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 /* rxi_syscall is currently not prototyped */
14 #include <afsconfig.h>
15 #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>
36 #if !defined(AFS_AIX_ENV) && !defined(AFS_NT40_ENV)
37 # include <sys/syscall.h>
39 #include <afs/afs_args.h>
40 #include <afs/afsutil.h>
42 #ifndef IPPORT_USERRESERVED
43 /* If in.h doesn't define this, define it anyway. Unfortunately, defining
44 this doesn't put the code into the kernel to restrict kernel assigned
45 port numbers to numbers below IPPORT_USERRESERVED... */
46 #define IPPORT_USERRESERVED 5000
49 #if defined(HAVE_LINUX_ERRQUEUE_H) && defined(ADAPT_PMTU)
50 #include <linux/types.h>
51 #include <linux/errqueue.h>
58 # include <sys/time.h>
61 # include "rx_globals.h"
63 #ifdef AFS_PTHREAD_ENV
67 * The rx_if_init_mutex mutex protects the following global variables:
71 afs_kmutex_t rx_if_init_mutex;
72 #define LOCK_IF_INIT MUTEX_ENTER(&rx_if_init_mutex)
73 #define UNLOCK_IF_INIT MUTEX_EXIT(&rx_if_init_mutex)
76 * The rx_if_mutex mutex protects the following global variables:
82 afs_kmutex_t rx_if_mutex;
83 #define LOCK_IF MUTEX_ENTER(&rx_if_mutex)
84 #define UNLOCK_IF MUTEX_EXIT(&rx_if_mutex)
87 #define UNLOCK_IF_INIT
90 #endif /* AFS_PTHREAD_ENV */
94 * Make a socket for receiving/sending IP packets. Set it into non-blocking
95 * and large buffering modes. If port isn't specified, the kernel will pick
96 * one. Returns the socket (>= 0) on success. Returns OSI_NULLSOCKET on
97 * failure. Port must be in network byte order.
100 rxi_GetHostUDPSocket(u_int ahost, u_short port)
103 osi_socket socketFd = OSI_NULLSOCKET;
104 struct sockaddr_in taddr;
105 char *name = "rxi_GetUDPSocket: ";
106 #ifdef AFS_LINUX22_ENV
107 #if defined(ADAPT_PMTU)
108 int pmtu=IP_PMTUDISC_WANT;
111 int pmtu=IP_PMTUDISC_DONT;
115 #if !defined(AFS_NT40_ENV)
116 if (ntohs(port) >= IPPORT_RESERVED && ntohs(port) < IPPORT_USERRESERVED) {
117 /* (osi_Msg "%s*WARNING* port number %d is not a reserved port number. Use port numbers above %d\n", name, port, IPPORT_USERRESERVED);
120 if (ntohs(port) > 0 && ntohs(port) < IPPORT_RESERVED && geteuid() != 0) {
122 "%sport number %d is a reserved port number which may only be used by root. Use port numbers above %d\n",
123 name, ntohs(port), IPPORT_USERRESERVED);
127 socketFd = socket(AF_INET, SOCK_DGRAM, 0);
135 rxi_xmit_init(socketFd);
136 #endif /* AFS_NT40_ENV */
138 taddr.sin_addr.s_addr = ahost;
139 taddr.sin_family = AF_INET;
140 taddr.sin_port = (u_short) port;
141 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
142 taddr.sin_len = sizeof(struct sockaddr_in);
144 #define MAX_RX_BINDS 10
145 for (binds = 0; binds < MAX_RX_BINDS; binds++) {
148 code = bind(socketFd, (struct sockaddr *)&taddr, sizeof(taddr));
154 (osi_Msg "%sbind failed\n", name);
157 #if !defined(AFS_NT40_ENV)
159 * Set close-on-exec on rx socket
161 fcntl(socketFd, F_SETFD, 1);
164 /* Use one of three different ways of getting a socket buffer expanded to
172 len2 = rx_UdpBufSize;
174 /* find the size closest to rx_UdpBufSize that will be accepted */
175 while (!greedy && len2 > len1) {
178 (socketFd, SOL_SOCKET, SO_RCVBUF, (char *)&len2,
184 /* but do not let it get smaller than 32K */
194 (socketFd, SOL_SOCKET, SO_SNDBUF, (char *)&len1,
198 (socketFd, SOL_SOCKET, SO_RCVBUF, (char *)&len2,
201 (osi_Msg "%s*WARNING* Unable to increase buffering on socket\n",
203 if (rx_stats_active) {
204 MUTEX_ENTER(&rx_stats_mutex);
205 rx_stats.socketGreedy = greedy;
206 MUTEX_EXIT(&rx_stats_mutex);
210 #ifdef AFS_LINUX22_ENV
211 setsockopt(socketFd, SOL_IP, IP_MTU_DISCOVER, &pmtu, sizeof(pmtu));
212 #if defined(ADAPT_PMTU)
213 setsockopt(socketFd, SOL_IP, IP_RECVERR, &recverr, sizeof(recverr));
216 if (rxi_Listen(socketFd) < 0) {
225 closesocket(socketFd);
231 return OSI_NULLSOCKET;
235 rxi_GetUDPSocket(u_short port)
237 return rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
241 osi_Panic(char *msg, ...)
245 (osi_Msg "Fatal Rx error: ");
254 * osi_AssertFailU() -- used by the osi_Assert() macro.
258 osi_AssertFailU(const char *expr, const char *file, int line)
260 osi_Panic("assertion failed: %s, file: %s, line: %d\n", expr,
264 #if defined(AFS_AIX32_ENV) && !defined(KERNEL)
266 static const char memZero;
268 osi_Alloc(afs_int32 x)
271 * 0-length allocs may return NULL ptr from malloc, so we special-case
272 * things so that NULL returned iff an error occurred
275 return (void *)&memZero;
280 osi_Free(void *x, afs_int32 size)
287 #endif /* defined(AFS_AIX32_ENV) && !defined(KERNEL) */
289 #define ADDRSPERSITE 16
292 static afs_uint32 rxi_NetAddrs[ADDRSPERSITE]; /* host order */
293 static int myNetMTUs[ADDRSPERSITE];
294 static int myNetMasks[ADDRSPERSITE];
295 static int myNetFlags[ADDRSPERSITE];
296 static u_int rxi_numNetAddrs;
297 static int Inited = 0;
299 #if defined(AFS_NT40_ENV)
303 /* The IP address list can change so we must query for it */
306 /* we don't want to use the loopback adapter which is first */
307 /* this is a bad bad hack */
308 if (rxi_numNetAddrs > 1)
309 return htonl(rxi_NetAddrs[1]);
310 else if (rxi_numNetAddrs > 0)
311 return htonl(rxi_NetAddrs[0]);
317 ** return number of addresses
318 ** and the addresses themselves in the buffer
319 ** maxSize - max number of interfaces to return.
322 rx_getAllAddr(afs_uint32 * buffer, int maxSize)
324 int count = 0, offset = 0;
326 /* The IP address list can change so we must query for it */
330 /* we don't want to use the loopback adapter which is first */
331 /* this is a bad bad hack.
332 * and doesn't hold true on Windows.
334 if ( rxi_numNetAddrs > 1 )
336 #endif /* AFS_DJGPP_ENV */
338 for (count = 0; offset < rxi_numNetAddrs && maxSize > 0;
339 count++, offset++, maxSize--)
340 buffer[count] = htonl(rxi_NetAddrs[offset]);
345 /* this function returns the total number of interface addresses
346 * the buffer has to be passed in by the caller. It also returns
347 * the matching interface mask and mtu. All values are returned
348 * in network byte order.
351 rx_getAllAddrMaskMtu(afs_uint32 addrBuffer[], afs_uint32 maskBuffer[],
352 afs_uint32 mtuBuffer[], int maxSize)
354 int count = 0, offset = 0;
356 /* The IP address list can change so we must query for it */
360 /* we don't want to use the loopback adapter which is first */
361 /* this is a bad bad hack.
362 * and doesn't hold true on Windows.
364 if ( rxi_numNetAddrs > 1 )
366 #endif /* AFS_DJGPP_ENV */
369 offset < rxi_numNetAddrs && maxSize > 0;
370 count++, offset++, maxSize--) {
371 addrBuffer[count] = htonl(rxi_NetAddrs[offset]);
372 maskBuffer[count] = htonl(myNetMasks[offset]);
373 mtuBuffer[count] = htonl(myNetMTUs[offset]);
380 extern int rxinit_status;
382 rxi_InitMorePackets(void) {
383 int npackets, ncbufs;
385 ncbufs = (rx_maxJumboRecvSize - RX_FIRSTBUFFERSIZE);
387 ncbufs = ncbufs / RX_CBUFFERSIZE;
388 npackets = rx_initSendWindow - 1;
389 rxi_MorePackets(npackets * (ncbufs + 1));
401 if (Inited < 2 && rxinit_status == 0) {
402 /* We couldn't initialize more packets earlier.
404 rxi_InitMorePackets();
414 rxi_numNetAddrs = ADDRSPERSITE;
415 (void)syscfg_GetIFInfo(&rxi_numNetAddrs, rxi_NetAddrs,
416 myNetMasks, myNetMTUs, myNetFlags);
418 for (i = 0; i < rxi_numNetAddrs; i++) {
419 rxsize = rxi_AdjustIfMTU(myNetMTUs[i] - RX_IPUDP_SIZE);
421 rxi_nRecvFrags * rxsize + (rxi_nRecvFrags - 1) * UDP_HDR_SIZE;
422 maxsize = rxi_AdjustMaxMTU(rxsize, maxsize);
423 if (rx_maxReceiveSize < maxsize) {
424 rx_maxReceiveSize = MIN(RX_MAX_PACKET_SIZE, maxsize);
426 MIN(rx_maxReceiveSize, rx_maxReceiveSizeUser);
433 * If rxinit_status is still set, rx_InitHost() has yet to be called
434 * and we therefore do not have any mutex locks initialized. As a
435 * result we cannot call rxi_MorePackets() without crashing.
440 rxi_InitMorePackets();
445 fudge_netmask(afs_uint32 addr)
451 else if (IN_CLASSB(addr))
453 else if (IN_CLASSC(addr))
463 #if !defined(AFS_AIX_ENV) && !defined(AFS_NT40_ENV) && !defined(AFS_LINUX20_ENV)
465 rxi_syscall(afs_uint32 a3, afs_uint32 a4, void *a5)
470 old = signal(SIGSYS, SIG_IGN);
472 #if defined(AFS_SGI_ENV)
473 rcode = afs_syscall(a3, a4, a5);
475 rcode = syscall(AFS_SYSCALL, 28 /* AFSCALL_CALL */ , a3, a4, a5);
476 #endif /* AFS_SGI_ENV */
482 #endif /* AFS_AIX_ENV */
491 struct ifreq ifs[ADDRSPERSITE];
494 char buf[BUFSIZ], *cp, *cplim;
496 struct sockaddr_in *a;
507 memset(rxi_NetAddrs, 0, sizeof(rxi_NetAddrs));
508 memset(myNetFlags, 0, sizeof(myNetFlags));
509 memset(myNetMTUs, 0, sizeof(myNetMTUs));
510 memset(myNetMasks, 0, sizeof(myNetMasks));
512 s = socket(AF_INET, SOCK_DGRAM, 0);
517 ifc.ifc_len = sizeof(buf);
521 ifc.ifc_len = sizeof(ifs);
522 ifc.ifc_buf = (caddr_t) & ifs[0];
523 memset(&ifs[0], 0, sizeof(ifs));
525 res = ioctl(s, SIOCGIFCONF, &ifc);
527 /* fputs(stderr, "ioctl error IFCONF\n"); */
534 #define size(p) MAX((p).sa_len, sizeof(p))
535 cplim = buf + ifc.ifc_len; /*skip over if's with big ifr_addr's */
536 for (cp = buf; cp < cplim;
537 cp += sizeof(ifr->ifr_name) + MAX(a->sin_len, sizeof(*a))) {
538 if (rxi_numNetAddrs >= ADDRSPERSITE)
541 ifr = (struct ifreq *)cp;
543 len = ifc.ifc_len / sizeof(struct ifreq);
544 if (len > ADDRSPERSITE)
547 for (i = 0; i < len; ++i) {
549 res = ioctl(s, SIOCGIFADDR, ifr);
552 /* fputs(stderr, "ioctl error IFADDR\n");
553 * perror(ifr->ifr_name); */
556 a = (struct sockaddr_in *)&ifr->ifr_addr;
557 if (a->sin_family != AF_INET)
559 rxi_NetAddrs[rxi_numNetAddrs] = ntohl(a->sin_addr.s_addr);
560 if (rxi_NetAddrs[rxi_numNetAddrs] == 0x7f000001) {
561 /* we don't really care about "localhost" */
564 for (j = 0; j < rxi_numNetAddrs; j++) {
565 if (rxi_NetAddrs[j] == rxi_NetAddrs[rxi_numNetAddrs])
568 if (j < rxi_numNetAddrs)
571 /* fprintf(stderr, "if %s addr=%x\n", ifr->ifr_name,
572 * rxi_NetAddrs[rxi_numNetAddrs]); */
575 res = ioctl(s, SIOCGIFFLAGS, ifr);
577 myNetFlags[rxi_numNetAddrs] = ifr->ifr_flags;
579 /* Handle aliased loopbacks as well. */
580 if (ifr->ifr_flags & IFF_LOOPBACK)
583 /* fprintf(stderr, "if %s flags=%x\n",
584 * ifr->ifr_name, ifr->ifr_flags); */
586 * fputs(stderr, "ioctl error IFFLAGS\n");
587 * perror(ifr->ifr_name); */
589 #endif /* SIOCGIFFLAGS */
591 #if !defined(AFS_AIX_ENV) && !defined(AFS_LINUX20_ENV)
592 /* this won't run on an AIX system w/o a cache manager */
593 rxi_syscallp = rxi_syscall;
596 /* If I refer to kernel extensions that aren't loaded on AIX, the
597 * program refuses to load and run, so I simply can't include the
598 * following code. Fortunately, AIX is the one operating system in
599 * which the subsequent ioctl works reliably. */
601 if ((*rxi_syscallp) (20 /*AFSOP_GETMTU */ ,
602 htonl(rxi_NetAddrs[rxi_numNetAddrs]),
603 &(myNetMTUs[rxi_numNetAddrs]))) {
604 /* fputs(stderr, "syscall error GETMTU\n");
605 * perror(ifr->ifr_name); */
606 myNetMTUs[rxi_numNetAddrs] = 0;
608 if ((*rxi_syscallp) (42 /*AFSOP_GETMASK */ ,
609 htonl(rxi_NetAddrs[rxi_numNetAddrs]),
610 &(myNetMasks[rxi_numNetAddrs]))) {
611 /* fputs(stderr, "syscall error GETMASK\n");
612 * perror(ifr->ifr_name); */
613 myNetMasks[rxi_numNetAddrs] = 0;
615 myNetMasks[rxi_numNetAddrs] =
616 ntohl(myNetMasks[rxi_numNetAddrs]);
617 /* fprintf(stderr, "if %s mask=0x%x\n",
618 * ifr->ifr_name, myNetMasks[rxi_numNetAddrs]); */
621 if (myNetMTUs[rxi_numNetAddrs] == 0) {
622 myNetMTUs[rxi_numNetAddrs] = OLD_MAX_PACKET_SIZE + RX_IPUDP_SIZE;
624 res = ioctl(s, SIOCGIFMTU, ifr);
625 if ((res == 0) && (ifr->ifr_metric > 128)) { /* sanity check */
626 myNetMTUs[rxi_numNetAddrs] = ifr->ifr_metric;
627 /* fprintf(stderr, "if %s mtu=%d\n",
628 * ifr->ifr_name, ifr->ifr_metric); */
630 /* fputs(stderr, "ioctl error IFMTU\n");
631 * perror(ifr->ifr_name); */
636 if (myNetMasks[rxi_numNetAddrs] == 0) {
637 myNetMasks[rxi_numNetAddrs] =
638 fudge_netmask(rxi_NetAddrs[rxi_numNetAddrs]);
639 #ifdef SIOCGIFNETMASK
640 res = ioctl(s, SIOCGIFNETMASK, ifr);
642 a = (struct sockaddr_in *)&ifr->ifr_addr;
643 myNetMasks[rxi_numNetAddrs] = ntohl(a->sin_addr.s_addr);
644 /* fprintf(stderr, "if %s subnetmask=0x%x\n",
645 * ifr->ifr_name, myNetMasks[rxi_numNetAddrs]); */
647 /* fputs(stderr, "ioctl error IFMASK\n");
648 * perror(ifr->ifr_name); */
653 if (rxi_NetAddrs[rxi_numNetAddrs] != 0x7f000001) { /* ignore lo0 */
656 rxi_nRecvFrags * (myNetMTUs[rxi_numNetAddrs] - RX_IP_SIZE);
657 maxsize -= UDP_HDR_SIZE; /* only the first frag has a UDP hdr */
658 if (rx_maxReceiveSize < maxsize)
659 rx_maxReceiveSize = MIN(RX_MAX_PACKET_SIZE, maxsize);
666 /* have to allocate at least enough to allow a single packet to reach its
667 * maximum size, so ReadPacket will work. Allocate enough for a couple
668 * of packets to do so, for good measure */
670 int npackets, ncbufs;
672 rx_maxJumboRecvSize =
673 RX_HEADER_SIZE + rxi_nDgramPackets * RX_JUMBOBUFFERSIZE +
674 (rxi_nDgramPackets - 1) * RX_JUMBOHEADERSIZE;
675 rx_maxJumboRecvSize = MAX(rx_maxJumboRecvSize, rx_maxReceiveSize);
676 ncbufs = (rx_maxJumboRecvSize - RX_FIRSTBUFFERSIZE);
678 ncbufs = ncbufs / RX_CBUFFERSIZE;
679 npackets = rx_initSendWindow - 1;
680 rxi_MorePackets(npackets * (ncbufs + 1));
684 #endif /* AFS_NT40_ENV */
686 /* Called from rxi_FindPeer, when initializing a clear rx_peer structure,
687 * to get interesting information.
688 * Curiously enough, the rx_peerHashTable_lock currently protects the
689 * Inited variable (and hence rx_GetIFInfo). When the fs suite uses
690 * pthreads, this issue will need to be revisited.
694 rxi_InitPeerParams(struct rx_peer *pp)
699 #if defined(ADAPT_PMTU) && defined(IP_MTU)
701 struct sockaddr_in addr;
710 * there's a race here since more than one thread could call
711 * rx_GetIFInfo. The race stops in rx_GetIFInfo.
719 /* try to second-guess IP, and identify which link is most likely to
720 * be used for traffic to/from this host. */
721 ppaddr = ntohl(pp->host);
725 pp->rateFlag = 2; /* start timing after two full packets */
726 /* I don't initialize these, because I presume they are bzero'd...
727 * pp->burstSize pp->burst pp->burstWait.sec pp->burstWait.usec
728 * pp->timeout.usec */
731 for (ix = 0; ix < rxi_numNetAddrs; ++ix) {
732 if ((rxi_NetAddrs[ix] & myNetMasks[ix]) == (ppaddr & myNetMasks[ix])) {
733 #ifdef IFF_POINTOPOINT
734 if (myNetFlags[ix] & IFF_POINTOPOINT)
736 #endif /* IFF_POINTOPOINT */
737 rxmtu = myNetMTUs[ix] - RX_IPUDP_SIZE;
738 if (rxmtu < RX_MIN_PACKET_SIZE)
739 rxmtu = RX_MIN_PACKET_SIZE;
740 if (pp->ifMTU < rxmtu)
741 pp->ifMTU = MIN(rx_MyMaxSendSize, rxmtu);
745 if (!pp->ifMTU) { /* not local */
747 pp->ifMTU = MIN(rx_MyMaxSendSize, RX_REMOTE_PACKET_SIZE);
749 #else /* ADAPT_MTU */
750 pp->rateFlag = 2; /* start timing after two full packets */
752 pp->ifMTU = MIN(rx_MyMaxSendSize, OLD_MAX_PACKET_SIZE);
753 #endif /* ADAPT_MTU */
754 #if defined(ADAPT_PMTU) && defined(IP_MTU)
755 sock=socket(PF_INET, SOCK_DGRAM, IPPROTO_UDP);
757 addr.sin_family = AF_INET;
758 addr.sin_addr.s_addr = pp->host;
759 addr.sin_port = pp->port;
760 if (connect(sock, (struct sockaddr *)&addr, sizeof(addr)) == 0) {
762 socklen_t s = sizeof(mtu);
763 if (getsockopt(sock, SOL_IP, IP_MTU, &mtu, &s)== 0) {
764 pp->ifMTU = MIN(mtu - RX_IPUDP_SIZE, pp->ifMTU);
770 pp->ifMTU = rxi_AdjustIfMTU(pp->ifMTU);
771 pp->maxMTU = OLD_MAX_PACKET_SIZE; /* for compatibility with old guys */
772 pp->natMTU = MIN((int)pp->ifMTU, OLD_MAX_PACKET_SIZE);
773 pp->maxDgramPackets =
774 MIN(rxi_nDgramPackets,
775 rxi_AdjustDgramPackets(rxi_nSendFrags, pp->ifMTU));
777 MIN(rxi_nDgramPackets,
778 rxi_AdjustDgramPackets(rxi_nSendFrags, pp->ifMTU));
779 pp->maxDgramPackets = 1;
780 /* Initialize slow start parameters */
781 pp->MTU = MIN(pp->natMTU, pp->maxMTU);
783 pp->nDgramPackets = 1;
787 /* Don't expose jumobgram internals. */
791 rx_maxReceiveSize = OLD_MAX_PACKET_SIZE;
792 rxi_nSendFrags = rxi_nRecvFrags = 1;
795 /* Override max MTU. If rx_SetNoJumbo is called, it must be
796 called before calling rx_SetMaxMTU since SetNoJumbo clobbers rx_maxReceiveSize */
798 rx_SetMaxMTU(int mtu)
800 rx_MyMaxSendSize = rx_maxReceiveSizeUser = rx_maxReceiveSize = mtu;
803 #if defined(HAVE_LINUX_ERRQUEUE_H) && defined(ADAPT_PMTU)
805 rxi_HandleSocketError(int socket)
808 struct cmsghdr *cmsg;
809 struct sock_extended_err *err;
810 struct sockaddr_in addr;
811 struct sockaddr *offender;
812 char controlmsgbuf[256];
816 msg.msg_name = &addr;
817 msg.msg_namelen = sizeof(addr);
820 msg.msg_control = controlmsgbuf;
821 msg.msg_controllen = 256;
823 code = recvmsg(socket, &msg, MSG_ERRQUEUE|MSG_DONTWAIT|MSG_TRUNC);
825 if (code < 0 || !(msg.msg_flags & MSG_ERRQUEUE))
828 for (cmsg = CMSG_FIRSTHDR(&msg); cmsg; cmsg = CMSG_NXTHDR(&msg, cmsg)) {
829 if ((char *)cmsg - controlmsgbuf > msg.msg_controllen - CMSG_SPACE(0) ||
830 (char *)cmsg - controlmsgbuf > msg.msg_controllen - CMSG_SPACE(cmsg->cmsg_len) ||
831 cmsg->cmsg_len == 0) {
835 if (cmsg->cmsg_level == SOL_IP && cmsg->cmsg_type == IP_RECVERR)
841 err =(struct sock_extended_err *) CMSG_DATA(cmsg);
843 if (err->ee_errno == EMSGSIZE && err->ee_info >= 68) {
844 rxi_SetPeerMtu(addr.sin_addr.s_addr, addr.sin_port,
845 err->ee_info - RX_IPUDP_SIZE);
847 /* other DEST_UNREACH's and TIME_EXCEEDED should be dealt with too */