1 /*- 2 * Copyright (c) 2015 Gleb Smirnoff <glebius@FreeBSD.org> 3 * Copyright (c) 2015 Adrian Chadd <adrian@FreeBSD.org> 4 * Copyright (c) 1982, 1986, 1988, 1993 5 * The Regents of the University of California. All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 3. Neither the name of the University nor the names of its contributors 16 * may be used to endorse or promote products derived from this software 17 * without specific prior written permission. 18 * 19 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 22 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 29 * SUCH DAMAGE. 30 * 31 * @(#)ip_input.c 8.2 (Berkeley) 1/4/94 32 */ 33 34 #include <sys/cdefs.h> 35 __FBSDID("$FreeBSD$"); 36 37 #include "opt_rss.h" 38 39 #include <sys/param.h> 40 #include <sys/systm.h> 41 #include <sys/eventhandler.h> 42 #include <sys/kernel.h> 43 #include <sys/hash.h> 44 #include <sys/mbuf.h> 45 #include <sys/malloc.h> 46 #include <sys/limits.h> 47 #include <sys/lock.h> 48 #include <sys/mutex.h> 49 #include <sys/sysctl.h> 50 #include <sys/socket.h> 51 52 #include <net/if.h> 53 #include <net/if_var.h> 54 #include <net/rss_config.h> 55 #include <net/netisr.h> 56 #include <net/vnet.h> 57 58 #include <netinet/in.h> 59 #include <netinet/ip.h> 60 #include <netinet/ip_var.h> 61 #include <netinet/in_rss.h> 62 #ifdef MAC 63 #include <security/mac/mac_framework.h> 64 #endif 65 66 SYSCTL_DECL(_net_inet_ip); 67 68 /* 69 * Reassembly headers are stored in hash buckets. 70 */ 71 #define IPREASS_NHASH_LOG2 10 72 #define IPREASS_NHASH (1 << IPREASS_NHASH_LOG2) 73 #define IPREASS_HMASK (IPREASS_NHASH - 1) 74 75 struct ipqbucket { 76 TAILQ_HEAD(ipqhead, ipq) head; 77 struct mtx lock; 78 int count; 79 }; 80 81 VNET_DEFINE_STATIC(struct ipqbucket, ipq[IPREASS_NHASH]); 82 #define V_ipq VNET(ipq) 83 VNET_DEFINE_STATIC(uint32_t, ipq_hashseed); 84 #define V_ipq_hashseed VNET(ipq_hashseed) 85 86 #define IPQ_LOCK(i) mtx_lock(&V_ipq[i].lock) 87 #define IPQ_TRYLOCK(i) mtx_trylock(&V_ipq[i].lock) 88 #define IPQ_UNLOCK(i) mtx_unlock(&V_ipq[i].lock) 89 #define IPQ_LOCK_ASSERT(i) mtx_assert(&V_ipq[i].lock, MA_OWNED) 90 91 VNET_DEFINE_STATIC(int, ipreass_maxbucketsize); 92 #define V_ipreass_maxbucketsize VNET(ipreass_maxbucketsize) 93 94 void ipreass_init(void); 95 void ipreass_drain(void); 96 void ipreass_slowtimo(void); 97 #ifdef VIMAGE 98 void ipreass_destroy(void); 99 #endif 100 static int sysctl_maxfragpackets(SYSCTL_HANDLER_ARGS); 101 static int sysctl_maxfragbucketsize(SYSCTL_HANDLER_ARGS); 102 static void ipreass_zone_change(void *); 103 static void ipreass_drain_tomax(void); 104 static void ipq_free(struct ipqbucket *, struct ipq *); 105 static struct ipq * ipq_reuse(int); 106 107 static inline void 108 ipq_timeout(struct ipqbucket *bucket, struct ipq *fp) 109 { 110 111 IPSTAT_ADD(ips_fragtimeout, fp->ipq_nfrags); 112 ipq_free(bucket, fp); 113 } 114 115 static inline void 116 ipq_drop(struct ipqbucket *bucket, struct ipq *fp) 117 { 118 119 IPSTAT_ADD(ips_fragdropped, fp->ipq_nfrags); 120 ipq_free(bucket, fp); 121 } 122 123 /* 124 * By default, limit the number of IP fragments across all reassembly 125 * queues to 1/32 of the total number of mbuf clusters. 126 * 127 * Limit the total number of reassembly queues per VNET to the 128 * IP fragment limit, but ensure the limit will not allow any bucket 129 * to grow above 100 items. (The bucket limit is 130 * IP_MAXFRAGPACKETS / (IPREASS_NHASH / 2), so the 50 is the correct 131 * multiplier to reach a 100-item limit.) 132 * The 100-item limit was chosen as brief testing seems to show that 133 * this produces "reasonable" performance on some subset of systems 134 * under DoS attack. 135 */ 136 #define IP_MAXFRAGS (nmbclusters / 32) 137 #define IP_MAXFRAGPACKETS (imin(IP_MAXFRAGS, IPREASS_NHASH * 50)) 138 139 static int maxfrags; 140 static volatile u_int nfrags; 141 SYSCTL_INT(_net_inet_ip, OID_AUTO, maxfrags, CTLFLAG_RW, 142 &maxfrags, 0, 143 "Maximum number of IPv4 fragments allowed across all reassembly queues"); 144 SYSCTL_UINT(_net_inet_ip, OID_AUTO, curfrags, CTLFLAG_RD, 145 __DEVOLATILE(u_int *, &nfrags), 0, 146 "Current number of IPv4 fragments across all reassembly queues"); 147 148 VNET_DEFINE_STATIC(uma_zone_t, ipq_zone); 149 #define V_ipq_zone VNET(ipq_zone) 150 SYSCTL_PROC(_net_inet_ip, OID_AUTO, maxfragpackets, CTLFLAG_VNET | 151 CTLTYPE_INT | CTLFLAG_RW, NULL, 0, sysctl_maxfragpackets, "I", 152 "Maximum number of IPv4 fragment reassembly queue entries"); 153 SYSCTL_UMA_CUR(_net_inet_ip, OID_AUTO, fragpackets, CTLFLAG_VNET, 154 &VNET_NAME(ipq_zone), 155 "Current number of IPv4 fragment reassembly queue entries"); 156 157 VNET_DEFINE_STATIC(int, noreass); 158 #define V_noreass VNET(noreass) 159 160 VNET_DEFINE_STATIC(int, maxfragsperpacket); 161 #define V_maxfragsperpacket VNET(maxfragsperpacket) 162 SYSCTL_INT(_net_inet_ip, OID_AUTO, maxfragsperpacket, CTLFLAG_VNET | CTLFLAG_RW, 163 &VNET_NAME(maxfragsperpacket), 0, 164 "Maximum number of IPv4 fragments allowed per packet"); 165 SYSCTL_PROC(_net_inet_ip, OID_AUTO, maxfragbucketsize, 166 CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, 167 sysctl_maxfragbucketsize, "I", 168 "Maximum number of IPv4 fragment reassembly queue entries per bucket"); 169 170 /* 171 * Take incoming datagram fragment and try to reassemble it into 172 * whole datagram. If the argument is the first fragment or one 173 * in between the function will return NULL and store the mbuf 174 * in the fragment chain. If the argument is the last fragment 175 * the packet will be reassembled and the pointer to the new 176 * mbuf returned for further processing. Only m_tags attached 177 * to the first packet/fragment are preserved. 178 * The IP header is *NOT* adjusted out of iplen. 179 */ 180 #define M_IP_FRAG M_PROTO9 181 struct mbuf * 182 ip_reass(struct mbuf *m) 183 { 184 struct ip *ip; 185 struct mbuf *p, *q, *nq, *t; 186 struct ipq *fp; 187 struct ifnet *srcifp; 188 struct ipqhead *head; 189 int i, hlen, next, tmpmax; 190 u_int8_t ecn, ecn0; 191 uint32_t hash, hashkey[3]; 192 #ifdef RSS 193 uint32_t rss_hash, rss_type; 194 #endif 195 196 /* 197 * If no reassembling or maxfragsperpacket are 0, 198 * never accept fragments. 199 * Also, drop packet if it would exceed the maximum 200 * number of fragments. 201 */ 202 tmpmax = maxfrags; 203 if (V_noreass == 1 || V_maxfragsperpacket == 0 || 204 (tmpmax >= 0 && atomic_load_int(&nfrags) >= (u_int)tmpmax)) { 205 IPSTAT_INC(ips_fragments); 206 IPSTAT_INC(ips_fragdropped); 207 m_freem(m); 208 return (NULL); 209 } 210 211 ip = mtod(m, struct ip *); 212 hlen = ip->ip_hl << 2; 213 214 /* 215 * Adjust ip_len to not reflect header, 216 * convert offset of this to bytes. 217 */ 218 ip->ip_len = htons(ntohs(ip->ip_len) - hlen); 219 /* 220 * Make sure that fragments have a data length 221 * that's a non-zero multiple of 8 bytes, unless 222 * this is the last fragment. 223 */ 224 if (ip->ip_len == htons(0) || 225 ((ip->ip_off & htons(IP_MF)) && (ntohs(ip->ip_len) & 0x7) != 0)) { 226 IPSTAT_INC(ips_toosmall); /* XXX */ 227 IPSTAT_INC(ips_fragdropped); 228 m_freem(m); 229 return (NULL); 230 } 231 if (ip->ip_off & htons(IP_MF)) 232 m->m_flags |= M_IP_FRAG; 233 else 234 m->m_flags &= ~M_IP_FRAG; 235 ip->ip_off = htons(ntohs(ip->ip_off) << 3); 236 237 /* 238 * Make sure the fragment lies within a packet of valid size. 239 */ 240 if (ntohs(ip->ip_len) + ntohs(ip->ip_off) > IP_MAXPACKET) { 241 IPSTAT_INC(ips_toolong); 242 IPSTAT_INC(ips_fragdropped); 243 m_freem(m); 244 return (NULL); 245 } 246 247 /* 248 * Store receive network interface pointer for later. 249 */ 250 srcifp = m->m_pkthdr.rcvif; 251 252 /* 253 * Attempt reassembly; if it succeeds, proceed. 254 * ip_reass() will return a different mbuf. 255 */ 256 IPSTAT_INC(ips_fragments); 257 m->m_pkthdr.PH_loc.ptr = ip; 258 259 /* 260 * Presence of header sizes in mbufs 261 * would confuse code below. 262 */ 263 m->m_data += hlen; 264 m->m_len -= hlen; 265 266 hashkey[0] = ip->ip_src.s_addr; 267 hashkey[1] = ip->ip_dst.s_addr; 268 hashkey[2] = (uint32_t)ip->ip_p << 16; 269 hashkey[2] += ip->ip_id; 270 hash = jenkins_hash32(hashkey, nitems(hashkey), V_ipq_hashseed); 271 hash &= IPREASS_HMASK; 272 head = &V_ipq[hash].head; 273 IPQ_LOCK(hash); 274 275 /* 276 * Look for queue of fragments 277 * of this datagram. 278 */ 279 TAILQ_FOREACH(fp, head, ipq_list) 280 if (ip->ip_id == fp->ipq_id && 281 ip->ip_src.s_addr == fp->ipq_src.s_addr && 282 ip->ip_dst.s_addr == fp->ipq_dst.s_addr && 283 #ifdef MAC 284 mac_ipq_match(m, fp) && 285 #endif 286 ip->ip_p == fp->ipq_p) 287 break; 288 /* 289 * If first fragment to arrive, create a reassembly queue. 290 */ 291 if (fp == NULL) { 292 if (V_ipq[hash].count < V_ipreass_maxbucketsize) 293 fp = uma_zalloc(V_ipq_zone, M_NOWAIT); 294 if (fp == NULL) 295 fp = ipq_reuse(hash); 296 if (fp == NULL) 297 goto dropfrag; 298 #ifdef MAC 299 if (mac_ipq_init(fp, M_NOWAIT) != 0) { 300 uma_zfree(V_ipq_zone, fp); 301 fp = NULL; 302 goto dropfrag; 303 } 304 mac_ipq_create(m, fp); 305 #endif 306 TAILQ_INSERT_HEAD(head, fp, ipq_list); 307 V_ipq[hash].count++; 308 fp->ipq_nfrags = 1; 309 atomic_add_int(&nfrags, 1); 310 fp->ipq_ttl = IPFRAGTTL; 311 fp->ipq_p = ip->ip_p; 312 fp->ipq_id = ip->ip_id; 313 fp->ipq_src = ip->ip_src; 314 fp->ipq_dst = ip->ip_dst; 315 fp->ipq_frags = m; 316 if (m->m_flags & M_IP_FRAG) 317 fp->ipq_maxoff = -1; 318 else 319 fp->ipq_maxoff = ntohs(ip->ip_off) + ntohs(ip->ip_len); 320 m->m_nextpkt = NULL; 321 goto done; 322 } else { 323 /* 324 * If we already saw the last fragment, make sure 325 * this fragment's offset looks sane. Otherwise, if 326 * this is the last fragment, record its endpoint. 327 */ 328 if (fp->ipq_maxoff > 0) { 329 i = ntohs(ip->ip_off) + ntohs(ip->ip_len); 330 if (((m->m_flags & M_IP_FRAG) && i >= fp->ipq_maxoff) || 331 ((m->m_flags & M_IP_FRAG) == 0 && 332 i != fp->ipq_maxoff)) { 333 fp = NULL; 334 goto dropfrag; 335 } 336 } else if ((m->m_flags & M_IP_FRAG) == 0) 337 fp->ipq_maxoff = ntohs(ip->ip_off) + ntohs(ip->ip_len); 338 fp->ipq_nfrags++; 339 atomic_add_int(&nfrags, 1); 340 #ifdef MAC 341 mac_ipq_update(m, fp); 342 #endif 343 } 344 345 #define GETIP(m) ((struct ip*)((m)->m_pkthdr.PH_loc.ptr)) 346 347 /* 348 * Handle ECN by comparing this segment with the first one; 349 * if CE is set, do not lose CE. 350 * drop if CE and not-ECT are mixed for the same packet. 351 */ 352 ecn = ip->ip_tos & IPTOS_ECN_MASK; 353 ecn0 = GETIP(fp->ipq_frags)->ip_tos & IPTOS_ECN_MASK; 354 if (ecn == IPTOS_ECN_CE) { 355 if (ecn0 == IPTOS_ECN_NOTECT) 356 goto dropfrag; 357 if (ecn0 != IPTOS_ECN_CE) 358 GETIP(fp->ipq_frags)->ip_tos |= IPTOS_ECN_CE; 359 } 360 if (ecn == IPTOS_ECN_NOTECT && ecn0 != IPTOS_ECN_NOTECT) 361 goto dropfrag; 362 363 /* 364 * Find a segment which begins after this one does. 365 */ 366 for (p = NULL, q = fp->ipq_frags; q; p = q, q = q->m_nextpkt) 367 if (ntohs(GETIP(q)->ip_off) > ntohs(ip->ip_off)) 368 break; 369 370 /* 371 * If there is a preceding segment, it may provide some of 372 * our data already. If so, drop the data from the incoming 373 * segment. If it provides all of our data, drop us, otherwise 374 * stick new segment in the proper place. 375 * 376 * If some of the data is dropped from the preceding 377 * segment, then it's checksum is invalidated. 378 */ 379 if (p) { 380 i = ntohs(GETIP(p)->ip_off) + ntohs(GETIP(p)->ip_len) - 381 ntohs(ip->ip_off); 382 if (i > 0) { 383 if (i >= ntohs(ip->ip_len)) 384 goto dropfrag; 385 m_adj(m, i); 386 m->m_pkthdr.csum_flags = 0; 387 ip->ip_off = htons(ntohs(ip->ip_off) + i); 388 ip->ip_len = htons(ntohs(ip->ip_len) - i); 389 } 390 m->m_nextpkt = p->m_nextpkt; 391 p->m_nextpkt = m; 392 } else { 393 m->m_nextpkt = fp->ipq_frags; 394 fp->ipq_frags = m; 395 } 396 397 /* 398 * While we overlap succeeding segments trim them or, 399 * if they are completely covered, dequeue them. 400 */ 401 for (; q != NULL && ntohs(ip->ip_off) + ntohs(ip->ip_len) > 402 ntohs(GETIP(q)->ip_off); q = nq) { 403 i = (ntohs(ip->ip_off) + ntohs(ip->ip_len)) - 404 ntohs(GETIP(q)->ip_off); 405 if (i < ntohs(GETIP(q)->ip_len)) { 406 GETIP(q)->ip_len = htons(ntohs(GETIP(q)->ip_len) - i); 407 GETIP(q)->ip_off = htons(ntohs(GETIP(q)->ip_off) + i); 408 m_adj(q, i); 409 q->m_pkthdr.csum_flags = 0; 410 break; 411 } 412 nq = q->m_nextpkt; 413 m->m_nextpkt = nq; 414 IPSTAT_INC(ips_fragdropped); 415 fp->ipq_nfrags--; 416 atomic_subtract_int(&nfrags, 1); 417 m_freem(q); 418 } 419 420 /* 421 * Check for complete reassembly and perform frag per packet 422 * limiting. 423 * 424 * Frag limiting is performed here so that the nth frag has 425 * a chance to complete the packet before we drop the packet. 426 * As a result, n+1 frags are actually allowed per packet, but 427 * only n will ever be stored. (n = maxfragsperpacket.) 428 * 429 */ 430 next = 0; 431 for (p = NULL, q = fp->ipq_frags; q; p = q, q = q->m_nextpkt) { 432 if (ntohs(GETIP(q)->ip_off) != next) { 433 if (fp->ipq_nfrags > V_maxfragsperpacket) 434 ipq_drop(&V_ipq[hash], fp); 435 goto done; 436 } 437 next += ntohs(GETIP(q)->ip_len); 438 } 439 /* Make sure the last packet didn't have the IP_MF flag */ 440 if (p->m_flags & M_IP_FRAG) { 441 if (fp->ipq_nfrags > V_maxfragsperpacket) 442 ipq_drop(&V_ipq[hash], fp); 443 goto done; 444 } 445 446 /* 447 * Reassembly is complete. Make sure the packet is a sane size. 448 */ 449 q = fp->ipq_frags; 450 ip = GETIP(q); 451 if (next + (ip->ip_hl << 2) > IP_MAXPACKET) { 452 IPSTAT_INC(ips_toolong); 453 ipq_drop(&V_ipq[hash], fp); 454 goto done; 455 } 456 457 /* 458 * Concatenate fragments. 459 */ 460 m = q; 461 t = m->m_next; 462 m->m_next = NULL; 463 m_cat(m, t); 464 nq = q->m_nextpkt; 465 q->m_nextpkt = NULL; 466 for (q = nq; q != NULL; q = nq) { 467 nq = q->m_nextpkt; 468 q->m_nextpkt = NULL; 469 m->m_pkthdr.csum_flags &= q->m_pkthdr.csum_flags; 470 m->m_pkthdr.csum_data += q->m_pkthdr.csum_data; 471 m_demote_pkthdr(q); 472 m_cat(m, q); 473 } 474 /* 475 * In order to do checksumming faster we do 'end-around carry' here 476 * (and not in for{} loop), though it implies we are not going to 477 * reassemble more than 64k fragments. 478 */ 479 while (m->m_pkthdr.csum_data & 0xffff0000) 480 m->m_pkthdr.csum_data = (m->m_pkthdr.csum_data & 0xffff) + 481 (m->m_pkthdr.csum_data >> 16); 482 atomic_subtract_int(&nfrags, fp->ipq_nfrags); 483 #ifdef MAC 484 mac_ipq_reassemble(fp, m); 485 mac_ipq_destroy(fp); 486 #endif 487 488 /* 489 * Create header for new ip packet by modifying header of first 490 * packet; dequeue and discard fragment reassembly header. 491 * Make header visible. 492 */ 493 ip->ip_len = htons((ip->ip_hl << 2) + next); 494 ip->ip_src = fp->ipq_src; 495 ip->ip_dst = fp->ipq_dst; 496 TAILQ_REMOVE(head, fp, ipq_list); 497 V_ipq[hash].count--; 498 uma_zfree(V_ipq_zone, fp); 499 m->m_len += (ip->ip_hl << 2); 500 m->m_data -= (ip->ip_hl << 2); 501 /* some debugging cruft by sklower, below, will go away soon */ 502 if (m->m_flags & M_PKTHDR) { /* XXX this should be done elsewhere */ 503 m_fixhdr(m); 504 /* set valid receive interface pointer */ 505 m->m_pkthdr.rcvif = srcifp; 506 } 507 IPSTAT_INC(ips_reassembled); 508 IPQ_UNLOCK(hash); 509 510 #ifdef RSS 511 /* 512 * Query the RSS layer for the flowid / flowtype for the 513 * mbuf payload. 514 * 515 * For now, just assume we have to calculate a new one. 516 * Later on we should check to see if the assigned flowid matches 517 * what RSS wants for the given IP protocol and if so, just keep it. 518 * 519 * We then queue into the relevant netisr so it can be dispatched 520 * to the correct CPU. 521 * 522 * Note - this may return 1, which means the flowid in the mbuf 523 * is correct for the configured RSS hash types and can be used. 524 */ 525 if (rss_mbuf_software_hash_v4(m, 0, &rss_hash, &rss_type) == 0) { 526 m->m_pkthdr.flowid = rss_hash; 527 M_HASHTYPE_SET(m, rss_type); 528 } 529 530 /* 531 * Queue/dispatch for reprocessing. 532 * 533 * Note: this is much slower than just handling the frame in the 534 * current receive context. It's likely worth investigating 535 * why this is. 536 */ 537 netisr_dispatch(NETISR_IP_DIRECT, m); 538 return (NULL); 539 #endif 540 541 /* Handle in-line */ 542 return (m); 543 544 dropfrag: 545 IPSTAT_INC(ips_fragdropped); 546 if (fp != NULL) { 547 fp->ipq_nfrags--; 548 atomic_subtract_int(&nfrags, 1); 549 } 550 m_freem(m); 551 done: 552 IPQ_UNLOCK(hash); 553 return (NULL); 554 555 #undef GETIP 556 } 557 558 /* 559 * Initialize IP reassembly structures. 560 */ 561 void 562 ipreass_init(void) 563 { 564 int max; 565 566 for (int i = 0; i < IPREASS_NHASH; i++) { 567 TAILQ_INIT(&V_ipq[i].head); 568 mtx_init(&V_ipq[i].lock, "IP reassembly", NULL, 569 MTX_DEF | MTX_DUPOK); 570 V_ipq[i].count = 0; 571 } 572 V_ipq_hashseed = arc4random(); 573 V_maxfragsperpacket = 16; 574 V_ipq_zone = uma_zcreate("ipq", sizeof(struct ipq), NULL, NULL, NULL, 575 NULL, UMA_ALIGN_PTR, 0); 576 max = IP_MAXFRAGPACKETS; 577 max = uma_zone_set_max(V_ipq_zone, max); 578 V_ipreass_maxbucketsize = imax(max / (IPREASS_NHASH / 2), 1); 579 580 if (IS_DEFAULT_VNET(curvnet)) { 581 maxfrags = IP_MAXFRAGS; 582 EVENTHANDLER_REGISTER(nmbclusters_change, ipreass_zone_change, 583 NULL, EVENTHANDLER_PRI_ANY); 584 } 585 } 586 587 /* 588 * If a timer expires on a reassembly queue, discard it. 589 */ 590 void 591 ipreass_slowtimo(void) 592 { 593 struct ipq *fp, *tmp; 594 595 for (int i = 0; i < IPREASS_NHASH; i++) { 596 IPQ_LOCK(i); 597 TAILQ_FOREACH_SAFE(fp, &V_ipq[i].head, ipq_list, tmp) 598 if (--fp->ipq_ttl == 0) 599 ipq_timeout(&V_ipq[i], fp); 600 IPQ_UNLOCK(i); 601 } 602 } 603 604 /* 605 * Drain off all datagram fragments. 606 */ 607 void 608 ipreass_drain(void) 609 { 610 611 for (int i = 0; i < IPREASS_NHASH; i++) { 612 IPQ_LOCK(i); 613 while(!TAILQ_EMPTY(&V_ipq[i].head)) 614 ipq_drop(&V_ipq[i], TAILQ_FIRST(&V_ipq[i].head)); 615 KASSERT(V_ipq[i].count == 0, 616 ("%s: V_ipq[%d] count %d (V_ipq=%p)", __func__, i, 617 V_ipq[i].count, V_ipq)); 618 IPQ_UNLOCK(i); 619 } 620 } 621 622 /* 623 * Drain off all datagram fragments belonging to 624 * the given network interface. 625 */ 626 static void 627 ipreass_cleanup(void *arg __unused, struct ifnet *ifp) 628 { 629 struct ipq *fp, *temp; 630 struct mbuf *m; 631 int i; 632 633 KASSERT(ifp != NULL, ("%s: ifp is NULL", __func__)); 634 635 CURVNET_SET_QUIET(ifp->if_vnet); 636 637 /* 638 * Skip processing if IPv4 reassembly is not initialised or 639 * torn down by ipreass_destroy(). 640 */ 641 if (V_ipq_zone == NULL) { 642 CURVNET_RESTORE(); 643 return; 644 } 645 646 for (i = 0; i < IPREASS_NHASH; i++) { 647 IPQ_LOCK(i); 648 /* Scan fragment list. */ 649 TAILQ_FOREACH_SAFE(fp, &V_ipq[i].head, ipq_list, temp) { 650 for (m = fp->ipq_frags; m != NULL; m = m->m_nextpkt) { 651 /* clear no longer valid rcvif pointer */ 652 if (m->m_pkthdr.rcvif == ifp) 653 m->m_pkthdr.rcvif = NULL; 654 } 655 } 656 IPQ_UNLOCK(i); 657 } 658 CURVNET_RESTORE(); 659 } 660 EVENTHANDLER_DEFINE(ifnet_departure_event, ipreass_cleanup, NULL, 0); 661 662 #ifdef VIMAGE 663 /* 664 * Destroy IP reassembly structures. 665 */ 666 void 667 ipreass_destroy(void) 668 { 669 670 ipreass_drain(); 671 uma_zdestroy(V_ipq_zone); 672 V_ipq_zone = NULL; 673 for (int i = 0; i < IPREASS_NHASH; i++) 674 mtx_destroy(&V_ipq[i].lock); 675 } 676 #endif 677 678 /* 679 * After maxnipq has been updated, propagate the change to UMA. The UMA zone 680 * max has slightly different semantics than the sysctl, for historical 681 * reasons. 682 */ 683 static void 684 ipreass_drain_tomax(void) 685 { 686 struct ipq *fp; 687 int target; 688 689 /* 690 * Make sure each bucket is under the new limit. If 691 * necessary, drop enough of the oldest elements from 692 * each bucket to get under the new limit. 693 */ 694 for (int i = 0; i < IPREASS_NHASH; i++) { 695 IPQ_LOCK(i); 696 while (V_ipq[i].count > V_ipreass_maxbucketsize && 697 (fp = TAILQ_LAST(&V_ipq[i].head, ipqhead)) != NULL) 698 ipq_timeout(&V_ipq[i], fp); 699 IPQ_UNLOCK(i); 700 } 701 702 /* 703 * If we are over the maximum number of fragments, 704 * drain off enough to get down to the new limit, 705 * stripping off last elements on queues. Every 706 * run we strip the oldest element from each bucket. 707 */ 708 target = uma_zone_get_max(V_ipq_zone); 709 while (uma_zone_get_cur(V_ipq_zone) > target) { 710 for (int i = 0; i < IPREASS_NHASH; i++) { 711 IPQ_LOCK(i); 712 fp = TAILQ_LAST(&V_ipq[i].head, ipqhead); 713 if (fp != NULL) 714 ipq_timeout(&V_ipq[i], fp); 715 IPQ_UNLOCK(i); 716 } 717 } 718 } 719 720 static void 721 ipreass_zone_change(void *tag) 722 { 723 VNET_ITERATOR_DECL(vnet_iter); 724 int max; 725 726 maxfrags = IP_MAXFRAGS; 727 max = IP_MAXFRAGPACKETS; 728 VNET_LIST_RLOCK_NOSLEEP(); 729 VNET_FOREACH(vnet_iter) { 730 CURVNET_SET(vnet_iter); 731 max = uma_zone_set_max(V_ipq_zone, max); 732 V_ipreass_maxbucketsize = imax(max / (IPREASS_NHASH / 2), 1); 733 ipreass_drain_tomax(); 734 CURVNET_RESTORE(); 735 } 736 VNET_LIST_RUNLOCK_NOSLEEP(); 737 } 738 739 /* 740 * Change the limit on the UMA zone, or disable the fragment allocation 741 * at all. Since 0 and -1 is a special values here, we need our own handler, 742 * instead of sysctl_handle_uma_zone_max(). 743 */ 744 static int 745 sysctl_maxfragpackets(SYSCTL_HANDLER_ARGS) 746 { 747 int error, max; 748 749 if (V_noreass == 0) { 750 max = uma_zone_get_max(V_ipq_zone); 751 if (max == 0) 752 max = -1; 753 } else 754 max = 0; 755 error = sysctl_handle_int(oidp, &max, 0, req); 756 if (error || !req->newptr) 757 return (error); 758 if (max > 0) { 759 /* 760 * XXXRW: Might be a good idea to sanity check the argument 761 * and place an extreme upper bound. 762 */ 763 max = uma_zone_set_max(V_ipq_zone, max); 764 V_ipreass_maxbucketsize = imax(max / (IPREASS_NHASH / 2), 1); 765 ipreass_drain_tomax(); 766 V_noreass = 0; 767 } else if (max == 0) { 768 V_noreass = 1; 769 ipreass_drain(); 770 } else if (max == -1) { 771 V_noreass = 0; 772 uma_zone_set_max(V_ipq_zone, 0); 773 V_ipreass_maxbucketsize = INT_MAX; 774 } else 775 return (EINVAL); 776 return (0); 777 } 778 779 /* 780 * Seek for old fragment queue header that can be reused. Try to 781 * reuse a header from currently locked hash bucket. 782 */ 783 static struct ipq * 784 ipq_reuse(int start) 785 { 786 struct ipq *fp; 787 int bucket, i; 788 789 IPQ_LOCK_ASSERT(start); 790 791 for (i = 0; i < IPREASS_NHASH; i++) { 792 bucket = (start + i) % IPREASS_NHASH; 793 if (bucket != start && IPQ_TRYLOCK(bucket) == 0) 794 continue; 795 fp = TAILQ_LAST(&V_ipq[bucket].head, ipqhead); 796 if (fp) { 797 struct mbuf *m; 798 799 IPSTAT_ADD(ips_fragtimeout, fp->ipq_nfrags); 800 atomic_subtract_int(&nfrags, fp->ipq_nfrags); 801 while (fp->ipq_frags) { 802 m = fp->ipq_frags; 803 fp->ipq_frags = m->m_nextpkt; 804 m_freem(m); 805 } 806 TAILQ_REMOVE(&V_ipq[bucket].head, fp, ipq_list); 807 V_ipq[bucket].count--; 808 if (bucket != start) 809 IPQ_UNLOCK(bucket); 810 break; 811 } 812 if (bucket != start) 813 IPQ_UNLOCK(bucket); 814 } 815 IPQ_LOCK_ASSERT(start); 816 return (fp); 817 } 818 819 /* 820 * Free a fragment reassembly header and all associated datagrams. 821 */ 822 static void 823 ipq_free(struct ipqbucket *bucket, struct ipq *fp) 824 { 825 struct mbuf *q; 826 827 atomic_subtract_int(&nfrags, fp->ipq_nfrags); 828 while (fp->ipq_frags) { 829 q = fp->ipq_frags; 830 fp->ipq_frags = q->m_nextpkt; 831 m_freem(q); 832 } 833 TAILQ_REMOVE(&bucket->head, fp, ipq_list); 834 bucket->count--; 835 uma_zfree(V_ipq_zone, fp); 836 } 837 838 /* 839 * Get or set the maximum number of reassembly queues per bucket. 840 */ 841 static int 842 sysctl_maxfragbucketsize(SYSCTL_HANDLER_ARGS) 843 { 844 int error, max; 845 846 max = V_ipreass_maxbucketsize; 847 error = sysctl_handle_int(oidp, &max, 0, req); 848 if (error || !req->newptr) 849 return (error); 850 if (max <= 0) 851 return (EINVAL); 852 V_ipreass_maxbucketsize = max; 853 ipreass_drain_tomax(); 854 return (0); 855 } 856