LT( h->p[father].key, h->p[son].key ) ) 262 break ; /* found right position */ 263 /* son smaller than father, swap and repeat */ 264 HEAP_SWAP(h->p[son], h->p[father], tmp) ; 265 SET_OFFSET(h, son); 266 son = father ; 267 } 268 SET_OFFSET(h, son); 269 return 0 ; 270} 271 272/* 273 * remove top element from heap, or obj if obj != NULL 274 */ 275static void 276heap_extract(struct dn_heap *h, void *obj) 277{ 278 int child, father, max = h->elements - 1 ; 279 280 if (max < 0) { 281 printf("warning, extract from empty heap 0x%p\n", h); 282 return ; 283 } 284 father = 0 ; /* default: move up smallest child */ 285 if (obj != NULL) { /* extract specific element, index is at offset */ 286 if (h->offset <= 0) 287 panic("*** heap_extract from middle not supported on this heap!!!\n"); 288 father = *((int *)((char *)obj + h->offset)) ; 289 if (father < 0 || father >= h->elements) { 290 printf("dummynet: heap_extract, father %d out of bound 0..%d\n", 291 father, h->elements); 292 panic("heap_extract"); 293 } 294 } 295 RESET_OFFSET(h, father); 296 child = HEAP_LEFT(father) ; /* left child */ 297 while (child <= max) { /* valid entry */ 298 if (child != max && DN_KEY_LT(h->p[child+1].key, h->p[child].key) ) 299 child = child+1 ; /* take right child, otherwise left */ 300 h->p[father] = h->p[child] ; 301 SET_OFFSET(h, father); 302 father = child ; 303 child = HEAP_LEFT(child) ; /* left child for next loop */ 304 } 305 h->elements-- ; 306 if (father != max) { 307 /* 308 * Fill hole with last entry and bubble up, reusing the insert code 309 */ 310 h->p[father] = h->p[max] ; 311 heap_insert(h, father, NULL); /* this one cannot fail */ 312 } 313} 314 315#if 0 316/* 317 * change object position and update references 318 * XXX this one is never used! 319 */ 320static void 321heap_move(struct dn_heap *h, dn_key new_key, void *object) 322{ 323 int temp; 324 int i ; 325 int max = h->elements-1 ; 326 struct dn_heap_entry buf ; 327 328 if (h->offset <= 0) 329 panic("cannot move items on this heap"); 330 331 i = *((int *)((char *)object + h->offset)); 332 if (DN_KEY_LT(new_key, h->p[i].key) ) { /* must move up */ 333 h->p[i].key = new_key ; 334 for (; i>0 && DN_KEY_LT(new_key, h->p[(temp = HEAP_FATHER(i))].key) ; 335 i = temp ) { /* bubble up */ 336 HEAP_SWAP(h->p[i], h->p[temp], buf) ; 337 SET_OFFSET(h, i); 338 } 339 } else { /* must move down */ 340 h->p[i].key = new_key ; 341 while ( (temp = HEAP_LEFT(i)) <= max ) { /* found left child */ 342 if ((temp != max) && DN_KEY_GT(h->p[temp].key, h->p[temp+1].key)) 343 temp++ ; /* select child with min key */ 344 if (DN_KEY_GT(new_key, h->p[temp].key)) { /* go down */ 345 HEAP_SWAP(h->p[i], h->p[temp], buf) ; 346 SET_OFFSET(h, i); 347 } else 348 break ; 349 i = temp ; 350 } 351 } 352 SET_OFFSET(h, i); 353} 354#endif /* heap_move, unused */ 355 356/* 357 * heapify() will reorganize data inside an array to maintain the 358 * heap property. It is needed when we delete a bunch of entries. 359 */ 360static void 361heapify(struct dn_heap *h) 362{ 363 int i ; 364 365 for (i = 0 ; i < h->elements ; i++ ) 366 heap_insert(h, i , NULL) ; 367} 368 369/* 370 * cleanup the heap and free data structure 371 */ 372static void 373heap_free(struct dn_heap *h) 374{ 375 if (h->size >0 ) 376 free(h->p, M_DUMMYNET); 377 bzero(h, sizeof(*h) ); 378} 379 380/* 381 * --- end of heap management functions --- 382 */ 383 384/* 385 * Scheduler functions: 386 * 387 * transmit_event() is called when the delay-line needs to enter 388 * the scheduler, either because of existing pkts getting ready, 389 * or new packets entering the queue. The event handled is the delivery 390 * time of the packet. 391 * 392 * ready_event() does something similar with fixed-rate queues, and the 393 * event handled is the finish time of the head pkt. 394 * 395 * wfq_ready_event() does something similar with WF2Q queues, and the 396 * event handled is the start time of the head pkt. 397 * 398 * In all cases, we make sure that the data structures are consistent 399 * before passing pkts out, because this might trigger recursive 400 * invocations of the procedures. 401 */ 402static void 403transmit_event(struct dn_pipe *pipe) 404{ 405 struct dn_pkt *pkt ; 406 407 while ( (pkt = pipe->head) && DN_KEY_LEQ(pkt->output_time, curr_time) ) { 408 /* 409 * first unlink, then call procedures, since ip_input() can invoke 410 * ip_output() and viceversa, thus causing nested calls 411 */ 412 pipe->head = DN_NEXT(pkt) ; 413 414 /* 415 * The actual mbuf is preceded by a struct dn_pkt, resembling an mbuf 416 * (NOT A REAL one, just a small block of malloc'ed memory) with 417 * m_type = MT_TAG, m_flags = PACKET_TAG_DUMMYNET 418 * dn_m (m_next) = actual mbuf to be processed by ip_input/output 419 * and some other fields. 420 * The block IS FREED HERE because it contains parameters passed 421 * to the called routine. 422 */ 423 switch (pkt->dn_dir) { 424 case DN_TO_IP_OUT:
| 28 */ 29 30#define DEB(x) 31#define DDB(x) x 32 33/* 34 * This module implements IP dummynet, a bandwidth limiter/delay emulator 35 * used in conjunction with the ipfw package. 36 * Description of the data structures used is in ip_dummynet.h 37 * Here you mainly find the following blocks of code: 38 * + variable declarations; 39 * + heap management functions; 40 * + scheduler and dummynet functions; 41 * + configuration and initialization. 42 * 43 * NOTA BENE: critical sections are protected by splimp()/splx() 44 * pairs. One would think that splnet() is enough as for most of 45 * the netinet code, but it is not so because when used with 46 * bridging, dummynet is invoked at splimp(). 47 * 48 * Most important Changes: 49 * 50 * 011004: KLDable 51 * 010124: Fixed WF2Q behaviour 52 * 010122: Fixed spl protection. 53 * 000601: WF2Q support 54 * 000106: large rewrite, use heaps to handle very many pipes. 55 * 980513: initial release 56 * 57 * include files marked with XXX are probably not needed 58 */ 59 60#include <sys/param.h> 61#include <sys/systm.h> 62#include <sys/malloc.h> 63#include <sys/mbuf.h> 64#include <sys/kernel.h> 65#include <sys/module.h> 66#include <sys/proc.h> 67#include <sys/socket.h> 68#include <sys/socketvar.h> 69#include <sys/time.h> 70#include <sys/sysctl.h> 71#include <net/if.h> 72#include <net/route.h> 73#include <netinet/in.h> 74#include <netinet/in_systm.h> 75#include <netinet/in_var.h> 76#include <netinet/ip.h> 77#include <netinet/ip_fw.h> 78#include <netinet/ip_dummynet.h> 79#include <netinet/ip_var.h> 80 81#include <netinet/if_ether.h> /* for struct arpcom */ 82#include <net/bridge.h> 83 84/* 85 * We keep a private variable for the simulation time, but we could 86 * probably use an existing one ("softticks" in sys/kern/kern_timer.c) 87 */ 88static dn_key curr_time = 0 ; /* current simulation time */ 89 90static int dn_hash_size = 64 ; /* default hash size */ 91 92/* statistics on number of queue searches and search steps */ 93static int searches, search_steps ; 94static int pipe_expire = 1 ; /* expire queue if empty */ 95static int dn_max_ratio = 16 ; /* max queues/buckets ratio */ 96 97static int red_lookup_depth = 256; /* RED - default lookup table depth */ 98static int red_avg_pkt_size = 512; /* RED - default medium packet size */ 99static int red_max_pkt_size = 1500; /* RED - default max packet size */ 100 101/* 102 * Three heaps contain queues and pipes that the scheduler handles: 103 * 104 * ready_heap contains all dn_flow_queue related to fixed-rate pipes. 105 * 106 * wfq_ready_heap contains the pipes associated with WF2Q flows 107 * 108 * extract_heap contains pipes associated with delay lines. 109 * 110 */ 111 112MALLOC_DEFINE(M_DUMMYNET, "dummynet", "dummynet heap"); 113 114static struct dn_heap ready_heap, extract_heap, wfq_ready_heap ; 115 116static int heap_init(struct dn_heap *h, int size) ; 117static int heap_insert (struct dn_heap *h, dn_key key1, void *p); 118static void heap_extract(struct dn_heap *h, void *obj); 119 120static void transmit_event(struct dn_pipe *pipe); 121static void ready_event(struct dn_flow_queue *q); 122 123static struct dn_pipe *all_pipes = NULL ; /* list of all pipes */ 124static struct dn_flow_set *all_flow_sets = NULL ;/* list of all flow_sets */ 125 126static struct callout_handle dn_timeout; 127 128#ifdef SYSCTL_NODE 129SYSCTL_NODE(_net_inet_ip, OID_AUTO, dummynet, 130 CTLFLAG_RW, 0, "Dummynet"); 131SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, hash_size, 132 CTLFLAG_RW, &dn_hash_size, 0, "Default hash table size"); 133SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, curr_time, 134 CTLFLAG_RD, &curr_time, 0, "Current tick"); 135SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, ready_heap, 136 CTLFLAG_RD, &ready_heap.size, 0, "Size of ready heap"); 137SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, extract_heap, 138 CTLFLAG_RD, &extract_heap.size, 0, "Size of extract heap"); 139SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, searches, 140 CTLFLAG_RD, &searches, 0, "Number of queue searches"); 141SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, search_steps, 142 CTLFLAG_RD, &search_steps, 0, "Number of queue search steps"); 143SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, expire, 144 CTLFLAG_RW, &pipe_expire, 0, "Expire queue if empty"); 145SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, max_chain_len, 146 CTLFLAG_RW, &dn_max_ratio, 0, 147 "Max ratio between dynamic queues and buckets"); 148SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_lookup_depth, 149 CTLFLAG_RD, &red_lookup_depth, 0, "Depth of RED lookup table"); 150SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_avg_pkt_size, 151 CTLFLAG_RD, &red_avg_pkt_size, 0, "RED Medium packet size"); 152SYSCTL_INT(_net_inet_ip_dummynet, OID_AUTO, red_max_pkt_size, 153 CTLFLAG_RD, &red_max_pkt_size, 0, "RED Max packet size"); 154#endif 155 156static int config_pipe(struct dn_pipe *p); 157static int ip_dn_ctl(struct sockopt *sopt); 158 159static void rt_unref(struct rtentry *); 160static void dummynet(void *); 161static void dummynet_flush(void); 162void dummynet_drain(void); 163static ip_dn_io_t dummynet_io; 164static void dn_rule_delete(void *); 165 166int if_tx_rdy(struct ifnet *ifp); 167 168static void 169rt_unref(struct rtentry *rt) 170{ 171 if (rt == NULL) 172 return ; 173 if (rt->rt_refcnt <= 0) 174 printf("-- warning, refcnt now %ld, decreasing\n", rt->rt_refcnt); 175 RTFREE(rt); 176} 177 178/* 179 * Heap management functions. 180 * 181 * In the heap, first node is element 0. Children of i are 2i+1 and 2i+2. 182 * Some macros help finding parent/children so we can optimize them. 183 * 184 * heap_init() is called to expand the heap when needed. 185 * Increment size in blocks of 16 entries. 186 * XXX failure to allocate a new element is a pretty bad failure 187 * as we basically stall a whole queue forever!! 188 * Returns 1 on error, 0 on success 189 */ 190#define HEAP_FATHER(x) ( ( (x) - 1 ) / 2 ) 191#define HEAP_LEFT(x) ( 2*(x) + 1 ) 192#define HEAP_IS_LEFT(x) ( (x) & 1 ) 193#define HEAP_RIGHT(x) ( 2*(x) + 2 ) 194#define HEAP_SWAP(a, b, buffer) { buffer = a ; a = b ; b = buffer ; } 195#define HEAP_INCREMENT 15 196 197static int 198heap_init(struct dn_heap *h, int new_size) 199{ 200 struct dn_heap_entry *p; 201 202 if (h->size >= new_size ) { 203 printf("heap_init, Bogus call, have %d want %d\n", 204 h->size, new_size); 205 return 0 ; 206 } 207 new_size = (new_size + HEAP_INCREMENT ) & ~HEAP_INCREMENT ; 208 p = malloc(new_size * sizeof(*p), M_DUMMYNET, M_DONTWAIT ); 209 if (p == NULL) { 210 printf(" heap_init, resize %d failed\n", new_size ); 211 return 1 ; /* error */ 212 } 213 if (h->size > 0) { 214 bcopy(h->p, p, h->size * sizeof(*p) ); 215 free(h->p, M_DUMMYNET); 216 } 217 h->p = p ; 218 h->size = new_size ; 219 return 0 ; 220} 221 222/* 223 * Insert element in heap. Normally, p != NULL, we insert p in 224 * a new position and bubble up. If p == NULL, then the element is 225 * already in place, and key is the position where to start the 226 * bubble-up. 227 * Returns 1 on failure (cannot allocate new heap entry) 228 * 229 * If offset > 0 the position (index, int) of the element in the heap is 230 * also stored in the element itself at the given offset in bytes. 231 */ 232#define SET_OFFSET(heap, node) \ 233 if (heap->offset > 0) \ 234 *((int *)((char *)(heap->p[node].object) + heap->offset)) = node ; 235/* 236 * RESET_OFFSET is used for sanity checks. It sets offset to an invalid value. 237 */ 238#define RESET_OFFSET(heap, node) \ 239 if (heap->offset > 0) \ 240 *((int *)((char *)(heap->p[node].object) + heap->offset)) = -1 ; 241static int 242heap_insert(struct dn_heap *h, dn_key key1, void *p) 243{ 244 int son = h->elements ; 245 246 if (p == NULL) /* data already there, set starting point */ 247 son = key1 ; 248 else { /* insert new element at the end, possibly resize */ 249 son = h->elements ; 250 if (son == h->size) /* need resize... */ 251 if (heap_init(h, h->elements+1) ) 252 return 1 ; /* failure... */ 253 h->p[son].object = p ; 254 h->p[son].key = key1 ; 255 h->elements++ ; 256 } 257 while (son > 0) { /* bubble up */ 258 int father = HEAP_FATHER(son) ; 259 struct dn_heap_entry tmp ; 260 261 if (DN_KEY_LT( h->p[father].key, h->p[son].key ) ) 262 break ; /* found right position */ 263 /* son smaller than father, swap and repeat */ 264 HEAP_SWAP(h->p[son], h->p[father], tmp) ; 265 SET_OFFSET(h, son); 266 son = father ; 267 } 268 SET_OFFSET(h, son); 269 return 0 ; 270} 271 272/* 273 * remove top element from heap, or obj if obj != NULL 274 */ 275static void 276heap_extract(struct dn_heap *h, void *obj) 277{ 278 int child, father, max = h->elements - 1 ; 279 280 if (max < 0) { 281 printf("warning, extract from empty heap 0x%p\n", h); 282 return ; 283 } 284 father = 0 ; /* default: move up smallest child */ 285 if (obj != NULL) { /* extract specific element, index is at offset */ 286 if (h->offset <= 0) 287 panic("*** heap_extract from middle not supported on this heap!!!\n"); 288 father = *((int *)((char *)obj + h->offset)) ; 289 if (father < 0 || father >= h->elements) { 290 printf("dummynet: heap_extract, father %d out of bound 0..%d\n", 291 father, h->elements); 292 panic("heap_extract"); 293 } 294 } 295 RESET_OFFSET(h, father); 296 child = HEAP_LEFT(father) ; /* left child */ 297 while (child <= max) { /* valid entry */ 298 if (child != max && DN_KEY_LT(h->p[child+1].key, h->p[child].key) ) 299 child = child+1 ; /* take right child, otherwise left */ 300 h->p[father] = h->p[child] ; 301 SET_OFFSET(h, father); 302 father = child ; 303 child = HEAP_LEFT(child) ; /* left child for next loop */ 304 } 305 h->elements-- ; 306 if (father != max) { 307 /* 308 * Fill hole with last entry and bubble up, reusing the insert code 309 */ 310 h->p[father] = h->p[max] ; 311 heap_insert(h, father, NULL); /* this one cannot fail */ 312 } 313} 314 315#if 0 316/* 317 * change object position and update references 318 * XXX this one is never used! 319 */ 320static void 321heap_move(struct dn_heap *h, dn_key new_key, void *object) 322{ 323 int temp; 324 int i ; 325 int max = h->elements-1 ; 326 struct dn_heap_entry buf ; 327 328 if (h->offset <= 0) 329 panic("cannot move items on this heap"); 330 331 i = *((int *)((char *)object + h->offset)); 332 if (DN_KEY_LT(new_key, h->p[i].key) ) { /* must move up */ 333 h->p[i].key = new_key ; 334 for (; i>0 && DN_KEY_LT(new_key, h->p[(temp = HEAP_FATHER(i))].key) ; 335 i = temp ) { /* bubble up */ 336 HEAP_SWAP(h->p[i], h->p[temp], buf) ; 337 SET_OFFSET(h, i); 338 } 339 } else { /* must move down */ 340 h->p[i].key = new_key ; 341 while ( (temp = HEAP_LEFT(i)) <= max ) { /* found left child */ 342 if ((temp != max) && DN_KEY_GT(h->p[temp].key, h->p[temp+1].key)) 343 temp++ ; /* select child with min key */ 344 if (DN_KEY_GT(new_key, h->p[temp].key)) { /* go down */ 345 HEAP_SWAP(h->p[i], h->p[temp], buf) ; 346 SET_OFFSET(h, i); 347 } else 348 break ; 349 i = temp ; 350 } 351 } 352 SET_OFFSET(h, i); 353} 354#endif /* heap_move, unused */ 355 356/* 357 * heapify() will reorganize data inside an array to maintain the 358 * heap property. It is needed when we delete a bunch of entries. 359 */ 360static void 361heapify(struct dn_heap *h) 362{ 363 int i ; 364 365 for (i = 0 ; i < h->elements ; i++ ) 366 heap_insert(h, i , NULL) ; 367} 368 369/* 370 * cleanup the heap and free data structure 371 */ 372static void 373heap_free(struct dn_heap *h) 374{ 375 if (h->size >0 ) 376 free(h->p, M_DUMMYNET); 377 bzero(h, sizeof(*h) ); 378} 379 380/* 381 * --- end of heap management functions --- 382 */ 383 384/* 385 * Scheduler functions: 386 * 387 * transmit_event() is called when the delay-line needs to enter 388 * the scheduler, either because of existing pkts getting ready, 389 * or new packets entering the queue. The event handled is the delivery 390 * time of the packet. 391 * 392 * ready_event() does something similar with fixed-rate queues, and the 393 * event handled is the finish time of the head pkt. 394 * 395 * wfq_ready_event() does something similar with WF2Q queues, and the 396 * event handled is the start time of the head pkt. 397 * 398 * In all cases, we make sure that the data structures are consistent 399 * before passing pkts out, because this might trigger recursive 400 * invocations of the procedures. 401 */ 402static void 403transmit_event(struct dn_pipe *pipe) 404{ 405 struct dn_pkt *pkt ; 406 407 while ( (pkt = pipe->head) && DN_KEY_LEQ(pkt->output_time, curr_time) ) { 408 /* 409 * first unlink, then call procedures, since ip_input() can invoke 410 * ip_output() and viceversa, thus causing nested calls 411 */ 412 pipe->head = DN_NEXT(pkt) ; 413 414 /* 415 * The actual mbuf is preceded by a struct dn_pkt, resembling an mbuf 416 * (NOT A REAL one, just a small block of malloc'ed memory) with 417 * m_type = MT_TAG, m_flags = PACKET_TAG_DUMMYNET 418 * dn_m (m_next) = actual mbuf to be processed by ip_input/output 419 * and some other fields. 420 * The block IS FREED HERE because it contains parameters passed 421 * to the called routine. 422 */ 423 switch (pkt->dn_dir) { 424 case DN_TO_IP_OUT:
|
426 rt_unref (pkt->ro.ro_rt) ; 427 break ; 428 429 case DN_TO_IP_IN : 430 ip_input((struct mbuf *)pkt) ; 431 break ; 432 433 case DN_TO_BDG_FWD : 434 if (!BDG_LOADED) { 435 /* somebody unloaded the bridge module. Drop pkt */ 436 printf("-- dropping bridged packet trapped in pipe--\n"); 437 m_freem(pkt->dn_m); 438 break; 439 } /* fallthrough */ 440 case DN_TO_ETH_DEMUX: 441 { 442 struct mbuf *m = (struct mbuf *)pkt ; 443 struct ether_header *eh; 444 445 if (pkt->dn_m->m_len < ETHER_HDR_LEN && 446 (pkt->dn_m = m_pullup(pkt->dn_m, ETHER_HDR_LEN)) == NULL) { 447 printf("dummynet/bridge: pullup fail, dropping pkt\n"); 448 break; 449 } 450 /* 451 * same as ether_input, make eh be a pointer into the mbuf 452 */ 453 eh = mtod(pkt->dn_m, struct ether_header *); 454 m_adj(pkt->dn_m, ETHER_HDR_LEN); 455 /* 456 * bdg_forward() wants a pointer to the pseudo-mbuf-header, but 457 * on return it will supply the pointer to the actual packet 458 * (originally pkt->dn_m, but could be something else now) if 459 * it has not consumed it. 460 */ 461 if (pkt->dn_dir == DN_TO_BDG_FWD) { 462 m = bdg_forward_ptr(m, eh, pkt->ifp); 463 if (m) 464 m_freem(m); 465 } else 466 ether_demux(NULL, eh, m); /* which consumes the mbuf */ 467 } 468 break ; 469 case DN_TO_ETH_OUT: 470 ether_output_frame(pkt->ifp, (struct mbuf *)pkt); 471 break; 472 473 default: 474 printf("dummynet: bad switch %d!\n", pkt->dn_dir); 475 m_freem(pkt->dn_m); 476 break ; 477 } 478 free(pkt, M_DUMMYNET); 479 } 480 /* if there are leftover packets, put into the heap for next event */ 481 if ( (pkt = pipe->head) ) 482 heap_insert(&extract_heap, pkt->output_time, pipe ) ; 483 /* XXX should check errors on heap_insert, by draining the 484 * whole pipe p and hoping in the future we are more successful 485 */ 486} 487 488/* 489 * the following macro computes how many ticks we have to wait 490 * before being able to transmit a packet. The credit is taken from 491 * either a pipe (WF2Q) or a flow_queue (per-flow queueing) 492 */ 493#define SET_TICKS(pkt, q, p) \ 494 (pkt->dn_m->m_pkthdr.len*8*hz - (q)->numbytes + p->bandwidth - 1 ) / \ 495 p->bandwidth ; 496 497/* 498 * extract pkt from queue, compute output time (could be now) 499 * and put into delay line (p_queue) 500 */ 501static void 502move_pkt(struct dn_pkt *pkt, struct dn_flow_queue *q, 503 struct dn_pipe *p, int len) 504{ 505 q->head = DN_NEXT(pkt) ; 506 q->len-- ; 507 q->len_bytes -= len ; 508 509 pkt->output_time = curr_time + p->delay ; 510 511 if (p->head == NULL) 512 p->head = pkt; 513 else 514 DN_NEXT(p->tail) = pkt; 515 p->tail = pkt; 516 DN_NEXT(p->tail) = NULL; 517} 518 519/* 520 * ready_event() is invoked every time the queue must enter the 521 * scheduler, either because the first packet arrives, or because 522 * a previously scheduled event fired. 523 * On invokation, drain as many pkts as possible (could be 0) and then 524 * if there are leftover packets reinsert the pkt in the scheduler. 525 */ 526static void 527ready_event(struct dn_flow_queue *q) 528{ 529 struct dn_pkt *pkt; 530 struct dn_pipe *p = q->fs->pipe ; 531 int p_was_empty ; 532 533 if (p == NULL) { 534 printf("ready_event- pipe is gone\n"); 535 return ; 536 } 537 p_was_empty = (p->head == NULL) ; 538 539 /* 540 * schedule fixed-rate queues linked to this pipe: 541 * Account for the bw accumulated since last scheduling, then 542 * drain as many pkts as allowed by q->numbytes and move to 543 * the delay line (in p) computing output time. 544 * bandwidth==0 (no limit) means we can drain the whole queue, 545 * setting len_scaled = 0 does the job. 546 */ 547 q->numbytes += ( curr_time - q->sched_time ) * p->bandwidth; 548 while ( (pkt = q->head) != NULL ) { 549 int len = pkt->dn_m->m_pkthdr.len; 550 int len_scaled = p->bandwidth ? len*8*hz : 0 ; 551 if (len_scaled > q->numbytes ) 552 break ; 553 q->numbytes -= len_scaled ; 554 move_pkt(pkt, q, p, len); 555 } 556 /* 557 * If we have more packets queued, schedule next ready event 558 * (can only occur when bandwidth != 0, otherwise we would have 559 * flushed the whole queue in the previous loop). 560 * To this purpose we record the current time and compute how many 561 * ticks to go for the finish time of the packet. 562 */ 563 if ( (pkt = q->head) != NULL ) { /* this implies bandwidth != 0 */ 564 dn_key t = SET_TICKS(pkt, q, p); /* ticks i have to wait */ 565 q->sched_time = curr_time ; 566 heap_insert(&ready_heap, curr_time + t, (void *)q ); 567 /* XXX should check errors on heap_insert, and drain the whole 568 * queue on error hoping next time we are luckier. 569 */ 570 } else /* RED needs to know when the queue becomes empty */ 571 q->q_time = curr_time; 572 /* 573 * If the delay line was empty call transmit_event(p) now. 574 * Otherwise, the scheduler will take care of it. 575 */ 576 if (p_was_empty) 577 transmit_event(p); 578} 579 580/* 581 * Called when we can transmit packets on WF2Q queues. Take pkts out of 582 * the queues at their start time, and enqueue into the delay line. 583 * Packets are drained until p->numbytes < 0. As long as 584 * len_scaled >= p->numbytes, the packet goes into the delay line 585 * with a deadline p->delay. For the last packet, if p->numbytes<0, 586 * there is an additional delay. 587 */ 588static void 589ready_event_wfq(struct dn_pipe *p) 590{ 591 int p_was_empty = (p->head == NULL) ; 592 struct dn_heap *sch = &(p->scheduler_heap); 593 struct dn_heap *neh = &(p->not_eligible_heap) ; 594 595 if (p->if_name[0] == 0) /* tx clock is simulated */ 596 p->numbytes += ( curr_time - p->sched_time ) * p->bandwidth; 597 else { /* tx clock is for real, the ifq must be empty or this is a NOP */ 598 if (p->ifp && p->ifp->if_snd.ifq_head != NULL) 599 return ; 600 else { 601 DEB(printf("pipe %d ready from %s --\n", 602 p->pipe_nr, p->if_name);) 603 } 604 } 605 606 /* 607 * While we have backlogged traffic AND credit, we need to do 608 * something on the queue. 609 */ 610 while ( p->numbytes >=0 && (sch->elements>0 || neh->elements >0) ) { 611 if (sch->elements > 0) { /* have some eligible pkts to send out */ 612 struct dn_flow_queue *q = sch->p[0].object ; 613 struct dn_pkt *pkt = q->head; 614 struct dn_flow_set *fs = q->fs; 615 u_int64_t len = pkt->dn_m->m_pkthdr.len; 616 int len_scaled = p->bandwidth ? len*8*hz : 0 ; 617 618 heap_extract(sch, NULL); /* remove queue from heap */ 619 p->numbytes -= len_scaled ; 620 move_pkt(pkt, q, p, len); 621 622 p->V += (len<<MY_M) / p->sum ; /* update V */ 623 q->S = q->F ; /* update start time */ 624 if (q->len == 0) { /* Flow not backlogged any more */ 625 fs->backlogged-- ; 626 heap_insert(&(p->idle_heap), q->F, q); 627 } else { /* still backlogged */ 628 /* 629 * update F and position in backlogged queue, then 630 * put flow in not_eligible_heap (we will fix this later). 631 */ 632 len = (q->head)->dn_m->m_pkthdr.len; 633 q->F += (len<<MY_M)/(u_int64_t) fs->weight ; 634 if (DN_KEY_LEQ(q->S, p->V)) 635 heap_insert(neh, q->S, q); 636 else 637 heap_insert(sch, q->F, q); 638 } 639 } 640 /* 641 * now compute V = max(V, min(S_i)). Remember that all elements in sch 642 * have by definition S_i <= V so if sch is not empty, V is surely 643 * the max and we must not update it. Conversely, if sch is empty 644 * we only need to look at neh. 645 */ 646 if (sch->elements == 0 && neh->elements > 0) 647 p->V = MAX64 ( p->V, neh->p[0].key ); 648 /* move from neh to sch any packets that have become eligible */ 649 while (neh->elements > 0 && DN_KEY_LEQ(neh->p[0].key, p->V) ) { 650 struct dn_flow_queue *q = neh->p[0].object ; 651 heap_extract(neh, NULL); 652 heap_insert(sch, q->F, q); 653 } 654 655 if (p->if_name[0] != '\0') {/* tx clock is from a real thing */ 656 p->numbytes = -1 ; /* mark not ready for I/O */ 657 break ; 658 } 659 } 660 if (sch->elements == 0 && neh->elements == 0 && p->numbytes >= 0 661 && p->idle_heap.elements > 0) { 662 /* 663 * no traffic and no events scheduled. We can get rid of idle-heap. 664 */ 665 int i ; 666 667 for (i = 0 ; i < p->idle_heap.elements ; i++) { 668 struct dn_flow_queue *q = p->idle_heap.p[i].object ; 669 670 q->F = 0 ; 671 q->S = q->F + 1 ; 672 } 673 p->sum = 0 ; 674 p->V = 0 ; 675 p->idle_heap.elements = 0 ; 676 } 677 /* 678 * If we are getting clocks from dummynet (not a real interface) and 679 * If we are under credit, schedule the next ready event. 680 * Also fix the delivery time of the last packet. 681 */ 682 if (p->if_name[0]==0 && p->numbytes < 0) { /* this implies bandwidth >0 */ 683 dn_key t=0 ; /* number of ticks i have to wait */ 684 685 if (p->bandwidth > 0) 686 t = ( p->bandwidth -1 - p->numbytes) / p->bandwidth ; 687 p->tail->output_time += t ; 688 p->sched_time = curr_time ; 689 heap_insert(&wfq_ready_heap, curr_time + t, (void *)p); 690 /* XXX should check errors on heap_insert, and drain the whole 691 * queue on error hoping next time we are luckier. 692 */ 693 } 694 /* 695 * If the delay line was empty call transmit_event(p) now. 696 * Otherwise, the scheduler will take care of it. 697 */ 698 if (p_was_empty) 699 transmit_event(p); 700} 701 702/* 703 * This is called once per tick, or HZ times per second. It is used to 704 * increment the current tick counter and schedule expired events. 705 */ 706static void 707dummynet(void * __unused unused) 708{ 709 void *p ; /* generic parameter to handler */ 710 struct dn_heap *h ; 711 int s ; 712 struct dn_heap *heaps[3]; 713 int i; 714 struct dn_pipe *pe ; 715 716 heaps[0] = &ready_heap ; /* fixed-rate queues */ 717 heaps[1] = &wfq_ready_heap ; /* wfq queues */ 718 heaps[2] = &extract_heap ; /* delay line */ 719 s = splimp(); /* see note on top, splnet() is not enough */ 720 curr_time++ ; 721 for (i=0; i < 3 ; i++) { 722 h = heaps[i]; 723 while (h->elements > 0 && DN_KEY_LEQ(h->p[0].key, curr_time) ) { 724 DDB(if (h->p[0].key > curr_time) 725 printf("-- dummynet: warning, heap %d is %d ticks late\n", 726 i, (int)(curr_time - h->p[0].key));) 727 p = h->p[0].object ; /* store a copy before heap_extract */ 728 heap_extract(h, NULL); /* need to extract before processing */ 729 if (i == 0) 730 ready_event(p) ; 731 else if (i == 1) { 732 struct dn_pipe *pipe = p; 733 if (pipe->if_name[0] != '\0') 734 printf("*** bad ready_event_wfq for pipe %s\n", 735 pipe->if_name); 736 else 737 ready_event_wfq(p) ; 738 } else 739 transmit_event(p); 740 } 741 } 742 /* sweep pipes trying to expire idle flow_queues */ 743 for (pe = all_pipes; pe ; pe = pe->next ) 744 if (pe->idle_heap.elements > 0 && 745 DN_KEY_LT(pe->idle_heap.p[0].key, pe->V) ) { 746 struct dn_flow_queue *q = pe->idle_heap.p[0].object ; 747 748 heap_extract(&(pe->idle_heap), NULL); 749 q->S = q->F + 1 ; /* mark timestamp as invalid */ 750 pe->sum -= q->fs->weight ; 751 } 752 splx(s); 753 dn_timeout = timeout(dummynet, NULL, 1); 754} 755 756/* 757 * called by an interface when tx_rdy occurs. 758 */ 759int 760if_tx_rdy(struct ifnet *ifp) 761{ 762 struct dn_pipe *p; 763 764 for (p = all_pipes; p ; p = p->next ) 765 if (p->ifp == ifp) 766 break ; 767 if (p == NULL) { 768 char buf[32]; 769 sprintf(buf, "%s%d",ifp->if_name, ifp->if_unit); 770 for (p = all_pipes; p ; p = p->next ) 771 if (!strcmp(p->if_name, buf) ) { 772 p->ifp = ifp ; 773 DEB(printf("++ tx rdy from %s (now found)\n", buf);) 774 break ; 775 } 776 } 777 if (p != NULL) { 778 DEB(printf("++ tx rdy from %s%d - qlen %d\n", ifp->if_name, 779 ifp->if_unit, ifp->if_snd.ifq_len);) 780 p->numbytes = 0 ; /* mark ready for I/O */ 781 ready_event_wfq(p); 782 } 783 return 0; 784} 785 786/* 787 * Unconditionally expire empty queues in case of shortage. 788 * Returns the number of queues freed. 789 */ 790static int 791expire_queues(struct dn_flow_set *fs) 792{ 793 struct dn_flow_queue *q, *prev ; 794 int i, initial_elements = fs->rq_elements ; 795 796 if (fs->last_expired == time_second) 797 return 0 ; 798 fs->last_expired = time_second ; 799 for (i = 0 ; i <= fs->rq_size ; i++) /* last one is overflow */ 800 for (prev=NULL, q = fs->rq[i] ; q != NULL ; ) 801 if (q->head != NULL || q->S != q->F+1) { 802 prev = q ; 803 q = q->next ; 804 } else { /* entry is idle, expire it */ 805 struct dn_flow_queue *old_q = q ; 806 807 if (prev != NULL) 808 prev->next = q = q->next ; 809 else 810 fs->rq[i] = q = q->next ; 811 fs->rq_elements-- ; 812 free(old_q, M_DUMMYNET); 813 } 814 return initial_elements - fs->rq_elements ; 815} 816 817/* 818 * If room, create a new queue and put at head of slot i; 819 * otherwise, create or use the default queue. 820 */ 821static struct dn_flow_queue * 822create_queue(struct dn_flow_set *fs, int i) 823{ 824 struct dn_flow_queue *q ; 825 826 if (fs->rq_elements > fs->rq_size * dn_max_ratio && 827 expire_queues(fs) == 0) { 828 /* 829 * No way to get room, use or create overflow queue. 830 */ 831 i = fs->rq_size ; 832 if ( fs->rq[i] != NULL ) 833 return fs->rq[i] ; 834 } 835 q = malloc(sizeof(*q), M_DUMMYNET, M_DONTWAIT | M_ZERO); 836 if (q == NULL) { 837 printf("sorry, cannot allocate queue for new flow\n"); 838 return NULL ; 839 } 840 q->fs = fs ; 841 q->hash_slot = i ; 842 q->next = fs->rq[i] ; 843 q->S = q->F + 1; /* hack - mark timestamp as invalid */ 844 fs->rq[i] = q ; 845 fs->rq_elements++ ; 846 return q ; 847} 848 849/* 850 * Given a flow_set and a pkt in last_pkt, find a matching queue 851 * after appropriate masking. The queue is moved to front 852 * so that further searches take less time. 853 */ 854static struct dn_flow_queue * 855find_queue(struct dn_flow_set *fs, struct ipfw_flow_id *id) 856{ 857 int i = 0 ; /* we need i and q for new allocations */ 858 struct dn_flow_queue *q, *prev; 859 860 if ( !(fs->flags_fs & DN_HAVE_FLOW_MASK) ) 861 q = fs->rq[0] ; 862 else { 863 /* first, do the masking */ 864 id->dst_ip &= fs->flow_mask.dst_ip ; 865 id->src_ip &= fs->flow_mask.src_ip ; 866 id->dst_port &= fs->flow_mask.dst_port ; 867 id->src_port &= fs->flow_mask.src_port ; 868 id->proto &= fs->flow_mask.proto ; 869 id->flags = 0 ; /* we don't care about this one */ 870 /* then, hash function */ 871 i = ( (id->dst_ip) & 0xffff ) ^ 872 ( (id->dst_ip >> 15) & 0xffff ) ^ 873 ( (id->src_ip << 1) & 0xffff ) ^ 874 ( (id->src_ip >> 16 ) & 0xffff ) ^ 875 (id->dst_port << 1) ^ (id->src_port) ^ 876 (id->proto ); 877 i = i % fs->rq_size ; 878 /* finally, scan the current list for a match */ 879 searches++ ; 880 for (prev=NULL, q = fs->rq[i] ; q ; ) { 881 search_steps++; 882 if (bcmp(id, &(q->id), sizeof(q->id) ) == 0) 883 break ; /* found */ 884 else if (pipe_expire && q->head == NULL && q->S == q->F+1 ) { 885 /* entry is idle and not in any heap, expire it */ 886 struct dn_flow_queue *old_q = q ; 887 888 if (prev != NULL) 889 prev->next = q = q->next ; 890 else 891 fs->rq[i] = q = q->next ; 892 fs->rq_elements-- ; 893 free(old_q, M_DUMMYNET); 894 continue ; 895 } 896 prev = q ; 897 q = q->next ; 898 } 899 if (q && prev != NULL) { /* found and not in front */ 900 prev->next = q->next ; 901 q->next = fs->rq[i] ; 902 fs->rq[i] = q ; 903 } 904 } 905 if (q == NULL) { /* no match, need to allocate a new entry */ 906 q = create_queue(fs, i); 907 if (q != NULL) 908 q->id = *id ; 909 } 910 return q ; 911} 912 913static int 914red_drops(struct dn_flow_set *fs, struct dn_flow_queue *q, int len) 915{ 916 /* 917 * RED algorithm 918 * 919 * RED calculates the average queue size (avg) using a low-pass filter 920 * with an exponential weighted (w_q) moving average: 921 * avg <- (1-w_q) * avg + w_q * q_size 922 * where q_size is the queue length (measured in bytes or * packets). 923 * 924 * If q_size == 0, we compute the idle time for the link, and set 925 * avg = (1 - w_q)^(idle/s) 926 * where s is the time needed for transmitting a medium-sized packet. 927 * 928 * Now, if avg < min_th the packet is enqueued. 929 * If avg > max_th the packet is dropped. Otherwise, the packet is 930 * dropped with probability P function of avg. 931 * 932 */ 933 934 int64_t p_b = 0; 935 /* queue in bytes or packets ? */ 936 u_int q_size = (fs->flags_fs & DN_QSIZE_IS_BYTES) ? q->len_bytes : q->len; 937 938 DEB(printf("\n%d q: %2u ", (int) curr_time, q_size);) 939 940 /* average queue size estimation */ 941 if (q_size != 0) { 942 /* 943 * queue is not empty, avg <- avg + (q_size - avg) * w_q 944 */ 945 int diff = SCALE(q_size) - q->avg; 946 int64_t v = SCALE_MUL((int64_t) diff, (int64_t) fs->w_q); 947 948 q->avg += (int) v; 949 } else { 950 /* 951 * queue is empty, find for how long the queue has been 952 * empty and use a lookup table for computing 953 * (1 - * w_q)^(idle_time/s) where s is the time to send a 954 * (small) packet. 955 * XXX check wraps... 956 */ 957 if (q->avg) { 958 u_int t = (curr_time - q->q_time) / fs->lookup_step; 959 960 q->avg = (t < fs->lookup_depth) ? 961 SCALE_MUL(q->avg, fs->w_q_lookup[t]) : 0; 962 } 963 } 964 DEB(printf("avg: %u ", SCALE_VAL(q->avg));) 965 966 /* should i drop ? */ 967 968 if (q->avg < fs->min_th) { 969 q->count = -1; 970 return 0; /* accept packet ; */ 971 } 972 if (q->avg >= fs->max_th) { /* average queue >= max threshold */ 973 if (fs->flags_fs & DN_IS_GENTLE_RED) { 974 /* 975 * According to Gentle-RED, if avg is greater than max_th the 976 * packet is dropped with a probability 977 * p_b = c_3 * avg - c_4 978 * where c_3 = (1 - max_p) / max_th, and c_4 = 1 - 2 * max_p 979 */ 980 p_b = SCALE_MUL((int64_t) fs->c_3, (int64_t) q->avg) - fs->c_4; 981 } else { 982 q->count = -1; 983 printf("- drop"); 984 return 1 ; 985 } 986 } else if (q->avg > fs->min_th) { 987 /* 988 * we compute p_b using the linear dropping function p_b = c_1 * 989 * avg - c_2, where c_1 = max_p / (max_th - min_th), and c_2 = 990 * max_p * min_th / (max_th - min_th) 991 */ 992 p_b = SCALE_MUL((int64_t) fs->c_1, (int64_t) q->avg) - fs->c_2; 993 } 994 if (fs->flags_fs & DN_QSIZE_IS_BYTES) 995 p_b = (p_b * len) / fs->max_pkt_size; 996 if (++q->count == 0) 997 q->random = random() & 0xffff; 998 else { 999 /* 1000 * q->count counts packets arrived since last drop, so a greater 1001 * value of q->count means a greater packet drop probability. 1002 */ 1003 if (SCALE_MUL(p_b, SCALE((int64_t) q->count)) > q->random) { 1004 q->count = 0; 1005 DEB(printf("- red drop");) 1006 /* after a drop we calculate a new random value */ 1007 q->random = random() & 0xffff; 1008 return 1; /* drop */ 1009 } 1010 } 1011 /* end of RED algorithm */ 1012 return 0 ; /* accept */ 1013} 1014 1015static __inline 1016struct dn_flow_set * 1017locate_flowset(int pipe_nr, struct ip_fw *rule) 1018{ 1019#if IPFW2 1020 struct dn_flow_set *fs; 1021 ipfw_insn *cmd = rule->cmd + rule->act_ofs; 1022 1023 if (cmd->opcode == O_LOG) 1024 cmd += F_LEN(cmd); 1025 fs = ((ipfw_insn_pipe *)cmd)->pipe_ptr; 1026 1027 if (fs != NULL) 1028 return fs; 1029 1030 if (cmd->opcode == O_QUEUE) 1031#else /* !IPFW2 */ 1032 struct dn_flow_set *fs = NULL ; 1033 1034 if ( (rule->fw_flg & IP_FW_F_COMMAND) == IP_FW_F_QUEUE ) 1035#endif /* !IPFW2 */ 1036 for (fs=all_flow_sets; fs && fs->fs_nr != pipe_nr; fs=fs->next) 1037 ; 1038 else { 1039 struct dn_pipe *p1; 1040 for (p1 = all_pipes; p1 && p1->pipe_nr != pipe_nr; p1 = p1->next) 1041 ; 1042 if (p1 != NULL) 1043 fs = &(p1->fs) ; 1044 } 1045 /* record for the future */ 1046#if IPFW2 1047 ((ipfw_insn_pipe *)cmd)->pipe_ptr = fs; 1048#else 1049 if (fs != NULL) 1050 rule->pipe_ptr = fs; 1051#endif 1052 return fs ; 1053} 1054 1055/* 1056 * dummynet hook for packets. Below 'pipe' is a pipe or a queue 1057 * depending on whether WF2Q or fixed bw is used. 1058 * 1059 * pipe_nr pipe or queue the packet is destined for. 1060 * dir where shall we send the packet after dummynet. 1061 * m the mbuf with the packet 1062 * ifp the 'ifp' parameter from the caller. 1063 * NULL in ip_input, destination interface in ip_output, 1064 * real_dst in bdg_forward 1065 * ro route parameter (only used in ip_output, NULL otherwise) 1066 * dst destination address, only used by ip_output 1067 * rule matching rule, in case of multiple passes 1068 * flags flags from the caller, only used in ip_output 1069 * 1070 */ 1071static int 1072dummynet_io(struct mbuf *m, int pipe_nr, int dir, struct ip_fw_args *fwa) 1073{ 1074 struct dn_pkt *pkt; 1075 struct dn_flow_set *fs; 1076 struct dn_pipe *pipe ; 1077 u_int64_t len = m->m_pkthdr.len ; 1078 struct dn_flow_queue *q = NULL ; 1079 int s = splimp(); 1080 int is_pipe; 1081#if IPFW2 1082 ipfw_insn *cmd = fwa->rule->cmd + fwa->rule->act_ofs; 1083 1084 if (cmd->opcode == O_LOG) 1085 cmd += F_LEN(cmd); 1086 is_pipe = (cmd->opcode == O_PIPE); 1087#else 1088 is_pipe = (fwa->rule->fw_flg & IP_FW_F_COMMAND) == IP_FW_F_PIPE; 1089#endif 1090 1091 pipe_nr &= 0xffff ; 1092 1093 /* 1094 * this is a dummynet rule, so we expect a O_PIPE or O_QUEUE rule 1095 */ 1096 fs = locate_flowset(pipe_nr, fwa->rule); 1097 if (fs == NULL) 1098 goto dropit ; /* this queue/pipe does not exist! */ 1099 pipe = fs->pipe ; 1100 if (pipe == NULL) { /* must be a queue, try find a matching pipe */ 1101 for (pipe = all_pipes; pipe && pipe->pipe_nr != fs->parent_nr; 1102 pipe = pipe->next) 1103 ; 1104 if (pipe != NULL) 1105 fs->pipe = pipe ; 1106 else { 1107 printf("No pipe %d for queue %d, drop pkt\n", 1108 fs->parent_nr, fs->fs_nr); 1109 goto dropit ; 1110 } 1111 } 1112 q = find_queue(fs, &(fwa->f_id)); 1113 if ( q == NULL ) 1114 goto dropit ; /* cannot allocate queue */ 1115 /* 1116 * update statistics, then check reasons to drop pkt 1117 */ 1118 q->tot_bytes += len ; 1119 q->tot_pkts++ ; 1120 if ( fs->plr && random() < fs->plr ) 1121 goto dropit ; /* random pkt drop */ 1122 if ( fs->flags_fs & DN_QSIZE_IS_BYTES) { 1123 if (q->len_bytes > fs->qsize) 1124 goto dropit ; /* queue size overflow */ 1125 } else { 1126 if (q->len >= fs->qsize) 1127 goto dropit ; /* queue count overflow */ 1128 } 1129 if ( fs->flags_fs & DN_IS_RED && red_drops(fs, q, len) ) 1130 goto dropit ; 1131 1132 /* XXX expensive to zero, see if we can remove it*/ 1133 pkt = (struct dn_pkt *)malloc(sizeof (*pkt), M_DUMMYNET, M_NOWAIT|M_ZERO); 1134 if ( pkt == NULL ) 1135 goto dropit ; /* cannot allocate packet header */ 1136 /* ok, i can handle the pkt now... */ 1137 /* build and enqueue packet + parameters */ 1138 pkt->hdr.mh_type = MT_TAG; 1139 pkt->hdr.mh_flags = PACKET_TAG_DUMMYNET; 1140 pkt->rule = fwa->rule ; 1141 DN_NEXT(pkt) = NULL; 1142 pkt->dn_m = m; 1143 pkt->dn_dir = dir ; 1144 1145 pkt->ifp = fwa->oif; 1146 if (dir == DN_TO_IP_OUT) { 1147 /* 1148 * We need to copy *ro because for ICMP pkts (and maybe others) 1149 * the caller passed a pointer into the stack; dst might also be 1150 * a pointer into *ro so it needs to be updated. 1151 */ 1152 pkt->ro = *(fwa->ro); 1153 if (fwa->ro->ro_rt) 1154 fwa->ro->ro_rt->rt_refcnt++ ; 1155 if (fwa->dst == (struct sockaddr_in *)&fwa->ro->ro_dst) /* dst points into ro */ 1156 fwa->dst = (struct sockaddr_in *)&(pkt->ro.ro_dst) ; 1157 1158 pkt->dn_dst = fwa->dst; 1159 pkt->flags = fwa->flags; 1160 } 1161 if (q->head == NULL) 1162 q->head = pkt; 1163 else 1164 DN_NEXT(q->tail) = pkt; 1165 q->tail = pkt; 1166 q->len++; 1167 q->len_bytes += len ; 1168 1169 if ( q->head != pkt ) /* flow was not idle, we are done */ 1170 goto done; 1171 /* 1172 * If we reach this point the flow was previously idle, so we need 1173 * to schedule it. This involves different actions for fixed-rate or 1174 * WF2Q queues. 1175 */ 1176 if (is_pipe) { 1177 /* 1178 * Fixed-rate queue: just insert into the ready_heap. 1179 */ 1180 dn_key t = 0 ; 1181 if (pipe->bandwidth) 1182 t = SET_TICKS(pkt, q, pipe); 1183 q->sched_time = curr_time ; 1184 if (t == 0) /* must process it now */ 1185 ready_event( q ); 1186 else 1187 heap_insert(&ready_heap, curr_time + t , q ); 1188 } else { 1189 /* 1190 * WF2Q. First, compute start time S: if the flow was idle (S=F+1) 1191 * set S to the virtual time V for the controlling pipe, and update 1192 * the sum of weights for the pipe; otherwise, remove flow from 1193 * idle_heap and set S to max(F,V). 1194 * Second, compute finish time F = S + len/weight. 1195 * Third, if pipe was idle, update V=max(S, V). 1196 * Fourth, count one more backlogged flow. 1197 */ 1198 if (DN_KEY_GT(q->S, q->F)) { /* means timestamps are invalid */ 1199 q->S = pipe->V ; 1200 pipe->sum += fs->weight ; /* add weight of new queue */ 1201 } else { 1202 heap_extract(&(pipe->idle_heap), q); 1203 q->S = MAX64(q->F, pipe->V ) ; 1204 } 1205 q->F = q->S + ( len<<MY_M )/(u_int64_t) fs->weight; 1206 1207 if (pipe->not_eligible_heap.elements == 0 && 1208 pipe->scheduler_heap.elements == 0) 1209 pipe->V = MAX64 ( q->S, pipe->V ); 1210 fs->backlogged++ ; 1211 /* 1212 * Look at eligibility. A flow is not eligibile if S>V (when 1213 * this happens, it means that there is some other flow already 1214 * scheduled for the same pipe, so the scheduler_heap cannot be 1215 * empty). If the flow is not eligible we just store it in the 1216 * not_eligible_heap. Otherwise, we store in the scheduler_heap 1217 * and possibly invoke ready_event_wfq() right now if there is 1218 * leftover credit. 1219 * Note that for all flows in scheduler_heap (SCH), S_i <= V, 1220 * and for all flows in not_eligible_heap (NEH), S_i > V . 1221 * So when we need to compute max( V, min(S_i) ) forall i in SCH+NEH, 1222 * we only need to look into NEH. 1223 */ 1224 if (DN_KEY_GT(q->S, pipe->V) ) { /* not eligible */ 1225 if (pipe->scheduler_heap.elements == 0) 1226 printf("++ ouch! not eligible but empty scheduler!\n"); 1227 heap_insert(&(pipe->not_eligible_heap), q->S, q); 1228 } else { 1229 heap_insert(&(pipe->scheduler_heap), q->F, q); 1230 if (pipe->numbytes >= 0) { /* pipe is idle */ 1231 if (pipe->scheduler_heap.elements != 1) 1232 printf("*** OUCH! pipe should have been idle!\n"); 1233 DEB(printf("Waking up pipe %d at %d\n", 1234 pipe->pipe_nr, (int)(q->F >> MY_M)); ) 1235 pipe->sched_time = curr_time ; 1236 ready_event_wfq(pipe); 1237 } 1238 } 1239 } 1240done: 1241 splx(s); 1242 return 0; 1243 1244dropit: 1245 splx(s); 1246 if (q) 1247 q->drops++ ; 1248 m_freem(m); 1249 return ( (fs && (fs->flags_fs & DN_NOERROR)) ? 0 : ENOBUFS); 1250} 1251 1252/* 1253 * Below, the rt_unref is only needed when (pkt->dn_dir == DN_TO_IP_OUT) 1254 * Doing this would probably save us the initial bzero of dn_pkt 1255 */ 1256#define DN_FREE_PKT(pkt) { \ 1257 struct dn_pkt *n = pkt ; \ 1258 rt_unref ( n->ro.ro_rt ) ; \ 1259 m_freem(n->dn_m); \ 1260 pkt = DN_NEXT(n) ; \ 1261 free(n, M_DUMMYNET) ; } 1262 1263/* 1264 * Dispose all packets and flow_queues on a flow_set. 1265 * If all=1, also remove red lookup table and other storage, 1266 * including the descriptor itself. 1267 * For the one in dn_pipe MUST also cleanup ready_heap... 1268 */ 1269static void 1270purge_flow_set(struct dn_flow_set *fs, int all) 1271{ 1272 struct dn_pkt *pkt ; 1273 struct dn_flow_queue *q, *qn ; 1274 int i ; 1275 1276 for (i = 0 ; i <= fs->rq_size ; i++ ) { 1277 for (q = fs->rq[i] ; q ; q = qn ) { 1278 for (pkt = q->head ; pkt ; ) 1279 DN_FREE_PKT(pkt) ; 1280 qn = q->next ; 1281 free(q, M_DUMMYNET); 1282 } 1283 fs->rq[i] = NULL ; 1284 } 1285 fs->rq_elements = 0 ; 1286 if (all) { 1287 /* RED - free lookup table */ 1288 if (fs->w_q_lookup) 1289 free(fs->w_q_lookup, M_DUMMYNET); 1290 if (fs->rq) 1291 free(fs->rq, M_DUMMYNET); 1292 /* if this fs is not part of a pipe, free it */ 1293 if (fs->pipe && fs != &(fs->pipe->fs) ) 1294 free(fs, M_DUMMYNET); 1295 } 1296} 1297 1298/* 1299 * Dispose all packets queued on a pipe (not a flow_set). 1300 * Also free all resources associated to a pipe, which is about 1301 * to be deleted. 1302 */ 1303static void 1304purge_pipe(struct dn_pipe *pipe) 1305{ 1306 struct dn_pkt *pkt ; 1307 1308 purge_flow_set( &(pipe->fs), 1 ); 1309 1310 for (pkt = pipe->head ; pkt ; ) 1311 DN_FREE_PKT(pkt) ; 1312 1313 heap_free( &(pipe->scheduler_heap) ); 1314 heap_free( &(pipe->not_eligible_heap) ); 1315 heap_free( &(pipe->idle_heap) ); 1316} 1317 1318/* 1319 * Delete all pipes and heaps returning memory. Must also 1320 * remove references from all ipfw rules to all pipes. 1321 */ 1322static void 1323dummynet_flush() 1324{ 1325 struct dn_pipe *curr_p, *p ; 1326 struct dn_flow_set *fs, *curr_fs; 1327 int s ; 1328 1329 s = splimp() ; 1330 1331 /* remove all references to pipes ...*/ 1332 flush_pipe_ptrs(NULL); 1333 /* prevent future matches... */ 1334 p = all_pipes ; 1335 all_pipes = NULL ; 1336 fs = all_flow_sets ; 1337 all_flow_sets = NULL ; 1338 /* and free heaps so we don't have unwanted events */ 1339 heap_free(&ready_heap); 1340 heap_free(&wfq_ready_heap); 1341 heap_free(&extract_heap); 1342 splx(s) ; 1343 /* 1344 * Now purge all queued pkts and delete all pipes 1345 */ 1346 /* scan and purge all flow_sets. */ 1347 for ( ; fs ; ) { 1348 curr_fs = fs ; 1349 fs = fs->next ; 1350 purge_flow_set(curr_fs, 1); 1351 } 1352 for ( ; p ; ) { 1353 purge_pipe(p); 1354 curr_p = p ; 1355 p = p->next ; 1356 free(curr_p, M_DUMMYNET); 1357 } 1358} 1359 1360 1361extern struct ip_fw *ip_fw_default_rule ; 1362static void 1363dn_rule_delete_fs(struct dn_flow_set *fs, void *r) 1364{ 1365 int i ; 1366 struct dn_flow_queue *q ; 1367 struct dn_pkt *pkt ; 1368 1369 for (i = 0 ; i <= fs->rq_size ; i++) /* last one is ovflow */ 1370 for (q = fs->rq[i] ; q ; q = q->next ) 1371 for (pkt = q->head ; pkt ; pkt = DN_NEXT(pkt) ) 1372 if (pkt->rule == r) 1373 pkt->rule = ip_fw_default_rule ; 1374} 1375/* 1376 * when a firewall rule is deleted, scan all queues and remove the flow-id 1377 * from packets matching this rule. 1378 */ 1379void 1380dn_rule_delete(void *r) 1381{ 1382 struct dn_pipe *p ; 1383 struct dn_pkt *pkt ; 1384 struct dn_flow_set *fs ; 1385 1386 /* 1387 * If the rule references a queue (dn_flow_set), then scan 1388 * the flow set, otherwise scan pipes. Should do either, but doing 1389 * both does not harm. 1390 */ 1391 for ( fs = all_flow_sets ; fs ; fs = fs->next ) 1392 dn_rule_delete_fs(fs, r); 1393 for ( p = all_pipes ; p ; p = p->next ) { 1394 fs = &(p->fs) ; 1395 dn_rule_delete_fs(fs, r); 1396 for (pkt = p->head ; pkt ; pkt = DN_NEXT(pkt) ) 1397 if (pkt->rule == r) 1398 pkt->rule = ip_fw_default_rule ; 1399 } 1400} 1401 1402/* 1403 * setup RED parameters 1404 */ 1405static int 1406config_red(struct dn_flow_set *p, struct dn_flow_set * x) 1407{ 1408 int i; 1409 1410 x->w_q = p->w_q; 1411 x->min_th = SCALE(p->min_th); 1412 x->max_th = SCALE(p->max_th); 1413 x->max_p = p->max_p; 1414 1415 x->c_1 = p->max_p / (p->max_th - p->min_th); 1416 x->c_2 = SCALE_MUL(x->c_1, SCALE(p->min_th)); 1417 if (x->flags_fs & DN_IS_GENTLE_RED) { 1418 x->c_3 = (SCALE(1) - p->max_p) / p->max_th; 1419 x->c_4 = (SCALE(1) - 2 * p->max_p); 1420 } 1421 1422 /* if the lookup table already exist, free and create it again */ 1423 if (x->w_q_lookup) { 1424 free(x->w_q_lookup, M_DUMMYNET); 1425 x->w_q_lookup = NULL ; 1426 } 1427 if (red_lookup_depth == 0) { 1428 printf("\nnet.inet.ip.dummynet.red_lookup_depth must be > 0"); 1429 free(x, M_DUMMYNET); 1430 return EINVAL; 1431 } 1432 x->lookup_depth = red_lookup_depth; 1433 x->w_q_lookup = (u_int *) malloc(x->lookup_depth * sizeof(int), 1434 M_DUMMYNET, M_DONTWAIT); 1435 if (x->w_q_lookup == NULL) { 1436 printf("sorry, cannot allocate red lookup table\n"); 1437 free(x, M_DUMMYNET); 1438 return ENOSPC; 1439 } 1440 1441 /* fill the lookup table with (1 - w_q)^x */ 1442 x->lookup_step = p->lookup_step ; 1443 x->lookup_weight = p->lookup_weight ; 1444 x->w_q_lookup[0] = SCALE(1) - x->w_q; 1445 for (i = 1; i < x->lookup_depth; i++) 1446 x->w_q_lookup[i] = SCALE_MUL(x->w_q_lookup[i - 1], x->lookup_weight); 1447 if (red_avg_pkt_size < 1) 1448 red_avg_pkt_size = 512 ; 1449 x->avg_pkt_size = red_avg_pkt_size ; 1450 if (red_max_pkt_size < 1) 1451 red_max_pkt_size = 1500 ; 1452 x->max_pkt_size = red_max_pkt_size ; 1453 return 0 ; 1454} 1455 1456static int 1457alloc_hash(struct dn_flow_set *x, struct dn_flow_set *pfs) 1458{ 1459 if (x->flags_fs & DN_HAVE_FLOW_MASK) { /* allocate some slots */ 1460 int l = pfs->rq_size; 1461 1462 if (l == 0) 1463 l = dn_hash_size; 1464 if (l < 4) 1465 l = 4; 1466 else if (l > DN_MAX_HASH_SIZE) 1467 l = DN_MAX_HASH_SIZE; 1468 x->rq_size = l; 1469 } else /* one is enough for null mask */ 1470 x->rq_size = 1; 1471 x->rq = malloc((1 + x->rq_size) * sizeof(struct dn_flow_queue *), 1472 M_DUMMYNET, M_DONTWAIT | M_ZERO); 1473 if (x->rq == NULL) { 1474 printf("sorry, cannot allocate queue\n"); 1475 return ENOSPC; 1476 } 1477 x->rq_elements = 0; 1478 return 0 ; 1479} 1480 1481static void 1482set_fs_parms(struct dn_flow_set *x, struct dn_flow_set *src) 1483{ 1484 x->flags_fs = src->flags_fs; 1485 x->qsize = src->qsize; 1486 x->plr = src->plr; 1487 x->flow_mask = src->flow_mask; 1488 if (x->flags_fs & DN_QSIZE_IS_BYTES) { 1489 if (x->qsize > 1024*1024) 1490 x->qsize = 1024*1024 ; 1491 } else { 1492 if (x->qsize == 0) 1493 x->qsize = 50 ; 1494 if (x->qsize > 100) 1495 x->qsize = 50 ; 1496 } 1497 /* configuring RED */ 1498 if ( x->flags_fs & DN_IS_RED ) 1499 config_red(src, x) ; /* XXX should check errors */ 1500} 1501 1502/* 1503 * setup pipe or queue parameters. 1504 */ 1505 1506static int 1507config_pipe(struct dn_pipe *p) 1508{ 1509 int s ; 1510 struct dn_flow_set *pfs = &(p->fs); 1511 1512 /* 1513 * The config program passes parameters as follows: 1514 * bw = bits/second (0 means no limits), 1515 * delay = ms, must be translated into ticks. 1516 * qsize = slots/bytes 1517 */ 1518 p->delay = ( p->delay * hz ) / 1000 ; 1519 /* We need either a pipe number or a flow_set number */ 1520 if (p->pipe_nr == 0 && pfs->fs_nr == 0) 1521 return EINVAL ; 1522 if (p->pipe_nr != 0 && pfs->fs_nr != 0) 1523 return EINVAL ; 1524 if (p->pipe_nr != 0) { /* this is a pipe */ 1525 struct dn_pipe *x, *a, *b; 1526 /* locate pipe */ 1527 for (a = NULL , b = all_pipes ; b && b->pipe_nr < p->pipe_nr ; 1528 a = b , b = b->next) ; 1529 1530 if (b == NULL || b->pipe_nr != p->pipe_nr) { /* new pipe */ 1531 x = malloc(sizeof(struct dn_pipe), M_DUMMYNET, M_DONTWAIT | M_ZERO); 1532 if (x == NULL) { 1533 printf("ip_dummynet.c: no memory for new pipe\n"); 1534 return ENOSPC; 1535 } 1536 x->pipe_nr = p->pipe_nr; 1537 x->fs.pipe = x ; 1538 /* idle_heap is the only one from which we extract from the middle. 1539 */ 1540 x->idle_heap.size = x->idle_heap.elements = 0 ; 1541 x->idle_heap.offset=OFFSET_OF(struct dn_flow_queue, heap_pos); 1542 } else 1543 x = b; 1544 1545 x->bandwidth = p->bandwidth ; 1546 x->numbytes = 0; /* just in case... */ 1547 bcopy(p->if_name, x->if_name, sizeof(p->if_name) ); 1548 x->ifp = NULL ; /* reset interface ptr */ 1549 x->delay = p->delay ; 1550 set_fs_parms(&(x->fs), pfs); 1551 1552 1553 if ( x->fs.rq == NULL ) { /* a new pipe */ 1554 s = alloc_hash(&(x->fs), pfs) ; 1555 if (s) { 1556 free(x, M_DUMMYNET); 1557 return s ; 1558 } 1559 s = splimp() ; 1560 x->next = b ; 1561 if (a == NULL) 1562 all_pipes = x ; 1563 else 1564 a->next = x ; 1565 splx(s); 1566 } 1567 } else { /* config queue */ 1568 struct dn_flow_set *x, *a, *b ; 1569 1570 /* locate flow_set */ 1571 for (a=NULL, b=all_flow_sets ; b && b->fs_nr < pfs->fs_nr ; 1572 a = b , b = b->next) ; 1573 1574 if (b == NULL || b->fs_nr != pfs->fs_nr) { /* new */ 1575 if (pfs->parent_nr == 0) /* need link to a pipe */ 1576 return EINVAL ; 1577 x = malloc(sizeof(struct dn_flow_set),M_DUMMYNET,M_DONTWAIT|M_ZERO); 1578 if (x == NULL) { 1579 printf("ip_dummynet.c: no memory for new flow_set\n"); 1580 return ENOSPC; 1581 } 1582 x->fs_nr = pfs->fs_nr; 1583 x->parent_nr = pfs->parent_nr; 1584 x->weight = pfs->weight ; 1585 if (x->weight == 0) 1586 x->weight = 1 ; 1587 else if (x->weight > 100) 1588 x->weight = 100 ; 1589 } else { 1590 /* Change parent pipe not allowed; must delete and recreate */ 1591 if (pfs->parent_nr != 0 && b->parent_nr != pfs->parent_nr) 1592 return EINVAL ; 1593 x = b; 1594 } 1595 set_fs_parms(x, pfs); 1596 1597 if ( x->rq == NULL ) { /* a new flow_set */ 1598 s = alloc_hash(x, pfs) ; 1599 if (s) { 1600 free(x, M_DUMMYNET); 1601 return s ; 1602 } 1603 s = splimp() ; 1604 x->next = b; 1605 if (a == NULL) 1606 all_flow_sets = x; 1607 else 1608 a->next = x; 1609 splx(s); 1610 } 1611 } 1612 return 0 ; 1613} 1614 1615/* 1616 * Helper function to remove from a heap queues which are linked to 1617 * a flow_set about to be deleted. 1618 */ 1619static void 1620fs_remove_from_heap(struct dn_heap *h, struct dn_flow_set *fs) 1621{ 1622 int i = 0, found = 0 ; 1623 for (; i < h->elements ;) 1624 if ( ((struct dn_flow_queue *)h->p[i].object)->fs == fs) { 1625 h->elements-- ; 1626 h->p[i] = h->p[h->elements] ; 1627 found++ ; 1628 } else 1629 i++ ; 1630 if (found) 1631 heapify(h); 1632} 1633 1634/* 1635 * helper function to remove a pipe from a heap (can be there at most once) 1636 */ 1637static void 1638pipe_remove_from_heap(struct dn_heap *h, struct dn_pipe *p) 1639{ 1640 if (h->elements > 0) { 1641 int i = 0 ; 1642 for (i=0; i < h->elements ; i++ ) { 1643 if (h->p[i].object == p) { /* found it */ 1644 h->elements-- ; 1645 h->p[i] = h->p[h->elements] ; 1646 heapify(h); 1647 break ; 1648 } 1649 } 1650 } 1651} 1652 1653/* 1654 * drain all queues. Called in case of severe mbuf shortage. 1655 */ 1656void 1657dummynet_drain() 1658{ 1659 struct dn_flow_set *fs; 1660 struct dn_pipe *p; 1661 struct dn_pkt *pkt; 1662 1663 heap_free(&ready_heap); 1664 heap_free(&wfq_ready_heap); 1665 heap_free(&extract_heap); 1666 /* remove all references to this pipe from flow_sets */ 1667 for (fs = all_flow_sets; fs; fs= fs->next ) 1668 purge_flow_set(fs, 0); 1669 1670 for (p = all_pipes; p; p= p->next ) { 1671 purge_flow_set(&(p->fs), 0); 1672 for (pkt = p->head ; pkt ; ) 1673 DN_FREE_PKT(pkt) ; 1674 p->head = p->tail = NULL ; 1675 } 1676} 1677 1678/* 1679 * Fully delete a pipe or a queue, cleaning up associated info. 1680 */ 1681static int 1682delete_pipe(struct dn_pipe *p) 1683{ 1684 int s ; 1685 1686 if (p->pipe_nr == 0 && p->fs.fs_nr == 0) 1687 return EINVAL ; 1688 if (p->pipe_nr != 0 && p->fs.fs_nr != 0) 1689 return EINVAL ; 1690 if (p->pipe_nr != 0) { /* this is an old-style pipe */ 1691 struct dn_pipe *a, *b; 1692 struct dn_flow_set *fs; 1693 1694 /* locate pipe */ 1695 for (a = NULL , b = all_pipes ; b && b->pipe_nr < p->pipe_nr ; 1696 a = b , b = b->next) ; 1697 if (b == NULL || (b->pipe_nr != p->pipe_nr) ) 1698 return EINVAL ; /* not found */ 1699 1700 s = splimp() ; 1701 1702 /* unlink from list of pipes */ 1703 if (a == NULL) 1704 all_pipes = b->next ; 1705 else 1706 a->next = b->next ; 1707 /* remove references to this pipe from the ip_fw rules. */ 1708 flush_pipe_ptrs(&(b->fs)); 1709 1710 /* remove all references to this pipe from flow_sets */ 1711 for (fs = all_flow_sets; fs; fs= fs->next ) 1712 if (fs->pipe == b) { 1713 printf("++ ref to pipe %d from fs %d\n", 1714 p->pipe_nr, fs->fs_nr); 1715 fs->pipe = NULL ; 1716 purge_flow_set(fs, 0); 1717 } 1718 fs_remove_from_heap(&ready_heap, &(b->fs)); 1719 purge_pipe(b); /* remove all data associated to this pipe */ 1720 /* remove reference to here from extract_heap and wfq_ready_heap */ 1721 pipe_remove_from_heap(&extract_heap, b); 1722 pipe_remove_from_heap(&wfq_ready_heap, b); 1723 splx(s); 1724 free(b, M_DUMMYNET); 1725 } else { /* this is a WF2Q queue (dn_flow_set) */ 1726 struct dn_flow_set *a, *b; 1727 1728 /* locate set */ 1729 for (a = NULL, b = all_flow_sets ; b && b->fs_nr < p->fs.fs_nr ; 1730 a = b , b = b->next) ; 1731 if (b == NULL || (b->fs_nr != p->fs.fs_nr) ) 1732 return EINVAL ; /* not found */ 1733 1734 s = splimp() ; 1735 if (a == NULL) 1736 all_flow_sets = b->next ; 1737 else 1738 a->next = b->next ; 1739 /* remove references to this flow_set from the ip_fw rules. */ 1740 flush_pipe_ptrs(b); 1741 1742 if (b->pipe != NULL) { 1743 /* Update total weight on parent pipe and cleanup parent heaps */ 1744 b->pipe->sum -= b->weight * b->backlogged ; 1745 fs_remove_from_heap(&(b->pipe->not_eligible_heap), b); 1746 fs_remove_from_heap(&(b->pipe->scheduler_heap), b); 1747#if 1 /* XXX should i remove from idle_heap as well ? */ 1748 fs_remove_from_heap(&(b->pipe->idle_heap), b); 1749#endif 1750 } 1751 purge_flow_set(b, 1); 1752 splx(s); 1753 } 1754 return 0 ; 1755} 1756 1757/* 1758 * helper function used to copy data from kernel in DUMMYNET_GET 1759 */ 1760static char * 1761dn_copy_set(struct dn_flow_set *set, char *bp) 1762{ 1763 int i, copied = 0 ; 1764 struct dn_flow_queue *q, *qp = (struct dn_flow_queue *)bp; 1765 1766 for (i = 0 ; i <= set->rq_size ; i++) 1767 for (q = set->rq[i] ; q ; q = q->next, qp++ ) { 1768 if (q->hash_slot != i) 1769 printf("++ at %d: wrong slot (have %d, " 1770 "should be %d)\n", copied, q->hash_slot, i); 1771 if (q->fs != set) 1772 printf("++ at %d: wrong fs ptr (have %p, should be %p)\n", 1773 i, q->fs, set); 1774 copied++ ; 1775 bcopy(q, qp, sizeof( *q ) ); 1776 /* cleanup pointers */ 1777 qp->next = NULL ; 1778 qp->head = qp->tail = NULL ; 1779 qp->fs = NULL ; 1780 } 1781 if (copied != set->rq_elements) 1782 printf("++ wrong count, have %d should be %d\n", 1783 copied, set->rq_elements); 1784 return (char *)qp ; 1785} 1786 1787static int 1788dummynet_get(struct sockopt *sopt) 1789{ 1790 char *buf, *bp ; /* bp is the "copy-pointer" */ 1791 size_t size ; 1792 struct dn_flow_set *set ; 1793 struct dn_pipe *p ; 1794 int s, error=0 ; 1795 1796 s = splimp(); 1797 /* 1798 * compute size of data structures: list of pipes and flow_sets. 1799 */ 1800 for (p = all_pipes, size = 0 ; p ; p = p->next ) 1801 size += sizeof( *p ) + 1802 p->fs.rq_elements * sizeof(struct dn_flow_queue); 1803 for (set = all_flow_sets ; set ; set = set->next ) 1804 size += sizeof ( *set ) + 1805 set->rq_elements * sizeof(struct dn_flow_queue); 1806 buf = malloc(size, M_TEMP, M_DONTWAIT); 1807 if (buf == 0) { 1808 splx(s); 1809 return ENOBUFS ; 1810 } 1811 for (p = all_pipes, bp = buf ; p ; p = p->next ) { 1812 struct dn_pipe *pipe_bp = (struct dn_pipe *)bp ; 1813 1814 /* 1815 * copy pipe descriptor into *bp, convert delay back to ms, 1816 * then copy the flow_set descriptor(s) one at a time. 1817 * After each flow_set, copy the queue descriptor it owns. 1818 */ 1819 bcopy(p, bp, sizeof( *p ) ); 1820 pipe_bp->delay = (pipe_bp->delay * 1000) / hz ; 1821 /* 1822 * XXX the following is a hack based on ->next being the 1823 * first field in dn_pipe and dn_flow_set. The correct 1824 * solution would be to move the dn_flow_set to the beginning 1825 * of struct dn_pipe. 1826 */ 1827 pipe_bp->next = (struct dn_pipe *)DN_IS_PIPE ; 1828 /* clean pointers */ 1829 pipe_bp->head = pipe_bp->tail = NULL ; 1830 pipe_bp->fs.next = NULL ; 1831 pipe_bp->fs.pipe = NULL ; 1832 pipe_bp->fs.rq = NULL ; 1833 1834 bp += sizeof( *p ) ; 1835 bp = dn_copy_set( &(p->fs), bp ); 1836 } 1837 for (set = all_flow_sets ; set ; set = set->next ) { 1838 struct dn_flow_set *fs_bp = (struct dn_flow_set *)bp ; 1839 bcopy(set, bp, sizeof( *set ) ); 1840 /* XXX same hack as above */ 1841 fs_bp->next = (struct dn_flow_set *)DN_IS_QUEUE ; 1842 fs_bp->pipe = NULL ; 1843 fs_bp->rq = NULL ; 1844 bp += sizeof( *set ) ; 1845 bp = dn_copy_set( set, bp ); 1846 } 1847 splx(s); 1848 error = sooptcopyout(sopt, buf, size); 1849 free(buf, M_TEMP); 1850 return error ; 1851} 1852 1853/* 1854 * Handler for the various dummynet socket options (get, flush, config, del) 1855 */ 1856static int 1857ip_dn_ctl(struct sockopt *sopt) 1858{ 1859 int error = 0 ; 1860 struct dn_pipe *p, tmp_pipe; 1861 1862 /* Disallow sets in really-really secure mode. */ 1863 if (sopt->sopt_dir == SOPT_SET) { 1864#if __FreeBSD_version >= 500034 1865 error = securelevel_ge(sopt->sopt_td->td_ucred, 3); 1866 if (error) 1867 return (error); 1868#else 1869 if (securelevel >= 3) 1870 return (EPERM); 1871#endif 1872 } 1873 1874 switch (sopt->sopt_name) { 1875 default : 1876 printf("ip_dn_ctl -- unknown option %d", sopt->sopt_name); 1877 return EINVAL ; 1878 1879 case IP_DUMMYNET_GET : 1880 error = dummynet_get(sopt); 1881 break ; 1882 1883 case IP_DUMMYNET_FLUSH : 1884 dummynet_flush() ; 1885 break ; 1886 1887 case IP_DUMMYNET_CONFIGURE : 1888 p = &tmp_pipe ; 1889 error = sooptcopyin(sopt, p, sizeof *p, sizeof *p); 1890 if (error) 1891 break ; 1892 error = config_pipe(p); 1893 break ; 1894 1895 case IP_DUMMYNET_DEL : /* remove a pipe or queue */ 1896 p = &tmp_pipe ; 1897 error = sooptcopyin(sopt, p, sizeof *p, sizeof *p); 1898 if (error) 1899 break ; 1900 1901 error = delete_pipe(p); 1902 break ; 1903 } 1904 return error ; 1905} 1906 1907static void 1908ip_dn_init(void) 1909{ 1910 printf("DUMMYNET initialized (011031)\n"); 1911 all_pipes = NULL ; 1912 all_flow_sets = NULL ; 1913 ready_heap.size = ready_heap.elements = 0 ; 1914 ready_heap.offset = 0 ; 1915 1916 wfq_ready_heap.size = wfq_ready_heap.elements = 0 ; 1917 wfq_ready_heap.offset = 0 ; 1918 1919 extract_heap.size = extract_heap.elements = 0 ; 1920 extract_heap.offset = 0 ; 1921 ip_dn_ctl_ptr = ip_dn_ctl; 1922 ip_dn_io_ptr = dummynet_io; 1923 ip_dn_ruledel_ptr = dn_rule_delete; 1924 bzero(&dn_timeout, sizeof(struct callout_handle)); 1925 dn_timeout = timeout(dummynet, NULL, 1); 1926} 1927 1928static int 1929dummynet_modevent(module_t mod, int type, void *data) 1930{ 1931 int s; 1932 switch (type) { 1933 case MOD_LOAD: 1934 s = splimp(); 1935 if (DUMMYNET_LOADED) { 1936 splx(s); 1937 printf("DUMMYNET already loaded\n"); 1938 return EEXIST ; 1939 } 1940 ip_dn_init(); 1941 splx(s); 1942 break; 1943 1944 case MOD_UNLOAD: 1945#if !defined(KLD_MODULE) 1946 printf("dummynet statically compiled, cannot unload\n"); 1947 return EINVAL ; 1948#else 1949 s = splimp(); 1950 untimeout(dummynet, NULL, dn_timeout); 1951 dummynet_flush(); 1952 ip_dn_ctl_ptr = NULL; 1953 ip_dn_io_ptr = NULL; 1954 ip_dn_ruledel_ptr = NULL; 1955 splx(s); 1956#endif 1957 break ; 1958 default: 1959 break ; 1960 } 1961 return 0 ; 1962} 1963 1964static moduledata_t dummynet_mod = { 1965 "dummynet", 1966 dummynet_modevent, 1967 NULL 1968}; 1969DECLARE_MODULE(dummynet, dummynet_mod, SI_SUB_PSEUDO, SI_ORDER_ANY); 1970MODULE_DEPEND(dummynet, ipfw, 1, 1, 1); 1971MODULE_VERSION(dummynet, 1);
| 426 rt_unref (pkt->ro.ro_rt) ; 427 break ; 428 429 case DN_TO_IP_IN : 430 ip_input((struct mbuf *)pkt) ; 431 break ; 432 433 case DN_TO_BDG_FWD : 434 if (!BDG_LOADED) { 435 /* somebody unloaded the bridge module. Drop pkt */ 436 printf("-- dropping bridged packet trapped in pipe--\n"); 437 m_freem(pkt->dn_m); 438 break; 439 } /* fallthrough */ 440 case DN_TO_ETH_DEMUX: 441 { 442 struct mbuf *m = (struct mbuf *)pkt ; 443 struct ether_header *eh; 444 445 if (pkt->dn_m->m_len < ETHER_HDR_LEN && 446 (pkt->dn_m = m_pullup(pkt->dn_m, ETHER_HDR_LEN)) == NULL) { 447 printf("dummynet/bridge: pullup fail, dropping pkt\n"); 448 break; 449 } 450 /* 451 * same as ether_input, make eh be a pointer into the mbuf 452 */ 453 eh = mtod(pkt->dn_m, struct ether_header *); 454 m_adj(pkt->dn_m, ETHER_HDR_LEN); 455 /* 456 * bdg_forward() wants a pointer to the pseudo-mbuf-header, but 457 * on return it will supply the pointer to the actual packet 458 * (originally pkt->dn_m, but could be something else now) if 459 * it has not consumed it. 460 */ 461 if (pkt->dn_dir == DN_TO_BDG_FWD) { 462 m = bdg_forward_ptr(m, eh, pkt->ifp); 463 if (m) 464 m_freem(m); 465 } else 466 ether_demux(NULL, eh, m); /* which consumes the mbuf */ 467 } 468 break ; 469 case DN_TO_ETH_OUT: 470 ether_output_frame(pkt->ifp, (struct mbuf *)pkt); 471 break; 472 473 default: 474 printf("dummynet: bad switch %d!\n", pkt->dn_dir); 475 m_freem(pkt->dn_m); 476 break ; 477 } 478 free(pkt, M_DUMMYNET); 479 } 480 /* if there are leftover packets, put into the heap for next event */ 481 if ( (pkt = pipe->head) ) 482 heap_insert(&extract_heap, pkt->output_time, pipe ) ; 483 /* XXX should check errors on heap_insert, by draining the 484 * whole pipe p and hoping in the future we are more successful 485 */ 486} 487 488/* 489 * the following macro computes how many ticks we have to wait 490 * before being able to transmit a packet. The credit is taken from 491 * either a pipe (WF2Q) or a flow_queue (per-flow queueing) 492 */ 493#define SET_TICKS(pkt, q, p) \ 494 (pkt->dn_m->m_pkthdr.len*8*hz - (q)->numbytes + p->bandwidth - 1 ) / \ 495 p->bandwidth ; 496 497/* 498 * extract pkt from queue, compute output time (could be now) 499 * and put into delay line (p_queue) 500 */ 501static void 502move_pkt(struct dn_pkt *pkt, struct dn_flow_queue *q, 503 struct dn_pipe *p, int len) 504{ 505 q->head = DN_NEXT(pkt) ; 506 q->len-- ; 507 q->len_bytes -= len ; 508 509 pkt->output_time = curr_time + p->delay ; 510 511 if (p->head == NULL) 512 p->head = pkt; 513 else 514 DN_NEXT(p->tail) = pkt; 515 p->tail = pkt; 516 DN_NEXT(p->tail) = NULL; 517} 518 519/* 520 * ready_event() is invoked every time the queue must enter the 521 * scheduler, either because the first packet arrives, or because 522 * a previously scheduled event fired. 523 * On invokation, drain as many pkts as possible (could be 0) and then 524 * if there are leftover packets reinsert the pkt in the scheduler. 525 */ 526static void 527ready_event(struct dn_flow_queue *q) 528{ 529 struct dn_pkt *pkt; 530 struct dn_pipe *p = q->fs->pipe ; 531 int p_was_empty ; 532 533 if (p == NULL) { 534 printf("ready_event- pipe is gone\n"); 535 return ; 536 } 537 p_was_empty = (p->head == NULL) ; 538 539 /* 540 * schedule fixed-rate queues linked to this pipe: 541 * Account for the bw accumulated since last scheduling, then 542 * drain as many pkts as allowed by q->numbytes and move to 543 * the delay line (in p) computing output time. 544 * bandwidth==0 (no limit) means we can drain the whole queue, 545 * setting len_scaled = 0 does the job. 546 */ 547 q->numbytes += ( curr_time - q->sched_time ) * p->bandwidth; 548 while ( (pkt = q->head) != NULL ) { 549 int len = pkt->dn_m->m_pkthdr.len; 550 int len_scaled = p->bandwidth ? len*8*hz : 0 ; 551 if (len_scaled > q->numbytes ) 552 break ; 553 q->numbytes -= len_scaled ; 554 move_pkt(pkt, q, p, len); 555 } 556 /* 557 * If we have more packets queued, schedule next ready event 558 * (can only occur when bandwidth != 0, otherwise we would have 559 * flushed the whole queue in the previous loop). 560 * To this purpose we record the current time and compute how many 561 * ticks to go for the finish time of the packet. 562 */ 563 if ( (pkt = q->head) != NULL ) { /* this implies bandwidth != 0 */ 564 dn_key t = SET_TICKS(pkt, q, p); /* ticks i have to wait */ 565 q->sched_time = curr_time ; 566 heap_insert(&ready_heap, curr_time + t, (void *)q ); 567 /* XXX should check errors on heap_insert, and drain the whole 568 * queue on error hoping next time we are luckier. 569 */ 570 } else /* RED needs to know when the queue becomes empty */ 571 q->q_time = curr_time; 572 /* 573 * If the delay line was empty call transmit_event(p) now. 574 * Otherwise, the scheduler will take care of it. 575 */ 576 if (p_was_empty) 577 transmit_event(p); 578} 579 580/* 581 * Called when we can transmit packets on WF2Q queues. Take pkts out of 582 * the queues at their start time, and enqueue into the delay line. 583 * Packets are drained until p->numbytes < 0. As long as 584 * len_scaled >= p->numbytes, the packet goes into the delay line 585 * with a deadline p->delay. For the last packet, if p->numbytes<0, 586 * there is an additional delay. 587 */ 588static void 589ready_event_wfq(struct dn_pipe *p) 590{ 591 int p_was_empty = (p->head == NULL) ; 592 struct dn_heap *sch = &(p->scheduler_heap); 593 struct dn_heap *neh = &(p->not_eligible_heap) ; 594 595 if (p->if_name[0] == 0) /* tx clock is simulated */ 596 p->numbytes += ( curr_time - p->sched_time ) * p->bandwidth; 597 else { /* tx clock is for real, the ifq must be empty or this is a NOP */ 598 if (p->ifp && p->ifp->if_snd.ifq_head != NULL) 599 return ; 600 else { 601 DEB(printf("pipe %d ready from %s --\n", 602 p->pipe_nr, p->if_name);) 603 } 604 } 605 606 /* 607 * While we have backlogged traffic AND credit, we need to do 608 * something on the queue. 609 */ 610 while ( p->numbytes >=0 && (sch->elements>0 || neh->elements >0) ) { 611 if (sch->elements > 0) { /* have some eligible pkts to send out */ 612 struct dn_flow_queue *q = sch->p[0].object ; 613 struct dn_pkt *pkt = q->head; 614 struct dn_flow_set *fs = q->fs; 615 u_int64_t len = pkt->dn_m->m_pkthdr.len; 616 int len_scaled = p->bandwidth ? len*8*hz : 0 ; 617 618 heap_extract(sch, NULL); /* remove queue from heap */ 619 p->numbytes -= len_scaled ; 620 move_pkt(pkt, q, p, len); 621 622 p->V += (len<<MY_M) / p->sum ; /* update V */ 623 q->S = q->F ; /* update start time */ 624 if (q->len == 0) { /* Flow not backlogged any more */ 625 fs->backlogged-- ; 626 heap_insert(&(p->idle_heap), q->F, q); 627 } else { /* still backlogged */ 628 /* 629 * update F and position in backlogged queue, then 630 * put flow in not_eligible_heap (we will fix this later). 631 */ 632 len = (q->head)->dn_m->m_pkthdr.len; 633 q->F += (len<<MY_M)/(u_int64_t) fs->weight ; 634 if (DN_KEY_LEQ(q->S, p->V)) 635 heap_insert(neh, q->S, q); 636 else 637 heap_insert(sch, q->F, q); 638 } 639 } 640 /* 641 * now compute V = max(V, min(S_i)). Remember that all elements in sch 642 * have by definition S_i <= V so if sch is not empty, V is surely 643 * the max and we must not update it. Conversely, if sch is empty 644 * we only need to look at neh. 645 */ 646 if (sch->elements == 0 && neh->elements > 0) 647 p->V = MAX64 ( p->V, neh->p[0].key ); 648 /* move from neh to sch any packets that have become eligible */ 649 while (neh->elements > 0 && DN_KEY_LEQ(neh->p[0].key, p->V) ) { 650 struct dn_flow_queue *q = neh->p[0].object ; 651 heap_extract(neh, NULL); 652 heap_insert(sch, q->F, q); 653 } 654 655 if (p->if_name[0] != '\0') {/* tx clock is from a real thing */ 656 p->numbytes = -1 ; /* mark not ready for I/O */ 657 break ; 658 } 659 } 660 if (sch->elements == 0 && neh->elements == 0 && p->numbytes >= 0 661 && p->idle_heap.elements > 0) { 662 /* 663 * no traffic and no events scheduled. We can get rid of idle-heap. 664 */ 665 int i ; 666 667 for (i = 0 ; i < p->idle_heap.elements ; i++) { 668 struct dn_flow_queue *q = p->idle_heap.p[i].object ; 669 670 q->F = 0 ; 671 q->S = q->F + 1 ; 672 } 673 p->sum = 0 ; 674 p->V = 0 ; 675 p->idle_heap.elements = 0 ; 676 } 677 /* 678 * If we are getting clocks from dummynet (not a real interface) and 679 * If we are under credit, schedule the next ready event. 680 * Also fix the delivery time of the last packet. 681 */ 682 if (p->if_name[0]==0 && p->numbytes < 0) { /* this implies bandwidth >0 */ 683 dn_key t=0 ; /* number of ticks i have to wait */ 684 685 if (p->bandwidth > 0) 686 t = ( p->bandwidth -1 - p->numbytes) / p->bandwidth ; 687 p->tail->output_time += t ; 688 p->sched_time = curr_time ; 689 heap_insert(&wfq_ready_heap, curr_time + t, (void *)p); 690 /* XXX should check errors on heap_insert, and drain the whole 691 * queue on error hoping next time we are luckier. 692 */ 693 } 694 /* 695 * If the delay line was empty call transmit_event(p) now. 696 * Otherwise, the scheduler will take care of it. 697 */ 698 if (p_was_empty) 699 transmit_event(p); 700} 701 702/* 703 * This is called once per tick, or HZ times per second. It is used to 704 * increment the current tick counter and schedule expired events. 705 */ 706static void 707dummynet(void * __unused unused) 708{ 709 void *p ; /* generic parameter to handler */ 710 struct dn_heap *h ; 711 int s ; 712 struct dn_heap *heaps[3]; 713 int i; 714 struct dn_pipe *pe ; 715 716 heaps[0] = &ready_heap ; /* fixed-rate queues */ 717 heaps[1] = &wfq_ready_heap ; /* wfq queues */ 718 heaps[2] = &extract_heap ; /* delay line */ 719 s = splimp(); /* see note on top, splnet() is not enough */ 720 curr_time++ ; 721 for (i=0; i < 3 ; i++) { 722 h = heaps[i]; 723 while (h->elements > 0 && DN_KEY_LEQ(h->p[0].key, curr_time) ) { 724 DDB(if (h->p[0].key > curr_time) 725 printf("-- dummynet: warning, heap %d is %d ticks late\n", 726 i, (int)(curr_time - h->p[0].key));) 727 p = h->p[0].object ; /* store a copy before heap_extract */ 728 heap_extract(h, NULL); /* need to extract before processing */ 729 if (i == 0) 730 ready_event(p) ; 731 else if (i == 1) { 732 struct dn_pipe *pipe = p; 733 if (pipe->if_name[0] != '\0') 734 printf("*** bad ready_event_wfq for pipe %s\n", 735 pipe->if_name); 736 else 737 ready_event_wfq(p) ; 738 } else 739 transmit_event(p); 740 } 741 } 742 /* sweep pipes trying to expire idle flow_queues */ 743 for (pe = all_pipes; pe ; pe = pe->next ) 744 if (pe->idle_heap.elements > 0 && 745 DN_KEY_LT(pe->idle_heap.p[0].key, pe->V) ) { 746 struct dn_flow_queue *q = pe->idle_heap.p[0].object ; 747 748 heap_extract(&(pe->idle_heap), NULL); 749 q->S = q->F + 1 ; /* mark timestamp as invalid */ 750 pe->sum -= q->fs->weight ; 751 } 752 splx(s); 753 dn_timeout = timeout(dummynet, NULL, 1); 754} 755 756/* 757 * called by an interface when tx_rdy occurs. 758 */ 759int 760if_tx_rdy(struct ifnet *ifp) 761{ 762 struct dn_pipe *p; 763 764 for (p = all_pipes; p ; p = p->next ) 765 if (p->ifp == ifp) 766 break ; 767 if (p == NULL) { 768 char buf[32]; 769 sprintf(buf, "%s%d",ifp->if_name, ifp->if_unit); 770 for (p = all_pipes; p ; p = p->next ) 771 if (!strcmp(p->if_name, buf) ) { 772 p->ifp = ifp ; 773 DEB(printf("++ tx rdy from %s (now found)\n", buf);) 774 break ; 775 } 776 } 777 if (p != NULL) { 778 DEB(printf("++ tx rdy from %s%d - qlen %d\n", ifp->if_name, 779 ifp->if_unit, ifp->if_snd.ifq_len);) 780 p->numbytes = 0 ; /* mark ready for I/O */ 781 ready_event_wfq(p); 782 } 783 return 0; 784} 785 786/* 787 * Unconditionally expire empty queues in case of shortage. 788 * Returns the number of queues freed. 789 */ 790static int 791expire_queues(struct dn_flow_set *fs) 792{ 793 struct dn_flow_queue *q, *prev ; 794 int i, initial_elements = fs->rq_elements ; 795 796 if (fs->last_expired == time_second) 797 return 0 ; 798 fs->last_expired = time_second ; 799 for (i = 0 ; i <= fs->rq_size ; i++) /* last one is overflow */ 800 for (prev=NULL, q = fs->rq[i] ; q != NULL ; ) 801 if (q->head != NULL || q->S != q->F+1) { 802 prev = q ; 803 q = q->next ; 804 } else { /* entry is idle, expire it */ 805 struct dn_flow_queue *old_q = q ; 806 807 if (prev != NULL) 808 prev->next = q = q->next ; 809 else 810 fs->rq[i] = q = q->next ; 811 fs->rq_elements-- ; 812 free(old_q, M_DUMMYNET); 813 } 814 return initial_elements - fs->rq_elements ; 815} 816 817/* 818 * If room, create a new queue and put at head of slot i; 819 * otherwise, create or use the default queue. 820 */ 821static struct dn_flow_queue * 822create_queue(struct dn_flow_set *fs, int i) 823{ 824 struct dn_flow_queue *q ; 825 826 if (fs->rq_elements > fs->rq_size * dn_max_ratio && 827 expire_queues(fs) == 0) { 828 /* 829 * No way to get room, use or create overflow queue. 830 */ 831 i = fs->rq_size ; 832 if ( fs->rq[i] != NULL ) 833 return fs->rq[i] ; 834 } 835 q = malloc(sizeof(*q), M_DUMMYNET, M_DONTWAIT | M_ZERO); 836 if (q == NULL) { 837 printf("sorry, cannot allocate queue for new flow\n"); 838 return NULL ; 839 } 840 q->fs = fs ; 841 q->hash_slot = i ; 842 q->next = fs->rq[i] ; 843 q->S = q->F + 1; /* hack - mark timestamp as invalid */ 844 fs->rq[i] = q ; 845 fs->rq_elements++ ; 846 return q ; 847} 848 849/* 850 * Given a flow_set and a pkt in last_pkt, find a matching queue 851 * after appropriate masking. The queue is moved to front 852 * so that further searches take less time. 853 */ 854static struct dn_flow_queue * 855find_queue(struct dn_flow_set *fs, struct ipfw_flow_id *id) 856{ 857 int i = 0 ; /* we need i and q for new allocations */ 858 struct dn_flow_queue *q, *prev; 859 860 if ( !(fs->flags_fs & DN_HAVE_FLOW_MASK) ) 861 q = fs->rq[0] ; 862 else { 863 /* first, do the masking */ 864 id->dst_ip &= fs->flow_mask.dst_ip ; 865 id->src_ip &= fs->flow_mask.src_ip ; 866 id->dst_port &= fs->flow_mask.dst_port ; 867 id->src_port &= fs->flow_mask.src_port ; 868 id->proto &= fs->flow_mask.proto ; 869 id->flags = 0 ; /* we don't care about this one */ 870 /* then, hash function */ 871 i = ( (id->dst_ip) & 0xffff ) ^ 872 ( (id->dst_ip >> 15) & 0xffff ) ^ 873 ( (id->src_ip << 1) & 0xffff ) ^ 874 ( (id->src_ip >> 16 ) & 0xffff ) ^ 875 (id->dst_port << 1) ^ (id->src_port) ^ 876 (id->proto ); 877 i = i % fs->rq_size ; 878 /* finally, scan the current list for a match */ 879 searches++ ; 880 for (prev=NULL, q = fs->rq[i] ; q ; ) { 881 search_steps++; 882 if (bcmp(id, &(q->id), sizeof(q->id) ) == 0) 883 break ; /* found */ 884 else if (pipe_expire && q->head == NULL && q->S == q->F+1 ) { 885 /* entry is idle and not in any heap, expire it */ 886 struct dn_flow_queue *old_q = q ; 887 888 if (prev != NULL) 889 prev->next = q = q->next ; 890 else 891 fs->rq[i] = q = q->next ; 892 fs->rq_elements-- ; 893 free(old_q, M_DUMMYNET); 894 continue ; 895 } 896 prev = q ; 897 q = q->next ; 898 } 899 if (q && prev != NULL) { /* found and not in front */ 900 prev->next = q->next ; 901 q->next = fs->rq[i] ; 902 fs->rq[i] = q ; 903 } 904 } 905 if (q == NULL) { /* no match, need to allocate a new entry */ 906 q = create_queue(fs, i); 907 if (q != NULL) 908 q->id = *id ; 909 } 910 return q ; 911} 912 913static int 914red_drops(struct dn_flow_set *fs, struct dn_flow_queue *q, int len) 915{ 916 /* 917 * RED algorithm 918 * 919 * RED calculates the average queue size (avg) using a low-pass filter 920 * with an exponential weighted (w_q) moving average: 921 * avg <- (1-w_q) * avg + w_q * q_size 922 * where q_size is the queue length (measured in bytes or * packets). 923 * 924 * If q_size == 0, we compute the idle time for the link, and set 925 * avg = (1 - w_q)^(idle/s) 926 * where s is the time needed for transmitting a medium-sized packet. 927 * 928 * Now, if avg < min_th the packet is enqueued. 929 * If avg > max_th the packet is dropped. Otherwise, the packet is 930 * dropped with probability P function of avg. 931 * 932 */ 933 934 int64_t p_b = 0; 935 /* queue in bytes or packets ? */ 936 u_int q_size = (fs->flags_fs & DN_QSIZE_IS_BYTES) ? q->len_bytes : q->len; 937 938 DEB(printf("\n%d q: %2u ", (int) curr_time, q_size);) 939 940 /* average queue size estimation */ 941 if (q_size != 0) { 942 /* 943 * queue is not empty, avg <- avg + (q_size - avg) * w_q 944 */ 945 int diff = SCALE(q_size) - q->avg; 946 int64_t v = SCALE_MUL((int64_t) diff, (int64_t) fs->w_q); 947 948 q->avg += (int) v; 949 } else { 950 /* 951 * queue is empty, find for how long the queue has been 952 * empty and use a lookup table for computing 953 * (1 - * w_q)^(idle_time/s) where s is the time to send a 954 * (small) packet. 955 * XXX check wraps... 956 */ 957 if (q->avg) { 958 u_int t = (curr_time - q->q_time) / fs->lookup_step; 959 960 q->avg = (t < fs->lookup_depth) ? 961 SCALE_MUL(q->avg, fs->w_q_lookup[t]) : 0; 962 } 963 } 964 DEB(printf("avg: %u ", SCALE_VAL(q->avg));) 965 966 /* should i drop ? */ 967 968 if (q->avg < fs->min_th) { 969 q->count = -1; 970 return 0; /* accept packet ; */ 971 } 972 if (q->avg >= fs->max_th) { /* average queue >= max threshold */ 973 if (fs->flags_fs & DN_IS_GENTLE_RED) { 974 /* 975 * According to Gentle-RED, if avg is greater than max_th the 976 * packet is dropped with a probability 977 * p_b = c_3 * avg - c_4 978 * where c_3 = (1 - max_p) / max_th, and c_4 = 1 - 2 * max_p 979 */ 980 p_b = SCALE_MUL((int64_t) fs->c_3, (int64_t) q->avg) - fs->c_4; 981 } else { 982 q->count = -1; 983 printf("- drop"); 984 return 1 ; 985 } 986 } else if (q->avg > fs->min_th) { 987 /* 988 * we compute p_b using the linear dropping function p_b = c_1 * 989 * avg - c_2, where c_1 = max_p / (max_th - min_th), and c_2 = 990 * max_p * min_th / (max_th - min_th) 991 */ 992 p_b = SCALE_MUL((int64_t) fs->c_1, (int64_t) q->avg) - fs->c_2; 993 } 994 if (fs->flags_fs & DN_QSIZE_IS_BYTES) 995 p_b = (p_b * len) / fs->max_pkt_size; 996 if (++q->count == 0) 997 q->random = random() & 0xffff; 998 else { 999 /* 1000 * q->count counts packets arrived since last drop, so a greater 1001 * value of q->count means a greater packet drop probability. 1002 */ 1003 if (SCALE_MUL(p_b, SCALE((int64_t) q->count)) > q->random) { 1004 q->count = 0; 1005 DEB(printf("- red drop");) 1006 /* after a drop we calculate a new random value */ 1007 q->random = random() & 0xffff; 1008 return 1; /* drop */ 1009 } 1010 } 1011 /* end of RED algorithm */ 1012 return 0 ; /* accept */ 1013} 1014 1015static __inline 1016struct dn_flow_set * 1017locate_flowset(int pipe_nr, struct ip_fw *rule) 1018{ 1019#if IPFW2 1020 struct dn_flow_set *fs; 1021 ipfw_insn *cmd = rule->cmd + rule->act_ofs; 1022 1023 if (cmd->opcode == O_LOG) 1024 cmd += F_LEN(cmd); 1025 fs = ((ipfw_insn_pipe *)cmd)->pipe_ptr; 1026 1027 if (fs != NULL) 1028 return fs; 1029 1030 if (cmd->opcode == O_QUEUE) 1031#else /* !IPFW2 */ 1032 struct dn_flow_set *fs = NULL ; 1033 1034 if ( (rule->fw_flg & IP_FW_F_COMMAND) == IP_FW_F_QUEUE ) 1035#endif /* !IPFW2 */ 1036 for (fs=all_flow_sets; fs && fs->fs_nr != pipe_nr; fs=fs->next) 1037 ; 1038 else { 1039 struct dn_pipe *p1; 1040 for (p1 = all_pipes; p1 && p1->pipe_nr != pipe_nr; p1 = p1->next) 1041 ; 1042 if (p1 != NULL) 1043 fs = &(p1->fs) ; 1044 } 1045 /* record for the future */ 1046#if IPFW2 1047 ((ipfw_insn_pipe *)cmd)->pipe_ptr = fs; 1048#else 1049 if (fs != NULL) 1050 rule->pipe_ptr = fs; 1051#endif 1052 return fs ; 1053} 1054 1055/* 1056 * dummynet hook for packets. Below 'pipe' is a pipe or a queue 1057 * depending on whether WF2Q or fixed bw is used. 1058 * 1059 * pipe_nr pipe or queue the packet is destined for. 1060 * dir where shall we send the packet after dummynet. 1061 * m the mbuf with the packet 1062 * ifp the 'ifp' parameter from the caller. 1063 * NULL in ip_input, destination interface in ip_output, 1064 * real_dst in bdg_forward 1065 * ro route parameter (only used in ip_output, NULL otherwise) 1066 * dst destination address, only used by ip_output 1067 * rule matching rule, in case of multiple passes 1068 * flags flags from the caller, only used in ip_output 1069 * 1070 */ 1071static int 1072dummynet_io(struct mbuf *m, int pipe_nr, int dir, struct ip_fw_args *fwa) 1073{ 1074 struct dn_pkt *pkt; 1075 struct dn_flow_set *fs; 1076 struct dn_pipe *pipe ; 1077 u_int64_t len = m->m_pkthdr.len ; 1078 struct dn_flow_queue *q = NULL ; 1079 int s = splimp(); 1080 int is_pipe; 1081#if IPFW2 1082 ipfw_insn *cmd = fwa->rule->cmd + fwa->rule->act_ofs; 1083 1084 if (cmd->opcode == O_LOG) 1085 cmd += F_LEN(cmd); 1086 is_pipe = (cmd->opcode == O_PIPE); 1087#else 1088 is_pipe = (fwa->rule->fw_flg & IP_FW_F_COMMAND) == IP_FW_F_PIPE; 1089#endif 1090 1091 pipe_nr &= 0xffff ; 1092 1093 /* 1094 * this is a dummynet rule, so we expect a O_PIPE or O_QUEUE rule 1095 */ 1096 fs = locate_flowset(pipe_nr, fwa->rule); 1097 if (fs == NULL) 1098 goto dropit ; /* this queue/pipe does not exist! */ 1099 pipe = fs->pipe ; 1100 if (pipe == NULL) { /* must be a queue, try find a matching pipe */ 1101 for (pipe = all_pipes; pipe && pipe->pipe_nr != fs->parent_nr; 1102 pipe = pipe->next) 1103 ; 1104 if (pipe != NULL) 1105 fs->pipe = pipe ; 1106 else { 1107 printf("No pipe %d for queue %d, drop pkt\n", 1108 fs->parent_nr, fs->fs_nr); 1109 goto dropit ; 1110 } 1111 } 1112 q = find_queue(fs, &(fwa->f_id)); 1113 if ( q == NULL ) 1114 goto dropit ; /* cannot allocate queue */ 1115 /* 1116 * update statistics, then check reasons to drop pkt 1117 */ 1118 q->tot_bytes += len ; 1119 q->tot_pkts++ ; 1120 if ( fs->plr && random() < fs->plr ) 1121 goto dropit ; /* random pkt drop */ 1122 if ( fs->flags_fs & DN_QSIZE_IS_BYTES) { 1123 if (q->len_bytes > fs->qsize) 1124 goto dropit ; /* queue size overflow */ 1125 } else { 1126 if (q->len >= fs->qsize) 1127 goto dropit ; /* queue count overflow */ 1128 } 1129 if ( fs->flags_fs & DN_IS_RED && red_drops(fs, q, len) ) 1130 goto dropit ; 1131 1132 /* XXX expensive to zero, see if we can remove it*/ 1133 pkt = (struct dn_pkt *)malloc(sizeof (*pkt), M_DUMMYNET, M_NOWAIT|M_ZERO); 1134 if ( pkt == NULL ) 1135 goto dropit ; /* cannot allocate packet header */ 1136 /* ok, i can handle the pkt now... */ 1137 /* build and enqueue packet + parameters */ 1138 pkt->hdr.mh_type = MT_TAG; 1139 pkt->hdr.mh_flags = PACKET_TAG_DUMMYNET; 1140 pkt->rule = fwa->rule ; 1141 DN_NEXT(pkt) = NULL; 1142 pkt->dn_m = m; 1143 pkt->dn_dir = dir ; 1144 1145 pkt->ifp = fwa->oif; 1146 if (dir == DN_TO_IP_OUT) { 1147 /* 1148 * We need to copy *ro because for ICMP pkts (and maybe others) 1149 * the caller passed a pointer into the stack; dst might also be 1150 * a pointer into *ro so it needs to be updated. 1151 */ 1152 pkt->ro = *(fwa->ro); 1153 if (fwa->ro->ro_rt) 1154 fwa->ro->ro_rt->rt_refcnt++ ; 1155 if (fwa->dst == (struct sockaddr_in *)&fwa->ro->ro_dst) /* dst points into ro */ 1156 fwa->dst = (struct sockaddr_in *)&(pkt->ro.ro_dst) ; 1157 1158 pkt->dn_dst = fwa->dst; 1159 pkt->flags = fwa->flags; 1160 } 1161 if (q->head == NULL) 1162 q->head = pkt; 1163 else 1164 DN_NEXT(q->tail) = pkt; 1165 q->tail = pkt; 1166 q->len++; 1167 q->len_bytes += len ; 1168 1169 if ( q->head != pkt ) /* flow was not idle, we are done */ 1170 goto done; 1171 /* 1172 * If we reach this point the flow was previously idle, so we need 1173 * to schedule it. This involves different actions for fixed-rate or 1174 * WF2Q queues. 1175 */ 1176 if (is_pipe) { 1177 /* 1178 * Fixed-rate queue: just insert into the ready_heap. 1179 */ 1180 dn_key t = 0 ; 1181 if (pipe->bandwidth) 1182 t = SET_TICKS(pkt, q, pipe); 1183 q->sched_time = curr_time ; 1184 if (t == 0) /* must process it now */ 1185 ready_event( q ); 1186 else 1187 heap_insert(&ready_heap, curr_time + t , q ); 1188 } else { 1189 /* 1190 * WF2Q. First, compute start time S: if the flow was idle (S=F+1) 1191 * set S to the virtual time V for the controlling pipe, and update 1192 * the sum of weights for the pipe; otherwise, remove flow from 1193 * idle_heap and set S to max(F,V). 1194 * Second, compute finish time F = S + len/weight. 1195 * Third, if pipe was idle, update V=max(S, V). 1196 * Fourth, count one more backlogged flow. 1197 */ 1198 if (DN_KEY_GT(q->S, q->F)) { /* means timestamps are invalid */ 1199 q->S = pipe->V ; 1200 pipe->sum += fs->weight ; /* add weight of new queue */ 1201 } else { 1202 heap_extract(&(pipe->idle_heap), q); 1203 q->S = MAX64(q->F, pipe->V ) ; 1204 } 1205 q->F = q->S + ( len<<MY_M )/(u_int64_t) fs->weight; 1206 1207 if (pipe->not_eligible_heap.elements == 0 && 1208 pipe->scheduler_heap.elements == 0) 1209 pipe->V = MAX64 ( q->S, pipe->V ); 1210 fs->backlogged++ ; 1211 /* 1212 * Look at eligibility. A flow is not eligibile if S>V (when 1213 * this happens, it means that there is some other flow already 1214 * scheduled for the same pipe, so the scheduler_heap cannot be 1215 * empty). If the flow is not eligible we just store it in the 1216 * not_eligible_heap. Otherwise, we store in the scheduler_heap 1217 * and possibly invoke ready_event_wfq() right now if there is 1218 * leftover credit. 1219 * Note that for all flows in scheduler_heap (SCH), S_i <= V, 1220 * and for all flows in not_eligible_heap (NEH), S_i > V . 1221 * So when we need to compute max( V, min(S_i) ) forall i in SCH+NEH, 1222 * we only need to look into NEH. 1223 */ 1224 if (DN_KEY_GT(q->S, pipe->V) ) { /* not eligible */ 1225 if (pipe->scheduler_heap.elements == 0) 1226 printf("++ ouch! not eligible but empty scheduler!\n"); 1227 heap_insert(&(pipe->not_eligible_heap), q->S, q); 1228 } else { 1229 heap_insert(&(pipe->scheduler_heap), q->F, q); 1230 if (pipe->numbytes >= 0) { /* pipe is idle */ 1231 if (pipe->scheduler_heap.elements != 1) 1232 printf("*** OUCH! pipe should have been idle!\n"); 1233 DEB(printf("Waking up pipe %d at %d\n", 1234 pipe->pipe_nr, (int)(q->F >> MY_M)); ) 1235 pipe->sched_time = curr_time ; 1236 ready_event_wfq(pipe); 1237 } 1238 } 1239 } 1240done: 1241 splx(s); 1242 return 0; 1243 1244dropit: 1245 splx(s); 1246 if (q) 1247 q->drops++ ; 1248 m_freem(m); 1249 return ( (fs && (fs->flags_fs & DN_NOERROR)) ? 0 : ENOBUFS); 1250} 1251 1252/* 1253 * Below, the rt_unref is only needed when (pkt->dn_dir == DN_TO_IP_OUT) 1254 * Doing this would probably save us the initial bzero of dn_pkt 1255 */ 1256#define DN_FREE_PKT(pkt) { \ 1257 struct dn_pkt *n = pkt ; \ 1258 rt_unref ( n->ro.ro_rt ) ; \ 1259 m_freem(n->dn_m); \ 1260 pkt = DN_NEXT(n) ; \ 1261 free(n, M_DUMMYNET) ; } 1262 1263/* 1264 * Dispose all packets and flow_queues on a flow_set. 1265 * If all=1, also remove red lookup table and other storage, 1266 * including the descriptor itself. 1267 * For the one in dn_pipe MUST also cleanup ready_heap... 1268 */ 1269static void 1270purge_flow_set(struct dn_flow_set *fs, int all) 1271{ 1272 struct dn_pkt *pkt ; 1273 struct dn_flow_queue *q, *qn ; 1274 int i ; 1275 1276 for (i = 0 ; i <= fs->rq_size ; i++ ) { 1277 for (q = fs->rq[i] ; q ; q = qn ) { 1278 for (pkt = q->head ; pkt ; ) 1279 DN_FREE_PKT(pkt) ; 1280 qn = q->next ; 1281 free(q, M_DUMMYNET); 1282 } 1283 fs->rq[i] = NULL ; 1284 } 1285 fs->rq_elements = 0 ; 1286 if (all) { 1287 /* RED - free lookup table */ 1288 if (fs->w_q_lookup) 1289 free(fs->w_q_lookup, M_DUMMYNET); 1290 if (fs->rq) 1291 free(fs->rq, M_DUMMYNET); 1292 /* if this fs is not part of a pipe, free it */ 1293 if (fs->pipe && fs != &(fs->pipe->fs) ) 1294 free(fs, M_DUMMYNET); 1295 } 1296} 1297 1298/* 1299 * Dispose all packets queued on a pipe (not a flow_set). 1300 * Also free all resources associated to a pipe, which is about 1301 * to be deleted. 1302 */ 1303static void 1304purge_pipe(struct dn_pipe *pipe) 1305{ 1306 struct dn_pkt *pkt ; 1307 1308 purge_flow_set( &(pipe->fs), 1 ); 1309 1310 for (pkt = pipe->head ; pkt ; ) 1311 DN_FREE_PKT(pkt) ; 1312 1313 heap_free( &(pipe->scheduler_heap) ); 1314 heap_free( &(pipe->not_eligible_heap) ); 1315 heap_free( &(pipe->idle_heap) ); 1316} 1317 1318/* 1319 * Delete all pipes and heaps returning memory. Must also 1320 * remove references from all ipfw rules to all pipes. 1321 */ 1322static void 1323dummynet_flush() 1324{ 1325 struct dn_pipe *curr_p, *p ; 1326 struct dn_flow_set *fs, *curr_fs; 1327 int s ; 1328 1329 s = splimp() ; 1330 1331 /* remove all references to pipes ...*/ 1332 flush_pipe_ptrs(NULL); 1333 /* prevent future matches... */ 1334 p = all_pipes ; 1335 all_pipes = NULL ; 1336 fs = all_flow_sets ; 1337 all_flow_sets = NULL ; 1338 /* and free heaps so we don't have unwanted events */ 1339 heap_free(&ready_heap); 1340 heap_free(&wfq_ready_heap); 1341 heap_free(&extract_heap); 1342 splx(s) ; 1343 /* 1344 * Now purge all queued pkts and delete all pipes 1345 */ 1346 /* scan and purge all flow_sets. */ 1347 for ( ; fs ; ) { 1348 curr_fs = fs ; 1349 fs = fs->next ; 1350 purge_flow_set(curr_fs, 1); 1351 } 1352 for ( ; p ; ) { 1353 purge_pipe(p); 1354 curr_p = p ; 1355 p = p->next ; 1356 free(curr_p, M_DUMMYNET); 1357 } 1358} 1359 1360 1361extern struct ip_fw *ip_fw_default_rule ; 1362static void 1363dn_rule_delete_fs(struct dn_flow_set *fs, void *r) 1364{ 1365 int i ; 1366 struct dn_flow_queue *q ; 1367 struct dn_pkt *pkt ; 1368 1369 for (i = 0 ; i <= fs->rq_size ; i++) /* last one is ovflow */ 1370 for (q = fs->rq[i] ; q ; q = q->next ) 1371 for (pkt = q->head ; pkt ; pkt = DN_NEXT(pkt) ) 1372 if (pkt->rule == r) 1373 pkt->rule = ip_fw_default_rule ; 1374} 1375/* 1376 * when a firewall rule is deleted, scan all queues and remove the flow-id 1377 * from packets matching this rule. 1378 */ 1379void 1380dn_rule_delete(void *r) 1381{ 1382 struct dn_pipe *p ; 1383 struct dn_pkt *pkt ; 1384 struct dn_flow_set *fs ; 1385 1386 /* 1387 * If the rule references a queue (dn_flow_set), then scan 1388 * the flow set, otherwise scan pipes. Should do either, but doing 1389 * both does not harm. 1390 */ 1391 for ( fs = all_flow_sets ; fs ; fs = fs->next ) 1392 dn_rule_delete_fs(fs, r); 1393 for ( p = all_pipes ; p ; p = p->next ) { 1394 fs = &(p->fs) ; 1395 dn_rule_delete_fs(fs, r); 1396 for (pkt = p->head ; pkt ; pkt = DN_NEXT(pkt) ) 1397 if (pkt->rule == r) 1398 pkt->rule = ip_fw_default_rule ; 1399 } 1400} 1401 1402/* 1403 * setup RED parameters 1404 */ 1405static int 1406config_red(struct dn_flow_set *p, struct dn_flow_set * x) 1407{ 1408 int i; 1409 1410 x->w_q = p->w_q; 1411 x->min_th = SCALE(p->min_th); 1412 x->max_th = SCALE(p->max_th); 1413 x->max_p = p->max_p; 1414 1415 x->c_1 = p->max_p / (p->max_th - p->min_th); 1416 x->c_2 = SCALE_MUL(x->c_1, SCALE(p->min_th)); 1417 if (x->flags_fs & DN_IS_GENTLE_RED) { 1418 x->c_3 = (SCALE(1) - p->max_p) / p->max_th; 1419 x->c_4 = (SCALE(1) - 2 * p->max_p); 1420 } 1421 1422 /* if the lookup table already exist, free and create it again */ 1423 if (x->w_q_lookup) { 1424 free(x->w_q_lookup, M_DUMMYNET); 1425 x->w_q_lookup = NULL ; 1426 } 1427 if (red_lookup_depth == 0) { 1428 printf("\nnet.inet.ip.dummynet.red_lookup_depth must be > 0"); 1429 free(x, M_DUMMYNET); 1430 return EINVAL; 1431 } 1432 x->lookup_depth = red_lookup_depth; 1433 x->w_q_lookup = (u_int *) malloc(x->lookup_depth * sizeof(int), 1434 M_DUMMYNET, M_DONTWAIT); 1435 if (x->w_q_lookup == NULL) { 1436 printf("sorry, cannot allocate red lookup table\n"); 1437 free(x, M_DUMMYNET); 1438 return ENOSPC; 1439 } 1440 1441 /* fill the lookup table with (1 - w_q)^x */ 1442 x->lookup_step = p->lookup_step ; 1443 x->lookup_weight = p->lookup_weight ; 1444 x->w_q_lookup[0] = SCALE(1) - x->w_q; 1445 for (i = 1; i < x->lookup_depth; i++) 1446 x->w_q_lookup[i] = SCALE_MUL(x->w_q_lookup[i - 1], x->lookup_weight); 1447 if (red_avg_pkt_size < 1) 1448 red_avg_pkt_size = 512 ; 1449 x->avg_pkt_size = red_avg_pkt_size ; 1450 if (red_max_pkt_size < 1) 1451 red_max_pkt_size = 1500 ; 1452 x->max_pkt_size = red_max_pkt_size ; 1453 return 0 ; 1454} 1455 1456static int 1457alloc_hash(struct dn_flow_set *x, struct dn_flow_set *pfs) 1458{ 1459 if (x->flags_fs & DN_HAVE_FLOW_MASK) { /* allocate some slots */ 1460 int l = pfs->rq_size; 1461 1462 if (l == 0) 1463 l = dn_hash_size; 1464 if (l < 4) 1465 l = 4; 1466 else if (l > DN_MAX_HASH_SIZE) 1467 l = DN_MAX_HASH_SIZE; 1468 x->rq_size = l; 1469 } else /* one is enough for null mask */ 1470 x->rq_size = 1; 1471 x->rq = malloc((1 + x->rq_size) * sizeof(struct dn_flow_queue *), 1472 M_DUMMYNET, M_DONTWAIT | M_ZERO); 1473 if (x->rq == NULL) { 1474 printf("sorry, cannot allocate queue\n"); 1475 return ENOSPC; 1476 } 1477 x->rq_elements = 0; 1478 return 0 ; 1479} 1480 1481static void 1482set_fs_parms(struct dn_flow_set *x, struct dn_flow_set *src) 1483{ 1484 x->flags_fs = src->flags_fs; 1485 x->qsize = src->qsize; 1486 x->plr = src->plr; 1487 x->flow_mask = src->flow_mask; 1488 if (x->flags_fs & DN_QSIZE_IS_BYTES) { 1489 if (x->qsize > 1024*1024) 1490 x->qsize = 1024*1024 ; 1491 } else { 1492 if (x->qsize == 0) 1493 x->qsize = 50 ; 1494 if (x->qsize > 100) 1495 x->qsize = 50 ; 1496 } 1497 /* configuring RED */ 1498 if ( x->flags_fs & DN_IS_RED ) 1499 config_red(src, x) ; /* XXX should check errors */ 1500} 1501 1502/* 1503 * setup pipe or queue parameters. 1504 */ 1505 1506static int 1507config_pipe(struct dn_pipe *p) 1508{ 1509 int s ; 1510 struct dn_flow_set *pfs = &(p->fs); 1511 1512 /* 1513 * The config program passes parameters as follows: 1514 * bw = bits/second (0 means no limits), 1515 * delay = ms, must be translated into ticks. 1516 * qsize = slots/bytes 1517 */ 1518 p->delay = ( p->delay * hz ) / 1000 ; 1519 /* We need either a pipe number or a flow_set number */ 1520 if (p->pipe_nr == 0 && pfs->fs_nr == 0) 1521 return EINVAL ; 1522 if (p->pipe_nr != 0 && pfs->fs_nr != 0) 1523 return EINVAL ; 1524 if (p->pipe_nr != 0) { /* this is a pipe */ 1525 struct dn_pipe *x, *a, *b; 1526 /* locate pipe */ 1527 for (a = NULL , b = all_pipes ; b && b->pipe_nr < p->pipe_nr ; 1528 a = b , b = b->next) ; 1529 1530 if (b == NULL || b->pipe_nr != p->pipe_nr) { /* new pipe */ 1531 x = malloc(sizeof(struct dn_pipe), M_DUMMYNET, M_DONTWAIT | M_ZERO); 1532 if (x == NULL) { 1533 printf("ip_dummynet.c: no memory for new pipe\n"); 1534 return ENOSPC; 1535 } 1536 x->pipe_nr = p->pipe_nr; 1537 x->fs.pipe = x ; 1538 /* idle_heap is the only one from which we extract from the middle. 1539 */ 1540 x->idle_heap.size = x->idle_heap.elements = 0 ; 1541 x->idle_heap.offset=OFFSET_OF(struct dn_flow_queue, heap_pos); 1542 } else 1543 x = b; 1544 1545 x->bandwidth = p->bandwidth ; 1546 x->numbytes = 0; /* just in case... */ 1547 bcopy(p->if_name, x->if_name, sizeof(p->if_name) ); 1548 x->ifp = NULL ; /* reset interface ptr */ 1549 x->delay = p->delay ; 1550 set_fs_parms(&(x->fs), pfs); 1551 1552 1553 if ( x->fs.rq == NULL ) { /* a new pipe */ 1554 s = alloc_hash(&(x->fs), pfs) ; 1555 if (s) { 1556 free(x, M_DUMMYNET); 1557 return s ; 1558 } 1559 s = splimp() ; 1560 x->next = b ; 1561 if (a == NULL) 1562 all_pipes = x ; 1563 else 1564 a->next = x ; 1565 splx(s); 1566 } 1567 } else { /* config queue */ 1568 struct dn_flow_set *x, *a, *b ; 1569 1570 /* locate flow_set */ 1571 for (a=NULL, b=all_flow_sets ; b && b->fs_nr < pfs->fs_nr ; 1572 a = b , b = b->next) ; 1573 1574 if (b == NULL || b->fs_nr != pfs->fs_nr) { /* new */ 1575 if (pfs->parent_nr == 0) /* need link to a pipe */ 1576 return EINVAL ; 1577 x = malloc(sizeof(struct dn_flow_set),M_DUMMYNET,M_DONTWAIT|M_ZERO); 1578 if (x == NULL) { 1579 printf("ip_dummynet.c: no memory for new flow_set\n"); 1580 return ENOSPC; 1581 } 1582 x->fs_nr = pfs->fs_nr; 1583 x->parent_nr = pfs->parent_nr; 1584 x->weight = pfs->weight ; 1585 if (x->weight == 0) 1586 x->weight = 1 ; 1587 else if (x->weight > 100) 1588 x->weight = 100 ; 1589 } else { 1590 /* Change parent pipe not allowed; must delete and recreate */ 1591 if (pfs->parent_nr != 0 && b->parent_nr != pfs->parent_nr) 1592 return EINVAL ; 1593 x = b; 1594 } 1595 set_fs_parms(x, pfs); 1596 1597 if ( x->rq == NULL ) { /* a new flow_set */ 1598 s = alloc_hash(x, pfs) ; 1599 if (s) { 1600 free(x, M_DUMMYNET); 1601 return s ; 1602 } 1603 s = splimp() ; 1604 x->next = b; 1605 if (a == NULL) 1606 all_flow_sets = x; 1607 else 1608 a->next = x; 1609 splx(s); 1610 } 1611 } 1612 return 0 ; 1613} 1614 1615/* 1616 * Helper function to remove from a heap queues which are linked to 1617 * a flow_set about to be deleted. 1618 */ 1619static void 1620fs_remove_from_heap(struct dn_heap *h, struct dn_flow_set *fs) 1621{ 1622 int i = 0, found = 0 ; 1623 for (; i < h->elements ;) 1624 if ( ((struct dn_flow_queue *)h->p[i].object)->fs == fs) { 1625 h->elements-- ; 1626 h->p[i] = h->p[h->elements] ; 1627 found++ ; 1628 } else 1629 i++ ; 1630 if (found) 1631 heapify(h); 1632} 1633 1634/* 1635 * helper function to remove a pipe from a heap (can be there at most once) 1636 */ 1637static void 1638pipe_remove_from_heap(struct dn_heap *h, struct dn_pipe *p) 1639{ 1640 if (h->elements > 0) { 1641 int i = 0 ; 1642 for (i=0; i < h->elements ; i++ ) { 1643 if (h->p[i].object == p) { /* found it */ 1644 h->elements-- ; 1645 h->p[i] = h->p[h->elements] ; 1646 heapify(h); 1647 break ; 1648 } 1649 } 1650 } 1651} 1652 1653/* 1654 * drain all queues. Called in case of severe mbuf shortage. 1655 */ 1656void 1657dummynet_drain() 1658{ 1659 struct dn_flow_set *fs; 1660 struct dn_pipe *p; 1661 struct dn_pkt *pkt; 1662 1663 heap_free(&ready_heap); 1664 heap_free(&wfq_ready_heap); 1665 heap_free(&extract_heap); 1666 /* remove all references to this pipe from flow_sets */ 1667 for (fs = all_flow_sets; fs; fs= fs->next ) 1668 purge_flow_set(fs, 0); 1669 1670 for (p = all_pipes; p; p= p->next ) { 1671 purge_flow_set(&(p->fs), 0); 1672 for (pkt = p->head ; pkt ; ) 1673 DN_FREE_PKT(pkt) ; 1674 p->head = p->tail = NULL ; 1675 } 1676} 1677 1678/* 1679 * Fully delete a pipe or a queue, cleaning up associated info. 1680 */ 1681static int 1682delete_pipe(struct dn_pipe *p) 1683{ 1684 int s ; 1685 1686 if (p->pipe_nr == 0 && p->fs.fs_nr == 0) 1687 return EINVAL ; 1688 if (p->pipe_nr != 0 && p->fs.fs_nr != 0) 1689 return EINVAL ; 1690 if (p->pipe_nr != 0) { /* this is an old-style pipe */ 1691 struct dn_pipe *a, *b; 1692 struct dn_flow_set *fs; 1693 1694 /* locate pipe */ 1695 for (a = NULL , b = all_pipes ; b && b->pipe_nr < p->pipe_nr ; 1696 a = b , b = b->next) ; 1697 if (b == NULL || (b->pipe_nr != p->pipe_nr) ) 1698 return EINVAL ; /* not found */ 1699 1700 s = splimp() ; 1701 1702 /* unlink from list of pipes */ 1703 if (a == NULL) 1704 all_pipes = b->next ; 1705 else 1706 a->next = b->next ; 1707 /* remove references to this pipe from the ip_fw rules. */ 1708 flush_pipe_ptrs(&(b->fs)); 1709 1710 /* remove all references to this pipe from flow_sets */ 1711 for (fs = all_flow_sets; fs; fs= fs->next ) 1712 if (fs->pipe == b) { 1713 printf("++ ref to pipe %d from fs %d\n", 1714 p->pipe_nr, fs->fs_nr); 1715 fs->pipe = NULL ; 1716 purge_flow_set(fs, 0); 1717 } 1718 fs_remove_from_heap(&ready_heap, &(b->fs)); 1719 purge_pipe(b); /* remove all data associated to this pipe */ 1720 /* remove reference to here from extract_heap and wfq_ready_heap */ 1721 pipe_remove_from_heap(&extract_heap, b); 1722 pipe_remove_from_heap(&wfq_ready_heap, b); 1723 splx(s); 1724 free(b, M_DUMMYNET); 1725 } else { /* this is a WF2Q queue (dn_flow_set) */ 1726 struct dn_flow_set *a, *b; 1727 1728 /* locate set */ 1729 for (a = NULL, b = all_flow_sets ; b && b->fs_nr < p->fs.fs_nr ; 1730 a = b , b = b->next) ; 1731 if (b == NULL || (b->fs_nr != p->fs.fs_nr) ) 1732 return EINVAL ; /* not found */ 1733 1734 s = splimp() ; 1735 if (a == NULL) 1736 all_flow_sets = b->next ; 1737 else 1738 a->next = b->next ; 1739 /* remove references to this flow_set from the ip_fw rules. */ 1740 flush_pipe_ptrs(b); 1741 1742 if (b->pipe != NULL) { 1743 /* Update total weight on parent pipe and cleanup parent heaps */ 1744 b->pipe->sum -= b->weight * b->backlogged ; 1745 fs_remove_from_heap(&(b->pipe->not_eligible_heap), b); 1746 fs_remove_from_heap(&(b->pipe->scheduler_heap), b); 1747#if 1 /* XXX should i remove from idle_heap as well ? */ 1748 fs_remove_from_heap(&(b->pipe->idle_heap), b); 1749#endif 1750 } 1751 purge_flow_set(b, 1); 1752 splx(s); 1753 } 1754 return 0 ; 1755} 1756 1757/* 1758 * helper function used to copy data from kernel in DUMMYNET_GET 1759 */ 1760static char * 1761dn_copy_set(struct dn_flow_set *set, char *bp) 1762{ 1763 int i, copied = 0 ; 1764 struct dn_flow_queue *q, *qp = (struct dn_flow_queue *)bp; 1765 1766 for (i = 0 ; i <= set->rq_size ; i++) 1767 for (q = set->rq[i] ; q ; q = q->next, qp++ ) { 1768 if (q->hash_slot != i) 1769 printf("++ at %d: wrong slot (have %d, " 1770 "should be %d)\n", copied, q->hash_slot, i); 1771 if (q->fs != set) 1772 printf("++ at %d: wrong fs ptr (have %p, should be %p)\n", 1773 i, q->fs, set); 1774 copied++ ; 1775 bcopy(q, qp, sizeof( *q ) ); 1776 /* cleanup pointers */ 1777 qp->next = NULL ; 1778 qp->head = qp->tail = NULL ; 1779 qp->fs = NULL ; 1780 } 1781 if (copied != set->rq_elements) 1782 printf("++ wrong count, have %d should be %d\n", 1783 copied, set->rq_elements); 1784 return (char *)qp ; 1785} 1786 1787static int 1788dummynet_get(struct sockopt *sopt) 1789{ 1790 char *buf, *bp ; /* bp is the "copy-pointer" */ 1791 size_t size ; 1792 struct dn_flow_set *set ; 1793 struct dn_pipe *p ; 1794 int s, error=0 ; 1795 1796 s = splimp(); 1797 /* 1798 * compute size of data structures: list of pipes and flow_sets. 1799 */ 1800 for (p = all_pipes, size = 0 ; p ; p = p->next ) 1801 size += sizeof( *p ) + 1802 p->fs.rq_elements * sizeof(struct dn_flow_queue); 1803 for (set = all_flow_sets ; set ; set = set->next ) 1804 size += sizeof ( *set ) + 1805 set->rq_elements * sizeof(struct dn_flow_queue); 1806 buf = malloc(size, M_TEMP, M_DONTWAIT); 1807 if (buf == 0) { 1808 splx(s); 1809 return ENOBUFS ; 1810 } 1811 for (p = all_pipes, bp = buf ; p ; p = p->next ) { 1812 struct dn_pipe *pipe_bp = (struct dn_pipe *)bp ; 1813 1814 /* 1815 * copy pipe descriptor into *bp, convert delay back to ms, 1816 * then copy the flow_set descriptor(s) one at a time. 1817 * After each flow_set, copy the queue descriptor it owns. 1818 */ 1819 bcopy(p, bp, sizeof( *p ) ); 1820 pipe_bp->delay = (pipe_bp->delay * 1000) / hz ; 1821 /* 1822 * XXX the following is a hack based on ->next being the 1823 * first field in dn_pipe and dn_flow_set. The correct 1824 * solution would be to move the dn_flow_set to the beginning 1825 * of struct dn_pipe. 1826 */ 1827 pipe_bp->next = (struct dn_pipe *)DN_IS_PIPE ; 1828 /* clean pointers */ 1829 pipe_bp->head = pipe_bp->tail = NULL ; 1830 pipe_bp->fs.next = NULL ; 1831 pipe_bp->fs.pipe = NULL ; 1832 pipe_bp->fs.rq = NULL ; 1833 1834 bp += sizeof( *p ) ; 1835 bp = dn_copy_set( &(p->fs), bp ); 1836 } 1837 for (set = all_flow_sets ; set ; set = set->next ) { 1838 struct dn_flow_set *fs_bp = (struct dn_flow_set *)bp ; 1839 bcopy(set, bp, sizeof( *set ) ); 1840 /* XXX same hack as above */ 1841 fs_bp->next = (struct dn_flow_set *)DN_IS_QUEUE ; 1842 fs_bp->pipe = NULL ; 1843 fs_bp->rq = NULL ; 1844 bp += sizeof( *set ) ; 1845 bp = dn_copy_set( set, bp ); 1846 } 1847 splx(s); 1848 error = sooptcopyout(sopt, buf, size); 1849 free(buf, M_TEMP); 1850 return error ; 1851} 1852 1853/* 1854 * Handler for the various dummynet socket options (get, flush, config, del) 1855 */ 1856static int 1857ip_dn_ctl(struct sockopt *sopt) 1858{ 1859 int error = 0 ; 1860 struct dn_pipe *p, tmp_pipe; 1861 1862 /* Disallow sets in really-really secure mode. */ 1863 if (sopt->sopt_dir == SOPT_SET) { 1864#if __FreeBSD_version >= 500034 1865 error = securelevel_ge(sopt->sopt_td->td_ucred, 3); 1866 if (error) 1867 return (error); 1868#else 1869 if (securelevel >= 3) 1870 return (EPERM); 1871#endif 1872 } 1873 1874 switch (sopt->sopt_name) { 1875 default : 1876 printf("ip_dn_ctl -- unknown option %d", sopt->sopt_name); 1877 return EINVAL ; 1878 1879 case IP_DUMMYNET_GET : 1880 error = dummynet_get(sopt); 1881 break ; 1882 1883 case IP_DUMMYNET_FLUSH : 1884 dummynet_flush() ; 1885 break ; 1886 1887 case IP_DUMMYNET_CONFIGURE : 1888 p = &tmp_pipe ; 1889 error = sooptcopyin(sopt, p, sizeof *p, sizeof *p); 1890 if (error) 1891 break ; 1892 error = config_pipe(p); 1893 break ; 1894 1895 case IP_DUMMYNET_DEL : /* remove a pipe or queue */ 1896 p = &tmp_pipe ; 1897 error = sooptcopyin(sopt, p, sizeof *p, sizeof *p); 1898 if (error) 1899 break ; 1900 1901 error = delete_pipe(p); 1902 break ; 1903 } 1904 return error ; 1905} 1906 1907static void 1908ip_dn_init(void) 1909{ 1910 printf("DUMMYNET initialized (011031)\n"); 1911 all_pipes = NULL ; 1912 all_flow_sets = NULL ; 1913 ready_heap.size = ready_heap.elements = 0 ; 1914 ready_heap.offset = 0 ; 1915 1916 wfq_ready_heap.size = wfq_ready_heap.elements = 0 ; 1917 wfq_ready_heap.offset = 0 ; 1918 1919 extract_heap.size = extract_heap.elements = 0 ; 1920 extract_heap.offset = 0 ; 1921 ip_dn_ctl_ptr = ip_dn_ctl; 1922 ip_dn_io_ptr = dummynet_io; 1923 ip_dn_ruledel_ptr = dn_rule_delete; 1924 bzero(&dn_timeout, sizeof(struct callout_handle)); 1925 dn_timeout = timeout(dummynet, NULL, 1); 1926} 1927 1928static int 1929dummynet_modevent(module_t mod, int type, void *data) 1930{ 1931 int s; 1932 switch (type) { 1933 case MOD_LOAD: 1934 s = splimp(); 1935 if (DUMMYNET_LOADED) { 1936 splx(s); 1937 printf("DUMMYNET already loaded\n"); 1938 return EEXIST ; 1939 } 1940 ip_dn_init(); 1941 splx(s); 1942 break; 1943 1944 case MOD_UNLOAD: 1945#if !defined(KLD_MODULE) 1946 printf("dummynet statically compiled, cannot unload\n"); 1947 return EINVAL ; 1948#else 1949 s = splimp(); 1950 untimeout(dummynet, NULL, dn_timeout); 1951 dummynet_flush(); 1952 ip_dn_ctl_ptr = NULL; 1953 ip_dn_io_ptr = NULL; 1954 ip_dn_ruledel_ptr = NULL; 1955 splx(s); 1956#endif 1957 break ; 1958 default: 1959 break ; 1960 } 1961 return 0 ; 1962} 1963 1964static moduledata_t dummynet_mod = { 1965 "dummynet", 1966 dummynet_modevent, 1967 NULL 1968}; 1969DECLARE_MODULE(dummynet, dummynet_mod, SI_SUB_PSEUDO, SI_ORDER_ANY); 1970MODULE_DEPEND(dummynet, ipfw, 1, 1, 1); 1971MODULE_VERSION(dummynet, 1);
|