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;
170 (socketFd, SOL_SOCKET, SO_RCVBUF, (char *)&len2,
173 len2 = 32766; /* fall back to old size... uh-oh! */
178 (socketFd, SOL_SOCKET, SO_SNDBUF, (char *)&len1,
182 (socketFd, SOL_SOCKET, SO_RCVBUF, (char *)&len2,
185 (osi_Msg "%s*WARNING* Unable to increase buffering on socket\n",
187 MUTEX_ENTER(&rx_stats_mutex);
188 rx_stats.socketGreedy = greedy;
189 MUTEX_EXIT(&rx_stats_mutex);
192 #ifdef AFS_LINUX22_ENV
193 setsockopt(socketFd, SOL_IP, IP_MTU_DISCOVER, &pmtu, sizeof(pmtu));
194 #if defined(ADAPT_PMTU)
195 setsockopt(socketFd, SOL_IP, IP_RECVERR, &recverr, sizeof(recverr));
198 if (rxi_Listen(socketFd) < 0) {
207 closesocket(socketFd);
213 return OSI_NULLSOCKET;
217 rxi_GetUDPSocket(u_short port)
219 return rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
223 osi_Panic(char *msg, ...)
227 (osi_Msg "Fatal Rx error: ");
236 * osi_AssertFailU() -- used by the osi_Assert() macro.
240 osi_AssertFailU(const char *expr, const char *file, int line)
242 osi_Panic("assertion failed: %s, file: %s, line: %d\n", expr,
246 #if defined(AFS_AIX32_ENV) && !defined(KERNEL)
248 static const char memZero;
250 osi_Alloc(afs_int32 x)
253 * 0-length allocs may return NULL ptr from malloc, so we special-case
254 * things so that NULL returned iff an error occurred
257 return (void *)&memZero;
262 osi_Free(void *x, afs_int32 size)
269 #endif /* defined(AFS_AIX32_ENV) && !defined(KERNEL) */
271 #define ADDRSPERSITE 16
274 static afs_uint32 rxi_NetAddrs[ADDRSPERSITE]; /* host order */
275 static int myNetMTUs[ADDRSPERSITE];
276 static int myNetMasks[ADDRSPERSITE];
277 static int myNetFlags[ADDRSPERSITE];
278 static u_int rxi_numNetAddrs;
279 static int Inited = 0;
281 #if defined(AFS_NT40_ENV)
285 /* The IP address list can change so we must query for it */
288 /* we don't want to use the loopback adapter which is first */
289 /* this is a bad bad hack */
290 if (rxi_numNetAddrs > 1)
291 return htonl(rxi_NetAddrs[1]);
292 else if (rxi_numNetAddrs > 0)
293 return htonl(rxi_NetAddrs[0]);
299 ** return number of addresses
300 ** and the addresses themselves in the buffer
301 ** maxSize - max number of interfaces to return.
304 rx_getAllAddr(afs_uint32 * buffer, int maxSize)
306 int count = 0, offset = 0;
308 /* The IP address list can change so we must query for it */
312 /* we don't want to use the loopback adapter which is first */
313 /* this is a bad bad hack.
314 * and doesn't hold true on Windows.
316 if ( rxi_numNetAddrs > 1 )
318 #endif /* AFS_DJGPP_ENV */
320 for (count = 0; offset < rxi_numNetAddrs && maxSize > 0;
321 count++, offset++, maxSize--)
322 buffer[count] = htonl(rxi_NetAddrs[offset]);
327 /* this function returns the total number of interface addresses
328 * the buffer has to be passed in by the caller. It also returns
329 * the matching interface mask and mtu. All values are returned
330 * in network byte order.
333 rx_getAllAddrMaskMtu(afs_uint32 addrBuffer[], afs_uint32 maskBuffer[],
334 afs_uint32 mtuBuffer[], int maxSize)
336 int count = 0, offset = 0;
338 /* The IP address list can change so we must query for it */
342 /* we don't want to use the loopback adapter which is first */
343 /* this is a bad bad hack.
344 * and doesn't hold true on Windows.
346 if ( rxi_numNetAddrs > 1 )
348 #endif /* AFS_DJGPP_ENV */
351 offset < rxi_numNetAddrs && maxSize > 0;
352 count++, offset++, maxSize--) {
353 addrBuffer[count] = htonl(rxi_NetAddrs[offset]);
354 maskBuffer[count] = htonl(myNetMasks[offset]);
355 mtuBuffer[count] = htonl(myNetMTUs[offset]);
362 extern int rxinit_status;
364 rxi_InitMorePackets(void) {
365 int npackets, ncbufs;
367 ncbufs = (rx_maxJumboRecvSize - RX_FIRSTBUFFERSIZE);
369 ncbufs = ncbufs / RX_CBUFFERSIZE;
370 npackets = rx_initSendWindow - 1;
371 rxi_MorePackets(npackets * (ncbufs + 1));
383 if (Inited < 2 && rxinit_status == 0) {
384 /* We couldn't initialize more packets earlier.
386 rxi_InitMorePackets();
396 rxi_numNetAddrs = ADDRSPERSITE;
397 (void)syscfg_GetIFInfo(&rxi_numNetAddrs, rxi_NetAddrs,
398 myNetMasks, myNetMTUs, myNetFlags);
400 for (i = 0; i < rxi_numNetAddrs; i++) {
401 rxsize = rxi_AdjustIfMTU(myNetMTUs[i] - RX_IPUDP_SIZE);
403 rxi_nRecvFrags * rxsize + (rxi_nRecvFrags - 1) * UDP_HDR_SIZE;
404 maxsize = rxi_AdjustMaxMTU(rxsize, maxsize);
405 if (rx_maxReceiveSize < maxsize) {
406 rx_maxReceiveSize = MIN(RX_MAX_PACKET_SIZE, maxsize);
408 MIN(rx_maxReceiveSize, rx_maxReceiveSizeUser);
415 * If rxinit_status is still set, rx_InitHost() has yet to be called
416 * and we therefore do not have any mutex locks initialized. As a
417 * result we cannot call rxi_MorePackets() without crashing.
422 rxi_InitMorePackets();
427 fudge_netmask(afs_uint32 addr)
433 else if (IN_CLASSB(addr))
435 else if (IN_CLASSC(addr))
445 #if !defined(AFS_AIX_ENV) && !defined(AFS_NT40_ENV) && !defined(AFS_LINUX20_ENV)
447 rxi_syscall(a3, a4, a5)
454 old = (void (*)())signal(SIGSYS, SIG_IGN);
456 #if defined(AFS_SGI_ENV)
457 rcode = afs_syscall(a3, a4, a5);
459 rcode = syscall(AFS_SYSCALL, 28 /* AFSCALL_CALL */ , a3, a4, a5);
460 #endif /* AFS_SGI_ENV */
466 #endif /* AFS_AIX_ENV */
475 struct ifreq ifs[ADDRSPERSITE];
478 char buf[BUFSIZ], *cp, *cplim;
480 struct sockaddr_in *a;
491 memset(rxi_NetAddrs, 0, sizeof(rxi_NetAddrs));
492 memset(myNetFlags, 0, sizeof(myNetFlags));
493 memset(myNetMTUs, 0, sizeof(myNetMTUs));
494 memset(myNetMasks, 0, sizeof(myNetMasks));
496 s = socket(AF_INET, SOCK_DGRAM, 0);
501 ifc.ifc_len = sizeof(buf);
505 ifc.ifc_len = sizeof(ifs);
506 ifc.ifc_buf = (caddr_t) & ifs[0];
507 memset(&ifs[0], 0, sizeof(ifs));
509 res = ioctl(s, SIOCGIFCONF, &ifc);
511 /* fputs(stderr, "ioctl error IFCONF\n"); */
518 #define size(p) MAX((p).sa_len, sizeof(p))
519 cplim = buf + ifc.ifc_len; /*skip over if's with big ifr_addr's */
520 for (cp = buf; cp < cplim;
521 cp += sizeof(ifr->ifr_name) + MAX(a->sin_len, sizeof(*a))) {
522 if (rxi_numNetAddrs >= ADDRSPERSITE)
525 ifr = (struct ifreq *)cp;
527 len = ifc.ifc_len / sizeof(struct ifreq);
528 if (len > ADDRSPERSITE)
531 for (i = 0; i < len; ++i) {
533 res = ioctl(s, SIOCGIFADDR, ifr);
536 /* fputs(stderr, "ioctl error IFADDR\n");
537 * perror(ifr->ifr_name); */
540 a = (struct sockaddr_in *)&ifr->ifr_addr;
541 if (a->sin_family != AF_INET)
543 rxi_NetAddrs[rxi_numNetAddrs] = ntohl(a->sin_addr.s_addr);
544 if (rxi_NetAddrs[rxi_numNetAddrs] == 0x7f000001) {
545 /* we don't really care about "localhost" */
548 for (j = 0; j < rxi_numNetAddrs; j++) {
549 if (rxi_NetAddrs[j] == rxi_NetAddrs[rxi_numNetAddrs])
552 if (j < rxi_numNetAddrs)
555 /* fprintf(stderr, "if %s addr=%x\n", ifr->ifr_name,
556 * rxi_NetAddrs[rxi_numNetAddrs]); */
559 res = ioctl(s, SIOCGIFFLAGS, ifr);
561 myNetFlags[rxi_numNetAddrs] = ifr->ifr_flags;
563 /* Handle aliased loopbacks as well. */
564 if (ifr->ifr_flags & IFF_LOOPBACK)
567 /* fprintf(stderr, "if %s flags=%x\n",
568 * ifr->ifr_name, ifr->ifr_flags); */
570 * fputs(stderr, "ioctl error IFFLAGS\n");
571 * perror(ifr->ifr_name); */
573 #endif /* SIOCGIFFLAGS */
575 #if !defined(AFS_AIX_ENV) && !defined(AFS_LINUX20_ENV)
576 /* this won't run on an AIX system w/o a cache manager */
577 rxi_syscallp = rxi_syscall;
580 /* If I refer to kernel extensions that aren't loaded on AIX, the
581 * program refuses to load and run, so I simply can't include the
582 * following code. Fortunately, AIX is the one operating system in
583 * which the subsequent ioctl works reliably. */
585 if ((*rxi_syscallp) (20 /*AFSOP_GETMTU */ ,
586 htonl(rxi_NetAddrs[rxi_numNetAddrs]),
587 &(myNetMTUs[rxi_numNetAddrs]))) {
588 /* fputs(stderr, "syscall error GETMTU\n");
589 * perror(ifr->ifr_name); */
590 myNetMTUs[rxi_numNetAddrs] = 0;
592 if ((*rxi_syscallp) (42 /*AFSOP_GETMASK */ ,
593 htonl(rxi_NetAddrs[rxi_numNetAddrs]),
594 &(myNetMasks[rxi_numNetAddrs]))) {
595 /* fputs(stderr, "syscall error GETMASK\n");
596 * perror(ifr->ifr_name); */
597 myNetMasks[rxi_numNetAddrs] = 0;
599 myNetMasks[rxi_numNetAddrs] =
600 ntohl(myNetMasks[rxi_numNetAddrs]);
601 /* fprintf(stderr, "if %s mask=0x%x\n",
602 * ifr->ifr_name, myNetMasks[rxi_numNetAddrs]); */
605 if (myNetMTUs[rxi_numNetAddrs] == 0) {
606 myNetMTUs[rxi_numNetAddrs] = OLD_MAX_PACKET_SIZE + RX_IPUDP_SIZE;
608 res = ioctl(s, SIOCGIFMTU, ifr);
609 if ((res == 0) && (ifr->ifr_metric > 128)) { /* sanity check */
610 myNetMTUs[rxi_numNetAddrs] = ifr->ifr_metric;
611 /* fprintf(stderr, "if %s mtu=%d\n",
612 * ifr->ifr_name, ifr->ifr_metric); */
614 /* fputs(stderr, "ioctl error IFMTU\n");
615 * perror(ifr->ifr_name); */
620 if (myNetMasks[rxi_numNetAddrs] == 0) {
621 myNetMasks[rxi_numNetAddrs] =
622 fudge_netmask(rxi_NetAddrs[rxi_numNetAddrs]);
623 #ifdef SIOCGIFNETMASK
624 res = ioctl(s, SIOCGIFNETMASK, ifr);
626 a = (struct sockaddr_in *)&ifr->ifr_addr;
627 myNetMasks[rxi_numNetAddrs] = ntohl(a->sin_addr.s_addr);
628 /* fprintf(stderr, "if %s subnetmask=0x%x\n",
629 * ifr->ifr_name, myNetMasks[rxi_numNetAddrs]); */
631 /* fputs(stderr, "ioctl error IFMASK\n");
632 * perror(ifr->ifr_name); */
637 if (rxi_NetAddrs[rxi_numNetAddrs] != 0x7f000001) { /* ignore lo0 */
640 rxi_nRecvFrags * (myNetMTUs[rxi_numNetAddrs] - RX_IP_SIZE);
641 maxsize -= UDP_HDR_SIZE; /* only the first frag has a UDP hdr */
642 if (rx_maxReceiveSize < maxsize)
643 rx_maxReceiveSize = MIN(RX_MAX_PACKET_SIZE, maxsize);
650 /* have to allocate at least enough to allow a single packet to reach its
651 * maximum size, so ReadPacket will work. Allocate enough for a couple
652 * of packets to do so, for good measure */
654 int npackets, ncbufs;
656 rx_maxJumboRecvSize =
657 RX_HEADER_SIZE + rxi_nDgramPackets * RX_JUMBOBUFFERSIZE +
658 (rxi_nDgramPackets - 1) * RX_JUMBOHEADERSIZE;
659 rx_maxJumboRecvSize = MAX(rx_maxJumboRecvSize, rx_maxReceiveSize);
660 ncbufs = (rx_maxJumboRecvSize - RX_FIRSTBUFFERSIZE);
662 ncbufs = ncbufs / RX_CBUFFERSIZE;
663 npackets = rx_initSendWindow - 1;
664 rxi_MorePackets(npackets * (ncbufs + 1));
668 #endif /* AFS_NT40_ENV */
670 /* Called from rxi_FindPeer, when initializing a clear rx_peer structure,
671 * to get interesting information.
672 * Curiously enough, the rx_peerHashTable_lock currently protects the
673 * Inited variable (and hence rx_GetIFInfo). When the fs suite uses
674 * pthreads, this issue will need to be revisited.
678 rxi_InitPeerParams(struct rx_peer *pp)
683 #if defined(ADAPT_PMTU) && defined(IP_MTU)
685 struct sockaddr_in addr;
692 * there's a race here since more than one thread could call
693 * rx_GetIFInfo. The race stops in rx_GetIFInfo.
701 /* try to second-guess IP, and identify which link is most likely to
702 * be used for traffic to/from this host. */
703 ppaddr = ntohl(pp->host);
707 pp->rateFlag = 2; /* start timing after two full packets */
708 /* I don't initialize these, because I presume they are bzero'd...
709 * pp->burstSize pp->burst pp->burstWait.sec pp->burstWait.usec
710 * pp->timeout.usec */
713 for (ix = 0; ix < rxi_numNetAddrs; ++ix) {
714 if ((rxi_NetAddrs[ix] & myNetMasks[ix]) == (ppaddr & myNetMasks[ix])) {
715 #ifdef IFF_POINTOPOINT
716 if (myNetFlags[ix] & IFF_POINTOPOINT)
718 #endif /* IFF_POINTOPOINT */
719 rxmtu = myNetMTUs[ix] - RX_IPUDP_SIZE;
720 if (rxmtu < RX_MIN_PACKET_SIZE)
721 rxmtu = RX_MIN_PACKET_SIZE;
722 if (pp->ifMTU < rxmtu)
723 pp->ifMTU = MIN(rx_MyMaxSendSize, rxmtu);
727 if (!pp->ifMTU) { /* not local */
729 pp->ifMTU = MIN(rx_MyMaxSendSize, RX_REMOTE_PACKET_SIZE);
731 #else /* ADAPT_MTU */
732 pp->rateFlag = 2; /* start timing after two full packets */
734 pp->ifMTU = MIN(rx_MyMaxSendSize, OLD_MAX_PACKET_SIZE);
735 #endif /* ADAPT_MTU */
736 #if defined(ADAPT_PMTU) && defined(IP_MTU)
737 sock=socket(PF_INET, SOCK_DGRAM, IPPROTO_UDP);
739 addr.sin_family = AF_INET;
740 addr.sin_addr.s_addr = pp->host;
741 addr.sin_port = pp->port;
742 if (connect(sock, (struct sockaddr *)&addr, sizeof(addr)) == 0) {
744 socklen_t s = sizeof(mtu);
745 if (getsockopt(sock, SOL_IP, IP_MTU, &mtu, &s)== 0) {
746 pp->ifMTU = MIN(mtu - RX_IPUDP_SIZE, pp->ifMTU);
752 pp->ifMTU = rxi_AdjustIfMTU(pp->ifMTU);
753 pp->maxMTU = OLD_MAX_PACKET_SIZE; /* for compatibility with old guys */
754 pp->natMTU = MIN((int)pp->ifMTU, OLD_MAX_PACKET_SIZE);
755 pp->maxDgramPackets =
756 MIN(rxi_nDgramPackets,
757 rxi_AdjustDgramPackets(rxi_nSendFrags, pp->ifMTU));
759 MIN(rxi_nDgramPackets,
760 rxi_AdjustDgramPackets(rxi_nSendFrags, pp->ifMTU));
761 pp->maxDgramPackets = 1;
762 /* Initialize slow start parameters */
763 pp->MTU = MIN(pp->natMTU, pp->maxMTU);
765 pp->nDgramPackets = 1;
769 /* Don't expose jumobgram internals. */
773 rx_maxReceiveSize = OLD_MAX_PACKET_SIZE;
774 rxi_nSendFrags = rxi_nRecvFrags = 1;
777 /* Override max MTU. If rx_SetNoJumbo is called, it must be
778 called before calling rx_SetMaxMTU since SetNoJumbo clobbers rx_maxReceiveSize */
780 rx_SetMaxMTU(int mtu)
782 rx_MyMaxSendSize = rx_maxReceiveSizeUser = rx_maxReceiveSize = mtu;
785 #if defined(HAVE_LINUX_ERRQUEUE_H) && defined(ADAPT_PMTU)
787 rxi_HandleSocketError(int socket)
790 struct cmsghdr *cmsg;
791 struct sock_extended_err *err;
792 struct sockaddr_in addr;
793 struct sockaddr *offender;
794 char controlmsgbuf[256];
798 msg.msg_name = &addr;
799 msg.msg_namelen = sizeof(addr);
802 msg.msg_control = controlmsgbuf;
803 msg.msg_controllen = 256;
805 code = recvmsg(socket, &msg, MSG_ERRQUEUE|MSG_DONTWAIT|MSG_TRUNC);
807 if (code < 0 || !(msg.msg_flags & MSG_ERRQUEUE))
810 for (cmsg = CMSG_FIRSTHDR(&msg); cmsg; cmsg = CMSG_NXTHDR(&msg, cmsg)) {
811 if ((char *)cmsg - controlmsgbuf > msg.msg_controllen - CMSG_SPACE(0) ||
812 (char *)cmsg - controlmsgbuf > msg.msg_controllen - CMSG_SPACE(cmsg->cmsg_len) ||
813 cmsg->cmsg_len == 0) {
817 if (cmsg->cmsg_level == SOL_IP && cmsg->cmsg_type == IP_RECVERR)
823 err =(struct sock_extended_err *) CMSG_DATA(cmsg);
825 if (err->ee_errno == EMSGSIZE && err->ee_info >= 68) {
826 rxi_SetPeerMtu(addr.sin_addr.s_addr, addr.sin_port,
827 err->ee_info - RX_IPUDP_SIZE);
829 /* other DEST_UNREACH's and TIME_EXCEEDED should be dealt with too */