1 /* 2 * Copyright (c) 2010 Riccardo Panicucci, Universita` di Pisa 3 * Copyright (c) 2000-2002 Luigi Rizzo, Universita` di Pisa 4 * All rights reserved 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25 * SUCH DAMAGE. 26 */ 27 28 /* 29 * $FreeBSD$ 30 */ 31 32 #ifdef _KERNEL 33 #include <sys/malloc.h> 34 #include <sys/socket.h> 35 #include <sys/socketvar.h> 36 #include <sys/kernel.h> 37 #include <sys/mbuf.h> 38 #include <sys/module.h> 39 #include <net/if.h> /* IFNAMSIZ */ 40 #include <netinet/in.h> 41 #include <netinet/ip_var.h> /* ipfw_rule_ref */ 42 #include <netinet/ip_fw.h> /* flow_id */ 43 #include <netinet/ip_dummynet.h> 44 #include <netpfil/ipfw/dn_heap.h> 45 #include <netpfil/ipfw/ip_dn_private.h> 46 #ifdef NEW_AQM 47 #include <netpfil/ipfw/dn_aqm.h> 48 #endif 49 #include <netpfil/ipfw/dn_sched.h> 50 #else 51 #include <dn_test.h> 52 #endif 53 54 #ifndef MAX64 55 #define MAX64(x,y) (( (int64_t) ( (y)-(x) )) > 0 ) ? (y) : (x) 56 #endif 57 58 /* 59 * timestamps are computed on 64 bit using fixed point arithmetic. 60 * LMAX_BITS, WMAX_BITS are the max number of bits for the packet len 61 * and sum of weights, respectively. FRAC_BITS is the number of 62 * fractional bits. We want FRAC_BITS >> WMAX_BITS to avoid too large 63 * errors when computing the inverse, FRAC_BITS < 32 so we can do 1/w 64 * using an unsigned 32-bit division, and to avoid wraparounds we need 65 * LMAX_BITS + WMAX_BITS + FRAC_BITS << 64 66 * As an example 67 * FRAC_BITS = 26, LMAX_BITS=14, WMAX_BITS = 19 68 */ 69 #ifndef FRAC_BITS 70 #define FRAC_BITS 28 /* shift for fixed point arithmetic */ 71 #define ONE_FP (1UL << FRAC_BITS) 72 #endif 73 74 /* 75 * Private information for the scheduler instance: 76 * sch_heap (key is Finish time) returns the next queue to serve 77 * ne_heap (key is Start time) stores not-eligible queues 78 * idle_heap (key=start/finish time) stores idle flows. It must 79 * support extract-from-middle. 80 * A flow is only in 1 of the three heaps. 81 * XXX todo: use a more efficient data structure, e.g. a tree sorted 82 * by F with min_subtree(S) in each node 83 */ 84 struct wf2qp_si { 85 struct dn_heap sch_heap; /* top extract - key Finish time */ 86 struct dn_heap ne_heap; /* top extract - key Start time */ 87 struct dn_heap idle_heap; /* random extract - key Start=Finish time */ 88 uint64_t V; /* virtual time */ 89 uint32_t inv_wsum; /* inverse of sum of weights */ 90 uint32_t wsum; /* sum of weights */ 91 }; 92 93 struct wf2qp_queue { 94 struct dn_queue _q; 95 uint64_t S, F; /* start time, finish time */ 96 uint32_t inv_w; /* ONE_FP / weight */ 97 int32_t heap_pos; /* position (index) of struct in heap */ 98 }; 99 100 /* 101 * This file implements a WF2Q+ scheduler as it has been in dummynet 102 * since 2000. 103 * The scheduler supports per-flow queues and has O(log N) complexity. 104 * 105 * WF2Q+ needs to drain entries from the idle heap so that we 106 * can keep the sum of weights up to date. We can do it whenever 107 * we get a chance, or periodically, or following some other 108 * strategy. The function idle_check() drains at most N elements 109 * from the idle heap. 110 */ 111 static void 112 idle_check(struct wf2qp_si *si, int n, int force) 113 { 114 struct dn_heap *h = &si->idle_heap; 115 while (n-- > 0 && h->elements > 0 && 116 (force || DN_KEY_LT(HEAP_TOP(h)->key, si->V))) { 117 struct dn_queue *q = HEAP_TOP(h)->object; 118 struct wf2qp_queue *alg_fq = (struct wf2qp_queue *)q; 119 120 heap_extract(h, NULL); 121 /* XXX to let the flowset delete the queue we should 122 * mark it as 'unused' by the scheduler. 123 */ 124 alg_fq->S = alg_fq->F + 1; /* Mark timestamp as invalid. */ 125 si->wsum -= q->fs->fs.par[0]; /* adjust sum of weights */ 126 if (si->wsum > 0) 127 si->inv_wsum = ONE_FP/si->wsum; 128 } 129 } 130 131 static int 132 wf2qp_enqueue(struct dn_sch_inst *_si, struct dn_queue *q, struct mbuf *m) 133 { 134 struct dn_fsk *fs = q->fs; 135 struct wf2qp_si *si = (struct wf2qp_si *)(_si + 1); 136 struct wf2qp_queue *alg_fq; 137 uint64_t len = m->m_pkthdr.len; 138 139 if (m != q->mq.head) { 140 if (dn_enqueue(q, m, 0)) /* packet was dropped */ 141 return 1; 142 if (m != q->mq.head) /* queue was already busy */ 143 return 0; 144 } 145 146 /* If reach this point, queue q was idle */ 147 alg_fq = (struct wf2qp_queue *)q; 148 149 if (DN_KEY_LT(alg_fq->F, alg_fq->S)) { 150 /* F<S means timestamps are invalid ->brand new queue. */ 151 alg_fq->S = si->V; /* init start time */ 152 si->wsum += fs->fs.par[0]; /* add weight of new queue. */ 153 si->inv_wsum = ONE_FP/si->wsum; 154 } else { /* if it was idle then it was in the idle heap */ 155 heap_extract(&si->idle_heap, q); 156 alg_fq->S = MAX64(alg_fq->F, si->V); /* compute new S */ 157 } 158 alg_fq->F = alg_fq->S + len * alg_fq->inv_w; 159 160 /* if nothing is backlogged, make sure this flow is eligible */ 161 if (si->ne_heap.elements == 0 && si->sch_heap.elements == 0) 162 si->V = MAX64(alg_fq->S, si->V); 163 164 /* 165 * Look at eligibility. A flow is not eligibile if S>V (when 166 * this happens, it means that there is some other flow already 167 * scheduled for the same pipe, so the sch_heap cannot be 168 * empty). If the flow is not eligible we just store it in the 169 * ne_heap. Otherwise, we store in the sch_heap. 170 * Note that for all flows in sch_heap (SCH), S_i <= V, 171 * and for all flows in ne_heap (NEH), S_i > V. 172 * So when we need to compute max(V, min(S_i)) forall i in 173 * SCH+NEH, we only need to look into NEH. 174 */ 175 if (DN_KEY_LT(si->V, alg_fq->S)) { 176 /* S>V means flow Not eligible. */ 177 if (si->sch_heap.elements == 0) 178 D("++ ouch! not eligible but empty scheduler!"); 179 heap_insert(&si->ne_heap, alg_fq->S, q); 180 } else { 181 heap_insert(&si->sch_heap, alg_fq->F, q); 182 } 183 return 0; 184 } 185 186 /* XXX invariant: sch > 0 || V >= min(S in neh) */ 187 static struct mbuf * 188 wf2qp_dequeue(struct dn_sch_inst *_si) 189 { 190 /* Access scheduler instance private data */ 191 struct wf2qp_si *si = (struct wf2qp_si *)(_si + 1); 192 struct mbuf *m; 193 struct dn_queue *q; 194 struct dn_heap *sch = &si->sch_heap; 195 struct dn_heap *neh = &si->ne_heap; 196 struct wf2qp_queue *alg_fq; 197 198 if (sch->elements == 0 && neh->elements == 0) { 199 /* we have nothing to do. We could kill the idle heap 200 * altogether and reset V 201 */ 202 idle_check(si, 0x7fffffff, 1); 203 si->V = 0; 204 si->wsum = 0; /* should be set already */ 205 return NULL; /* quick return if nothing to do */ 206 } 207 idle_check(si, 1, 0); /* drain something from the idle heap */ 208 209 /* make sure at least one element is eligible, bumping V 210 * and moving entries that have become eligible. 211 * We need to repeat the first part twice, before and 212 * after extracting the candidate, or enqueue() will 213 * find the data structure in a wrong state. 214 */ 215 m = NULL; 216 for(;;) { 217 /* 218 * Compute V = max(V, min(S_i)). Remember that all elements 219 * in sch have by definition S_i <= V so if sch is not empty, 220 * V is surely the max and we must not update it. Conversely, 221 * if sch is empty we only need to look at neh. 222 * We don't need to move the queues, as it will be done at the 223 * next enqueue 224 */ 225 if (sch->elements == 0 && neh->elements > 0) { 226 si->V = MAX64(si->V, HEAP_TOP(neh)->key); 227 } 228 while (neh->elements > 0 && 229 DN_KEY_LEQ(HEAP_TOP(neh)->key, si->V)) { 230 q = HEAP_TOP(neh)->object; 231 alg_fq = (struct wf2qp_queue *)q; 232 heap_extract(neh, NULL); 233 heap_insert(sch, alg_fq->F, q); 234 } 235 if (m) /* pkt found in previous iteration */ 236 break; 237 /* ok we have at least one eligible pkt */ 238 q = HEAP_TOP(sch)->object; 239 alg_fq = (struct wf2qp_queue *)q; 240 m = dn_dequeue(q); 241 heap_extract(sch, NULL); /* Remove queue from heap. */ 242 si->V += (uint64_t)(m->m_pkthdr.len) * si->inv_wsum; 243 alg_fq->S = alg_fq->F; /* Update start time. */ 244 if (q->mq.head == 0) { /* not backlogged any more. */ 245 heap_insert(&si->idle_heap, alg_fq->F, q); 246 } else { /* Still backlogged. */ 247 /* Update F, store in neh or sch */ 248 uint64_t len = q->mq.head->m_pkthdr.len; 249 alg_fq->F += len * alg_fq->inv_w; 250 if (DN_KEY_LEQ(alg_fq->S, si->V)) { 251 heap_insert(sch, alg_fq->F, q); 252 } else { 253 heap_insert(neh, alg_fq->S, q); 254 } 255 } 256 } 257 return m; 258 } 259 260 static int 261 wf2qp_new_sched(struct dn_sch_inst *_si) 262 { 263 struct wf2qp_si *si = (struct wf2qp_si *)(_si + 1); 264 int ofs = offsetof(struct wf2qp_queue, heap_pos); 265 266 /* all heaps support extract from middle */ 267 if (heap_init(&si->idle_heap, 16, ofs) || 268 heap_init(&si->sch_heap, 16, ofs) || 269 heap_init(&si->ne_heap, 16, ofs)) { 270 heap_free(&si->ne_heap); 271 heap_free(&si->sch_heap); 272 heap_free(&si->idle_heap); 273 return ENOMEM; 274 } 275 return 0; 276 } 277 278 static int 279 wf2qp_free_sched(struct dn_sch_inst *_si) 280 { 281 struct wf2qp_si *si = (struct wf2qp_si *)(_si + 1); 282 283 heap_free(&si->sch_heap); 284 heap_free(&si->ne_heap); 285 heap_free(&si->idle_heap); 286 287 return 0; 288 } 289 290 static int 291 wf2qp_new_fsk(struct dn_fsk *fs) 292 { 293 ipdn_bound_var(&fs->fs.par[0], 1, 294 1, 100, "WF2Q+ weight"); 295 return 0; 296 } 297 298 static int 299 wf2qp_new_queue(struct dn_queue *_q) 300 { 301 struct wf2qp_queue *q = (struct wf2qp_queue *)_q; 302 303 _q->ni.oid.subtype = DN_SCHED_WF2QP; 304 q->F = 0; /* not strictly necessary */ 305 q->S = q->F + 1; /* mark timestamp as invalid. */ 306 q->inv_w = ONE_FP / _q->fs->fs.par[0]; 307 if (_q->mq.head != NULL) { 308 wf2qp_enqueue(_q->_si, _q, _q->mq.head); 309 } 310 return 0; 311 } 312 313 /* 314 * Called when the infrastructure removes a queue (e.g. flowset 315 * is reconfigured). Nothing to do if we did not 'own' the queue, 316 * otherwise remove it from the right heap and adjust the sum 317 * of weights. 318 */ 319 static int 320 wf2qp_free_queue(struct dn_queue *q) 321 { 322 struct wf2qp_queue *alg_fq = (struct wf2qp_queue *)q; 323 struct wf2qp_si *si = (struct wf2qp_si *)(q->_si + 1); 324 325 if (alg_fq->S >= alg_fq->F + 1) 326 return 0; /* nothing to do, not in any heap */ 327 si->wsum -= q->fs->fs.par[0]; 328 if (si->wsum > 0) 329 si->inv_wsum = ONE_FP/si->wsum; 330 331 /* extract from the heap. XXX TODO we may need to adjust V 332 * to make sure the invariants hold. 333 */ 334 if (q->mq.head == NULL) { 335 heap_extract(&si->idle_heap, q); 336 } else if (DN_KEY_LT(si->V, alg_fq->S)) { 337 heap_extract(&si->ne_heap, q); 338 } else { 339 heap_extract(&si->sch_heap, q); 340 } 341 return 0; 342 } 343 344 /* 345 * WF2Q+ scheduler descriptor 346 * contains the type of the scheduler, the name, the size of the 347 * structures and function pointers. 348 */ 349 static struct dn_alg wf2qp_desc = { 350 _SI( .type = ) DN_SCHED_WF2QP, 351 _SI( .name = ) "WF2Q+", 352 _SI( .flags = ) DN_MULTIQUEUE, 353 354 /* we need extra space in the si and the queue */ 355 _SI( .schk_datalen = ) 0, 356 _SI( .si_datalen = ) sizeof(struct wf2qp_si), 357 _SI( .q_datalen = ) sizeof(struct wf2qp_queue) - 358 sizeof(struct dn_queue), 359 360 _SI( .enqueue = ) wf2qp_enqueue, 361 _SI( .dequeue = ) wf2qp_dequeue, 362 363 _SI( .config = ) NULL, 364 _SI( .destroy = ) NULL, 365 _SI( .new_sched = ) wf2qp_new_sched, 366 _SI( .free_sched = ) wf2qp_free_sched, 367 368 _SI( .new_fsk = ) wf2qp_new_fsk, 369 _SI( .free_fsk = ) NULL, 370 371 _SI( .new_queue = ) wf2qp_new_queue, 372 _SI( .free_queue = ) wf2qp_free_queue, 373 #ifdef NEW_AQM 374 _SI( .getconfig = ) NULL, 375 #endif 376 377 }; 378 379 380 DECLARE_DNSCHED_MODULE(dn_wf2qp, &wf2qp_desc); 381