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>
22 # include <WINNT/syscfg.h>
26 #if !defined(AFS_AIX_ENV) && !defined(AFS_NT40_ENV)
27 # include <sys/syscall.h>
29 #include <afs/afs_args.h>
30 #include <afs/afsutil.h>
32 #ifndef IPPORT_USERRESERVED
33 /* If in.h doesn't define this, define it anyway. Unfortunately, defining
34 this doesn't put the code into the kernel to restrict kernel assigned
35 port numbers to numbers below IPPORT_USERRESERVED... */
36 #define IPPORT_USERRESERVED 5000
39 #if defined(HAVE_LINUX_ERRQUEUE_H) && defined(ADAPT_PMTU)
40 #include <linux/types.h>
41 #include <linux/errqueue.h>
48 #include "rx_atomic.h"
49 #include "rx_globals.h"
51 #ifdef AFS_PTHREAD_ENV
54 * The rx_if_init_mutex mutex protects the following global variables:
58 afs_kmutex_t rx_if_init_mutex;
59 #define LOCK_IF_INIT MUTEX_ENTER(&rx_if_init_mutex)
60 #define UNLOCK_IF_INIT MUTEX_EXIT(&rx_if_init_mutex)
63 * The rx_if_mutex mutex protects the following global variables:
69 afs_kmutex_t rx_if_mutex;
70 #define LOCK_IF MUTEX_ENTER(&rx_if_mutex)
71 #define UNLOCK_IF MUTEX_EXIT(&rx_if_mutex)
74 #define UNLOCK_IF_INIT
77 #endif /* AFS_PTHREAD_ENV */
81 * Make a socket for receiving/sending IP packets. Set it into non-blocking
82 * and large buffering modes. If port isn't specified, the kernel will pick
83 * one. Returns the socket (>= 0) on success. Returns OSI_NULLSOCKET on
84 * failure. Port must be in network byte order.
87 rxi_GetHostUDPSocket(u_int ahost, u_short port)
90 osi_socket socketFd = OSI_NULLSOCKET;
91 struct sockaddr_in taddr;
92 char *name = "rxi_GetUDPSocket: ";
93 #ifdef AFS_LINUX22_ENV
94 #if defined(ADAPT_PMTU)
95 int pmtu=IP_PMTUDISC_WANT;
98 int pmtu=IP_PMTUDISC_DONT;
102 #if !defined(AFS_NT40_ENV)
103 if (ntohs(port) >= IPPORT_RESERVED && ntohs(port) < IPPORT_USERRESERVED) {
104 /* (osi_Msg "%s*WARNING* port number %d is not a reserved port number. Use port numbers above %d\n", name, port, IPPORT_USERRESERVED);
107 if (ntohs(port) > 0 && ntohs(port) < IPPORT_RESERVED && geteuid() != 0) {
109 "%sport number %d is a reserved port number which may only be used by root. Use port numbers above %d\n",
110 name, ntohs(port), IPPORT_USERRESERVED);
114 socketFd = socket(AF_INET, SOCK_DGRAM, IPPROTO_UDP);
116 if (socketFd == OSI_NULLSOCKET) {
118 fprintf(stderr, "socket() failed with error %u\n", WSAGetLastError());
126 rxi_xmit_init(socketFd);
127 #endif /* AFS_NT40_ENV */
129 taddr.sin_addr.s_addr = ahost;
130 taddr.sin_family = AF_INET;
131 taddr.sin_port = (u_short) port;
132 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
133 taddr.sin_len = sizeof(struct sockaddr_in);
135 #define MAX_RX_BINDS 10
136 for (binds = 0; binds < MAX_RX_BINDS; binds++) {
139 code = bind(socketFd, (struct sockaddr *)&taddr, sizeof(taddr));
143 (osi_Msg "%sbind failed\n", name);
146 #if !defined(AFS_NT40_ENV)
148 * Set close-on-exec on rx socket
150 fcntl(socketFd, F_SETFD, 1);
153 /* Use one of three different ways of getting a socket buffer expanded to
161 len2 = rx_UdpBufSize;
163 /* find the size closest to rx_UdpBufSize that will be accepted */
164 while (!greedy && len2 > len1) {
167 (socketFd, SOL_SOCKET, SO_RCVBUF, (char *)&len2,
173 /* but do not let it get smaller than 32K */
183 (socketFd, SOL_SOCKET, SO_SNDBUF, (char *)&len1,
187 (socketFd, SOL_SOCKET, SO_RCVBUF, (char *)&len2,
190 (osi_Msg "%s*WARNING* Unable to increase buffering on socket\n",
193 rx_atomic_set(&rx_stats.socketGreedy, greedy);
196 #ifdef AFS_LINUX22_ENV
197 setsockopt(socketFd, SOL_IP, IP_MTU_DISCOVER, &pmtu, sizeof(pmtu));
198 #if defined(ADAPT_PMTU)
199 setsockopt(socketFd, SOL_IP, IP_RECVERR, &recverr, sizeof(recverr));
202 if (rxi_Listen(socketFd) < 0) {
210 if (socketFd != OSI_NULLSOCKET)
211 closesocket(socketFd);
213 if (socketFd != OSI_NULLSOCKET)
217 return OSI_NULLSOCKET;
221 rxi_GetUDPSocket(u_short port)
223 return rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
227 osi_Panic(char *msg, ...)
231 (osi_Msg "Fatal Rx error: ");
240 * osi_AssertFailU() -- used by the osi_Assert() macro.
244 osi_AssertFailU(const char *expr, const char *file, int line)
246 osi_Panic("assertion failed: %s, file: %s, line: %d\n", expr,
250 #if defined(AFS_AIX32_ENV) && !defined(KERNEL)
252 static const char memZero;
254 osi_Alloc(afs_int32 x)
257 * 0-length allocs may return NULL ptr from malloc, so we special-case
258 * things so that NULL returned iff an error occurred
261 return (void *)&memZero;
266 osi_Free(void *x, afs_int32 size)
273 #endif /* defined(AFS_AIX32_ENV) && !defined(KERNEL) */
275 #define ADDRSPERSITE 16
278 static afs_uint32 rxi_NetAddrs[ADDRSPERSITE]; /* host order */
279 static int myNetMTUs[ADDRSPERSITE];
280 static int myNetMasks[ADDRSPERSITE];
281 static int myNetFlags[ADDRSPERSITE];
282 static u_int rxi_numNetAddrs;
283 static int Inited = 0;
285 #if defined(AFS_NT40_ENV)
289 /* The IP address list can change so we must query for it */
292 /* we don't want to use the loopback adapter which is first */
293 /* this is a bad bad hack */
294 if (rxi_numNetAddrs > 1)
295 return htonl(rxi_NetAddrs[1]);
296 else if (rxi_numNetAddrs > 0)
297 return htonl(rxi_NetAddrs[0]);
303 ** return number of addresses
304 ** and the addresses themselves in the buffer
305 ** maxSize - max number of interfaces to return.
308 rx_getAllAddr(afs_uint32 * buffer, int maxSize)
310 int count = 0, offset = 0;
312 /* The IP address list can change so we must query for it */
315 for (count = 0; offset < rxi_numNetAddrs && maxSize > 0;
316 count++, offset++, maxSize--)
317 buffer[count] = htonl(rxi_NetAddrs[offset]);
322 /* this function returns the total number of interface addresses
323 * the buffer has to be passed in by the caller. It also returns
324 * the matching interface mask and mtu. All values are returned
325 * in network byte order.
328 rx_getAllAddrMaskMtu(afs_uint32 addrBuffer[], afs_uint32 maskBuffer[],
329 afs_uint32 mtuBuffer[], int maxSize)
331 int count = 0, offset = 0;
333 /* The IP address list can change so we must query for it */
337 offset < rxi_numNetAddrs && maxSize > 0;
338 count++, offset++, maxSize--) {
339 addrBuffer[count] = htonl(rxi_NetAddrs[offset]);
340 maskBuffer[count] = htonl(myNetMasks[offset]);
341 mtuBuffer[count] = htonl(myNetMTUs[offset]);
348 extern int rxinit_status;
350 rxi_InitMorePackets(void) {
351 int npackets, ncbufs;
353 ncbufs = (rx_maxJumboRecvSize - RX_FIRSTBUFFERSIZE);
355 ncbufs = ncbufs / RX_CBUFFERSIZE;
356 npackets = rx_initSendWindow - 1;
357 rxi_MorePackets(npackets * (ncbufs + 1));
369 if (Inited < 2 && rxinit_status == 0) {
370 /* We couldn't initialize more packets earlier.
372 rxi_InitMorePackets();
382 rxi_numNetAddrs = ADDRSPERSITE;
383 (void)syscfg_GetIFInfo(&rxi_numNetAddrs, rxi_NetAddrs,
384 myNetMasks, myNetMTUs, myNetFlags);
386 for (i = 0; i < rxi_numNetAddrs; i++) {
387 rxsize = rxi_AdjustIfMTU(myNetMTUs[i] - RX_IPUDP_SIZE);
389 rxi_nRecvFrags * rxsize + (rxi_nRecvFrags - 1) * UDP_HDR_SIZE;
390 maxsize = rxi_AdjustMaxMTU(rxsize, maxsize);
391 if (rx_maxReceiveSize > maxsize) {
392 rx_maxReceiveSize = MIN(RX_MAX_PACKET_SIZE, maxsize);
394 MIN(rx_maxReceiveSize, rx_maxReceiveSizeUser);
396 if (rx_MyMaxSendSize > maxsize) {
397 rx_MyMaxSendSize = MIN(RX_MAX_PACKET_SIZE, maxsize);
403 * If rxinit_status is still set, rx_InitHost() has yet to be called
404 * and we therefore do not have any mutex locks initialized. As a
405 * result we cannot call rxi_MorePackets() without crashing.
410 rxi_InitMorePackets();
415 fudge_netmask(afs_uint32 addr)
421 else if (IN_CLASSB(addr))
423 else if (IN_CLASSC(addr))
433 #if !defined(AFS_AIX_ENV) && !defined(AFS_NT40_ENV) && !defined(AFS_LINUX20_ENV)
435 rxi_syscall(afs_uint32 a3, afs_uint32 a4, void *a5)
440 old = signal(SIGSYS, SIG_IGN);
442 #if defined(AFS_SGI_ENV)
443 rcode = afs_syscall(AFS_SYSCALL, 28, a3, a4, a5);
445 rcode = syscall(AFS_SYSCALL, 28 /* AFSCALL_CALL */ , a3, a4, a5);
446 #endif /* AFS_SGI_ENV */
452 #endif /* AFS_AIX_ENV */
461 struct ifreq ifs[ADDRSPERSITE];
464 char buf[BUFSIZ], *cp, *cplim;
466 struct sockaddr_in *a;
477 memset(rxi_NetAddrs, 0, sizeof(rxi_NetAddrs));
478 memset(myNetFlags, 0, sizeof(myNetFlags));
479 memset(myNetMTUs, 0, sizeof(myNetMTUs));
480 memset(myNetMasks, 0, sizeof(myNetMasks));
482 s = socket(AF_INET, SOCK_DGRAM, IPPROTO_UDP);
483 if (s == OSI_NULLSOCKET)
486 ifc.ifc_len = sizeof(buf);
490 ifc.ifc_len = sizeof(ifs);
491 ifc.ifc_buf = (caddr_t) & ifs[0];
492 memset(&ifs[0], 0, sizeof(ifs));
494 res = ioctl(s, SIOCGIFCONF, &ifc);
496 /* fputs(stderr, "ioctl error IFCONF\n"); */
503 #define size(p) MAX((p).sa_len, sizeof(p))
504 cplim = buf + ifc.ifc_len; /*skip over if's with big ifr_addr's */
505 for (cp = buf; cp < cplim;
506 cp += sizeof(ifr->ifr_name) + MAX(a->sin_len, sizeof(*a))) {
507 if (rxi_numNetAddrs >= ADDRSPERSITE)
510 ifr = (struct ifreq *)cp;
512 len = ifc.ifc_len / sizeof(struct ifreq);
513 if (len > ADDRSPERSITE)
516 for (i = 0; i < len; ++i) {
518 res = ioctl(s, SIOCGIFADDR, ifr);
521 /* fputs(stderr, "ioctl error IFADDR\n");
522 * perror(ifr->ifr_name); */
525 a = (struct sockaddr_in *)&ifr->ifr_addr;
526 if (a->sin_family != AF_INET)
528 rxi_NetAddrs[rxi_numNetAddrs] = ntohl(a->sin_addr.s_addr);
529 if (rx_IsLoopbackAddr(rxi_NetAddrs[rxi_numNetAddrs])) {
530 /* we don't really care about "localhost" */
533 for (j = 0; j < rxi_numNetAddrs; j++) {
534 if (rxi_NetAddrs[j] == rxi_NetAddrs[rxi_numNetAddrs])
537 if (j < rxi_numNetAddrs)
540 /* fprintf(stderr, "if %s addr=%x\n", ifr->ifr_name,
541 * rxi_NetAddrs[rxi_numNetAddrs]); */
544 res = ioctl(s, SIOCGIFFLAGS, ifr);
546 myNetFlags[rxi_numNetAddrs] = ifr->ifr_flags;
548 /* Handle aliased loopbacks as well. */
549 if (ifr->ifr_flags & IFF_LOOPBACK)
552 /* fprintf(stderr, "if %s flags=%x\n",
553 * ifr->ifr_name, ifr->ifr_flags); */
555 * fputs(stderr, "ioctl error IFFLAGS\n");
556 * perror(ifr->ifr_name); */
558 #endif /* SIOCGIFFLAGS */
560 #if !defined(AFS_AIX_ENV) && !defined(AFS_LINUX20_ENV)
561 /* this won't run on an AIX system w/o a cache manager */
562 rxi_syscallp = rxi_syscall;
565 /* If I refer to kernel extensions that aren't loaded on AIX, the
566 * program refuses to load and run, so I simply can't include the
567 * following code. Fortunately, AIX is the one operating system in
568 * which the subsequent ioctl works reliably. */
570 if ((*rxi_syscallp) (20 /*AFSOP_GETMTU */ ,
571 htonl(rxi_NetAddrs[rxi_numNetAddrs]),
572 &(myNetMTUs[rxi_numNetAddrs]))) {
573 /* fputs(stderr, "syscall error GETMTU\n");
574 * perror(ifr->ifr_name); */
575 myNetMTUs[rxi_numNetAddrs] = 0;
577 if ((*rxi_syscallp) (42 /*AFSOP_GETMASK */ ,
578 htonl(rxi_NetAddrs[rxi_numNetAddrs]),
579 &(myNetMasks[rxi_numNetAddrs]))) {
580 /* fputs(stderr, "syscall error GETMASK\n");
581 * perror(ifr->ifr_name); */
582 myNetMasks[rxi_numNetAddrs] = 0;
584 myNetMasks[rxi_numNetAddrs] =
585 ntohl(myNetMasks[rxi_numNetAddrs]);
586 /* fprintf(stderr, "if %s mask=0x%x\n",
587 * ifr->ifr_name, myNetMasks[rxi_numNetAddrs]); */
590 if (myNetMTUs[rxi_numNetAddrs] == 0) {
591 myNetMTUs[rxi_numNetAddrs] = OLD_MAX_PACKET_SIZE + RX_IPUDP_SIZE;
593 res = ioctl(s, SIOCGIFMTU, ifr);
594 if ((res == 0) && (ifr->ifr_metric > 128)) { /* sanity check */
595 myNetMTUs[rxi_numNetAddrs] = ifr->ifr_metric;
596 /* fprintf(stderr, "if %s mtu=%d\n",
597 * ifr->ifr_name, ifr->ifr_metric); */
599 /* fputs(stderr, "ioctl error IFMTU\n");
600 * perror(ifr->ifr_name); */
605 if (myNetMasks[rxi_numNetAddrs] == 0) {
606 myNetMasks[rxi_numNetAddrs] =
607 fudge_netmask(rxi_NetAddrs[rxi_numNetAddrs]);
608 #ifdef SIOCGIFNETMASK
609 res = ioctl(s, SIOCGIFNETMASK, ifr);
611 a = (struct sockaddr_in *)&ifr->ifr_addr;
612 myNetMasks[rxi_numNetAddrs] = ntohl(a->sin_addr.s_addr);
613 /* fprintf(stderr, "if %s subnetmask=0x%x\n",
614 * ifr->ifr_name, myNetMasks[rxi_numNetAddrs]); */
616 /* fputs(stderr, "ioctl error IFMASK\n");
617 * perror(ifr->ifr_name); */
622 if (!rx_IsLoopbackAddr(rxi_NetAddrs[rxi_numNetAddrs])) { /* ignore lo0 */
625 rxi_nRecvFrags * (myNetMTUs[rxi_numNetAddrs] - RX_IP_SIZE);
626 maxsize -= UDP_HDR_SIZE; /* only the first frag has a UDP hdr */
627 if (rx_maxReceiveSize < maxsize)
628 rx_maxReceiveSize = MIN(RX_MAX_PACKET_SIZE, maxsize);
635 /* have to allocate at least enough to allow a single packet to reach its
636 * maximum size, so ReadPacket will work. Allocate enough for a couple
637 * of packets to do so, for good measure */
639 int npackets, ncbufs;
641 rx_maxJumboRecvSize =
642 RX_HEADER_SIZE + rxi_nDgramPackets * RX_JUMBOBUFFERSIZE +
643 (rxi_nDgramPackets - 1) * RX_JUMBOHEADERSIZE;
644 rx_maxJumboRecvSize = MAX(rx_maxJumboRecvSize, rx_maxReceiveSize);
645 ncbufs = (rx_maxJumboRecvSize - RX_FIRSTBUFFERSIZE);
647 ncbufs = ncbufs / RX_CBUFFERSIZE;
648 npackets = rx_initSendWindow - 1;
649 rxi_MorePackets(npackets * (ncbufs + 1));
653 #endif /* AFS_NT40_ENV */
655 /* Called from rxi_FindPeer, when initializing a clear rx_peer structure,
656 * to get interesting information.
657 * Curiously enough, the rx_peerHashTable_lock currently protects the
658 * Inited variable (and hence rx_GetIFInfo). When the fs suite uses
659 * pthreads, this issue will need to be revisited.
663 rxi_InitPeerParams(struct rx_peer *pp)
668 #if defined(ADAPT_PMTU) && defined(IP_MTU)
670 struct sockaddr_in addr;
677 * there's a race here since more than one thread could call
678 * rx_GetIFInfo. The race stops in rx_GetIFInfo.
686 /* try to second-guess IP, and identify which link is most likely to
687 * be used for traffic to/from this host. */
688 ppaddr = ntohl(pp->host);
691 rx_rto_setPeerTimeoutSecs(pp, 2);
692 /* I don't initialize these, because I presume they are bzero'd...
693 * pp->burstSize pp->burst pp->burstWait.sec pp->burstWait.usec
697 for (ix = 0; ix < rxi_numNetAddrs; ++ix) {
698 if ((rxi_NetAddrs[ix] & myNetMasks[ix]) == (ppaddr & myNetMasks[ix])) {
699 #ifdef IFF_POINTOPOINT
700 if (myNetFlags[ix] & IFF_POINTOPOINT)
701 rx_rto_setPeerTimeoutSecs(pp, 4);
702 #endif /* IFF_POINTOPOINT */
704 rxmtu = myNetMTUs[ix] - RX_IPUDP_SIZE;
705 if (rxmtu < RX_MIN_PACKET_SIZE)
706 rxmtu = RX_MIN_PACKET_SIZE;
707 if (pp->ifMTU < rxmtu)
708 pp->ifMTU = MIN(rx_MyMaxSendSize, rxmtu);
712 if (!pp->ifMTU) { /* not local */
713 rx_rto_setPeerTimeoutSecs(pp, 3);
714 pp->ifMTU = MIN(rx_MyMaxSendSize, RX_REMOTE_PACKET_SIZE);
716 #else /* ADAPT_MTU */
717 rx_rto_setPeerTimeoutSecs(pp, 2);
718 pp->ifMTU = MIN(rx_MyMaxSendSize, OLD_MAX_PACKET_SIZE);
719 #endif /* ADAPT_MTU */
720 #if defined(ADAPT_PMTU) && defined(IP_MTU)
721 sock=socket(PF_INET, SOCK_DGRAM, IPPROTO_UDP);
722 if (sock != OSI_NULLSOCKET) {
723 addr.sin_family = AF_INET;
724 addr.sin_addr.s_addr = pp->host;
725 addr.sin_port = pp->port;
726 if (connect(sock, (struct sockaddr *)&addr, sizeof(addr)) == 0) {
728 socklen_t s = sizeof(mtu);
729 if (getsockopt(sock, SOL_IP, IP_MTU, &mtu, &s)== 0) {
730 pp->ifMTU = MIN(mtu - RX_IPUDP_SIZE, pp->ifMTU);
740 pp->ifMTU = rxi_AdjustIfMTU(pp->ifMTU);
741 pp->maxMTU = OLD_MAX_PACKET_SIZE; /* for compatibility with old guys */
742 pp->natMTU = MIN((int)pp->ifMTU, OLD_MAX_PACKET_SIZE);
743 pp->maxDgramPackets =
744 MIN(rxi_nDgramPackets,
745 rxi_AdjustDgramPackets(rxi_nSendFrags, pp->ifMTU));
747 MIN(rxi_nDgramPackets,
748 rxi_AdjustDgramPackets(rxi_nSendFrags, pp->ifMTU));
749 pp->maxDgramPackets = 1;
750 /* Initialize slow start parameters */
751 pp->MTU = MIN(pp->natMTU, pp->maxMTU);
753 pp->nDgramPackets = 1;
757 /* Don't expose jumobgram internals. */
761 rx_maxReceiveSize = OLD_MAX_PACKET_SIZE;
762 rxi_nSendFrags = rxi_nRecvFrags = 1;
765 /* Override max MTU. If rx_SetNoJumbo is called, it must be
766 called before calling rx_SetMaxMTU since SetNoJumbo clobbers rx_maxReceiveSize */
768 rx_SetMaxMTU(int mtu)
770 rx_MyMaxSendSize = rx_maxReceiveSizeUser = rx_maxReceiveSize = mtu;
773 #if defined(ADAPT_PMTU)
775 rxi_HandleSocketError(int socket)
778 #if defined(HAVE_LINUX_ERRQUEUE_H)
780 struct cmsghdr *cmsg;
781 struct sock_extended_err *err;
782 struct sockaddr_in addr;
783 char controlmsgbuf[256];
786 msg.msg_name = &addr;
787 msg.msg_namelen = sizeof(addr);
790 msg.msg_control = controlmsgbuf;
791 msg.msg_controllen = 256;
793 code = recvmsg(socket, &msg, MSG_ERRQUEUE|MSG_DONTWAIT|MSG_TRUNC);
795 if (code < 0 || !(msg.msg_flags & MSG_ERRQUEUE))
798 for (cmsg = CMSG_FIRSTHDR(&msg); cmsg; cmsg = CMSG_NXTHDR(&msg, cmsg)) {
799 if ((char *)cmsg - controlmsgbuf > msg.msg_controllen - CMSG_SPACE(0) ||
800 (char *)cmsg - controlmsgbuf > msg.msg_controllen - CMSG_SPACE(cmsg->cmsg_len) ||
801 cmsg->cmsg_len == 0) {
805 if (cmsg->cmsg_level == SOL_IP && cmsg->cmsg_type == IP_RECVERR)
811 err =(struct sock_extended_err *) CMSG_DATA(cmsg);
813 if (err->ee_errno == EMSGSIZE && err->ee_info >= 68) {
814 rxi_SetPeerMtu(NULL, addr.sin_addr.s_addr, addr.sin_port,
815 err->ee_info - RX_IPUDP_SIZE);
817 /* other DEST_UNREACH's and TIME_EXCEEDED should be dealt with too */