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