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 /* This controls the size of an fd_set; it must be defined early before
13 * the system headers define that type and the macros that operate on it.
14 * Its value should be as large as the maximum file descriptor limit we
15 * are likely to run into on any platform. Right now, that is 65536
16 * which is the default hard fd limit on Solaris 9 */
18 #define FD_SETSIZE 65536
21 #include <afsconfig.h>
22 #include <afs/param.h>
25 #include <sys/types.h> /* fd_set on older platforms */
29 # include <winsock2.h>
31 # include <unistd.h> /* select() prototype */
32 # include <sys/time.h> /* struct timeval, select() prototype */
34 # include <sys/select.h> /* fd_set on newer platforms */
36 # include <sys/socket.h>
37 # include <sys/file.h>
39 # include <sys/stat.h>
40 # include <netinet/in.h>
42 # include <sys/ioctl.h>
43 # include <sys/time.h>
47 #include "rx_atomic.h"
48 #include "rx_globals.h"
52 #define MAXTHREADNAMELENGTH 64
54 int debugSelectFailure; /* # of times select failed */
57 * Sleep on the unique wait channel provided.
62 LWP_WaitProcess(addr);
66 * Wakeup any threads on the channel provided.
67 * They may be woken up spuriously, and must check any conditions.
70 rxi_Wakeup(void *addr)
72 LWP_NoYieldSignal(addr);
75 PROCESS rx_listenerPid = 0; /* LWP process id of socket listener process */
76 static void* rx_ListenerProc(void *dummy);
79 * Delay the current thread the specified number of seconds.
87 static int quitListening = 0;
89 /* This routine will kill the listener thread, if it exists. */
91 rxi_StopListener(void)
94 rxi_ReScheduleEvents();
97 /* This routine will get called by the event package whenever a new,
98 earlier than others, event is posted. If the Listener process
99 is blocked in selects, this will unblock it. It also can be called
100 to force a new trip through the rxi_Listener select loop when the set
101 of file descriptors it should be listening to changes... */
103 rxi_ReScheduleEvents(void)
106 IOMGR_Cancel(rx_listenerPid);
110 rxi_InitializeThreadSupport(void)
114 LWP_InitializeProcessSupport(LWP_NORMAL_PRIORITY, &junk);
116 FD_ZERO(&rx_selectMask);
120 rxi_StartServerProc(void *(*proc) (void *), int stacksize)
123 static int number = 0;
126 sprintf(name, "srv_%d", ++number);
127 LWP_CreateProcess(proc, stacksize, RX_PROCESS_PRIORITY, NULL,
128 "rx_ServerProc", &scratchPid);
130 (*registerProgram) (scratchPid, name);
134 rxi_StartListener(void)
136 /* Priority of listener should be high, so it can keep conns alive */
137 #define RX_LIST_STACK 24000
138 LWP_CreateProcess(rx_ListenerProc, RX_LIST_STACK, LWP_MAX_PRIORITY,
139 NULL, "rx_Listener", &rx_listenerPid);
141 (*registerProgram) (rx_listenerPid, "listener");
144 /* The main loop which listens to the net for datagrams, and handles timeouts
145 and retransmissions, etc. It also is responsible for scheduling the
146 execution of pending events (in conjunction with event.c).
148 Note interaction of nextPollTime and lastPollWorked. The idea is that if rx is not
149 keeping up with the incoming stream of packets (because there are threads that
150 are interfering with its running sufficiently often), rx does a polling select
151 before doing a real IOMGR_Select system call. Doing a real select means that
152 we don't have to let other processes run before processing more packets.
154 So, our algorithm is that if the last poll on the file descriptor found useful data, or
155 we're at the time nextPollTime (which is advanced so that it occurs every 3 or 4 seconds),
156 then we try the polling select before the IOMGR_Select. If we eventually catch up
157 (which we can tell by the polling select returning no input packets ready), then we
158 don't do a polling select again until several seconds later (via nextPollTime mechanism).
162 rxi_ListenerProc(fd_set * rfds, int *tnop, struct rx_call **newcallp)
166 struct rx_packet *p = (struct rx_packet *)0;
169 afs_int32 nextPollTime; /* time to next poll FD before sleeping */
170 int lastPollWorked, doingPoll; /* true iff last poll was useful */
171 struct timeval tv, *tvp;
177 char name[MAXTHREADNAMELENGTH] = "srv_0";
182 code = LWP_CurrentProcess(&pid);
184 osi_Panic("rxi_Listener: Can't get my pid.\n");
186 rx_listenerPid = pid;
188 (*swapNameProgram) (pid, "listener", &name[0]);
191 /* See if a check for additional packets was issued */
194 /* Grab a new packet only if necessary (otherwise re-use the old one) */
196 rxi_RestoreDataBufs(p);
198 if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_RECEIVE)))
199 osi_Panic("rxi_ListenerProc: no packets!"); /* Shouldn't happen */
201 /* Wait for the next event time or a packet to arrive. */
202 /* event_RaiseEvents schedules any events whose time has come and
203 * then atomically computes the time to the next event, guaranteeing
204 * that this is positive. If there is no next event, it returns 0 */
206 if (!rxevent_RaiseEvents(&cv))
209 /* It's important to copy cv to tv, because the 4.3 documentation
210 * for select threatens that *tv may be updated after a select, in
211 * future editions of the system, to indicate how much of the time
212 * period has elapsed. So we shouldn't rely on tv not being altered. */
213 tv.tv_sec = cv.sec; /* Time to next event */
214 tv.tv_usec = cv.usec;
218 rx_atomic_inc(&rx_stats.selects);
220 *rfds = rx_selectMask;
222 if (lastPollWorked || nextPollTime < clock_Sec()) {
223 /* we're catching up, or haven't tried to for a few seconds */
225 nextPollTime = clock_Sec() + 4; /* try again in 4 seconds no matter what */
226 tv.tv_sec = tv.tv_usec = 0; /* make sure we poll */
228 code = select((int)(rx_maxSocketNumber + 1), rfds, 0, 0, tvp);
231 code = IOMGR_Select((int)(rx_maxSocketNumber + 1), rfds, 0, 0, tvp);
233 lastPollWorked = 0; /* default is that it didn't find anything */
237 LWP_DestroyProcess(pid);
243 * If it was a timer interrupt then we can assume that
244 * the time has advanced by roughly the value of the
245 * previous timeout, and that there is now at least
251 /* select or IOMGR_Select returned failure */
252 debugSelectFailure++; /* update debugging counter */
257 * IOMGR_Cancel is invoked whenever a new event is
258 * posted that is earlier than any existing events.
259 * So we re-evaluate the time, and then go back to
266 /* Packets have arrived, presumably:
267 * If it wasn't a timer interrupt, then no event should have
268 * timed out yet (well some event may have, but only just...), so
269 * we don't bother looking to see if any have timed out, but just
270 * go directly to reading the data packets
276 for (i = 0; p && i < rfds->fd_count; i++) {
277 socket = rfds->fd_array[i];
278 if (rxi_ReadPacket(socket, p, &host, &port)) {
280 p = rxi_ReceivePacket(p, socket, host, port, tnop,
282 if (newcallp && *newcallp) {
286 if (swapNameProgram) {
287 (*swapNameProgram) (rx_listenerPid, name, 0);
295 for (socket = rx_minSocketNumber;
296 p && socket <= rx_maxSocketNumber; socket++) {
297 if (!FD_ISSET(socket, rfds))
299 if (rxi_ReadPacket(socket, p, &host, &port)) {
300 p = rxi_ReceivePacket(p, socket, host, port, tnop,
302 if (newcallp && *newcallp) {
306 if (swapNameProgram) {
307 (*swapNameProgram) (rx_listenerPid, name, 0);
321 /* This is the listener process request loop. The listener process loop
322 * becomes a server thread when rxi_ListenerProc returns, and stays
323 * server thread until rxi_ServerProc returns. */
325 rx_ListenerProc(void *dummy)
329 struct rx_call *newcall;
332 if (!(rfds = IOMGR_AllocFDSet())) {
333 osi_Panic("rx_ListenerProc: no fd_sets!\n");
339 rxi_ListenerProc(rfds, &threadID, &newcall);
340 /* osi_Assert(threadID != -1); */
341 /* osi_Assert(newcall != NULL); */
342 sock = OSI_NULLSOCKET;
343 rxi_ServerProc(threadID, newcall, &sock);
344 /* osi_Assert(sock != OSI_NULLSOCKET); */
350 /* This is the server process request loop. The server process loop
351 * becomes a listener thread when rxi_ServerProc returns, and stays
352 * listener thread until rxi_ListenerProc returns. */
354 rx_ServerProc(void * unused)
358 struct rx_call *newcall = NULL;
361 if (!(rfds = IOMGR_AllocFDSet())) {
362 osi_Panic("rxi_ListenerProc: no fd_sets!\n");
365 rxi_MorePackets(rx_maxReceiveWindow + 2); /* alloc more packets */
366 rxi_dataQuota += rx_initSendWindow; /* Reserve some pkts for hard times */
367 /* threadID is used for making decisions in GetCall. Get it by bumping
368 * number of threads handling incoming calls */
369 threadID = rxi_availProcs++;
372 sock = OSI_NULLSOCKET;
373 rxi_ServerProc(threadID, newcall, &sock);
374 /* osi_Assert(sock != OSI_NULLSOCKET); */
376 rxi_ListenerProc(rfds, &threadID, &newcall);
377 /* osi_Assert(threadID != -1); */
378 /* osi_Assert(newcall != NULL); */
385 * Called from GetUDPSocket.
386 * Called from a single thread at startup.
387 * Returns 0 on success; -1 on failure.
390 rxi_Listen(osi_socket sock)
394 * Put the socket into non-blocking mode so that rx_Listener
395 * can do a polling read before entering select
397 if (fcntl(sock, F_SETFL, FNDELAY) == -1) {
399 (osi_Msg "rxi_Listen: unable to set non-blocking mode on socket\n");
403 if (sock > FD_SETSIZE - 1) {
404 (osi_Msg "rxi_Listen: socket descriptor > (FD_SETSIZE-1) = %d\n",
410 FD_SET(sock, &rx_selectMask);
411 if (sock > rx_maxSocketNumber)
412 rx_maxSocketNumber = sock;
413 if (sock < rx_minSocketNumber)
414 rx_minSocketNumber = sock;
422 rxi_Recvmsg(osi_socket socket, struct msghdr *msg_p, int flags)
424 #if defined(HAVE_LINUX_ERRQUEUE_H) && defined(ADAPT_PMTU)
425 while((rxi_HandleSocketError(socket)) > 0)
428 return recvmsg(socket, msg_p, flags);
432 * Simulate a blocking sendmsg on the non-blocking socket.
433 * It's non blocking because it was set that way for recvmsg.
436 rxi_Sendmsg(osi_socket socket, struct msghdr *msg_p, int flags)
438 fd_set *sfds = (fd_set *) 0;
439 while (sendmsg(socket, msg_p, flags) == -1) {
442 rx_atomic_inc(&rx_stats.sendSelects);
445 if (!(sfds = IOMGR_AllocFDSet())) {
446 (osi_Msg "rx failed to alloc fd_set: ");
447 perror("rx_sendmsg");
450 FD_SET(socket, sfds);
452 #if defined(HAVE_LINUX_ERRQUEUE_H) && defined(ADAPT_PMTU)
453 while((rxi_HandleSocketError(socket)) > 0)
457 if (WSAGetLastError())
458 #elif defined(AFS_LINUX22_ENV)
459 /* linux unfortunately returns ECONNREFUSED if the target port
460 * is no longer in use */
461 /* and EAGAIN if a UDP checksum is incorrect */
462 if (errno != EWOULDBLOCK && errno != ENOBUFS && errno != ECONNREFUSED
465 if (errno != EWOULDBLOCK && errno != ENOBUFS)
468 (osi_Msg "rx failed to send packet: ");
469 perror("rx_sendmsg");
474 if (WSAGetLastError() > 0)
475 return -WSAGetLastError();
479 while ((err = select(socket + 1, 0, sfds, 0, 0)) != 1) {
480 if (err >= 0 || errno != EINTR)
481 osi_Panic("rxi_sendmsg: select error %d.%d", err, errno);
485 IOMGR_FreeFDSet(sfds);