bsd/net/classq/classq_fq_codel.c

   1 /*
   2  * Copyright (c) 2016-2017 Apple Inc. All rights reserved.
   3  *
   4  * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
   5  *
   6  * This file contains Original Code and/or Modifications of Original Code
   7  * as defined in and that are subject to the Apple Public Source License
   8  * Version 2.0 (the 'License'). You may not use this file except in
   9  * compliance with the License. The rights granted to you under the License
  10  * may not be used to create, or enable the creation or redistribution of,
  11  * unlawful or unlicensed copies of an Apple operating system, or to
  12  * circumvent, violate, or enable the circumvention or violation of, any
  13  * terms of an Apple operating system software license agreement.
  14  *
  15  * Please obtain a copy of the License at
  16  * http://www.opensource.apple.com/apsl/ and read it before using this file.
  17  *
  18  * The Original Code and all software distributed under the License are
  19  * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
  20  * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
  21  * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
  22  * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
  23  * Please see the License for the specific language governing rights and
  24  * limitations under the License.
  25  *
  26  * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
  27  */
  28
  29 #include <sys/cdefs.h>
  30 #include <sys/param.h>
  31 #include <sys/mbuf.h>
  32 #include <sys/socket.h>
  33 #include <sys/sockio.h>
  34 #include <sys/systm.h>
  35 #include <sys/sysctl.h>
  36 #include <sys/syslog.h>
  37 #include <sys/proc.h>
  38 #include <sys/errno.h>
  39 #include <sys/kernel.h>
  40 #include <sys/kauth.h>
  41 #include <kern/zalloc.h>
  42 #include <netinet/in.h>
  43
  44 #include <net/classq/classq.h>
  45 #include <net/classq/if_classq.h>
  46 #include <net/pktsched/pktsched.h>
  47 #include <net/pktsched/pktsched_fq_codel.h>
  48 #include <net/classq/classq_fq_codel.h>
  49
  50 static uint32_t flowq_size;                     /* size of flowq */
  51 static struct mcache *flowq_cache = NULL;       /* mcache for flowq */
  52
  53 #define FQ_ZONE_MAX     (32 * 1024)     /* across all interfaces */
  54
  55 #define DTYPE_NODROP    0       /* no drop */
  56 #define DTYPE_FORCED    1       /* a "forced" drop */
  57 #define DTYPE_EARLY     2       /* an "unforced" (early) drop */
  58
  59 void
  60 fq_codel_init(void)
  61 {
  62         if (flowq_cache != NULL) {
  63                 return;
  64         }
  65
  66         flowq_size = sizeof(fq_t);
  67         flowq_cache = mcache_create("fq.flowq", flowq_size, sizeof(uint64_t),
  68             0, MCR_SLEEP);
  69         if (flowq_cache == NULL) {
  70                 panic("%s: failed to allocate flowq_cache", __func__);
  71                 /* NOTREACHED */
  72         }
  73 }
  74
  75 void
  76 fq_codel_reap_caches(boolean_t purge)
  77 {
  78         mcache_reap_now(flowq_cache, purge);
  79 }
  80
  81 fq_t *
  82 fq_alloc(classq_pkt_type_t ptype)
  83 {
  84         fq_t *fq = NULL;
  85         fq = mcache_alloc(flowq_cache, MCR_SLEEP);
  86         if (fq == NULL) {
  87                 log(LOG_ERR, "%s: unable to allocate from flowq_cache\n");
  88                 return NULL;
  89         }
  90
  91         bzero(fq, flowq_size);
  92         fq->fq_ptype = ptype;
  93         if (ptype == QP_MBUF) {
  94                 MBUFQ_INIT(&fq->fq_mbufq);
  95         }
  96         return fq;
  97 }
  98
  99 void
 100 fq_destroy(fq_t *fq)
 101 {
 102         VERIFY(fq_empty(fq));
 103         VERIFY(!(fq->fq_flags & (FQF_NEW_FLOW | FQF_OLD_FLOW)));
 104         VERIFY(fq->fq_bytes == 0);
 105         mcache_free(flowq_cache, fq);
 106 }
 107
 108 static void
 109 fq_detect_dequeue_stall(fq_if_t *fqs, fq_t *flowq, fq_if_classq_t *fq_cl,
 110     u_int64_t *now)
 111 {
 112         u_int64_t maxgetqtime;
 113         if (FQ_IS_DELAYHIGH(flowq) || flowq->fq_getqtime == 0 ||
 114             fq_empty(flowq) ||
 115             flowq->fq_bytes < FQ_MIN_FC_THRESHOLD_BYTES) {
 116                 return;
 117         }
 118         maxgetqtime = flowq->fq_getqtime + fqs->fqs_update_interval;
 119         if ((*now) > maxgetqtime) {
 120                 /*
 121                  * there was no dequeue in an update interval worth of
 122                  * time. It means that the queue is stalled.
 123                  */
 124                 FQ_SET_DELAY_HIGH(flowq);
 125                 fq_cl->fcl_stat.fcl_dequeue_stall++;
 126         }
 127 }
 128
 129 void
 130 fq_head_drop(fq_if_t *fqs, fq_t *fq)
 131 {
 132         pktsched_pkt_t pkt;
 133         uint32_t *pkt_flags;
 134         uint64_t *pkt_timestamp;
 135         struct ifclassq *ifq = fqs->fqs_ifq;
 136
 137         _PKTSCHED_PKT_INIT(&pkt);
 138         if (fq_getq_flow_internal(fqs, fq, &pkt) == NULL) {
 139                 return;
 140         }
 141
 142         pktsched_get_pkt_vars(&pkt, &pkt_flags, &pkt_timestamp, NULL, NULL,
 143             NULL, NULL);
 144
 145         *pkt_timestamp = 0;
 146         if (pkt.pktsched_ptype == QP_MBUF) {
 147                 *pkt_flags &= ~PKTF_PRIV_GUARDED;
 148         }
 149
 150         IFCQ_DROP_ADD(ifq, 1, pktsched_get_pkt_len(&pkt));
 151         IFCQ_CONVERT_LOCK(ifq);
 152         pktsched_free_pkt(&pkt);
 153 }
 154
 155 int
 156 fq_addq(fq_if_t *fqs, pktsched_pkt_t *pkt, fq_if_classq_t *fq_cl)
 157 {
 158         int droptype = DTYPE_NODROP, fc_adv = 0, ret = CLASSQEQ_SUCCESS;
 159         u_int64_t now;
 160         fq_t *fq = NULL;
 161         uint64_t *pkt_timestamp;
 162         uint32_t *pkt_flags;
 163         uint32_t pkt_flowid, pkt_tx_start_seq;
 164         uint8_t pkt_proto, pkt_flowsrc;
 165
 166         pktsched_get_pkt_vars(pkt, &pkt_flags, &pkt_timestamp, &pkt_flowid,
 167             &pkt_flowsrc, &pkt_proto, &pkt_tx_start_seq);
 168
 169         if (pkt->pktsched_ptype == QP_MBUF) {
 170                 /* See comments in <rdar://problem/14040693> */
 171                 VERIFY(!(*pkt_flags & PKTF_PRIV_GUARDED));
 172                 *pkt_flags |= PKTF_PRIV_GUARDED;
 173         }
 174
 175         if (*pkt_timestamp > 0) {
 176                 now = *pkt_timestamp;
 177         } else {
 178                 struct timespec now_ts;
 179                 nanouptime(&now_ts);
 180                 now = (now_ts.tv_sec * NSEC_PER_SEC) + now_ts.tv_nsec;
 181                 *pkt_timestamp = now;
 182         }
 183
 184         /* find the flowq for this packet */
 185         fq = fq_if_hash_pkt(fqs, pkt_flowid, pktsched_get_pkt_svc(pkt),
 186             now, TRUE, pkt->pktsched_ptype);
 187         if (fq == NULL) {
 188                 /* drop the packet if we could not allocate a flow queue */
 189                 fq_cl->fcl_stat.fcl_drop_memfailure++;
 190                 IFCQ_CONVERT_LOCK(fqs->fqs_ifq);
 191                 return CLASSQEQ_DROP;
 192         }
 193         VERIFY(fq->fq_ptype == pkt->pktsched_ptype);
 194
 195         fq_detect_dequeue_stall(fqs, fq, fq_cl, &now);
 196
 197         if (FQ_IS_DELAYHIGH(fq)) {
 198                 if ((fq->fq_flags & FQF_FLOWCTL_CAPABLE) &&
 199                     (*pkt_flags & PKTF_FLOW_ADV)) {
 200                         fc_adv = 1;
 201                         /*
 202                          * If the flow is suspended or it is not
 203                          * TCP, drop the packet
 204                          */
 205                         if (pkt_proto != IPPROTO_TCP) {
 206                                 droptype = DTYPE_EARLY;
 207                                 fq_cl->fcl_stat.fcl_drop_early++;
 208                         }
 209                 } else {
 210                         /*
 211                          * Need to drop a packet, instead of dropping this
 212                          * one, try to drop from the head of the queue
 213                          */
 214                         if (!fq_empty(fq)) {
 215                                 fq_head_drop(fqs, fq);
 216                                 droptype = DTYPE_NODROP;
 217                         } else {
 218                                 droptype = DTYPE_EARLY;
 219                         }
 220                         fq_cl->fcl_stat.fcl_drop_early++;
 221                 }
 222         }
 223
 224         /* Set the return code correctly */
 225         if (fc_adv == 1 && droptype != DTYPE_FORCED) {
 226                 if (fq_if_add_fcentry(fqs, pkt, pkt_flowid, pkt_flowsrc,
 227                     fq_cl)) {
 228                         fq->fq_flags |= FQF_FLOWCTL_ON;
 229                         /* deliver flow control advisory error */
 230                         if (droptype == DTYPE_NODROP) {
 231                                 ret = CLASSQEQ_SUCCESS_FC;
 232                         } else {
 233                                 /* dropped due to flow control */
 234                                 ret = CLASSQEQ_DROP_FC;
 235                         }
 236                 } else {
 237                         /*
 238                          * if we could not flow control the flow, it is
 239                          * better to drop
 240                          */
 241                         droptype = DTYPE_FORCED;
 242                         ret = CLASSQEQ_DROP_FC;
 243                         fq_cl->fcl_stat.fcl_flow_control_fail++;
 244                 }
 245         }
 246
 247         /*
 248          * If the queue length hits the queue limit, drop a packet from the
 249          * front of the queue for a flow with maximum number of bytes. This
 250          * will penalize heavy and unresponsive flows. It will also avoid a
 251          * tail drop.
 252          */
 253         if (droptype == DTYPE_NODROP && fq_if_at_drop_limit(fqs)) {
 254                 if (fqs->fqs_large_flow == fq) {
 255                         /*
 256                          * Drop from the head of the current fq. Since a
 257                          * new packet will be added to the tail, it is ok
 258                          * to leave fq in place.
 259                          */
 260                         fq_head_drop(fqs, fq);
 261                 } else {
 262                         if (fqs->fqs_large_flow == NULL) {
 263                                 droptype = DTYPE_FORCED;
 264                                 fq_cl->fcl_stat.fcl_drop_overflow++;
 265                                 ret = CLASSQEQ_DROP;
 266
 267                                 /*
 268                                  * if this fq was freshly created and there
 269                                  * is nothing to enqueue, free it
 270                                  */
 271                                 if (fq_empty(fq) && !(fq->fq_flags &
 272                                     (FQF_NEW_FLOW | FQF_OLD_FLOW))) {
 273                                         fq_if_destroy_flow(fqs, fq_cl, fq);
 274                                         fq = NULL;
 275                                 }
 276                         } else {
 277                                 fq_if_drop_packet(fqs);
 278                         }
 279                 }
 280         }
 281
 282         if (droptype == DTYPE_NODROP) {
 283                 uint32_t pkt_len = pktsched_get_pkt_len(pkt);
 284                 fq_enqueue(fq, pkt->pktsched_pkt);
 285                 fq->fq_bytes += pkt_len;
 286                 fq_cl->fcl_stat.fcl_byte_cnt += pkt_len;
 287                 fq_cl->fcl_stat.fcl_pkt_cnt++;
 288
 289                 /*
 290                  * check if this queue will qualify to be the next
 291                  * victim queue
 292                  */
 293                 fq_if_is_flow_heavy(fqs, fq);
 294         } else {
 295                 IFCQ_CONVERT_LOCK(fqs->fqs_ifq);
 296                 return (ret != CLASSQEQ_SUCCESS) ? ret : CLASSQEQ_DROP;
 297         }
 298
 299         /*
 300          * If the queue is not currently active, add it to the end of new
 301          * flows list for that service class.
 302          */
 303         if ((fq->fq_flags & (FQF_NEW_FLOW | FQF_OLD_FLOW)) == 0) {
 304                 VERIFY(STAILQ_NEXT(fq, fq_actlink) == NULL);
 305                 STAILQ_INSERT_TAIL(&fq_cl->fcl_new_flows, fq, fq_actlink);
 306                 fq->fq_flags |= FQF_NEW_FLOW;
 307
 308                 fq_cl->fcl_stat.fcl_newflows_cnt++;
 309
 310                 fq->fq_deficit = fq_cl->fcl_quantum;
 311         }
 312         return ret;
 313 }
 314
 315 void *
 316 fq_getq_flow_internal(fq_if_t *fqs, fq_t *fq, pktsched_pkt_t *pkt)
 317 {
 318         void *p;
 319         uint32_t plen;
 320         fq_if_classq_t *fq_cl;
 321         struct ifclassq *ifq = fqs->fqs_ifq;
 322
 323         fq_dequeue(fq, p);
 324         if (p == NULL) {
 325                 return NULL;
 326         }
 327
 328         pktsched_pkt_encap(pkt, fq->fq_ptype, p);
 329         plen = pktsched_get_pkt_len(pkt);
 330
 331         VERIFY(fq->fq_bytes >= plen);
 332         fq->fq_bytes -= plen;
 333
 334         fq_cl = &fqs->fqs_classq[fq->fq_sc_index];
 335         fq_cl->fcl_stat.fcl_byte_cnt -= plen;
 336         fq_cl->fcl_stat.fcl_pkt_cnt--;
 337         IFCQ_DEC_LEN(ifq);
 338         IFCQ_DEC_BYTES(ifq, plen);
 339
 340         /* Reset getqtime so that we don't count idle times */
 341         if (fq_empty(fq)) {
 342                 fq->fq_getqtime = 0;
 343         }
 344
 345         return p;
 346 }
 347
 348 void *
 349 fq_getq_flow(fq_if_t *fqs, fq_t *fq, pktsched_pkt_t *pkt)
 350 {
 351         void *p;
 352         fq_if_classq_t *fq_cl;
 353         u_int64_t now;
 354         int64_t qdelay = 0;
 355         struct timespec now_ts;
 356         uint32_t *pkt_flags, pkt_tx_start_seq;
 357         uint64_t *pkt_timestamp;
 358
 359         p = fq_getq_flow_internal(fqs, fq, pkt);
 360         if (p == NULL) {
 361                 return NULL;
 362         }
 363
 364         pktsched_get_pkt_vars(pkt, &pkt_flags, &pkt_timestamp, NULL, NULL,
 365             NULL, &pkt_tx_start_seq);
 366
 367         nanouptime(&now_ts);
 368         now = (now_ts.tv_sec * NSEC_PER_SEC) + now_ts.tv_nsec;
 369
 370         /* this will compute qdelay in nanoseconds */
 371         if (now > *pkt_timestamp) {
 372                 qdelay = now - *pkt_timestamp;
 373         }
 374         fq_cl = &fqs->fqs_classq[fq->fq_sc_index];
 375
 376         if (fq->fq_min_qdelay == 0 ||
 377             (qdelay > 0 && (u_int64_t)qdelay < fq->fq_min_qdelay)) {
 378                 fq->fq_min_qdelay = qdelay;
 379         }
 380         if (now >= fq->fq_updatetime) {
 381                 if (fq->fq_min_qdelay > fqs->fqs_target_qdelay) {
 382                         if (!FQ_IS_DELAYHIGH(fq)) {
 383                                 FQ_SET_DELAY_HIGH(fq);
 384                         }
 385                 } else {
 386                         FQ_CLEAR_DELAY_HIGH(fq);
 387                 }
 388
 389
 390                 /* Reset measured queue delay and update time */
 391                 fq->fq_updatetime = now + fqs->fqs_update_interval;
 392                 fq->fq_min_qdelay = 0;
 393         }
 394         if (!FQ_IS_DELAYHIGH(fq) || fq_empty(fq)) {
 395                 FQ_CLEAR_DELAY_HIGH(fq);
 396                 if (fq->fq_flags & FQF_FLOWCTL_ON) {
 397                         fq_if_flow_feedback(fqs, fq, fq_cl);
 398                 }
 399         }
 400
 401         if (fq_empty(fq)) {
 402                 /* Reset getqtime so that we don't count idle times */
 403                 fq->fq_getqtime = 0;
 404         } else {
 405                 fq->fq_getqtime = now;
 406         }
 407         fq_if_is_flow_heavy(fqs, fq);
 408
 409         *pkt_timestamp = 0;
 410         if (pkt->pktsched_ptype == QP_MBUF) {
 411                 *pkt_flags &= ~PKTF_PRIV_GUARDED;
 412         }
 413
 414         return p;
 415 }