2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* rx_user.c contains routines specific to the user space UNIX implementation of rx */
12 #include <afsconfig.h>
13 #include <afs/param.h>
18 # include <sys/types.h>
23 # include <WINNT/syscfg.h>
25 # include <sys/socket.h>
26 # include <sys/file.h>
28 # include <sys/stat.h>
29 # include <netinet/in.h>
30 # include <sys/time.h>
32 # include <sys/ioctl.h>
35 #if !defined(AFS_AIX_ENV) && !defined(AFS_NT40_ENV)
36 # include <sys/syscall.h>
38 #include <afs/afs_args.h>
39 #include <afs/afsutil.h>
41 #ifndef IPPORT_USERRESERVED
42 /* If in.h doesn't define this, define it anyway. Unfortunately, defining
43 this doesn't put the code into the kernel to restrict kernel assigned
44 port numbers to numbers below IPPORT_USERRESERVED... */
45 #define IPPORT_USERRESERVED 5000
49 # include <sys/time.h>
51 # include "rx_internal.h"
53 # include "rx_globals.h"
55 #ifdef AFS_PTHREAD_ENV
59 * The rx_if_init_mutex mutex protects the following global variables:
63 pthread_mutex_t rx_if_init_mutex;
64 #define LOCK_IF_INIT assert(pthread_mutex_lock(&rx_if_init_mutex)==0)
65 #define UNLOCK_IF_INIT assert(pthread_mutex_unlock(&rx_if_init_mutex)==0)
68 * The rx_if_mutex mutex protects the following global variables:
74 pthread_mutex_t rx_if_mutex;
75 #define LOCK_IF assert(pthread_mutex_lock(&rx_if_mutex)==0)
76 #define UNLOCK_IF assert(pthread_mutex_unlock(&rx_if_mutex)==0)
79 #define UNLOCK_IF_INIT
82 #endif /* AFS_PTHREAD_ENV */
86 * Make a socket for receiving/sending IP packets. Set it into non-blocking
87 * and large buffering modes. If port isn't specified, the kernel will pick
88 * one. Returns the socket (>= 0) on success. Returns OSI_NULLSOCKET on
89 * failure. Port must be in network byte order.
92 rxi_GetHostUDPSocket(u_int ahost, u_short port)
95 osi_socket socketFd = OSI_NULLSOCKET;
96 struct sockaddr_in taddr;
97 char *name = "rxi_GetUDPSocket: ";
98 #ifdef AFS_LINUX22_ENV
99 #if defined(ADAPT_PMTU)
100 int pmtu=IP_PMTUDISC_WANT;
103 int pmtu=IP_PMTUDISC_DONT;
106 #if defined(HAVE_LINUX_ERRQUEUE_H) && defined(ADAPT_PMTU)
107 #include <linux/types.h>
108 #include <linux/errqueue.h>
114 #if !defined(AFS_NT40_ENV)
115 if (ntohs(port) >= IPPORT_RESERVED && ntohs(port) < IPPORT_USERRESERVED) {
116 /* (osi_Msg "%s*WARNING* port number %d is not a reserved port number. Use port numbers above %d\n", name, port, IPPORT_USERRESERVED);
119 if (ntohs(port) > 0 && ntohs(port) < IPPORT_RESERVED && geteuid() != 0) {
121 "%sport number %d is a reserved port number which may only be used by root. Use port numbers above %d\n",
122 name, ntohs(port), IPPORT_USERRESERVED);
126 socketFd = socket(AF_INET, SOCK_DGRAM, 0);
133 taddr.sin_addr.s_addr = ahost;
134 taddr.sin_family = AF_INET;
135 taddr.sin_port = (u_short) port;
136 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
137 taddr.sin_len = sizeof(struct sockaddr_in);
139 #define MAX_RX_BINDS 10
140 for (binds = 0; binds < MAX_RX_BINDS; binds++) {
143 code = bind(socketFd, (struct sockaddr *)&taddr, sizeof(taddr));
149 (osi_Msg "%sbind failed\n", name);
152 #if !defined(AFS_NT40_ENV)
154 * Set close-on-exec on rx socket
156 fcntl(socketFd, F_SETFD, 1);
159 /* Use one of three different ways of getting a socket buffer expanded to
167 len2 = rx_UdpBufSize;
169 /* find the size closest to rx_UdpBufSize that will be accepted */
170 while (!greedy && len2 > len1) {
173 (socketFd, SOL_SOCKET, SO_RCVBUF, (char *)&len2,
179 /* but do not let it get smaller than 32K */
189 (socketFd, SOL_SOCKET, SO_SNDBUF, (char *)&len1,
193 (socketFd, SOL_SOCKET, SO_RCVBUF, (char *)&len2,
196 (osi_Msg "%s*WARNING* Unable to increase buffering on socket\n",
198 MUTEX_ENTER(&rx_stats_mutex);
199 rx_stats.socketGreedy = greedy;
200 MUTEX_EXIT(&rx_stats_mutex);
203 #ifdef AFS_LINUX22_ENV
204 setsockopt(socketFd, SOL_IP, IP_MTU_DISCOVER, &pmtu, sizeof(pmtu));
205 #if defined(ADAPT_PMTU)
206 setsockopt(socketFd, SOL_IP, IP_RECVERR, &recverr, sizeof(recverr));
209 if (rxi_Listen(socketFd) < 0) {
218 closesocket(socketFd);
224 return OSI_NULLSOCKET;
228 rxi_GetUDPSocket(u_short port)
230 return rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
234 osi_Panic(char *msg, ...)
238 (osi_Msg "Fatal Rx error: ");
247 * osi_AssertFailU() -- used by the osi_Assert() macro.
251 osi_AssertFailU(const char *expr, const char *file, int line)
253 osi_Panic("assertion failed: %s, file: %s, line: %d\n", expr,
257 #if defined(AFS_AIX32_ENV) && !defined(KERNEL)
259 static const char memZero;
261 osi_Alloc(afs_int32 x)
264 * 0-length allocs may return NULL ptr from malloc, so we special-case
265 * things so that NULL returned iff an error occurred
268 return (void *)&memZero;
273 osi_Free(void *x, afs_int32 size)
280 #endif /* defined(AFS_AIX32_ENV) && !defined(KERNEL) */
282 #define ADDRSPERSITE 16
285 static afs_uint32 rxi_NetAddrs[ADDRSPERSITE]; /* host order */
286 static int myNetMTUs[ADDRSPERSITE];
287 static int myNetMasks[ADDRSPERSITE];
288 static int myNetFlags[ADDRSPERSITE];
289 static u_int rxi_numNetAddrs;
290 static int Inited = 0;
292 #if defined(AFS_NT40_ENV)
296 /* The IP address list can change so we must query for it */
299 /* we don't want to use the loopback adapter which is first */
300 /* this is a bad bad hack */
301 if (rxi_numNetAddrs > 1)
302 return htonl(rxi_NetAddrs[1]);
303 else if (rxi_numNetAddrs > 0)
304 return htonl(rxi_NetAddrs[0]);
310 ** return number of addresses
311 ** and the addresses themselves in the buffer
312 ** maxSize - max number of interfaces to return.
315 rx_getAllAddr(afs_uint32 * buffer, int maxSize)
317 int count = 0, offset = 0;
319 /* The IP address list can change so we must query for it */
323 /* we don't want to use the loopback adapter which is first */
324 /* this is a bad bad hack.
325 * and doesn't hold true on Windows.
327 if ( rxi_numNetAddrs > 1 )
329 #endif /* AFS_DJGPP_ENV */
331 for (count = 0; offset < rxi_numNetAddrs && maxSize > 0;
332 count++, offset++, maxSize--)
333 buffer[count] = htonl(rxi_NetAddrs[offset]);
338 /* this function returns the total number of interface addresses
339 * the buffer has to be passed in by the caller. It also returns
340 * the matching interface mask and mtu. All values are returned
341 * in network byte order.
344 rx_getAllAddrMaskMtu(afs_uint32 addrBuffer[], afs_uint32 maskBuffer[],
345 afs_uint32 mtuBuffer[], int maxSize)
347 int count = 0, offset = 0;
349 /* The IP address list can change so we must query for it */
353 /* we don't want to use the loopback adapter which is first */
354 /* this is a bad bad hack.
355 * and doesn't hold true on Windows.
357 if ( rxi_numNetAddrs > 1 )
359 #endif /* AFS_DJGPP_ENV */
362 offset < rxi_numNetAddrs && maxSize > 0;
363 count++, offset++, maxSize--) {
364 addrBuffer[count] = htonl(rxi_NetAddrs[offset]);
365 maskBuffer[count] = htonl(myNetMasks[offset]);
366 mtuBuffer[count] = htonl(myNetMTUs[offset]);
373 extern int rxinit_status;
375 rxi_InitMorePackets(void) {
376 int npackets, ncbufs;
378 ncbufs = (rx_maxJumboRecvSize - RX_FIRSTBUFFERSIZE);
380 ncbufs = ncbufs / RX_CBUFFERSIZE;
381 npackets = rx_initSendWindow - 1;
382 rxi_MorePackets(npackets * (ncbufs + 1));
394 if (Inited < 2 && rxinit_status == 0) {
395 /* We couldn't initialize more packets earlier.
397 rxi_InitMorePackets();
407 rxi_numNetAddrs = ADDRSPERSITE;
408 (void)syscfg_GetIFInfo(&rxi_numNetAddrs, rxi_NetAddrs,
409 myNetMasks, myNetMTUs, myNetFlags);
411 for (i = 0; i < rxi_numNetAddrs; i++) {
412 rxsize = rxi_AdjustIfMTU(myNetMTUs[i] - RX_IPUDP_SIZE);
414 rxi_nRecvFrags * rxsize + (rxi_nRecvFrags - 1) * UDP_HDR_SIZE;
415 maxsize = rxi_AdjustMaxMTU(rxsize, maxsize);
416 if (rx_maxReceiveSize < maxsize) {
417 rx_maxReceiveSize = MIN(RX_MAX_PACKET_SIZE, maxsize);
419 MIN(rx_maxReceiveSize, rx_maxReceiveSizeUser);
426 * If rxinit_status is still set, rx_InitHost() has yet to be called
427 * and we therefore do not have any mutex locks initialized. As a
428 * result we cannot call rxi_MorePackets() without crashing.
433 rxi_InitMorePackets();
438 fudge_netmask(afs_uint32 addr)
444 else if (IN_CLASSB(addr))
446 else if (IN_CLASSC(addr))
456 #if !defined(AFS_AIX_ENV) && !defined(AFS_NT40_ENV) && !defined(AFS_LINUX20_ENV)
458 rxi_syscall(a3, a4, a5)
465 old = (void (*)())signal(SIGSYS, SIG_IGN);
467 #if defined(AFS_SGI_ENV)
468 rcode = afs_syscall(a3, a4, a5);
470 rcode = syscall(AFS_SYSCALL, 28 /* AFSCALL_CALL */ , a3, a4, a5);
471 #endif /* AFS_SGI_ENV */
477 #endif /* AFS_AIX_ENV */
486 struct ifreq ifs[ADDRSPERSITE];
489 char buf[BUFSIZ], *cp, *cplim;
491 struct sockaddr_in *a;
502 memset(rxi_NetAddrs, 0, sizeof(rxi_NetAddrs));
503 memset(myNetFlags, 0, sizeof(myNetFlags));
504 memset(myNetMTUs, 0, sizeof(myNetMTUs));
505 memset(myNetMasks, 0, sizeof(myNetMasks));
507 s = socket(AF_INET, SOCK_DGRAM, 0);
512 ifc.ifc_len = sizeof(buf);
516 ifc.ifc_len = sizeof(ifs);
517 ifc.ifc_buf = (caddr_t) & ifs[0];
518 memset(&ifs[0], 0, sizeof(ifs));
520 res = ioctl(s, SIOCGIFCONF, &ifc);
522 /* fputs(stderr, "ioctl error IFCONF\n"); */
529 #define size(p) MAX((p).sa_len, sizeof(p))
530 cplim = buf + ifc.ifc_len; /*skip over if's with big ifr_addr's */
531 for (cp = buf; cp < cplim;
532 cp += sizeof(ifr->ifr_name) + MAX(a->sin_len, sizeof(*a))) {
533 if (rxi_numNetAddrs >= ADDRSPERSITE)
536 ifr = (struct ifreq *)cp;
538 len = ifc.ifc_len / sizeof(struct ifreq);
539 if (len > ADDRSPERSITE)
542 for (i = 0; i < len; ++i) {
544 res = ioctl(s, SIOCGIFADDR, ifr);
547 /* fputs(stderr, "ioctl error IFADDR\n");
548 * perror(ifr->ifr_name); */
551 a = (struct sockaddr_in *)&ifr->ifr_addr;
552 if (a->sin_family != AF_INET)
554 rxi_NetAddrs[rxi_numNetAddrs] = ntohl(a->sin_addr.s_addr);
555 if (rxi_NetAddrs[rxi_numNetAddrs] == 0x7f000001) {
556 /* we don't really care about "localhost" */
559 for (j = 0; j < rxi_numNetAddrs; j++) {
560 if (rxi_NetAddrs[j] == rxi_NetAddrs[rxi_numNetAddrs])
563 if (j < rxi_numNetAddrs)
566 /* fprintf(stderr, "if %s addr=%x\n", ifr->ifr_name,
567 * rxi_NetAddrs[rxi_numNetAddrs]); */
570 res = ioctl(s, SIOCGIFFLAGS, ifr);
572 myNetFlags[rxi_numNetAddrs] = ifr->ifr_flags;
574 /* Handle aliased loopbacks as well. */
575 if (ifr->ifr_flags & IFF_LOOPBACK)
578 /* fprintf(stderr, "if %s flags=%x\n",
579 * ifr->ifr_name, ifr->ifr_flags); */
581 * fputs(stderr, "ioctl error IFFLAGS\n");
582 * perror(ifr->ifr_name); */
584 #endif /* SIOCGIFFLAGS */
586 #if !defined(AFS_AIX_ENV) && !defined(AFS_LINUX20_ENV)
587 /* this won't run on an AIX system w/o a cache manager */
588 rxi_syscallp = rxi_syscall;
591 /* If I refer to kernel extensions that aren't loaded on AIX, the
592 * program refuses to load and run, so I simply can't include the
593 * following code. Fortunately, AIX is the one operating system in
594 * which the subsequent ioctl works reliably. */
596 if ((*rxi_syscallp) (20 /*AFSOP_GETMTU */ ,
597 htonl(rxi_NetAddrs[rxi_numNetAddrs]),
598 &(myNetMTUs[rxi_numNetAddrs]))) {
599 /* fputs(stderr, "syscall error GETMTU\n");
600 * perror(ifr->ifr_name); */
601 myNetMTUs[rxi_numNetAddrs] = 0;
603 if ((*rxi_syscallp) (42 /*AFSOP_GETMASK */ ,
604 htonl(rxi_NetAddrs[rxi_numNetAddrs]),
605 &(myNetMasks[rxi_numNetAddrs]))) {
606 /* fputs(stderr, "syscall error GETMASK\n");
607 * perror(ifr->ifr_name); */
608 myNetMasks[rxi_numNetAddrs] = 0;
610 myNetMasks[rxi_numNetAddrs] =
611 ntohl(myNetMasks[rxi_numNetAddrs]);
612 /* fprintf(stderr, "if %s mask=0x%x\n",
613 * ifr->ifr_name, myNetMasks[rxi_numNetAddrs]); */
616 if (myNetMTUs[rxi_numNetAddrs] == 0) {
617 myNetMTUs[rxi_numNetAddrs] = OLD_MAX_PACKET_SIZE + RX_IPUDP_SIZE;
619 res = ioctl(s, SIOCGIFMTU, ifr);
620 if ((res == 0) && (ifr->ifr_metric > 128)) { /* sanity check */
621 myNetMTUs[rxi_numNetAddrs] = ifr->ifr_metric;
622 /* fprintf(stderr, "if %s mtu=%d\n",
623 * ifr->ifr_name, ifr->ifr_metric); */
625 /* fputs(stderr, "ioctl error IFMTU\n");
626 * perror(ifr->ifr_name); */
631 if (myNetMasks[rxi_numNetAddrs] == 0) {
632 myNetMasks[rxi_numNetAddrs] =
633 fudge_netmask(rxi_NetAddrs[rxi_numNetAddrs]);
634 #ifdef SIOCGIFNETMASK
635 res = ioctl(s, SIOCGIFNETMASK, ifr);
637 a = (struct sockaddr_in *)&ifr->ifr_addr;
638 myNetMasks[rxi_numNetAddrs] = ntohl(a->sin_addr.s_addr);
639 /* fprintf(stderr, "if %s subnetmask=0x%x\n",
640 * ifr->ifr_name, myNetMasks[rxi_numNetAddrs]); */
642 /* fputs(stderr, "ioctl error IFMASK\n");
643 * perror(ifr->ifr_name); */
648 if (rxi_NetAddrs[rxi_numNetAddrs] != 0x7f000001) { /* ignore lo0 */
651 rxi_nRecvFrags * (myNetMTUs[rxi_numNetAddrs] - RX_IP_SIZE);
652 maxsize -= UDP_HDR_SIZE; /* only the first frag has a UDP hdr */
653 if (rx_maxReceiveSize < maxsize)
654 rx_maxReceiveSize = MIN(RX_MAX_PACKET_SIZE, maxsize);
661 /* have to allocate at least enough to allow a single packet to reach its
662 * maximum size, so ReadPacket will work. Allocate enough for a couple
663 * of packets to do so, for good measure */
665 int npackets, ncbufs;
667 rx_maxJumboRecvSize =
668 RX_HEADER_SIZE + rxi_nDgramPackets * RX_JUMBOBUFFERSIZE +
669 (rxi_nDgramPackets - 1) * RX_JUMBOHEADERSIZE;
670 rx_maxJumboRecvSize = MAX(rx_maxJumboRecvSize, rx_maxReceiveSize);
671 ncbufs = (rx_maxJumboRecvSize - RX_FIRSTBUFFERSIZE);
673 ncbufs = ncbufs / RX_CBUFFERSIZE;
674 npackets = rx_initSendWindow - 1;
675 rxi_MorePackets(npackets * (ncbufs + 1));
679 #endif /* AFS_NT40_ENV */
681 /* Called from rxi_FindPeer, when initializing a clear rx_peer structure,
682 * to get interesting information.
683 * Curiously enough, the rx_peerHashTable_lock currently protects the
684 * Inited variable (and hence rx_GetIFInfo). When the fs suite uses
685 * pthreads, this issue will need to be revisited.
689 rxi_InitPeerParams(struct rx_peer *pp)
694 #if defined(ADAPT_PMTU) && defined(IP_MTU)
696 struct sockaddr_in addr;
703 * there's a race here since more than one thread could call
704 * rx_GetIFInfo. The race stops in rx_GetIFInfo.
712 /* try to second-guess IP, and identify which link is most likely to
713 * be used for traffic to/from this host. */
714 ppaddr = ntohl(pp->host);
718 pp->rateFlag = 2; /* start timing after two full packets */
719 /* I don't initialize these, because I presume they are bzero'd...
720 * pp->burstSize pp->burst pp->burstWait.sec pp->burstWait.usec
721 * pp->timeout.usec */
724 for (ix = 0; ix < rxi_numNetAddrs; ++ix) {
725 if ((rxi_NetAddrs[ix] & myNetMasks[ix]) == (ppaddr & myNetMasks[ix])) {
726 #ifdef IFF_POINTOPOINT
727 if (myNetFlags[ix] & IFF_POINTOPOINT)
729 #endif /* IFF_POINTOPOINT */
730 rxmtu = myNetMTUs[ix] - RX_IPUDP_SIZE;
731 if (rxmtu < RX_MIN_PACKET_SIZE)
732 rxmtu = RX_MIN_PACKET_SIZE;
733 if (pp->ifMTU < rxmtu)
734 pp->ifMTU = MIN(rx_MyMaxSendSize, rxmtu);
738 if (!pp->ifMTU) { /* not local */
740 pp->ifMTU = MIN(rx_MyMaxSendSize, RX_REMOTE_PACKET_SIZE);
742 #else /* ADAPT_MTU */
743 pp->rateFlag = 2; /* start timing after two full packets */
745 pp->ifMTU = MIN(rx_MyMaxSendSize, OLD_MAX_PACKET_SIZE);
746 #endif /* ADAPT_MTU */
747 #if defined(ADAPT_PMTU) && defined(IP_MTU)
748 sock=socket(PF_INET, SOCK_DGRAM, IPPROTO_UDP);
750 addr.sin_family = AF_INET;
751 addr.sin_addr.s_addr = pp->host;
752 addr.sin_port = pp->port;
753 if (connect(sock, (struct sockaddr *)&addr, sizeof(addr)) == 0) {
755 socklen_t s = sizeof(mtu);
756 if (getsockopt(sock, SOL_IP, IP_MTU, &mtu, &s)== 0) {
757 pp->ifMTU = MIN(mtu - RX_IPUDP_SIZE, pp->ifMTU);
763 pp->ifMTU = rxi_AdjustIfMTU(pp->ifMTU);
764 pp->maxMTU = OLD_MAX_PACKET_SIZE; /* for compatibility with old guys */
765 pp->natMTU = MIN((int)pp->ifMTU, OLD_MAX_PACKET_SIZE);
766 pp->maxDgramPackets =
767 MIN(rxi_nDgramPackets,
768 rxi_AdjustDgramPackets(rxi_nSendFrags, pp->ifMTU));
770 MIN(rxi_nDgramPackets,
771 rxi_AdjustDgramPackets(rxi_nSendFrags, pp->ifMTU));
772 pp->maxDgramPackets = 1;
773 /* Initialize slow start parameters */
774 pp->MTU = MIN(pp->natMTU, pp->maxMTU);
776 pp->nDgramPackets = 1;
780 /* Don't expose jumobgram internals. */
784 rx_maxReceiveSize = OLD_MAX_PACKET_SIZE;
785 rxi_nSendFrags = rxi_nRecvFrags = 1;
788 /* Override max MTU. If rx_SetNoJumbo is called, it must be
789 called before calling rx_SetMaxMTU since SetNoJumbo clobbers rx_maxReceiveSize */
791 rx_SetMaxMTU(int mtu)
793 rx_MyMaxSendSize = rx_maxReceiveSizeUser = rx_maxReceiveSize = mtu;
796 #if defined(HAVE_LINUX_ERRQUEUE_H) && defined(ADAPT_PMTU)
798 rxi_HandleSocketError(int socket)
801 struct cmsghdr *cmsg;
802 struct sock_extended_err *err;
803 struct sockaddr_in addr;
804 struct sockaddr *offender;
805 char controlmsgbuf[256];
809 msg.msg_name = &addr;
810 msg.msg_namelen = sizeof(addr);
813 msg.msg_control = controlmsgbuf;
814 msg.msg_controllen = 256;
816 code = recvmsg(socket, &msg, MSG_ERRQUEUE|MSG_DONTWAIT|MSG_TRUNC);
818 if (code < 0 || !(msg.msg_flags & MSG_ERRQUEUE))
821 for (cmsg = CMSG_FIRSTHDR(&msg); cmsg; cmsg = CMSG_NXTHDR(&msg, cmsg)) {
822 if ((char *)cmsg - controlmsgbuf > msg.msg_controllen - CMSG_SPACE(0) ||
823 (char *)cmsg - controlmsgbuf > msg.msg_controllen - CMSG_SPACE(cmsg->cmsg_len) ||
824 cmsg->cmsg_len == 0) {
828 if (cmsg->cmsg_level == SOL_IP && cmsg->cmsg_type == IP_RECVERR)
834 err =(struct sock_extended_err *) CMSG_DATA(cmsg);
836 if (err->ee_errno == EMSGSIZE && err->ee_info >= 68) {
837 rxi_SetPeerMtu(addr.sin_addr.s_addr, addr.sin_port,
838 err->ee_info - RX_IPUDP_SIZE);
840 /* other DEST_UNREACH's and TIME_EXCEEDED should be dealt with too */