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: Extended Remote Procedure Call */
12 #include <afsconfig.h>
14 #include "afs/param.h"
16 #include <afs/param.h>
23 #include "afs/sysincludes.h"
24 #include "afsincludes.h"
30 #include <net/net_globals.h>
31 #endif /* AFS_OSF_ENV */
32 #ifdef AFS_LINUX20_ENV
35 #include "netinet/in.h"
37 #include "inet/common.h"
39 #include "inet/ip_ire.h"
41 #include "afs/afs_args.h"
42 #include "afs/afs_osi.h"
43 #ifdef RX_KERNEL_TRACE
44 #include "rx_kcommon.h"
46 #if (defined(AFS_AUX_ENV) || defined(AFS_AIX_ENV))
50 #undef RXDEBUG /* turn off debugging */
52 #if defined(AFS_SGI_ENV)
53 #include "sys/debug.h"
62 #endif /* AFS_OSF_ENV */
64 #include "afs/sysincludes.h"
65 #include "afsincludes.h"
68 #include "rx_kmutex.h"
69 #include "rx_kernel.h"
73 #include "rx_globals.h"
75 #define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
76 #define AFSOP_STOP_AFS 211 /* Stop AFS process */
77 #define AFSOP_STOP_BKG 212 /* Stop BKG process */
79 extern afs_int32 afs_termState;
81 #include "sys/lockl.h"
82 #include "sys/lock_def.h"
83 #endif /* AFS_AIX41_ENV */
84 # include "rxgen_consts.h"
86 # include <sys/types.h>
92 # include <afs/afsutil.h>
93 # include <WINNT\afsreg.h>
95 # include <sys/socket.h>
96 # include <sys/file.h>
98 # include <sys/stat.h>
99 # include <netinet/in.h>
100 # include <sys/time.h>
103 # include "rx_user.h"
104 # include "rx_clock.h"
105 # include "rx_queue.h"
106 # include "rx_globals.h"
107 # include "rx_trace.h"
108 # include <afs/rxgen_consts.h>
111 int (*registerProgram) () = 0;
112 int (*swapNameProgram) () = 0;
114 /* Local static routines */
115 static void rxi_DestroyConnectionNoLock(register struct rx_connection *conn);
116 #ifdef RX_ENABLE_LOCKS
117 static void rxi_SetAcksInTransmitQueue(register struct rx_call *call);
120 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
122 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
123 afs_int32 rxi_start_in_error;
125 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
128 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
129 * currently allocated within rx. This number is used to allocate the
130 * memory required to return the statistics when queried.
133 static unsigned int rxi_rpc_peer_stat_cnt;
136 * rxi_rpc_process_stat_cnt counts the total number of local process stat
137 * structures currently allocated within rx. The number is used to allocate
138 * the memory required to return the statistics when queried.
141 static unsigned int rxi_rpc_process_stat_cnt;
143 #if !defined(offsetof)
144 #include <stddef.h> /* for definition of offsetof() */
147 #ifdef AFS_PTHREAD_ENV
151 * Use procedural initialization of mutexes/condition variables
155 extern pthread_mutex_t rx_stats_mutex;
156 extern pthread_mutex_t des_init_mutex;
157 extern pthread_mutex_t des_random_mutex;
158 extern pthread_mutex_t rx_clock_mutex;
159 extern pthread_mutex_t rxi_connCacheMutex;
160 extern pthread_mutex_t rx_event_mutex;
161 extern pthread_mutex_t osi_malloc_mutex;
162 extern pthread_mutex_t event_handler_mutex;
163 extern pthread_mutex_t listener_mutex;
164 extern pthread_mutex_t rx_if_init_mutex;
165 extern pthread_mutex_t rx_if_mutex;
166 extern pthread_mutex_t rxkad_client_uid_mutex;
167 extern pthread_mutex_t rxkad_random_mutex;
169 extern pthread_cond_t rx_event_handler_cond;
170 extern pthread_cond_t rx_listener_cond;
172 static pthread_mutex_t epoch_mutex;
173 static pthread_mutex_t rx_init_mutex;
174 static pthread_mutex_t rx_debug_mutex;
177 rxi_InitPthread(void)
179 assert(pthread_mutex_init(&rx_clock_mutex, (const pthread_mutexattr_t *)0)
181 assert(pthread_mutex_init(&rx_stats_mutex, (const pthread_mutexattr_t *)0)
183 assert(pthread_mutex_init
184 (&rxi_connCacheMutex, (const pthread_mutexattr_t *)0) == 0);
185 assert(pthread_mutex_init(&rx_init_mutex, (const pthread_mutexattr_t *)0)
187 assert(pthread_mutex_init(&epoch_mutex, (const pthread_mutexattr_t *)0) ==
189 assert(pthread_mutex_init(&rx_event_mutex, (const pthread_mutexattr_t *)0)
191 assert(pthread_mutex_init(&des_init_mutex, (const pthread_mutexattr_t *)0)
193 assert(pthread_mutex_init
194 (&des_random_mutex, (const pthread_mutexattr_t *)0) == 0);
195 assert(pthread_mutex_init
196 (&osi_malloc_mutex, (const pthread_mutexattr_t *)0) == 0);
197 assert(pthread_mutex_init
198 (&event_handler_mutex, (const pthread_mutexattr_t *)0) == 0);
199 assert(pthread_mutex_init(&listener_mutex, (const pthread_mutexattr_t *)0)
201 assert(pthread_mutex_init
202 (&rx_if_init_mutex, (const pthread_mutexattr_t *)0) == 0);
203 assert(pthread_mutex_init(&rx_if_mutex, (const pthread_mutexattr_t *)0) ==
205 assert(pthread_mutex_init
206 (&rxkad_client_uid_mutex, (const pthread_mutexattr_t *)0) == 0);
207 assert(pthread_mutex_init
208 (&rxkad_random_mutex, (const pthread_mutexattr_t *)0) == 0);
209 assert(pthread_mutex_init(&rx_debug_mutex, (const pthread_mutexattr_t *)0)
212 assert(pthread_cond_init
213 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
214 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
216 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
217 assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
219 rxkad_global_stats_init();
222 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
223 #define INIT_PTHREAD_LOCKS \
224 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
226 * The rx_stats_mutex mutex protects the following global variables:
231 * rxi_lowConnRefCount
232 * rxi_lowPeerRefCount
241 #define INIT_PTHREAD_LOCKS
245 /* Variables for handling the minProcs implementation. availProcs gives the
246 * number of threads available in the pool at this moment (not counting dudes
247 * executing right now). totalMin gives the total number of procs required
248 * for handling all minProcs requests. minDeficit is a dynamic variable
249 * tracking the # of procs required to satisfy all of the remaining minProcs
251 * For fine grain locking to work, the quota check and the reservation of
252 * a server thread has to come while rxi_availProcs and rxi_minDeficit
253 * are locked. To this end, the code has been modified under #ifdef
254 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
255 * same time. A new function, ReturnToServerPool() returns the allocation.
257 * A call can be on several queue's (but only one at a time). When
258 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
259 * that no one else is touching the queue. To this end, we store the address
260 * of the queue lock in the call structure (under the call lock) when we
261 * put the call on a queue, and we clear the call_queue_lock when the
262 * call is removed from a queue (once the call lock has been obtained).
263 * This allows rxi_ResetCall to safely synchronize with others wishing
264 * to manipulate the queue.
267 #ifdef RX_ENABLE_LOCKS
268 static afs_kmutex_t rx_rpc_stats;
269 void rxi_StartUnlocked();
272 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
273 ** pretty good that the next packet coming in is from the same connection
274 ** as the last packet, since we're send multiple packets in a transmit window.
276 struct rx_connection *rxLastConn = 0;
278 #ifdef RX_ENABLE_LOCKS
279 /* The locking hierarchy for rx fine grain locking is composed of these
282 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
283 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
284 * call->lock - locks call data fields.
285 * These are independent of each other:
286 * rx_freeCallQueue_lock
291 * serverQueueEntry->lock
293 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
294 * peer->lock - locks peer data fields.
295 * conn_data_lock - that more than one thread is not updating a conn data
296 * field at the same time.
304 * Do we need a lock to protect the peer field in the conn structure?
305 * conn->peer was previously a constant for all intents and so has no
306 * lock protecting this field. The multihomed client delta introduced
307 * a RX code change : change the peer field in the connection structure
308 * to that remote inetrface from which the last packet for this
309 * connection was sent out. This may become an issue if further changes
312 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
313 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
315 /* rxdb_fileID is used to identify the lock location, along with line#. */
316 static int rxdb_fileID = RXDB_FILE_RX;
317 #endif /* RX_LOCKS_DB */
318 #else /* RX_ENABLE_LOCKS */
319 #define SET_CALL_QUEUE_LOCK(C, L)
320 #define CLEAR_CALL_QUEUE_LOCK(C)
321 #endif /* RX_ENABLE_LOCKS */
322 struct rx_serverQueueEntry *rx_waitForPacket = 0;
323 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
325 /* ------------Exported Interfaces------------- */
327 /* This function allows rxkad to set the epoch to a suitably random number
328 * which rx_NewConnection will use in the future. The principle purpose is to
329 * get rxnull connections to use the same epoch as the rxkad connections do, at
330 * least once the first rxkad connection is established. This is important now
331 * that the host/port addresses aren't used in FindConnection: the uniqueness
332 * of epoch/cid matters and the start time won't do. */
334 #ifdef AFS_PTHREAD_ENV
336 * This mutex protects the following global variables:
340 #define LOCK_EPOCH assert(pthread_mutex_lock(&epoch_mutex)==0)
341 #define UNLOCK_EPOCH assert(pthread_mutex_unlock(&epoch_mutex)==0)
345 #endif /* AFS_PTHREAD_ENV */
348 rx_SetEpoch(afs_uint32 epoch)
355 /* Initialize rx. A port number may be mentioned, in which case this
356 * becomes the default port number for any service installed later.
357 * If 0 is provided for the port number, a random port will be chosen
358 * by the kernel. Whether this will ever overlap anything in
359 * /etc/services is anybody's guess... Returns 0 on success, -1 on
364 int rxinit_status = 1;
365 #ifdef AFS_PTHREAD_ENV
367 * This mutex protects the following global variables:
371 #define LOCK_RX_INIT assert(pthread_mutex_lock(&rx_init_mutex)==0)
372 #define UNLOCK_RX_INIT assert(pthread_mutex_unlock(&rx_init_mutex)==0)
375 #define UNLOCK_RX_INIT
379 rx_InitHost(u_int host, u_int port)
386 char *htable, *ptable;
393 if (rxinit_status == 0) {
394 tmp_status = rxinit_status;
396 return tmp_status; /* Already started; return previous error code. */
402 if (afs_winsockInit() < 0)
408 * Initialize anything necessary to provide a non-premptive threading
411 rxi_InitializeThreadSupport();
414 /* Allocate and initialize a socket for client and perhaps server
417 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
418 if (rx_socket == OSI_NULLSOCKET) {
422 #ifdef RX_ENABLE_LOCKS
425 #endif /* RX_LOCKS_DB */
426 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
427 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
428 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
429 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
430 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
432 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
434 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
436 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
438 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
440 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
442 #if defined(KERNEL) && defined(AFS_HPUX110_ENV)
444 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
445 #endif /* KERNEL && AFS_HPUX110_ENV */
446 #endif /* RX_ENABLE_LOCKS */
449 rx_connDeadTime = 12;
450 rx_tranquil = 0; /* reset flag */
451 memset((char *)&rx_stats, 0, sizeof(struct rx_stats));
453 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
454 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
455 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
456 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
457 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
458 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
460 /* Malloc up a bunch of packets & buffers */
462 queue_Init(&rx_freePacketQueue);
463 rxi_NeedMorePackets = FALSE;
464 #ifdef RX_ENABLE_TSFPQ
465 rx_nPackets = 0; /* in TSFPQ version, rx_nPackets is managed by rxi_MorePackets* */
466 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
467 #else /* RX_ENABLE_TSFPQ */
468 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
469 rxi_MorePackets(rx_nPackets);
470 #endif /* RX_ENABLE_TSFPQ */
477 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
478 tv.tv_sec = clock_now.sec;
479 tv.tv_usec = clock_now.usec;
480 srand((unsigned int)tv.tv_usec);
487 #if defined(KERNEL) && !defined(UKERNEL)
488 /* Really, this should never happen in a real kernel */
491 struct sockaddr_in addr;
492 int addrlen = sizeof(addr);
493 if (getsockname((int)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
497 rx_port = addr.sin_port;
500 rx_stats.minRtt.sec = 9999999;
502 rx_SetEpoch(tv.tv_sec | 0x80000000);
504 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
505 * will provide a randomer value. */
507 MUTEX_ENTER(&rx_stats_mutex);
508 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
509 MUTEX_EXIT(&rx_stats_mutex);
510 /* *Slightly* random start time for the cid. This is just to help
511 * out with the hashing function at the peer */
512 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
513 rx_connHashTable = (struct rx_connection **)htable;
514 rx_peerHashTable = (struct rx_peer **)ptable;
516 rx_lastAckDelay.sec = 0;
517 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
518 rx_hardAckDelay.sec = 0;
519 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
520 rx_softAckDelay.sec = 0;
521 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
523 rxevent_Init(20, rxi_ReScheduleEvents);
525 /* Initialize various global queues */
526 queue_Init(&rx_idleServerQueue);
527 queue_Init(&rx_incomingCallQueue);
528 queue_Init(&rx_freeCallQueue);
530 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
531 /* Initialize our list of usable IP addresses. */
535 /* Start listener process (exact function is dependent on the
536 * implementation environment--kernel or user space) */
540 tmp_status = rxinit_status = 0;
548 return rx_InitHost(htonl(INADDR_ANY), port);
551 /* called with unincremented nRequestsRunning to see if it is OK to start
552 * a new thread in this service. Could be "no" for two reasons: over the
553 * max quota, or would prevent others from reaching their min quota.
555 #ifdef RX_ENABLE_LOCKS
556 /* This verion of QuotaOK reserves quota if it's ok while the
557 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
560 QuotaOK(register struct rx_service *aservice)
562 /* check if over max quota */
563 if (aservice->nRequestsRunning >= aservice->maxProcs) {
567 /* under min quota, we're OK */
568 /* otherwise, can use only if there are enough to allow everyone
569 * to go to their min quota after this guy starts.
571 MUTEX_ENTER(&rx_stats_mutex);
572 if ((aservice->nRequestsRunning < aservice->minProcs)
573 || (rxi_availProcs > rxi_minDeficit)) {
574 aservice->nRequestsRunning++;
575 /* just started call in minProcs pool, need fewer to maintain
577 if (aservice->nRequestsRunning <= aservice->minProcs)
580 MUTEX_EXIT(&rx_stats_mutex);
583 MUTEX_EXIT(&rx_stats_mutex);
589 ReturnToServerPool(register struct rx_service *aservice)
591 aservice->nRequestsRunning--;
592 MUTEX_ENTER(&rx_stats_mutex);
593 if (aservice->nRequestsRunning < aservice->minProcs)
596 MUTEX_EXIT(&rx_stats_mutex);
599 #else /* RX_ENABLE_LOCKS */
601 QuotaOK(register struct rx_service *aservice)
604 /* under min quota, we're OK */
605 if (aservice->nRequestsRunning < aservice->minProcs)
608 /* check if over max quota */
609 if (aservice->nRequestsRunning >= aservice->maxProcs)
612 /* otherwise, can use only if there are enough to allow everyone
613 * to go to their min quota after this guy starts.
615 if (rxi_availProcs > rxi_minDeficit)
619 #endif /* RX_ENABLE_LOCKS */
622 /* Called by rx_StartServer to start up lwp's to service calls.
623 NExistingProcs gives the number of procs already existing, and which
624 therefore needn't be created. */
626 rxi_StartServerProcs(int nExistingProcs)
628 register struct rx_service *service;
633 /* For each service, reserve N processes, where N is the "minimum"
634 * number of processes that MUST be able to execute a request in parallel,
635 * at any time, for that process. Also compute the maximum difference
636 * between any service's maximum number of processes that can run
637 * (i.e. the maximum number that ever will be run, and a guarantee
638 * that this number will run if other services aren't running), and its
639 * minimum number. The result is the extra number of processes that
640 * we need in order to provide the latter guarantee */
641 for (i = 0; i < RX_MAX_SERVICES; i++) {
643 service = rx_services[i];
644 if (service == (struct rx_service *)0)
646 nProcs += service->minProcs;
647 diff = service->maxProcs - service->minProcs;
651 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
652 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
653 for (i = 0; i < nProcs; i++) {
654 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
660 /* This routine is only required on Windows */
662 rx_StartClientThread(void)
664 #ifdef AFS_PTHREAD_ENV
666 pid = pthread_self();
667 #endif /* AFS_PTHREAD_ENV */
669 #endif /* AFS_NT40_ENV */
671 /* This routine must be called if any services are exported. If the
672 * donateMe flag is set, the calling process is donated to the server
675 rx_StartServer(int donateMe)
677 register struct rx_service *service;
683 /* Start server processes, if necessary (exact function is dependent
684 * on the implementation environment--kernel or user space). DonateMe
685 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
686 * case, one less new proc will be created rx_StartServerProcs.
688 rxi_StartServerProcs(donateMe);
690 /* count up the # of threads in minProcs, and add set the min deficit to
691 * be that value, too.
693 for (i = 0; i < RX_MAX_SERVICES; i++) {
694 service = rx_services[i];
695 if (service == (struct rx_service *)0)
697 MUTEX_ENTER(&rx_stats_mutex);
698 rxi_totalMin += service->minProcs;
699 /* below works even if a thread is running, since minDeficit would
700 * still have been decremented and later re-incremented.
702 rxi_minDeficit += service->minProcs;
703 MUTEX_EXIT(&rx_stats_mutex);
706 /* Turn on reaping of idle server connections */
707 rxi_ReapConnections();
716 #ifdef AFS_PTHREAD_ENV
718 pid = (pid_t) pthread_self();
719 #else /* AFS_PTHREAD_ENV */
721 LWP_CurrentProcess(&pid);
722 #endif /* AFS_PTHREAD_ENV */
724 sprintf(name, "srv_%d", ++nProcs);
726 (*registerProgram) (pid, name);
728 #endif /* AFS_NT40_ENV */
729 rx_ServerProc(NULL); /* Never returns */
731 #ifdef RX_ENABLE_TSFPQ
732 /* no use leaving packets around in this thread's local queue if
733 * it isn't getting donated to the server thread pool.
735 rxi_FlushLocalPacketsTSFPQ();
736 #endif /* RX_ENABLE_TSFPQ */
740 /* Create a new client connection to the specified service, using the
741 * specified security object to implement the security model for this
743 struct rx_connection *
744 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
745 register struct rx_securityClass *securityObject,
746 int serviceSecurityIndex)
749 afs_int32 cid, cix, nclones;
750 register struct rx_connection *conn, *tconn, *ptconn;
755 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", ntohl(shost), ntohs(sport), sservice, securityObject, serviceSecurityIndex));
758 nclones = rx_max_clones_per_connection;
760 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
761 * the case of kmem_alloc? */
764 MUTEX_ENTER(&rx_connHashTable_lock);
766 /* send in the clones */
767 for(cix = 0; cix <= nclones; ++cix) {
770 tconn = rxi_AllocConnection();
771 tconn->type = RX_CLIENT_CONNECTION;
772 tconn->epoch = rx_epoch;
773 tconn->peer = rxi_FindPeer(shost, sport, 0, 1);
774 tconn->serviceId = sservice;
775 tconn->securityObject = securityObject;
776 tconn->securityData = (void *) 0;
777 tconn->securityIndex = serviceSecurityIndex;
778 tconn->ackRate = RX_FAST_ACK_RATE;
779 tconn->nSpecific = 0;
780 tconn->specific = NULL;
781 tconn->challengeEvent = NULL;
782 tconn->delayedAbortEvent = NULL;
783 tconn->abortCount = 0;
785 for (i = 0; i < RX_MAXCALLS; i++) {
786 tconn->twind[i] = rx_initSendWindow;
787 tconn->rwind[i] = rx_initReceiveWindow;
790 tconn->next_clone = 0;
791 tconn->nclones = nclones;
792 rx_SetConnDeadTime(tconn, rx_connDeadTime);
797 tconn->flags |= RX_CLONED_CONNECTION;
798 tconn->parent = conn;
799 ptconn->next_clone = tconn;
802 /* generic connection setup */
803 #ifdef RX_ENABLE_LOCKS
804 MUTEX_INIT(&tconn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
805 MUTEX_INIT(&tconn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
806 CV_INIT(&tconn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
808 cid = (rx_nextCid += RX_MAXCALLS);
810 RXS_NewConnection(securityObject, tconn);
812 CONN_HASH(shost, sport, tconn->cid, tconn->epoch,
813 RX_CLIENT_CONNECTION);
814 tconn->refCount++; /* no lock required since only this thread knows */
815 tconn->next = rx_connHashTable[hashindex];
816 rx_connHashTable[hashindex] = tconn;
817 rx_MutexIncrement(rx_stats.nClientConns, rx_stats_mutex);
820 MUTEX_EXIT(&rx_connHashTable_lock);
826 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
828 /* The idea is to set the dead time to a value that allows several
829 * keepalives to be dropped without timing out the connection. */
830 struct rx_connection *tconn;
833 tconn->secondsUntilDead = MAX(seconds, 6);
834 tconn->secondsUntilPing = tconn->secondsUntilDead / 6;
835 } while(tconn->next_clone && (tconn = tconn->next_clone));
838 int rxi_lowPeerRefCount = 0;
839 int rxi_lowConnRefCount = 0;
842 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
843 * NOTE: must not be called with rx_connHashTable_lock held.
846 rxi_CleanupConnection(struct rx_connection *conn)
848 /* Notify the service exporter, if requested, that this connection
849 * is being destroyed */
850 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
851 (*conn->service->destroyConnProc) (conn);
853 /* Notify the security module that this connection is being destroyed */
854 RXS_DestroyConnection(conn->securityObject, conn);
856 /* If this is the last connection using the rx_peer struct, set its
857 * idle time to now. rxi_ReapConnections will reap it if it's still
858 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
860 MUTEX_ENTER(&rx_peerHashTable_lock);
861 if (conn->peer->refCount < 2) {
862 conn->peer->idleWhen = clock_Sec();
863 if (conn->peer->refCount < 1) {
864 conn->peer->refCount = 1;
865 MUTEX_ENTER(&rx_stats_mutex);
866 rxi_lowPeerRefCount++;
867 MUTEX_EXIT(&rx_stats_mutex);
870 conn->peer->refCount--;
871 MUTEX_EXIT(&rx_peerHashTable_lock);
873 if (conn->type == RX_SERVER_CONNECTION)
874 rx_MutexDecrement(rx_stats.nServerConns, rx_stats_mutex);
876 rx_MutexDecrement(rx_stats.nClientConns, rx_stats_mutex);
878 if (conn->specific) {
880 for (i = 0; i < conn->nSpecific; i++) {
881 if (conn->specific[i] && rxi_keyCreate_destructor[i])
882 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
883 conn->specific[i] = NULL;
885 free(conn->specific);
887 conn->specific = NULL;
891 MUTEX_DESTROY(&conn->conn_call_lock);
892 MUTEX_DESTROY(&conn->conn_data_lock);
893 CV_DESTROY(&conn->conn_call_cv);
895 rxi_FreeConnection(conn);
898 /* Destroy the specified connection */
900 rxi_DestroyConnection(register struct rx_connection *conn)
902 register struct rx_connection *tconn, *dtconn;
904 MUTEX_ENTER(&rx_connHashTable_lock);
906 if(!(conn->flags & RX_CLONED_CONNECTION)) {
907 tconn = conn->next_clone;
908 conn->next_clone = 0; /* once */
912 tconn = tconn->next_clone;
913 rxi_DestroyConnectionNoLock(dtconn);
915 if (dtconn == rx_connCleanup_list) {
916 rx_connCleanup_list = rx_connCleanup_list->next;
917 MUTEX_EXIT(&rx_connHashTable_lock);
918 /* rxi_CleanupConnection will free tconn */
919 rxi_CleanupConnection(dtconn);
920 MUTEX_ENTER(&rx_connHashTable_lock);
927 rxi_DestroyConnectionNoLock(conn);
928 /* conn should be at the head of the cleanup list */
929 if (conn == rx_connCleanup_list) {
930 rx_connCleanup_list = rx_connCleanup_list->next;
931 MUTEX_EXIT(&rx_connHashTable_lock);
932 rxi_CleanupConnection(conn);
934 #ifdef RX_ENABLE_LOCKS
936 MUTEX_EXIT(&rx_connHashTable_lock);
938 #endif /* RX_ENABLE_LOCKS */
942 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
944 register struct rx_connection **conn_ptr;
945 register int havecalls = 0;
946 struct rx_packet *packet;
953 MUTEX_ENTER(&conn->conn_data_lock);
954 if (conn->refCount > 0)
957 MUTEX_ENTER(&rx_stats_mutex);
958 rxi_lowConnRefCount++;
959 MUTEX_EXIT(&rx_stats_mutex);
962 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
963 /* Busy; wait till the last guy before proceeding */
964 MUTEX_EXIT(&conn->conn_data_lock);
969 /* If the client previously called rx_NewCall, but it is still
970 * waiting, treat this as a running call, and wait to destroy the
971 * connection later when the call completes. */
972 if ((conn->type == RX_CLIENT_CONNECTION)
973 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
974 conn->flags |= RX_CONN_DESTROY_ME;
975 MUTEX_EXIT(&conn->conn_data_lock);
979 MUTEX_EXIT(&conn->conn_data_lock);
981 /* Check for extant references to this connection */
982 for (i = 0; i < RX_MAXCALLS; i++) {
983 register struct rx_call *call = conn->call[i];
986 if (conn->type == RX_CLIENT_CONNECTION) {
987 MUTEX_ENTER(&call->lock);
988 if (call->delayedAckEvent) {
989 /* Push the final acknowledgment out now--there
990 * won't be a subsequent call to acknowledge the
991 * last reply packets */
992 rxevent_Cancel(call->delayedAckEvent, call,
993 RX_CALL_REFCOUNT_DELAY);
994 if (call->state == RX_STATE_PRECALL
995 || call->state == RX_STATE_ACTIVE) {
996 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
998 rxi_AckAll(NULL, call, 0);
1001 MUTEX_EXIT(&call->lock);
1005 #ifdef RX_ENABLE_LOCKS
1007 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1008 MUTEX_EXIT(&conn->conn_data_lock);
1010 /* Someone is accessing a packet right now. */
1014 #endif /* RX_ENABLE_LOCKS */
1017 /* Don't destroy the connection if there are any call
1018 * structures still in use */
1019 MUTEX_ENTER(&conn->conn_data_lock);
1020 conn->flags |= RX_CONN_DESTROY_ME;
1021 MUTEX_EXIT(&conn->conn_data_lock);
1026 if (conn->delayedAbortEvent) {
1027 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
1028 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1030 MUTEX_ENTER(&conn->conn_data_lock);
1031 rxi_SendConnectionAbort(conn, packet, 0, 1);
1032 MUTEX_EXIT(&conn->conn_data_lock);
1033 rxi_FreePacket(packet);
1037 /* Remove from connection hash table before proceeding */
1039 &rx_connHashTable[CONN_HASH
1040 (peer->host, peer->port, conn->cid, conn->epoch,
1042 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1043 if (*conn_ptr == conn) {
1044 *conn_ptr = conn->next;
1048 /* if the conn that we are destroying was the last connection, then we
1049 * clear rxLastConn as well */
1050 if (rxLastConn == conn)
1053 /* Make sure the connection is completely reset before deleting it. */
1054 /* get rid of pending events that could zap us later */
1055 if (conn->challengeEvent)
1056 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1057 if (conn->checkReachEvent)
1058 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1060 /* Add the connection to the list of destroyed connections that
1061 * need to be cleaned up. This is necessary to avoid deadlocks
1062 * in the routines we call to inform others that this connection is
1063 * being destroyed. */
1064 conn->next = rx_connCleanup_list;
1065 rx_connCleanup_list = conn;
1068 /* Externally available version */
1070 rx_DestroyConnection(register struct rx_connection *conn)
1075 rxi_DestroyConnection(conn);
1080 rx_GetConnection(register struct rx_connection *conn)
1085 MUTEX_ENTER(&conn->conn_data_lock);
1087 MUTEX_EXIT(&conn->conn_data_lock);
1091 /* Wait for the transmit queue to no longer be busy.
1092 * requires the call->lock to be held */
1093 static void rxi_WaitforTQBusy(struct rx_call *call) {
1094 while (call->flags & RX_CALL_TQ_BUSY) {
1095 call->flags |= RX_CALL_TQ_WAIT;
1097 #ifdef RX_ENABLE_LOCKS
1098 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1099 CV_WAIT(&call->cv_tq, &call->lock);
1100 #else /* RX_ENABLE_LOCKS */
1101 osi_rxSleep(&call->tq);
1102 #endif /* RX_ENABLE_LOCKS */
1104 if (call->tqWaiters == 0) {
1105 call->flags &= ~RX_CALL_TQ_WAIT;
1109 /* Start a new rx remote procedure call, on the specified connection.
1110 * If wait is set to 1, wait for a free call channel; otherwise return
1111 * 0. Maxtime gives the maximum number of seconds this call may take,
1112 * after rx_NewCall returns. After this time interval, a call to any
1113 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1114 * For fine grain locking, we hold the conn_call_lock in order to
1115 * to ensure that we don't get signalle after we found a call in an active
1116 * state and before we go to sleep.
1119 rx_NewCall(register struct rx_connection *conn)
1122 register struct rx_call *call;
1123 register struct rx_connection *tconn;
1124 struct clock queueTime;
1128 dpf(("rx_NewCall(conn %x)\n", conn));
1131 clock_GetTime(&queueTime);
1132 MUTEX_ENTER(&conn->conn_call_lock);
1135 * Check if there are others waiting for a new call.
1136 * If so, let them go first to avoid starving them.
1137 * This is a fairly simple scheme, and might not be
1138 * a complete solution for large numbers of waiters.
1140 * makeCallWaiters keeps track of the number of
1141 * threads waiting to make calls and the
1142 * RX_CONN_MAKECALL_WAITING flag bit is used to
1143 * indicate that there are indeed calls waiting.
1144 * The flag is set when the waiter is incremented.
1145 * It is only cleared in rx_EndCall when
1146 * makeCallWaiters is 0. This prevents us from
1147 * accidently destroying the connection while it
1148 * is potentially about to be used.
1150 MUTEX_ENTER(&conn->conn_data_lock);
1151 if (conn->makeCallWaiters) {
1152 conn->flags |= RX_CONN_MAKECALL_WAITING;
1153 conn->makeCallWaiters++;
1154 MUTEX_EXIT(&conn->conn_data_lock);
1156 #ifdef RX_ENABLE_LOCKS
1157 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1161 MUTEX_ENTER(&conn->conn_data_lock);
1162 conn->makeCallWaiters--;
1164 MUTEX_EXIT(&conn->conn_data_lock);
1166 /* search for next free call on this connection or
1167 * its clones, if any */
1171 for (i = 0; i < RX_MAXCALLS; i++) {
1172 call = tconn->call[i];
1174 MUTEX_ENTER(&call->lock);
1175 if (call->state == RX_STATE_DALLY) {
1176 rxi_ResetCall(call, 0);
1177 (*call->callNumber)++;
1180 MUTEX_EXIT(&call->lock);
1182 call = rxi_NewCall(tconn, i);
1185 } /* for i < RX_MAXCALLS */
1186 } while (tconn->next_clone && (tconn = tconn->next_clone));
1190 if (i < RX_MAXCALLS) {
1194 /* to be here, all available calls for this connection (and all
1195 * its clones) must be in use */
1197 MUTEX_ENTER(&conn->conn_data_lock);
1198 conn->flags |= RX_CONN_MAKECALL_WAITING;
1199 conn->makeCallWaiters++;
1200 MUTEX_EXIT(&conn->conn_data_lock);
1202 #ifdef RX_ENABLE_LOCKS
1203 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1207 MUTEX_ENTER(&conn->conn_data_lock);
1208 conn->makeCallWaiters--;
1209 MUTEX_EXIT(&conn->conn_data_lock);
1212 * Wake up anyone else who might be giving us a chance to
1213 * run (see code above that avoids resource starvation).
1215 #ifdef RX_ENABLE_LOCKS
1216 CV_BROADCAST(&conn->conn_call_cv);
1221 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1223 /* Client is initially in send mode */
1224 call->state = RX_STATE_ACTIVE;
1225 call->error = conn->error;
1227 call->mode = RX_MODE_ERROR;
1229 call->mode = RX_MODE_SENDING;
1231 /* remember start time for call in case we have hard dead time limit */
1232 call->queueTime = queueTime;
1233 clock_GetTime(&call->startTime);
1234 hzero(call->bytesSent);
1235 hzero(call->bytesRcvd);
1237 /* Turn on busy protocol. */
1238 rxi_KeepAliveOn(call);
1240 MUTEX_EXIT(&call->lock);
1241 MUTEX_EXIT(&conn->conn_call_lock);
1244 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1245 /* Now, if TQ wasn't cleared earlier, do it now. */
1246 MUTEX_ENTER(&call->lock);
1247 rxi_WaitforTQBusy(call);
1248 if (call->flags & RX_CALL_TQ_CLEARME) {
1249 rxi_ClearTransmitQueue(call, 0);
1250 queue_Init(&call->tq);
1252 MUTEX_EXIT(&call->lock);
1253 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1255 dpf(("rx_NewCall(call %x)\n", call));
1260 rxi_HasActiveCalls(register struct rx_connection *aconn)
1263 register struct rx_call *tcall;
1267 for (i = 0; i < RX_MAXCALLS; i++) {
1268 if ((tcall = aconn->call[i])) {
1269 if ((tcall->state == RX_STATE_ACTIVE)
1270 || (tcall->state == RX_STATE_PRECALL)) {
1281 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1282 register afs_int32 * aint32s)
1285 register struct rx_call *tcall;
1289 for (i = 0; i < RX_MAXCALLS; i++) {
1290 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1291 aint32s[i] = aconn->callNumber[i] + 1;
1293 aint32s[i] = aconn->callNumber[i];
1300 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1301 register afs_int32 * aint32s)
1304 register struct rx_call *tcall;
1308 for (i = 0; i < RX_MAXCALLS; i++) {
1309 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1310 aconn->callNumber[i] = aint32s[i] - 1;
1312 aconn->callNumber[i] = aint32s[i];
1318 /* Advertise a new service. A service is named locally by a UDP port
1319 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1322 char *serviceName; Name for identification purposes (e.g. the
1323 service name might be used for probing for
1326 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1327 char *serviceName, struct rx_securityClass **securityObjects,
1328 int nSecurityObjects,
1329 afs_int32(*serviceProc) (struct rx_call * acall))
1331 osi_socket socket = OSI_NULLSOCKET;
1332 register struct rx_service *tservice;
1338 if (serviceId == 0) {
1340 "rx_NewService: service id for service %s is not non-zero.\n",
1347 "rx_NewService: A non-zero port must be specified on this call if a non-zero port was not provided at Rx initialization (service %s).\n",
1355 tservice = rxi_AllocService();
1357 for (i = 0; i < RX_MAX_SERVICES; i++) {
1358 register struct rx_service *service = rx_services[i];
1360 if (port == service->servicePort && host == service->serviceHost) {
1361 if (service->serviceId == serviceId) {
1362 /* The identical service has already been
1363 * installed; if the caller was intending to
1364 * change the security classes used by this
1365 * service, he/she loses. */
1367 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1368 serviceName, serviceId, service->serviceName);
1370 rxi_FreeService(tservice);
1373 /* Different service, same port: re-use the socket
1374 * which is bound to the same port */
1375 socket = service->socket;
1378 if (socket == OSI_NULLSOCKET) {
1379 /* If we don't already have a socket (from another
1380 * service on same port) get a new one */
1381 socket = rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
1382 if (socket == OSI_NULLSOCKET) {
1384 rxi_FreeService(tservice);
1389 service->socket = socket;
1390 service->serviceHost = host;
1391 service->servicePort = port;
1392 service->serviceId = serviceId;
1393 service->serviceName = serviceName;
1394 service->nSecurityObjects = nSecurityObjects;
1395 service->securityObjects = securityObjects;
1396 service->minProcs = 0;
1397 service->maxProcs = 1;
1398 service->idleDeadTime = 60;
1399 service->idleDeadErr = 0;
1400 service->connDeadTime = rx_connDeadTime;
1401 service->executeRequestProc = serviceProc;
1402 service->checkReach = 0;
1403 rx_services[i] = service; /* not visible until now */
1409 rxi_FreeService(tservice);
1410 (osi_Msg "rx_NewService: cannot support > %d services\n",
1415 /* Set configuration options for all of a service's security objects */
1418 rx_SetSecurityConfiguration(struct rx_service *service,
1419 rx_securityConfigVariables type,
1423 for (i = 0; i<service->nSecurityObjects; i++) {
1424 if (service->securityObjects[i]) {
1425 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1433 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1434 struct rx_securityClass **securityObjects, int nSecurityObjects,
1435 afs_int32(*serviceProc) (struct rx_call * acall))
1437 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1440 /* Generic request processing loop. This routine should be called
1441 * by the implementation dependent rx_ServerProc. If socketp is
1442 * non-null, it will be set to the file descriptor that this thread
1443 * is now listening on. If socketp is null, this routine will never
1446 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1448 register struct rx_call *call;
1449 register afs_int32 code;
1450 register struct rx_service *tservice = NULL;
1457 call = rx_GetCall(threadID, tservice, socketp);
1458 if (socketp && *socketp != OSI_NULLSOCKET) {
1459 /* We are now a listener thread */
1464 /* if server is restarting( typically smooth shutdown) then do not
1465 * allow any new calls.
1468 if (rx_tranquil && (call != NULL)) {
1472 MUTEX_ENTER(&call->lock);
1474 rxi_CallError(call, RX_RESTARTING);
1475 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1477 MUTEX_EXIT(&call->lock);
1481 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1482 #ifdef RX_ENABLE_LOCKS
1484 #endif /* RX_ENABLE_LOCKS */
1485 afs_termState = AFSOP_STOP_AFS;
1486 afs_osi_Wakeup(&afs_termState);
1487 #ifdef RX_ENABLE_LOCKS
1489 #endif /* RX_ENABLE_LOCKS */
1494 tservice = call->conn->service;
1496 if (tservice->beforeProc)
1497 (*tservice->beforeProc) (call);
1499 code = call->conn->service->executeRequestProc(call);
1501 if (tservice->afterProc)
1502 (*tservice->afterProc) (call, code);
1504 rx_EndCall(call, code);
1505 MUTEX_ENTER(&rx_stats_mutex);
1507 MUTEX_EXIT(&rx_stats_mutex);
1513 rx_WakeupServerProcs(void)
1515 struct rx_serverQueueEntry *np, *tqp;
1519 MUTEX_ENTER(&rx_serverPool_lock);
1521 #ifdef RX_ENABLE_LOCKS
1522 if (rx_waitForPacket)
1523 CV_BROADCAST(&rx_waitForPacket->cv);
1524 #else /* RX_ENABLE_LOCKS */
1525 if (rx_waitForPacket)
1526 osi_rxWakeup(rx_waitForPacket);
1527 #endif /* RX_ENABLE_LOCKS */
1528 MUTEX_ENTER(&freeSQEList_lock);
1529 for (np = rx_FreeSQEList; np; np = tqp) {
1530 tqp = *(struct rx_serverQueueEntry **)np;
1531 #ifdef RX_ENABLE_LOCKS
1532 CV_BROADCAST(&np->cv);
1533 #else /* RX_ENABLE_LOCKS */
1535 #endif /* RX_ENABLE_LOCKS */
1537 MUTEX_EXIT(&freeSQEList_lock);
1538 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1539 #ifdef RX_ENABLE_LOCKS
1540 CV_BROADCAST(&np->cv);
1541 #else /* RX_ENABLE_LOCKS */
1543 #endif /* RX_ENABLE_LOCKS */
1545 MUTEX_EXIT(&rx_serverPool_lock);
1550 * One thing that seems to happen is that all the server threads get
1551 * tied up on some empty or slow call, and then a whole bunch of calls
1552 * arrive at once, using up the packet pool, so now there are more
1553 * empty calls. The most critical resources here are server threads
1554 * and the free packet pool. The "doreclaim" code seems to help in
1555 * general. I think that eventually we arrive in this state: there
1556 * are lots of pending calls which do have all their packets present,
1557 * so they won't be reclaimed, are multi-packet calls, so they won't
1558 * be scheduled until later, and thus are tying up most of the free
1559 * packet pool for a very long time.
1561 * 1. schedule multi-packet calls if all the packets are present.
1562 * Probably CPU-bound operation, useful to return packets to pool.
1563 * Do what if there is a full window, but the last packet isn't here?
1564 * 3. preserve one thread which *only* runs "best" calls, otherwise
1565 * it sleeps and waits for that type of call.
1566 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1567 * the current dataquota business is badly broken. The quota isn't adjusted
1568 * to reflect how many packets are presently queued for a running call.
1569 * So, when we schedule a queued call with a full window of packets queued
1570 * up for it, that *should* free up a window full of packets for other 2d-class
1571 * calls to be able to use from the packet pool. But it doesn't.
1573 * NB. Most of the time, this code doesn't run -- since idle server threads
1574 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1575 * as a new call arrives.
1577 /* Sleep until a call arrives. Returns a pointer to the call, ready
1578 * for an rx_Read. */
1579 #ifdef RX_ENABLE_LOCKS
1581 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1583 struct rx_serverQueueEntry *sq;
1584 register struct rx_call *call = (struct rx_call *)0;
1585 struct rx_service *service = NULL;
1588 MUTEX_ENTER(&freeSQEList_lock);
1590 if ((sq = rx_FreeSQEList)) {
1591 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1592 MUTEX_EXIT(&freeSQEList_lock);
1593 } else { /* otherwise allocate a new one and return that */
1594 MUTEX_EXIT(&freeSQEList_lock);
1595 sq = (struct rx_serverQueueEntry *)
1596 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1597 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1598 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1601 MUTEX_ENTER(&rx_serverPool_lock);
1602 if (cur_service != NULL) {
1603 ReturnToServerPool(cur_service);
1606 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1607 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1609 /* Scan for eligible incoming calls. A call is not eligible
1610 * if the maximum number of calls for its service type are
1611 * already executing */
1612 /* One thread will process calls FCFS (to prevent starvation),
1613 * while the other threads may run ahead looking for calls which
1614 * have all their input data available immediately. This helps
1615 * keep threads from blocking, waiting for data from the client. */
1616 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1617 service = tcall->conn->service;
1618 if (!QuotaOK(service)) {
1621 if (tno == rxi_fcfs_thread_num
1622 || !tcall->queue_item_header.next) {
1623 /* If we're the fcfs thread , then we'll just use
1624 * this call. If we haven't been able to find an optimal
1625 * choice, and we're at the end of the list, then use a
1626 * 2d choice if one has been identified. Otherwise... */
1627 call = (choice2 ? choice2 : tcall);
1628 service = call->conn->service;
1629 } else if (!queue_IsEmpty(&tcall->rq)) {
1630 struct rx_packet *rp;
1631 rp = queue_First(&tcall->rq, rx_packet);
1632 if (rp->header.seq == 1) {
1634 || (rp->header.flags & RX_LAST_PACKET)) {
1636 } else if (rxi_2dchoice && !choice2
1637 && !(tcall->flags & RX_CALL_CLEARED)
1638 && (tcall->rprev > rxi_HardAckRate)) {
1647 ReturnToServerPool(service);
1654 MUTEX_EXIT(&rx_serverPool_lock);
1655 MUTEX_ENTER(&call->lock);
1657 if (call->flags & RX_CALL_WAIT_PROC) {
1658 call->flags &= ~RX_CALL_WAIT_PROC;
1659 MUTEX_ENTER(&rx_stats_mutex);
1661 MUTEX_EXIT(&rx_stats_mutex);
1664 if (call->state != RX_STATE_PRECALL || call->error) {
1665 MUTEX_EXIT(&call->lock);
1666 MUTEX_ENTER(&rx_serverPool_lock);
1667 ReturnToServerPool(service);
1672 if (queue_IsEmpty(&call->rq)
1673 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1674 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1676 CLEAR_CALL_QUEUE_LOCK(call);
1679 /* If there are no eligible incoming calls, add this process
1680 * to the idle server queue, to wait for one */
1684 *socketp = OSI_NULLSOCKET;
1686 sq->socketp = socketp;
1687 queue_Append(&rx_idleServerQueue, sq);
1688 #ifndef AFS_AIX41_ENV
1689 rx_waitForPacket = sq;
1691 rx_waitingForPacket = sq;
1692 #endif /* AFS_AIX41_ENV */
1694 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1696 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1697 MUTEX_EXIT(&rx_serverPool_lock);
1698 return (struct rx_call *)0;
1701 } while (!(call = sq->newcall)
1702 && !(socketp && *socketp != OSI_NULLSOCKET));
1703 MUTEX_EXIT(&rx_serverPool_lock);
1705 MUTEX_ENTER(&call->lock);
1711 MUTEX_ENTER(&freeSQEList_lock);
1712 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1713 rx_FreeSQEList = sq;
1714 MUTEX_EXIT(&freeSQEList_lock);
1717 clock_GetTime(&call->startTime);
1718 call->state = RX_STATE_ACTIVE;
1719 call->mode = RX_MODE_RECEIVING;
1720 #ifdef RX_KERNEL_TRACE
1721 if (ICL_SETACTIVE(afs_iclSetp)) {
1722 int glockOwner = ISAFS_GLOCK();
1725 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1726 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1733 rxi_calltrace(RX_CALL_START, call);
1734 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1735 call->conn->service->servicePort, call->conn->service->serviceId,
1738 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1739 MUTEX_EXIT(&call->lock);
1741 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1746 #else /* RX_ENABLE_LOCKS */
1748 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1750 struct rx_serverQueueEntry *sq;
1751 register struct rx_call *call = (struct rx_call *)0, *choice2;
1752 struct rx_service *service = NULL;
1756 MUTEX_ENTER(&freeSQEList_lock);
1758 if ((sq = rx_FreeSQEList)) {
1759 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1760 MUTEX_EXIT(&freeSQEList_lock);
1761 } else { /* otherwise allocate a new one and return that */
1762 MUTEX_EXIT(&freeSQEList_lock);
1763 sq = (struct rx_serverQueueEntry *)
1764 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1765 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1766 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1768 MUTEX_ENTER(&sq->lock);
1770 if (cur_service != NULL) {
1771 cur_service->nRequestsRunning--;
1772 if (cur_service->nRequestsRunning < cur_service->minProcs)
1776 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1777 register struct rx_call *tcall, *ncall;
1778 /* Scan for eligible incoming calls. A call is not eligible
1779 * if the maximum number of calls for its service type are
1780 * already executing */
1781 /* One thread will process calls FCFS (to prevent starvation),
1782 * while the other threads may run ahead looking for calls which
1783 * have all their input data available immediately. This helps
1784 * keep threads from blocking, waiting for data from the client. */
1785 choice2 = (struct rx_call *)0;
1786 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1787 service = tcall->conn->service;
1788 if (QuotaOK(service)) {
1789 if (tno == rxi_fcfs_thread_num
1790 || !tcall->queue_item_header.next) {
1791 /* If we're the fcfs thread, then we'll just use
1792 * this call. If we haven't been able to find an optimal
1793 * choice, and we're at the end of the list, then use a
1794 * 2d choice if one has been identified. Otherwise... */
1795 call = (choice2 ? choice2 : tcall);
1796 service = call->conn->service;
1797 } else if (!queue_IsEmpty(&tcall->rq)) {
1798 struct rx_packet *rp;
1799 rp = queue_First(&tcall->rq, rx_packet);
1800 if (rp->header.seq == 1
1802 || (rp->header.flags & RX_LAST_PACKET))) {
1804 } else if (rxi_2dchoice && !choice2
1805 && !(tcall->flags & RX_CALL_CLEARED)
1806 && (tcall->rprev > rxi_HardAckRate)) {
1819 /* we can't schedule a call if there's no data!!! */
1820 /* send an ack if there's no data, if we're missing the
1821 * first packet, or we're missing something between first
1822 * and last -- there's a "hole" in the incoming data. */
1823 if (queue_IsEmpty(&call->rq)
1824 || queue_First(&call->rq, rx_packet)->header.seq != 1
1825 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1826 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1828 call->flags &= (~RX_CALL_WAIT_PROC);
1829 service->nRequestsRunning++;
1830 /* just started call in minProcs pool, need fewer to maintain
1832 if (service->nRequestsRunning <= service->minProcs)
1836 /* MUTEX_EXIT(&call->lock); */
1838 /* If there are no eligible incoming calls, add this process
1839 * to the idle server queue, to wait for one */
1842 *socketp = OSI_NULLSOCKET;
1844 sq->socketp = socketp;
1845 queue_Append(&rx_idleServerQueue, sq);
1849 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1851 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1852 return (struct rx_call *)0;
1855 } while (!(call = sq->newcall)
1856 && !(socketp && *socketp != OSI_NULLSOCKET));
1858 MUTEX_EXIT(&sq->lock);
1860 MUTEX_ENTER(&freeSQEList_lock);
1861 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1862 rx_FreeSQEList = sq;
1863 MUTEX_EXIT(&freeSQEList_lock);
1866 clock_GetTime(&call->startTime);
1867 call->state = RX_STATE_ACTIVE;
1868 call->mode = RX_MODE_RECEIVING;
1869 #ifdef RX_KERNEL_TRACE
1870 if (ICL_SETACTIVE(afs_iclSetp)) {
1871 int glockOwner = ISAFS_GLOCK();
1874 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1875 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1882 rxi_calltrace(RX_CALL_START, call);
1883 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1884 call->conn->service->servicePort, call->conn->service->serviceId,
1887 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1894 #endif /* RX_ENABLE_LOCKS */
1898 /* Establish a procedure to be called when a packet arrives for a
1899 * call. This routine will be called at most once after each call,
1900 * and will also be called if there is an error condition on the or
1901 * the call is complete. Used by multi rx to build a selection
1902 * function which determines which of several calls is likely to be a
1903 * good one to read from.
1904 * NOTE: the way this is currently implemented it is probably only a
1905 * good idea to (1) use it immediately after a newcall (clients only)
1906 * and (2) only use it once. Other uses currently void your warranty
1909 rx_SetArrivalProc(register struct rx_call *call,
1910 register void (*proc) (register struct rx_call * call,
1912 register int index),
1913 register void * handle, register int arg)
1915 call->arrivalProc = proc;
1916 call->arrivalProcHandle = handle;
1917 call->arrivalProcArg = arg;
1920 /* Call is finished (possibly prematurely). Return rc to the peer, if
1921 * appropriate, and return the final error code from the conversation
1925 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1927 register struct rx_connection *conn = call->conn;
1928 register struct rx_service *service;
1934 dpf(("rx_EndCall(call %x rc %d error %d abortCode %d)\n", call, rc, call->error, call->abortCode));
1937 MUTEX_ENTER(&call->lock);
1939 if (rc == 0 && call->error == 0) {
1940 call->abortCode = 0;
1941 call->abortCount = 0;
1944 call->arrivalProc = (void (*)())0;
1945 if (rc && call->error == 0) {
1946 rxi_CallError(call, rc);
1947 /* Send an abort message to the peer if this error code has
1948 * only just been set. If it was set previously, assume the
1949 * peer has already been sent the error code or will request it
1951 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1953 if (conn->type == RX_SERVER_CONNECTION) {
1954 /* Make sure reply or at least dummy reply is sent */
1955 if (call->mode == RX_MODE_RECEIVING) {
1956 rxi_WriteProc(call, 0, 0);
1958 if (call->mode == RX_MODE_SENDING) {
1959 rxi_FlushWrite(call);
1961 service = conn->service;
1962 rxi_calltrace(RX_CALL_END, call);
1963 /* Call goes to hold state until reply packets are acknowledged */
1964 if (call->tfirst + call->nSoftAcked < call->tnext) {
1965 call->state = RX_STATE_HOLD;
1967 call->state = RX_STATE_DALLY;
1968 rxi_ClearTransmitQueue(call, 0);
1969 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1970 rxevent_Cancel(call->keepAliveEvent, call,
1971 RX_CALL_REFCOUNT_ALIVE);
1973 } else { /* Client connection */
1975 /* Make sure server receives input packets, in the case where
1976 * no reply arguments are expected */
1977 if ((call->mode == RX_MODE_SENDING)
1978 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1979 (void)rxi_ReadProc(call, &dummy, 1);
1982 /* If we had an outstanding delayed ack, be nice to the server
1983 * and force-send it now.
1985 if (call->delayedAckEvent) {
1986 rxevent_Cancel(call->delayedAckEvent, call,
1987 RX_CALL_REFCOUNT_DELAY);
1988 call->delayedAckEvent = NULL;
1989 rxi_SendDelayedAck(NULL, call, NULL);
1992 /* We need to release the call lock since it's lower than the
1993 * conn_call_lock and we don't want to hold the conn_call_lock
1994 * over the rx_ReadProc call. The conn_call_lock needs to be held
1995 * here for the case where rx_NewCall is perusing the calls on
1996 * the connection structure. We don't want to signal until
1997 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1998 * have checked this call, found it active and by the time it
1999 * goes to sleep, will have missed the signal.
2001 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
2002 * there are threads waiting to use the conn object.
2004 MUTEX_EXIT(&call->lock);
2005 MUTEX_ENTER(&conn->conn_call_lock);
2006 MUTEX_ENTER(&call->lock);
2007 MUTEX_ENTER(&conn->conn_data_lock);
2008 conn->flags |= RX_CONN_BUSY;
2009 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2010 if (conn->makeCallWaiters == 0)
2011 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
2012 MUTEX_EXIT(&conn->conn_data_lock);
2013 #ifdef RX_ENABLE_LOCKS
2014 CV_BROADCAST(&conn->conn_call_cv);
2019 #ifdef RX_ENABLE_LOCKS
2021 MUTEX_EXIT(&conn->conn_data_lock);
2023 #endif /* RX_ENABLE_LOCKS */
2024 call->state = RX_STATE_DALLY;
2026 error = call->error;
2028 /* currentPacket, nLeft, and NFree must be zeroed here, because
2029 * ResetCall cannot: ResetCall may be called at splnet(), in the
2030 * kernel version, and may interrupt the macros rx_Read or
2031 * rx_Write, which run at normal priority for efficiency. */
2032 if (call->currentPacket) {
2033 queue_Prepend(&call->iovq, call->currentPacket);
2034 call->currentPacket = (struct rx_packet *)0;
2037 call->nLeft = call->nFree = call->curlen = 0;
2039 /* Free any packets from the last call to ReadvProc/WritevProc */
2040 rxi_FreePackets(0, &call->iovq);
2042 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2043 MUTEX_EXIT(&call->lock);
2044 if (conn->type == RX_CLIENT_CONNECTION) {
2045 MUTEX_EXIT(&conn->conn_call_lock);
2046 conn->flags &= ~RX_CONN_BUSY;
2050 * Map errors to the local host's errno.h format.
2052 error = ntoh_syserr_conv(error);
2056 #if !defined(KERNEL)
2058 /* Call this routine when shutting down a server or client (especially
2059 * clients). This will allow Rx to gracefully garbage collect server
2060 * connections, and reduce the number of retries that a server might
2061 * make to a dead client.
2062 * This is not quite right, since some calls may still be ongoing and
2063 * we can't lock them to destroy them. */
2067 register struct rx_connection **conn_ptr, **conn_end;
2071 if (rxinit_status == 1) {
2073 return; /* Already shutdown. */
2075 rxi_DeleteCachedConnections();
2076 if (rx_connHashTable) {
2077 MUTEX_ENTER(&rx_connHashTable_lock);
2078 for (conn_ptr = &rx_connHashTable[0], conn_end =
2079 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2081 struct rx_connection *conn, *next;
2082 for (conn = *conn_ptr; conn; conn = next) {
2084 if (conn->type == RX_CLIENT_CONNECTION) {
2085 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2087 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2088 #ifdef RX_ENABLE_LOCKS
2089 rxi_DestroyConnectionNoLock(conn);
2090 #else /* RX_ENABLE_LOCKS */
2091 rxi_DestroyConnection(conn);
2092 #endif /* RX_ENABLE_LOCKS */
2096 #ifdef RX_ENABLE_LOCKS
2097 while (rx_connCleanup_list) {
2098 struct rx_connection *conn;
2099 conn = rx_connCleanup_list;
2100 rx_connCleanup_list = rx_connCleanup_list->next;
2101 MUTEX_EXIT(&rx_connHashTable_lock);
2102 rxi_CleanupConnection(conn);
2103 MUTEX_ENTER(&rx_connHashTable_lock);
2105 MUTEX_EXIT(&rx_connHashTable_lock);
2106 #endif /* RX_ENABLE_LOCKS */
2111 afs_winsockCleanup();
2119 /* if we wakeup packet waiter too often, can get in loop with two
2120 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2122 rxi_PacketsUnWait(void)
2124 if (!rx_waitingForPackets) {
2128 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2129 return; /* still over quota */
2132 rx_waitingForPackets = 0;
2133 #ifdef RX_ENABLE_LOCKS
2134 CV_BROADCAST(&rx_waitingForPackets_cv);
2136 osi_rxWakeup(&rx_waitingForPackets);
2142 /* ------------------Internal interfaces------------------------- */
2144 /* Return this process's service structure for the
2145 * specified socket and service */
2147 rxi_FindService(register osi_socket socket, register u_short serviceId)
2149 register struct rx_service **sp;
2150 for (sp = &rx_services[0]; *sp; sp++) {
2151 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2157 /* Allocate a call structure, for the indicated channel of the
2158 * supplied connection. The mode and state of the call must be set by
2159 * the caller. Returns the call with mutex locked. */
2161 rxi_NewCall(register struct rx_connection *conn, register int channel)
2163 register struct rx_call *call;
2164 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2165 register struct rx_call *cp; /* Call pointer temp */
2166 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2167 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2169 dpf(("rxi_NewCall(conn %x, channel %d)\n", conn, channel));
2171 /* Grab an existing call structure, or allocate a new one.
2172 * Existing call structures are assumed to have been left reset by
2174 MUTEX_ENTER(&rx_freeCallQueue_lock);
2176 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2178 * EXCEPT that the TQ might not yet be cleared out.
2179 * Skip over those with in-use TQs.
2182 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2183 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2189 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2190 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2191 call = queue_First(&rx_freeCallQueue, rx_call);
2192 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2194 rx_MutexDecrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2195 MUTEX_EXIT(&rx_freeCallQueue_lock);
2196 MUTEX_ENTER(&call->lock);
2197 CLEAR_CALL_QUEUE_LOCK(call);
2198 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2199 /* Now, if TQ wasn't cleared earlier, do it now. */
2200 if (call->flags & RX_CALL_TQ_CLEARME) {
2201 rxi_ClearTransmitQueue(call, 0);
2202 queue_Init(&call->tq);
2204 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2205 /* Bind the call to its connection structure */
2207 rxi_ResetCall(call, 1);
2209 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2211 MUTEX_EXIT(&rx_freeCallQueue_lock);
2212 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2213 MUTEX_ENTER(&call->lock);
2214 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2215 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2216 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2218 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2219 /* Initialize once-only items */
2220 queue_Init(&call->tq);
2221 queue_Init(&call->rq);
2222 queue_Init(&call->iovq);
2223 /* Bind the call to its connection structure (prereq for reset) */
2225 rxi_ResetCall(call, 1);
2227 call->channel = channel;
2228 call->callNumber = &conn->callNumber[channel];
2229 call->rwind = conn->rwind[channel];
2230 call->twind = conn->twind[channel];
2231 /* Note that the next expected call number is retained (in
2232 * conn->callNumber[i]), even if we reallocate the call structure
2234 conn->call[channel] = call;
2235 /* if the channel's never been used (== 0), we should start at 1, otherwise
2236 * the call number is valid from the last time this channel was used */
2237 if (*call->callNumber == 0)
2238 *call->callNumber = 1;
2243 /* A call has been inactive long enough that so we can throw away
2244 * state, including the call structure, which is placed on the call
2246 * Call is locked upon entry.
2247 * haveCTLock set if called from rxi_ReapConnections
2249 #ifdef RX_ENABLE_LOCKS
2251 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2252 #else /* RX_ENABLE_LOCKS */
2254 rxi_FreeCall(register struct rx_call *call)
2255 #endif /* RX_ENABLE_LOCKS */
2257 register int channel = call->channel;
2258 register struct rx_connection *conn = call->conn;
2261 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2262 (*call->callNumber)++;
2263 rxi_ResetCall(call, 0);
2264 call->conn->call[channel] = (struct rx_call *)0;
2266 MUTEX_ENTER(&rx_freeCallQueue_lock);
2267 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2268 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2269 /* A call may be free even though its transmit queue is still in use.
2270 * Since we search the call list from head to tail, put busy calls at
2271 * the head of the list, and idle calls at the tail.
2273 if (call->flags & RX_CALL_TQ_BUSY)
2274 queue_Prepend(&rx_freeCallQueue, call);
2276 queue_Append(&rx_freeCallQueue, call);
2277 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2278 queue_Append(&rx_freeCallQueue, call);
2279 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2280 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2281 MUTEX_EXIT(&rx_freeCallQueue_lock);
2283 /* Destroy the connection if it was previously slated for
2284 * destruction, i.e. the Rx client code previously called
2285 * rx_DestroyConnection (client connections), or
2286 * rxi_ReapConnections called the same routine (server
2287 * connections). Only do this, however, if there are no
2288 * outstanding calls. Note that for fine grain locking, there appears
2289 * to be a deadlock in that rxi_FreeCall has a call locked and
2290 * DestroyConnectionNoLock locks each call in the conn. But note a
2291 * few lines up where we have removed this call from the conn.
2292 * If someone else destroys a connection, they either have no
2293 * call lock held or are going through this section of code.
2295 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2296 MUTEX_ENTER(&conn->conn_data_lock);
2298 MUTEX_EXIT(&conn->conn_data_lock);
2299 #ifdef RX_ENABLE_LOCKS
2301 rxi_DestroyConnectionNoLock(conn);
2303 rxi_DestroyConnection(conn);
2304 #else /* RX_ENABLE_LOCKS */
2305 rxi_DestroyConnection(conn);
2306 #endif /* RX_ENABLE_LOCKS */
2310 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2312 rxi_Alloc(register size_t size)
2316 rx_MutexAdd1Increment2(rxi_Allocsize, (afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2317 p = (char *)osi_Alloc(size);
2320 osi_Panic("rxi_Alloc error");
2326 rxi_Free(void *addr, register size_t size)
2328 rx_MutexAdd1Decrement2(rxi_Allocsize, -(afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2329 osi_Free(addr, size);
2333 rxi_SetPeerMtu(register afs_uint32 host, register afs_uint32 port, int mtu)
2335 struct rx_peer **peer_ptr, **peer_end;
2338 MUTEX_ENTER(&rx_peerHashTable_lock);
2340 for (peer_ptr = &rx_peerHashTable[0], peer_end =
2341 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
2343 struct rx_peer *peer, *next;
2344 for (peer = *peer_ptr; peer; peer = next) {
2346 if (host == peer->host) {
2347 MUTEX_ENTER(&peer->peer_lock);
2348 peer->ifMTU=MIN(mtu, peer->ifMTU);
2349 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2350 MUTEX_EXIT(&peer->peer_lock);
2355 struct rx_peer *peer, *next;
2356 hashIndex = PEER_HASH(host, port);
2357 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2358 if ((peer->host == host) && (peer->port == port)) {
2359 MUTEX_ENTER(&peer->peer_lock);
2360 peer->ifMTU=MIN(mtu, peer->ifMTU);
2361 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2362 MUTEX_EXIT(&peer->peer_lock);
2366 MUTEX_EXIT(&rx_peerHashTable_lock);
2369 /* Find the peer process represented by the supplied (host,port)
2370 * combination. If there is no appropriate active peer structure, a
2371 * new one will be allocated and initialized
2372 * The origPeer, if set, is a pointer to a peer structure on which the
2373 * refcount will be be decremented. This is used to replace the peer
2374 * structure hanging off a connection structure */
2376 rxi_FindPeer(register afs_uint32 host, register u_short port,
2377 struct rx_peer *origPeer, int create)
2379 register struct rx_peer *pp;
2381 hashIndex = PEER_HASH(host, port);
2382 MUTEX_ENTER(&rx_peerHashTable_lock);
2383 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2384 if ((pp->host == host) && (pp->port == port))
2389 pp = rxi_AllocPeer(); /* This bzero's *pp */
2390 pp->host = host; /* set here or in InitPeerParams is zero */
2392 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2393 queue_Init(&pp->congestionQueue);
2394 queue_Init(&pp->rpcStats);
2395 pp->next = rx_peerHashTable[hashIndex];
2396 rx_peerHashTable[hashIndex] = pp;
2397 rxi_InitPeerParams(pp);
2398 rx_MutexIncrement(rx_stats.nPeerStructs, rx_stats_mutex);
2405 origPeer->refCount--;
2406 MUTEX_EXIT(&rx_peerHashTable_lock);
2411 /* Find the connection at (host, port) started at epoch, and with the
2412 * given connection id. Creates the server connection if necessary.
2413 * The type specifies whether a client connection or a server
2414 * connection is desired. In both cases, (host, port) specify the
2415 * peer's (host, pair) pair. Client connections are not made
2416 * automatically by this routine. The parameter socket gives the
2417 * socket descriptor on which the packet was received. This is used,
2418 * in the case of server connections, to check that *new* connections
2419 * come via a valid (port, serviceId). Finally, the securityIndex
2420 * parameter must match the existing index for the connection. If a
2421 * server connection is created, it will be created using the supplied
2422 * index, if the index is valid for this service */
2423 struct rx_connection *
2424 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2425 register u_short port, u_short serviceId, afs_uint32 cid,
2426 afs_uint32 epoch, int type, u_int securityIndex)
2428 int hashindex, flag, i;
2429 register struct rx_connection *conn;
2430 hashindex = CONN_HASH(host, port, cid, epoch, type);
2431 MUTEX_ENTER(&rx_connHashTable_lock);
2432 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2433 rx_connHashTable[hashindex],
2436 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2437 && (epoch == conn->epoch)) {
2438 register struct rx_peer *pp = conn->peer;
2439 if (securityIndex != conn->securityIndex) {
2440 /* this isn't supposed to happen, but someone could forge a packet
2441 * like this, and there seems to be some CM bug that makes this
2442 * happen from time to time -- in which case, the fileserver
2444 MUTEX_EXIT(&rx_connHashTable_lock);
2445 return (struct rx_connection *)0;
2447 if (pp->host == host && pp->port == port)
2449 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2451 /* So what happens when it's a callback connection? */
2452 if ( /*type == RX_CLIENT_CONNECTION && */
2453 (conn->epoch & 0x80000000))
2457 /* the connection rxLastConn that was used the last time is not the
2458 ** one we are looking for now. Hence, start searching in the hash */
2460 conn = rx_connHashTable[hashindex];
2465 struct rx_service *service;
2466 if (type == RX_CLIENT_CONNECTION) {
2467 MUTEX_EXIT(&rx_connHashTable_lock);
2468 return (struct rx_connection *)0;
2470 service = rxi_FindService(socket, serviceId);
2471 if (!service || (securityIndex >= service->nSecurityObjects)
2472 || (service->securityObjects[securityIndex] == 0)) {
2473 MUTEX_EXIT(&rx_connHashTable_lock);
2474 return (struct rx_connection *)0;
2476 conn = rxi_AllocConnection(); /* This bzero's the connection */
2477 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2478 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2479 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2480 conn->next = rx_connHashTable[hashindex];
2481 rx_connHashTable[hashindex] = conn;
2482 conn->peer = rxi_FindPeer(host, port, 0, 1);
2483 conn->type = RX_SERVER_CONNECTION;
2484 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2485 conn->epoch = epoch;
2486 conn->cid = cid & RX_CIDMASK;
2487 /* conn->serial = conn->lastSerial = 0; */
2488 /* conn->timeout = 0; */
2489 conn->ackRate = RX_FAST_ACK_RATE;
2490 conn->service = service;
2491 conn->serviceId = serviceId;
2492 conn->securityIndex = securityIndex;
2493 conn->securityObject = service->securityObjects[securityIndex];
2494 conn->nSpecific = 0;
2495 conn->specific = NULL;
2496 rx_SetConnDeadTime(conn, service->connDeadTime);
2497 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2498 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2499 for (i = 0; i < RX_MAXCALLS; i++) {
2500 conn->twind[i] = rx_initSendWindow;
2501 conn->rwind[i] = rx_initReceiveWindow;
2503 /* Notify security object of the new connection */
2504 RXS_NewConnection(conn->securityObject, conn);
2505 /* XXXX Connection timeout? */
2506 if (service->newConnProc)
2507 (*service->newConnProc) (conn);
2508 rx_MutexIncrement(rx_stats.nServerConns, rx_stats_mutex);
2511 MUTEX_ENTER(&conn->conn_data_lock);
2513 MUTEX_EXIT(&conn->conn_data_lock);
2515 rxLastConn = conn; /* store this connection as the last conn used */
2516 MUTEX_EXIT(&rx_connHashTable_lock);
2520 /* There are two packet tracing routines available for testing and monitoring
2521 * Rx. One is called just after every packet is received and the other is
2522 * called just before every packet is sent. Received packets, have had their
2523 * headers decoded, and packets to be sent have not yet had their headers
2524 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2525 * containing the network address. Both can be modified. The return value, if
2526 * non-zero, indicates that the packet should be dropped. */
2528 int (*rx_justReceived) () = 0;
2529 int (*rx_almostSent) () = 0;
2531 /* A packet has been received off the interface. Np is the packet, socket is
2532 * the socket number it was received from (useful in determining which service
2533 * this packet corresponds to), and (host, port) reflect the host,port of the
2534 * sender. This call returns the packet to the caller if it is finished with
2535 * it, rather than de-allocating it, just as a small performance hack */
2538 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2539 afs_uint32 host, u_short port, int *tnop,
2540 struct rx_call **newcallp)
2542 register struct rx_call *call;
2543 register struct rx_connection *conn;
2545 afs_uint32 currentCallNumber;
2551 struct rx_packet *tnp;
2554 /* We don't print out the packet until now because (1) the time may not be
2555 * accurate enough until now in the lwp implementation (rx_Listener only gets
2556 * the time after the packet is read) and (2) from a protocol point of view,
2557 * this is the first time the packet has been seen */
2558 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2559 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2560 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2561 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2562 np->header.epoch, np->header.cid, np->header.callNumber,
2563 np->header.seq, np->header.flags, np));
2566 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2567 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2570 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2571 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2574 /* If an input tracer function is defined, call it with the packet and
2575 * network address. Note this function may modify its arguments. */
2576 if (rx_justReceived) {
2577 struct sockaddr_in addr;
2579 addr.sin_family = AF_INET;
2580 addr.sin_port = port;
2581 addr.sin_addr.s_addr = host;
2582 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2583 addr.sin_len = sizeof(addr);
2584 #endif /* AFS_OSF_ENV */
2585 drop = (*rx_justReceived) (np, &addr);
2586 /* drop packet if return value is non-zero */
2589 port = addr.sin_port; /* in case fcn changed addr */
2590 host = addr.sin_addr.s_addr;
2594 /* If packet was not sent by the client, then *we* must be the client */
2595 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2596 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2598 /* Find the connection (or fabricate one, if we're the server & if
2599 * necessary) associated with this packet */
2601 rxi_FindConnection(socket, host, port, np->header.serviceId,
2602 np->header.cid, np->header.epoch, type,
2603 np->header.securityIndex);
2606 /* If no connection found or fabricated, just ignore the packet.
2607 * (An argument could be made for sending an abort packet for
2612 MUTEX_ENTER(&conn->conn_data_lock);
2613 if (conn->maxSerial < np->header.serial)
2614 conn->maxSerial = np->header.serial;
2615 MUTEX_EXIT(&conn->conn_data_lock);
2617 /* If the connection is in an error state, send an abort packet and ignore
2618 * the incoming packet */
2620 /* Don't respond to an abort packet--we don't want loops! */
2621 MUTEX_ENTER(&conn->conn_data_lock);
2622 if (np->header.type != RX_PACKET_TYPE_ABORT)
2623 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2625 MUTEX_EXIT(&conn->conn_data_lock);
2629 /* Check for connection-only requests (i.e. not call specific). */
2630 if (np->header.callNumber == 0) {
2631 switch (np->header.type) {
2632 case RX_PACKET_TYPE_ABORT: {
2633 /* What if the supplied error is zero? */
2634 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2635 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2636 rxi_ConnectionError(conn, errcode);
2637 MUTEX_ENTER(&conn->conn_data_lock);
2639 MUTEX_EXIT(&conn->conn_data_lock);
2642 case RX_PACKET_TYPE_CHALLENGE:
2643 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2644 MUTEX_ENTER(&conn->conn_data_lock);
2646 MUTEX_EXIT(&conn->conn_data_lock);
2648 case RX_PACKET_TYPE_RESPONSE:
2649 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2650 MUTEX_ENTER(&conn->conn_data_lock);
2652 MUTEX_EXIT(&conn->conn_data_lock);
2654 case RX_PACKET_TYPE_PARAMS:
2655 case RX_PACKET_TYPE_PARAMS + 1:
2656 case RX_PACKET_TYPE_PARAMS + 2:
2657 /* ignore these packet types for now */
2658 MUTEX_ENTER(&conn->conn_data_lock);
2660 MUTEX_EXIT(&conn->conn_data_lock);
2665 /* Should not reach here, unless the peer is broken: send an
2667 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2668 MUTEX_ENTER(&conn->conn_data_lock);
2669 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2671 MUTEX_EXIT(&conn->conn_data_lock);
2676 channel = np->header.cid & RX_CHANNELMASK;
2677 call = conn->call[channel];
2678 #ifdef RX_ENABLE_LOCKS
2680 MUTEX_ENTER(&call->lock);
2681 /* Test to see if call struct is still attached to conn. */
2682 if (call != conn->call[channel]) {
2684 MUTEX_EXIT(&call->lock);
2685 if (type == RX_SERVER_CONNECTION) {
2686 call = conn->call[channel];
2687 /* If we started with no call attached and there is one now,
2688 * another thread is also running this routine and has gotten
2689 * the connection channel. We should drop this packet in the tests
2690 * below. If there was a call on this connection and it's now
2691 * gone, then we'll be making a new call below.
2692 * If there was previously a call and it's now different then
2693 * the old call was freed and another thread running this routine
2694 * has created a call on this channel. One of these two threads
2695 * has a packet for the old call and the code below handles those
2699 MUTEX_ENTER(&call->lock);
2701 /* This packet can't be for this call. If the new call address is
2702 * 0 then no call is running on this channel. If there is a call
2703 * then, since this is a client connection we're getting data for
2704 * it must be for the previous call.
2706 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2707 MUTEX_ENTER(&conn->conn_data_lock);
2709 MUTEX_EXIT(&conn->conn_data_lock);
2714 currentCallNumber = conn->callNumber[channel];
2716 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2717 if (np->header.callNumber < currentCallNumber) {
2718 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2719 #ifdef RX_ENABLE_LOCKS
2721 MUTEX_EXIT(&call->lock);
2723 MUTEX_ENTER(&conn->conn_data_lock);
2725 MUTEX_EXIT(&conn->conn_data_lock);
2729 MUTEX_ENTER(&conn->conn_call_lock);
2730 call = rxi_NewCall(conn, channel);
2731 MUTEX_EXIT(&conn->conn_call_lock);
2732 *call->callNumber = np->header.callNumber;
2733 if (np->header.callNumber == 0)
2734 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %lx resend %d.%0.3d len %d", np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port), np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq, np->header.flags, (unsigned long)np, np->retryTime.sec, np->retryTime.usec / 1000, np->length));
2736 call->state = RX_STATE_PRECALL;
2737 clock_GetTime(&call->queueTime);
2738 hzero(call->bytesSent);
2739 hzero(call->bytesRcvd);
2741 * If the number of queued calls exceeds the overload
2742 * threshold then abort this call.
2744 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2745 struct rx_packet *tp;
2747 rxi_CallError(call, rx_BusyError);
2748 tp = rxi_SendCallAbort(call, np, 1, 0);
2749 MUTEX_EXIT(&call->lock);
2750 MUTEX_ENTER(&conn->conn_data_lock);
2752 MUTEX_EXIT(&conn->conn_data_lock);
2753 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2756 rxi_KeepAliveOn(call);
2757 } else if (np->header.callNumber != currentCallNumber) {
2758 /* Wait until the transmit queue is idle before deciding
2759 * whether to reset the current call. Chances are that the
2760 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2763 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2764 while ((call->state == RX_STATE_ACTIVE)
2765 && (call->flags & RX_CALL_TQ_BUSY)) {
2766 call->flags |= RX_CALL_TQ_WAIT;
2768 #ifdef RX_ENABLE_LOCKS
2769 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2770 CV_WAIT(&call->cv_tq, &call->lock);
2771 #else /* RX_ENABLE_LOCKS */
2772 osi_rxSleep(&call->tq);
2773 #endif /* RX_ENABLE_LOCKS */
2775 if (call->tqWaiters == 0)
2776 call->flags &= ~RX_CALL_TQ_WAIT;
2778 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2779 /* If the new call cannot be taken right now send a busy and set
2780 * the error condition in this call, so that it terminates as
2781 * quickly as possible */
2782 if (call->state == RX_STATE_ACTIVE) {
2783 struct rx_packet *tp;
2785 rxi_CallError(call, RX_CALL_DEAD);
2786 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2788 MUTEX_EXIT(&call->lock);
2789 MUTEX_ENTER(&conn->conn_data_lock);
2791 MUTEX_EXIT(&conn->conn_data_lock);
2794 rxi_ResetCall(call, 0);
2795 *call->callNumber = np->header.callNumber;
2796 if (np->header.callNumber == 0)
2797 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %lx resend %d.%0.3d len %d", np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port), np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq, np->header.flags, (unsigned long)np, np->retryTime.sec, np->retryTime.usec / 1000, np->length));
2799 call->state = RX_STATE_PRECALL;
2800 clock_GetTime(&call->queueTime);
2801 hzero(call->bytesSent);
2802 hzero(call->bytesRcvd);
2804 * If the number of queued calls exceeds the overload
2805 * threshold then abort this call.
2807 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2808 struct rx_packet *tp;
2810 rxi_CallError(call, rx_BusyError);
2811 tp = rxi_SendCallAbort(call, np, 1, 0);
2812 MUTEX_EXIT(&call->lock);
2813 MUTEX_ENTER(&conn->conn_data_lock);
2815 MUTEX_EXIT(&conn->conn_data_lock);
2816 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2819 rxi_KeepAliveOn(call);
2821 /* Continuing call; do nothing here. */
2823 } else { /* we're the client */
2824 /* Ignore all incoming acknowledgements for calls in DALLY state */
2825 if (call && (call->state == RX_STATE_DALLY)
2826 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2827 rx_MutexIncrement(rx_stats.ignorePacketDally, rx_stats_mutex);
2828 #ifdef RX_ENABLE_LOCKS
2830 MUTEX_EXIT(&call->lock);
2833 MUTEX_ENTER(&conn->conn_data_lock);
2835 MUTEX_EXIT(&conn->conn_data_lock);
2839 /* Ignore anything that's not relevant to the current call. If there
2840 * isn't a current call, then no packet is relevant. */
2841 if (!call || (np->header.callNumber != currentCallNumber)) {
2842 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2843 #ifdef RX_ENABLE_LOCKS
2845 MUTEX_EXIT(&call->lock);
2848 MUTEX_ENTER(&conn->conn_data_lock);
2850 MUTEX_EXIT(&conn->conn_data_lock);
2853 /* If the service security object index stamped in the packet does not
2854 * match the connection's security index, ignore the packet */
2855 if (np->header.securityIndex != conn->securityIndex) {
2856 #ifdef RX_ENABLE_LOCKS
2857 MUTEX_EXIT(&call->lock);
2859 MUTEX_ENTER(&conn->conn_data_lock);
2861 MUTEX_EXIT(&conn->conn_data_lock);
2865 /* If we're receiving the response, then all transmit packets are
2866 * implicitly acknowledged. Get rid of them. */
2867 if (np->header.type == RX_PACKET_TYPE_DATA) {
2868 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2869 /* XXX Hack. Because we must release the global rx lock when
2870 * sending packets (osi_NetSend) we drop all acks while we're
2871 * traversing the tq in rxi_Start sending packets out because
2872 * packets may move to the freePacketQueue as result of being here!
2873 * So we drop these packets until we're safely out of the
2874 * traversing. Really ugly!
2875 * For fine grain RX locking, we set the acked field in the
2876 * packets and let rxi_Start remove them from the transmit queue.
2878 if (call->flags & RX_CALL_TQ_BUSY) {
2879 #ifdef RX_ENABLE_LOCKS
2880 rxi_SetAcksInTransmitQueue(call);
2883 return np; /* xmitting; drop packet */
2886 rxi_ClearTransmitQueue(call, 0);
2888 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2889 rxi_ClearTransmitQueue(call, 0);
2890 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2892 if (np->header.type == RX_PACKET_TYPE_ACK) {
2893 /* now check to see if this is an ack packet acknowledging that the
2894 * server actually *lost* some hard-acked data. If this happens we
2895 * ignore this packet, as it may indicate that the server restarted in
2896 * the middle of a call. It is also possible that this is an old ack
2897 * packet. We don't abort the connection in this case, because this
2898 * *might* just be an old ack packet. The right way to detect a server
2899 * restart in the midst of a call is to notice that the server epoch
2901 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2902 * XXX unacknowledged. I think that this is off-by-one, but
2903 * XXX I don't dare change it just yet, since it will
2904 * XXX interact badly with the server-restart detection
2905 * XXX code in receiveackpacket. */
2906 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2907 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2908 MUTEX_EXIT(&call->lock);
2909 MUTEX_ENTER(&conn->conn_data_lock);
2911 MUTEX_EXIT(&conn->conn_data_lock);
2915 } /* else not a data packet */
2918 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2919 /* Set remote user defined status from packet */
2920 call->remoteStatus = np->header.userStatus;
2922 /* Note the gap between the expected next packet and the actual
2923 * packet that arrived, when the new packet has a smaller serial number
2924 * than expected. Rioses frequently reorder packets all by themselves,
2925 * so this will be quite important with very large window sizes.
2926 * Skew is checked against 0 here to avoid any dependence on the type of
2927 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2929 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2930 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2931 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2933 MUTEX_ENTER(&conn->conn_data_lock);
2934 skew = conn->lastSerial - np->header.serial;
2935 conn->lastSerial = np->header.serial;
2936 MUTEX_EXIT(&conn->conn_data_lock);
2938 register struct rx_peer *peer;
2940 if (skew > peer->inPacketSkew) {
2941 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2943 peer->inPacketSkew = skew;
2947 /* Now do packet type-specific processing */
2948 switch (np->header.type) {
2949 case RX_PACKET_TYPE_DATA:
2950 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2953 case RX_PACKET_TYPE_ACK:
2954 /* Respond immediately to ack packets requesting acknowledgement
2956 if (np->header.flags & RX_REQUEST_ACK) {
2958 (void)rxi_SendCallAbort(call, 0, 1, 0);
2960 (void)rxi_SendAck(call, 0, np->header.serial,
2961 RX_ACK_PING_RESPONSE, 1);
2963 np = rxi_ReceiveAckPacket(call, np, 1);
2965 case RX_PACKET_TYPE_ABORT: {
2966 /* An abort packet: reset the call, passing the error up to the user. */
2967 /* What if error is zero? */
2968 /* What if the error is -1? the application will treat it as a timeout. */
2969 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
2970 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
2971 rxi_CallError(call, errdata);
2972 MUTEX_EXIT(&call->lock);
2973 MUTEX_ENTER(&conn->conn_data_lock);
2975 MUTEX_EXIT(&conn->conn_data_lock);
2976 return np; /* xmitting; drop packet */
2978 case RX_PACKET_TYPE_BUSY:
2981 case RX_PACKET_TYPE_ACKALL:
2982 /* All packets acknowledged, so we can drop all packets previously
2983 * readied for sending */
2984 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2985 /* XXX Hack. We because we can't release the global rx lock when
2986 * sending packets (osi_NetSend) we drop all ack pkts while we're
2987 * traversing the tq in rxi_Start sending packets out because
2988 * packets may move to the freePacketQueue as result of being
2989 * here! So we drop these packets until we're safely out of the
2990 * traversing. Really ugly!
2991 * For fine grain RX locking, we set the acked field in the packets
2992 * and let rxi_Start remove the packets from the transmit queue.
2994 if (call->flags & RX_CALL_TQ_BUSY) {
2995 #ifdef RX_ENABLE_LOCKS
2996 rxi_SetAcksInTransmitQueue(call);
2998 #else /* RX_ENABLE_LOCKS */
2999 MUTEX_EXIT(&call->lock);
3000 MUTEX_ENTER(&conn->conn_data_lock);
3002 MUTEX_EXIT(&conn->conn_data_lock);
3003 return np; /* xmitting; drop packet */
3004 #endif /* RX_ENABLE_LOCKS */
3006 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3007 rxi_ClearTransmitQueue(call, 0);
3008 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
3011 /* Should not reach here, unless the peer is broken: send an abort
3013 rxi_CallError(call, RX_PROTOCOL_ERROR);
3014 np = rxi_SendCallAbort(call, np, 1, 0);
3017 /* Note when this last legitimate packet was received, for keep-alive
3018 * processing. Note, we delay getting the time until now in the hope that
3019 * the packet will be delivered to the user before any get time is required
3020 * (if not, then the time won't actually be re-evaluated here). */
3021 call->lastReceiveTime = clock_Sec();
3022 MUTEX_EXIT(&call->lock);
3023 MUTEX_ENTER(&conn->conn_data_lock);
3025 MUTEX_EXIT(&conn->conn_data_lock);
3029 /* return true if this is an "interesting" connection from the point of view
3030 of someone trying to debug the system */
3032 rxi_IsConnInteresting(struct rx_connection *aconn)
3035 register struct rx_call *tcall;
3037 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3039 for (i = 0; i < RX_MAXCALLS; i++) {
3040 tcall = aconn->call[i];
3042 if ((tcall->state == RX_STATE_PRECALL)
3043 || (tcall->state == RX_STATE_ACTIVE))
3045 if ((tcall->mode == RX_MODE_SENDING)
3046 || (tcall->mode == RX_MODE_RECEIVING))
3054 /* if this is one of the last few packets AND it wouldn't be used by the
3055 receiving call to immediately satisfy a read request, then drop it on
3056 the floor, since accepting it might prevent a lock-holding thread from
3057 making progress in its reading. If a call has been cleared while in
3058 the precall state then ignore all subsequent packets until the call
3059 is assigned to a thread. */
3062 TooLow(struct rx_packet *ap, struct rx_call *acall)
3065 MUTEX_ENTER(&rx_stats_mutex);
3066 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3067 && (acall->state == RX_STATE_PRECALL))
3068 || ((rx_nFreePackets < rxi_dataQuota + 2)
3069 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3070 && (acall->flags & RX_CALL_READER_WAIT)))) {
3073 MUTEX_EXIT(&rx_stats_mutex);
3079 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
3080 struct rx_call *acall)
3082 struct rx_call *call = acall;
3083 struct clock when, now;
3086 MUTEX_ENTER(&conn->conn_data_lock);
3087 conn->checkReachEvent = NULL;
3088 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3091 MUTEX_EXIT(&conn->conn_data_lock);
3095 MUTEX_ENTER(&conn->conn_call_lock);
3096 MUTEX_ENTER(&conn->conn_data_lock);
3097 for (i = 0; i < RX_MAXCALLS; i++) {
3098 struct rx_call *tc = conn->call[i];
3099 if (tc && tc->state == RX_STATE_PRECALL) {
3105 /* Indicate that rxi_CheckReachEvent is no longer running by
3106 * clearing the flag. Must be atomic under conn_data_lock to
3107 * avoid a new call slipping by: rxi_CheckConnReach holds
3108 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3110 conn->flags &= ~RX_CONN_ATTACHWAIT;
3111 MUTEX_EXIT(&conn->conn_data_lock);
3112 MUTEX_EXIT(&conn->conn_call_lock);
3117 MUTEX_ENTER(&call->lock);
3118 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3120 MUTEX_EXIT(&call->lock);
3122 clock_GetTime(&now);
3124 when.sec += RX_CHECKREACH_TIMEOUT;
3125 MUTEX_ENTER(&conn->conn_data_lock);
3126 if (!conn->checkReachEvent) {
3128 conn->checkReachEvent =
3129 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3132 MUTEX_EXIT(&conn->conn_data_lock);
3138 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3140 struct rx_service *service = conn->service;
3141 struct rx_peer *peer = conn->peer;
3142 afs_uint32 now, lastReach;
3144 if (service->checkReach == 0)
3148 MUTEX_ENTER(&peer->peer_lock);
3149 lastReach = peer->lastReachTime;
3150 MUTEX_EXIT(&peer->peer_lock);
3151 if (now - lastReach < RX_CHECKREACH_TTL)
3154 MUTEX_ENTER(&conn->conn_data_lock);
3155 if (conn->flags & RX_CONN_ATTACHWAIT) {
3156 MUTEX_EXIT(&conn->conn_data_lock);
3159 conn->flags |= RX_CONN_ATTACHWAIT;
3160 MUTEX_EXIT(&conn->conn_data_lock);
3161 if (!conn->checkReachEvent)
3162 rxi_CheckReachEvent(NULL, conn, call);
3167 /* try to attach call, if authentication is complete */
3169 TryAttach(register struct rx_call *acall, register osi_socket socket,
3170 register int *tnop, register struct rx_call **newcallp,
3173 struct rx_connection *conn = acall->conn;
3175 if (conn->type == RX_SERVER_CONNECTION
3176 && acall->state == RX_STATE_PRECALL) {
3177 /* Don't attach until we have any req'd. authentication. */
3178 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3179 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3180 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3181 /* Note: this does not necessarily succeed; there
3182 * may not any proc available
3185 rxi_ChallengeOn(acall->conn);
3190 /* A data packet has been received off the interface. This packet is
3191 * appropriate to the call (the call is in the right state, etc.). This
3192 * routine can return a packet to the caller, for re-use */
3195 rxi_ReceiveDataPacket(register struct rx_call *call,
3196 register struct rx_packet *np, int istack,
3197 osi_socket socket, afs_uint32 host, u_short port,
3198 int *tnop, struct rx_call **newcallp)
3200 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3204 afs_uint32 seq, serial, flags;
3206 struct rx_packet *tnp;
3207 struct clock when, now;
3208 rx_MutexIncrement(rx_stats.dataPacketsRead, rx_stats_mutex);
3211 /* If there are no packet buffers, drop this new packet, unless we can find
3212 * packet buffers from inactive calls */
3214 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3215 MUTEX_ENTER(&rx_freePktQ_lock);
3216 rxi_NeedMorePackets = TRUE;
3217 MUTEX_EXIT(&rx_freePktQ_lock);
3218 rx_MutexIncrement(rx_stats.noPacketBuffersOnRead, rx_stats_mutex);
3219 call->rprev = np->header.serial;
3220 rxi_calltrace(RX_TRACE_DROP, call);
3221 dpf(("packet %x dropped on receipt - quota problems", np));
3223 rxi_ClearReceiveQueue(call);
3224 clock_GetTime(&now);
3226 clock_Add(&when, &rx_softAckDelay);
3227 if (!call->delayedAckEvent
3228 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3229 rxevent_Cancel(call->delayedAckEvent, call,
3230 RX_CALL_REFCOUNT_DELAY);
3231 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3232 call->delayedAckEvent =
3233 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3235 /* we've damaged this call already, might as well do it in. */
3241 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3242 * packet is one of several packets transmitted as a single
3243 * datagram. Do not send any soft or hard acks until all packets
3244 * in a jumbogram have been processed. Send negative acks right away.
3246 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3247 /* tnp is non-null when there are more packets in the
3248 * current jumbo gram */
3255 seq = np->header.seq;
3256 serial = np->header.serial;
3257 flags = np->header.flags;
3259 /* If the call is in an error state, send an abort message */
3261 return rxi_SendCallAbort(call, np, istack, 0);
3263 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3264 * AFS 3.5 jumbogram. */
3265 if (flags & RX_JUMBO_PACKET) {
3266 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3271 if (np->header.spare != 0) {
3272 MUTEX_ENTER(&call->conn->conn_data_lock);
3273 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3274 MUTEX_EXIT(&call->conn->conn_data_lock);
3277 /* The usual case is that this is the expected next packet */
3278 if (seq == call->rnext) {
3280 /* Check to make sure it is not a duplicate of one already queued */
3281 if (queue_IsNotEmpty(&call->rq)
3282 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3283 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3284 dpf(("packet %x dropped on receipt - duplicate", np));
3285 rxevent_Cancel(call->delayedAckEvent, call,
3286 RX_CALL_REFCOUNT_DELAY);
3287 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3293 /* It's the next packet. Stick it on the receive queue
3294 * for this call. Set newPackets to make sure we wake
3295 * the reader once all packets have been processed */
3296 queue_Prepend(&call->rq, np);
3298 np = NULL; /* We can't use this anymore */
3301 /* If an ack is requested then set a flag to make sure we
3302 * send an acknowledgement for this packet */
3303 if (flags & RX_REQUEST_ACK) {
3304 ackNeeded = RX_ACK_REQUESTED;
3307 /* Keep track of whether we have received the last packet */
3308 if (flags & RX_LAST_PACKET) {
3309 call->flags |= RX_CALL_HAVE_LAST;
3313 /* Check whether we have all of the packets for this call */
3314 if (call->flags & RX_CALL_HAVE_LAST) {
3315 afs_uint32 tseq; /* temporary sequence number */
3316 struct rx_packet *tp; /* Temporary packet pointer */
3317 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3319 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3320 if (tseq != tp->header.seq)
3322 if (tp->header.flags & RX_LAST_PACKET) {
3323 call->flags |= RX_CALL_RECEIVE_DONE;
3330 /* Provide asynchronous notification for those who want it
3331 * (e.g. multi rx) */
3332 if (call->arrivalProc) {
3333 (*call->arrivalProc) (call, call->arrivalProcHandle,
3334 call->arrivalProcArg);
3335 call->arrivalProc = (void (*)())0;
3338 /* Update last packet received */
3341 /* If there is no server process serving this call, grab
3342 * one, if available. We only need to do this once. If a
3343 * server thread is available, this thread becomes a server
3344 * thread and the server thread becomes a listener thread. */
3346 TryAttach(call, socket, tnop, newcallp, 0);
3349 /* This is not the expected next packet. */
3351 /* Determine whether this is a new or old packet, and if it's
3352 * a new one, whether it fits into the current receive window.
3353 * Also figure out whether the packet was delivered in sequence.
3354 * We use the prev variable to determine whether the new packet
3355 * is the successor of its immediate predecessor in the
3356 * receive queue, and the missing flag to determine whether
3357 * any of this packets predecessors are missing. */
3359 afs_uint32 prev; /* "Previous packet" sequence number */
3360 struct rx_packet *tp; /* Temporary packet pointer */
3361 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3362 int missing; /* Are any predecessors missing? */
3364 /* If the new packet's sequence number has been sent to the
3365 * application already, then this is a duplicate */
3366 if (seq < call->rnext) {
3367 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3368 rxevent_Cancel(call->delayedAckEvent, call,
3369 RX_CALL_REFCOUNT_DELAY);
3370 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3376 /* If the sequence number is greater than what can be
3377 * accomodated by the current window, then send a negative
3378 * acknowledge and drop the packet */
3379 if ((call->rnext + call->rwind) <= seq) {
3380 rxevent_Cancel(call->delayedAckEvent, call,
3381 RX_CALL_REFCOUNT_DELAY);
3382 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3389 /* Look for the packet in the queue of old received packets */
3390 for (prev = call->rnext - 1, missing =
3391 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3392 /*Check for duplicate packet */
3393 if (seq == tp->header.seq) {
3394 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3395 rxevent_Cancel(call->delayedAckEvent, call,
3396 RX_CALL_REFCOUNT_DELAY);
3397 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3403 /* If we find a higher sequence packet, break out and
3404 * insert the new packet here. */
3405 if (seq < tp->header.seq)
3407 /* Check for missing packet */
3408 if (tp->header.seq != prev + 1) {
3412 prev = tp->header.seq;
3415 /* Keep track of whether we have received the last packet. */
3416 if (flags & RX_LAST_PACKET) {
3417 call->flags |= RX_CALL_HAVE_LAST;
3420 /* It's within the window: add it to the the receive queue.
3421 * tp is left by the previous loop either pointing at the
3422 * packet before which to insert the new packet, or at the
3423 * queue head if the queue is empty or the packet should be
3425 queue_InsertBefore(tp, np);
3429 /* Check whether we have all of the packets for this call */
3430 if ((call->flags & RX_CALL_HAVE_LAST)
3431 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3432 afs_uint32 tseq; /* temporary sequence number */
3435 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3436 if (tseq != tp->header.seq)
3438 if (tp->header.flags & RX_LAST_PACKET) {
3439 call->flags |= RX_CALL_RECEIVE_DONE;
3446 /* We need to send an ack of the packet is out of sequence,
3447 * or if an ack was requested by the peer. */
3448 if (seq != prev + 1 || missing) {
3449 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3450 } else if (flags & RX_REQUEST_ACK) {
3451 ackNeeded = RX_ACK_REQUESTED;
3454 /* Acknowledge the last packet for each call */
3455 if (flags & RX_LAST_PACKET) {
3466 * If the receiver is waiting for an iovec, fill the iovec
3467 * using the data from the receive queue */
3468 if (call->flags & RX_CALL_IOVEC_WAIT) {
3469 didHardAck = rxi_FillReadVec(call, serial);
3470 /* the call may have been aborted */
3479 /* Wakeup the reader if any */
3480 if ((call->flags & RX_CALL_READER_WAIT)
3481 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3482 || (call->iovNext >= call->iovMax)
3483 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3484 call->flags &= ~RX_CALL_READER_WAIT;
3485 #ifdef RX_ENABLE_LOCKS
3486 CV_BROADCAST(&call->cv_rq);
3488 osi_rxWakeup(&call->rq);
3494 * Send an ack when requested by the peer, or once every
3495 * rxi_SoftAckRate packets until the last packet has been
3496 * received. Always send a soft ack for the last packet in
3497 * the server's reply. */
3499 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3500 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3501 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3502 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3503 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3504 } else if (call->nSoftAcks) {
3505 clock_GetTime(&now);
3507 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3508 clock_Add(&when, &rx_lastAckDelay);
3510 clock_Add(&when, &rx_softAckDelay);
3512 if (!call->delayedAckEvent
3513 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3514 rxevent_Cancel(call->delayedAckEvent, call,
3515 RX_CALL_REFCOUNT_DELAY);
3516 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3517 call->delayedAckEvent =
3518 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3520 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3521 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3528 static void rxi_ComputeRate();
3532 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3534 struct rx_peer *peer = conn->peer;
3536 MUTEX_ENTER(&peer->peer_lock);
3537 peer->lastReachTime = clock_Sec();
3538 MUTEX_EXIT(&peer->peer_lock);
3540 MUTEX_ENTER(&conn->conn_data_lock);
3541 if (conn->flags & RX_CONN_ATTACHWAIT) {
3544 conn->flags &= ~RX_CONN_ATTACHWAIT;
3545 MUTEX_EXIT(&conn->conn_data_lock);
3547 for (i = 0; i < RX_MAXCALLS; i++) {
3548 struct rx_call *call = conn->call[i];
3551 MUTEX_ENTER(&call->lock);
3552 /* tnop can be null if newcallp is null */
3553 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3555 MUTEX_EXIT(&call->lock);
3559 MUTEX_EXIT(&conn->conn_data_lock);
3563 rx_ack_reason(int reason)
3566 case RX_ACK_REQUESTED:
3568 case RX_ACK_DUPLICATE:
3570 case RX_ACK_OUT_OF_SEQUENCE:
3572 case RX_ACK_EXCEEDS_WINDOW:
3574 case RX_ACK_NOSPACE:
3578 case RX_ACK_PING_RESPONSE:
3590 /* rxi_ComputePeerNetStats
3592 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3593 * estimates (like RTT and throughput) based on ack packets. Caller
3594 * must ensure that the packet in question is the right one (i.e.
3595 * serial number matches).
3598 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3599 struct rx_ackPacket *ap, struct rx_packet *np)
3601 struct rx_peer *peer = call->conn->peer;
3603 /* Use RTT if not delayed by client. */
3604 if (ap->reason != RX_ACK_DELAY)
3605 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3607 rxi_ComputeRate(peer, call, p, np, ap->reason);
3611 /* The real smarts of the whole thing. */
3613 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3616 struct rx_ackPacket *ap;
3618 register struct rx_packet *tp;
3619 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3620 register struct rx_connection *conn = call->conn;
3621 struct rx_peer *peer = conn->peer;
3624 /* because there are CM's that are bogus, sending weird values for this. */
3625 afs_uint32 skew = 0;
3630 int newAckCount = 0;
3631 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3632 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3634 rx_MutexIncrement(rx_stats.ackPacketsRead, rx_stats_mutex);
3635 ap = (struct rx_ackPacket *)rx_DataOf(np);
3636 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3638 return np; /* truncated ack packet */
3640 /* depends on ack packet struct */
3641 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3642 first = ntohl(ap->firstPacket);
3643 serial = ntohl(ap->serial);
3644 /* temporarily disabled -- needs to degrade over time
3645 * skew = ntohs(ap->maxSkew); */
3647 /* Ignore ack packets received out of order */
3648 if (first < call->tfirst) {
3652 if (np->header.flags & RX_SLOW_START_OK) {
3653 call->flags |= RX_CALL_SLOW_START_OK;
3656 if (ap->reason == RX_ACK_PING_RESPONSE)
3657 rxi_UpdatePeerReach(conn, call);
3661 if (rxdebug_active) {
3665 len = _snprintf(msg, sizeof(msg),
3666 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3667 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3668 ntohl(ap->serial), ntohl(ap->previousPacket),
3669 (unsigned int)np->header.seq, (unsigned int)skew,
3670 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3674 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3675 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3679 OutputDebugString(msg);
3681 #else /* AFS_NT40_ENV */
3684 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3685 ap->reason, ntohl(ap->previousPacket),
3686 (unsigned int)np->header.seq, (unsigned int)serial,
3687 (unsigned int)skew, ntohl(ap->firstPacket));
3690 for (offset = 0; offset < nAcks; offset++)
3691 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3696 #endif /* AFS_NT40_ENV */
3699 /* Update the outgoing packet skew value to the latest value of
3700 * the peer's incoming packet skew value. The ack packet, of
3701 * course, could arrive out of order, but that won't affect things
3703 MUTEX_ENTER(&peer->peer_lock);
3704 peer->outPacketSkew = skew;
3706 /* Check for packets that no longer need to be transmitted, and
3707 * discard them. This only applies to packets positively
3708 * acknowledged as having been sent to the peer's upper level.
3709 * All other packets must be retained. So only packets with
3710 * sequence numbers < ap->firstPacket are candidates. */
3711 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3712 if (tp->header.seq >= first)
3714 call->tfirst = tp->header.seq + 1;
3716 && (tp->header.serial == serial || tp->firstSerial == serial))
3717 rxi_ComputePeerNetStats(call, tp, ap, np);
3718 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3721 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3722 /* XXX Hack. Because we have to release the global rx lock when sending
3723 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3724 * in rxi_Start sending packets out because packets may move to the
3725 * freePacketQueue as result of being here! So we drop these packets until
3726 * we're safely out of the traversing. Really ugly!
3727 * To make it even uglier, if we're using fine grain locking, we can
3728 * set the ack bits in the packets and have rxi_Start remove the packets
3729 * when it's done transmitting.
3731 if (call->flags & RX_CALL_TQ_BUSY) {
3732 #ifdef RX_ENABLE_LOCKS
3733 tp->flags |= RX_PKTFLAG_ACKED;
3734 call->flags |= RX_CALL_TQ_SOME_ACKED;
3735 #else /* RX_ENABLE_LOCKS */
3737 #endif /* RX_ENABLE_LOCKS */
3739 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3742 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3747 /* Give rate detector a chance to respond to ping requests */
3748 if (ap->reason == RX_ACK_PING_RESPONSE) {
3749 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3753 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3755 /* Now go through explicit acks/nacks and record the results in
3756 * the waiting packets. These are packets that can't be released
3757 * yet, even with a positive acknowledge. This positive
3758 * acknowledge only means the packet has been received by the
3759 * peer, not that it will be retained long enough to be sent to
3760 * the peer's upper level. In addition, reset the transmit timers
3761 * of any missing packets (those packets that must be missing
3762 * because this packet was out of sequence) */
3764 call->nSoftAcked = 0;
3765 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3766 /* Update round trip time if the ack was stimulated on receipt
3768 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3769 #ifdef RX_ENABLE_LOCKS
3770 if (tp->header.seq >= first)
3771 #endif /* RX_ENABLE_LOCKS */
3772 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3774 && (tp->header.serial == serial || tp->firstSerial == serial))
3775 rxi_ComputePeerNetStats(call, tp, ap, np);
3777 /* Set the acknowledge flag per packet based on the
3778 * information in the ack packet. An acknowlegded packet can
3779 * be downgraded when the server has discarded a packet it
3780 * soacked previously, or when an ack packet is received
3781 * out of sequence. */
3782 if (tp->header.seq < first) {
3783 /* Implicit ack information */
3784 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3787 tp->flags |= RX_PKTFLAG_ACKED;
3788 } else if (tp->header.seq < first + nAcks) {
3789 /* Explicit ack information: set it in the packet appropriately */
3790 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3791 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3793 tp->flags |= RX_PKTFLAG_ACKED;
3800 } else /* RX_ACK_TYPE_NACK */ {
3801 tp->flags &= ~RX_PKTFLAG_ACKED;
3805 tp->flags &= ~RX_PKTFLAG_ACKED;
3809 /* If packet isn't yet acked, and it has been transmitted at least
3810 * once, reset retransmit time using latest timeout
3811 * ie, this should readjust the retransmit timer for all outstanding
3812 * packets... So we don't just retransmit when we should know better*/
3814 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3815 tp->retryTime = tp->timeSent;
3816 clock_Add(&tp->retryTime, &peer->timeout);
3817 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3818 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3822 /* If the window has been extended by this acknowledge packet,
3823 * then wakeup a sender waiting in alloc for window space, or try
3824 * sending packets now, if he's been sitting on packets due to
3825 * lack of window space */
3826 if (call->tnext < (call->tfirst + call->twind)) {
3827 #ifdef RX_ENABLE_LOCKS
3828 CV_SIGNAL(&call->cv_twind);
3830 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3831 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3832 osi_rxWakeup(&call->twind);
3835 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3836 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3840 /* if the ack packet has a receivelen field hanging off it,
3841 * update our state */
3842 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3845 /* If the ack packet has a "recommended" size that is less than
3846 * what I am using now, reduce my size to match */
3847 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3848 (int)sizeof(afs_int32), &tSize);
3849 tSize = (afs_uint32) ntohl(tSize);
3850 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3852 /* Get the maximum packet size to send to this peer */
3853 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
3855 tSize = (afs_uint32) ntohl(tSize);
3856 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3857 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3859 /* sanity check - peer might have restarted with different params.
3860 * If peer says "send less", dammit, send less... Peer should never
3861 * be unable to accept packets of the size that prior AFS versions would
3862 * send without asking. */
3863 if (peer->maxMTU != tSize) {
3864 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
3866 peer->maxMTU = tSize;
3867 peer->MTU = MIN(tSize, peer->MTU);
3868 call->MTU = MIN(call->MTU, tSize);
3871 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3874 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3875 (int)sizeof(afs_int32), &tSize);
3876 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3877 if (tSize < call->twind) { /* smaller than our send */
3878 call->twind = tSize; /* window, we must send less... */
3879 call->ssthresh = MIN(call->twind, call->ssthresh);
3880 call->conn->twind[call->channel] = call->twind;
3883 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3884 * network MTU confused with the loopback MTU. Calculate the
3885 * maximum MTU here for use in the slow start code below.
3887 maxMTU = peer->maxMTU;
3888 /* Did peer restart with older RX version? */
3889 if (peer->maxDgramPackets > 1) {
3890 peer->maxDgramPackets = 1;
3892 } else if (np->length >=
3893 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3896 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3897 sizeof(afs_int32), &tSize);
3898 tSize = (afs_uint32) ntohl(tSize);
3900 * As of AFS 3.5 we set the send window to match the receive window.
3902 if (tSize < call->twind) {
3903 call->twind = tSize;
3904 call->conn->twind[call->channel] = call->twind;
3905 call->ssthresh = MIN(call->twind, call->ssthresh);
3906 } else if (tSize > call->twind) {
3907 call->twind = tSize;
3908 call->conn->twind[call->channel] = call->twind;
3912 * As of AFS 3.5, a jumbogram is more than one fixed size
3913 * packet transmitted in a single UDP datagram. If the remote
3914 * MTU is smaller than our local MTU then never send a datagram
3915 * larger than the natural MTU.
3918 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3919 sizeof(afs_int32), &tSize);
3920 maxDgramPackets = (afs_uint32) ntohl(tSize);
3921 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3922 maxDgramPackets = MIN(maxDgramPackets, peer->ifDgramPackets);
3923 if (peer->natMTU < peer->ifMTU)
3924 maxDgramPackets = MIN(maxDgramPackets, rxi_AdjustDgramPackets(1, peer->natMTU));
3925 if (maxDgramPackets > 1) {
3926 peer->maxDgramPackets = maxDgramPackets;
3927 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3929 peer->maxDgramPackets = 1;
3930 call->MTU = peer->natMTU;
3932 } else if (peer->maxDgramPackets > 1) {
3933 /* Restarted with lower version of RX */
3934 peer->maxDgramPackets = 1;
3936 } else if (peer->maxDgramPackets > 1
3937 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3938 /* Restarted with lower version of RX */
3939 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3940 peer->natMTU = OLD_MAX_PACKET_SIZE;
3941 peer->MTU = OLD_MAX_PACKET_SIZE;
3942 peer->maxDgramPackets = 1;
3943 peer->nDgramPackets = 1;
3945 call->MTU = OLD_MAX_PACKET_SIZE;
3950 * Calculate how many datagrams were successfully received after
3951 * the first missing packet and adjust the negative ack counter
3956 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3957 if (call->nNacks < nNacked) {
3958 call->nNacks = nNacked;
3961 call->nAcks += newAckCount;
3965 if (call->flags & RX_CALL_FAST_RECOVER) {
3967 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3969 call->flags &= ~RX_CALL_FAST_RECOVER;
3970 call->cwind = call->nextCwind;
3971 call->nextCwind = 0;
3974 call->nCwindAcks = 0;
3975 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3976 /* Three negative acks in a row trigger congestion recovery */
3977 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3978 MUTEX_EXIT(&peer->peer_lock);
3979 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3980 /* someone else is waiting to start recovery */
3983 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3984 rxi_WaitforTQBusy(call);
3985 MUTEX_ENTER(&peer->peer_lock);
3986 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3987 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3988 call->flags |= RX_CALL_FAST_RECOVER;
3989 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3991 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3992 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3993 call->nextCwind = call->ssthresh;
3996 peer->MTU = call->MTU;
3997 peer->cwind = call->nextCwind;
3998 peer->nDgramPackets = call->nDgramPackets;
4000 call->congestSeq = peer->congestSeq;
4001 /* Reset the resend times on the packets that were nacked
4002 * so we will retransmit as soon as the window permits*/
4003 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4005 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4006 clock_Zero(&tp->retryTime);
4008 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4013 /* If cwind is smaller than ssthresh, then increase
4014 * the window one packet for each ack we receive (exponential
4016 * If cwind is greater than or equal to ssthresh then increase
4017 * the congestion window by one packet for each cwind acks we
4018 * receive (linear growth). */
4019 if (call->cwind < call->ssthresh) {
4021 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4022 call->nCwindAcks = 0;
4024 call->nCwindAcks += newAckCount;
4025 if (call->nCwindAcks >= call->cwind) {
4026 call->nCwindAcks = 0;
4027 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4031 * If we have received several acknowledgements in a row then
4032 * it is time to increase the size of our datagrams
4034 if ((int)call->nAcks > rx_nDgramThreshold) {
4035 if (peer->maxDgramPackets > 1) {
4036 if (call->nDgramPackets < peer->maxDgramPackets) {
4037 call->nDgramPackets++;
4039 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4040 } else if (call->MTU < peer->maxMTU) {
4041 call->MTU += peer->natMTU;
4042 call->MTU = MIN(call->MTU, peer->maxMTU);
4048 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4050 /* Servers need to hold the call until all response packets have
4051 * been acknowledged. Soft acks are good enough since clients
4052 * are not allowed to clear their receive queues. */
4053 if (call->state == RX_STATE_HOLD
4054 && call->tfirst + call->nSoftAcked >= call->tnext) {
4055 call->state = RX_STATE_DALLY;
4056 rxi_ClearTransmitQueue(call, 0);
4057 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4058 } else if (!queue_IsEmpty(&call->tq)) {
4059 rxi_Start(0, call, 0, istack);
4064 /* Received a response to a challenge packet */
4066 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
4067 register struct rx_packet *np, int istack)
4071 /* Ignore the packet if we're the client */
4072 if (conn->type == RX_CLIENT_CONNECTION)
4075 /* If already authenticated, ignore the packet (it's probably a retry) */
4076 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4079 /* Otherwise, have the security object evaluate the response packet */
4080 error = RXS_CheckResponse(conn->securityObject, conn, np);
4082 /* If the response is invalid, reset the connection, sending
4083 * an abort to the peer */
4087 rxi_ConnectionError(conn, error);
4088 MUTEX_ENTER(&conn->conn_data_lock);
4089 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4090 MUTEX_EXIT(&conn->conn_data_lock);
4093 /* If the response is valid, any calls waiting to attach
4094 * servers can now do so */
4097 for (i = 0; i < RX_MAXCALLS; i++) {
4098 struct rx_call *call = conn->call[i];
4100 MUTEX_ENTER(&call->lock);
4101 if (call->state == RX_STATE_PRECALL)
4102 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4103 /* tnop can be null if newcallp is null */
4104 MUTEX_EXIT(&call->lock);
4108 /* Update the peer reachability information, just in case
4109 * some calls went into attach-wait while we were waiting
4110 * for authentication..
4112 rxi_UpdatePeerReach(conn, NULL);
4117 /* A client has received an authentication challenge: the security
4118 * object is asked to cough up a respectable response packet to send
4119 * back to the server. The server is responsible for retrying the
4120 * challenge if it fails to get a response. */
4123 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
4124 register struct rx_packet *np, int istack)
4128 /* Ignore the challenge if we're the server */
4129 if (conn->type == RX_SERVER_CONNECTION)
4132 /* Ignore the challenge if the connection is otherwise idle; someone's
4133 * trying to use us as an oracle. */
4134 if (!rxi_HasActiveCalls(conn))
4137 /* Send the security object the challenge packet. It is expected to fill
4138 * in the response. */
4139 error = RXS_GetResponse(conn->securityObject, conn, np);
4141 /* If the security object is unable to return a valid response, reset the
4142 * connection and send an abort to the peer. Otherwise send the response
4143 * packet to the peer connection. */
4145 rxi_ConnectionError(conn, error);
4146 MUTEX_ENTER(&conn->conn_data_lock);
4147 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4148 MUTEX_EXIT(&conn->conn_data_lock);
4150 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4151 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4157 /* Find an available server process to service the current request in
4158 * the given call structure. If one isn't available, queue up this
4159 * call so it eventually gets one */
4161 rxi_AttachServerProc(register struct rx_call *call,
4162 register osi_socket socket, register int *tnop,
4163 register struct rx_call **newcallp)
4165 register struct rx_serverQueueEntry *sq;
4166 register struct rx_service *service = call->conn->service;
4167 register int haveQuota = 0;
4169 /* May already be attached */
4170 if (call->state == RX_STATE_ACTIVE)
4173 MUTEX_ENTER(&rx_serverPool_lock);
4175 haveQuota = QuotaOK(service);
4176 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4177 /* If there are no processes available to service this call,
4178 * put the call on the incoming call queue (unless it's
4179 * already on the queue).
4181 #ifdef RX_ENABLE_LOCKS
4183 ReturnToServerPool(service);
4184 #endif /* RX_ENABLE_LOCKS */
4186 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4187 call->flags |= RX_CALL_WAIT_PROC;
4188 MUTEX_ENTER(&rx_stats_mutex);
4191 MUTEX_EXIT(&rx_stats_mutex);
4192 rxi_calltrace(RX_CALL_ARRIVAL, call);
4193 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4194 queue_Append(&rx_incomingCallQueue, call);
4197 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4199 /* If hot threads are enabled, and both newcallp and sq->socketp
4200 * are non-null, then this thread will process the call, and the
4201 * idle server thread will start listening on this threads socket.
4204 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4207 *sq->socketp = socket;
4208 clock_GetTime(&call->startTime);
4209 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4213 if (call->flags & RX_CALL_WAIT_PROC) {
4214 /* Conservative: I don't think this should happen */
4215 call->flags &= ~RX_CALL_WAIT_PROC;
4216 if (queue_IsOnQueue(call)) {
4218 MUTEX_ENTER(&rx_stats_mutex);
4220 MUTEX_EXIT(&rx_stats_mutex);
4223 call->state = RX_STATE_ACTIVE;
4224 call->mode = RX_MODE_RECEIVING;
4225 #ifdef RX_KERNEL_TRACE
4227 int glockOwner = ISAFS_GLOCK();
4230 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4231 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4237 if (call->flags & RX_CALL_CLEARED) {
4238 /* send an ack now to start the packet flow up again */
4239 call->flags &= ~RX_CALL_CLEARED;
4240 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4242 #ifdef RX_ENABLE_LOCKS
4245 service->nRequestsRunning++;
4246 if (service->nRequestsRunning <= service->minProcs)
4252 MUTEX_EXIT(&rx_serverPool_lock);
4255 /* Delay the sending of an acknowledge event for a short while, while
4256 * a new call is being prepared (in the case of a client) or a reply
4257 * is being prepared (in the case of a server). Rather than sending
4258 * an ack packet, an ACKALL packet is sent. */
4260 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4262 #ifdef RX_ENABLE_LOCKS
4264 MUTEX_ENTER(&call->lock);
4265 call->delayedAckEvent = NULL;
4266 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4268 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4269 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4271 MUTEX_EXIT(&call->lock);
4272 #else /* RX_ENABLE_LOCKS */
4274 call->delayedAckEvent = NULL;
4275 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4276 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4277 #endif /* RX_ENABLE_LOCKS */
4281 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4284 #ifdef RX_ENABLE_LOCKS
4286 MUTEX_ENTER(&call->lock);
4287 if (event == call->delayedAckEvent)
4288 call->delayedAckEvent = NULL;
4289 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4291 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4293 MUTEX_EXIT(&call->lock);
4294 #else /* RX_ENABLE_LOCKS */
4296 call->delayedAckEvent = NULL;
4297 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4298 #endif /* RX_ENABLE_LOCKS */
4302 #ifdef RX_ENABLE_LOCKS
4303 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4304 * clearing them out.
4307 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4309 register struct rx_packet *p, *tp;
4312 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4313 p->flags |= RX_PKTFLAG_ACKED;
4317 call->flags |= RX_CALL_TQ_CLEARME;
4318 call->flags |= RX_CALL_TQ_SOME_ACKED;
4321 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4322 call->tfirst = call->tnext;
4323 call->nSoftAcked = 0;
4325 if (call->flags & RX_CALL_FAST_RECOVER) {
4326 call->flags &= ~RX_CALL_FAST_RECOVER;
4327 call->cwind = call->nextCwind;
4328 call->nextCwind = 0;
4331 CV_SIGNAL(&call->cv_twind);
4333 #endif /* RX_ENABLE_LOCKS */
4335 /* Clear out the transmit queue for the current call (all packets have
4336 * been received by peer) */
4338 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4340 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4341 register struct rx_packet *p, *tp;
4343 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4345 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4346 p->flags |= RX_PKTFLAG_ACKED;
4350 call->flags |= RX_CALL_TQ_CLEARME;
4351 call->flags |= RX_CALL_TQ_SOME_ACKED;
4354 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4355 rxi_FreePackets(0, &call->tq);
4356 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4357 call->flags &= ~RX_CALL_TQ_CLEARME;
4359 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4361 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4362 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4363 call->nSoftAcked = 0;
4365 if (call->flags & RX_CALL_FAST_RECOVER) {
4366 call->flags &= ~RX_CALL_FAST_RECOVER;
4367 call->cwind = call->nextCwind;
4369 #ifdef RX_ENABLE_LOCKS
4370 CV_SIGNAL(&call->cv_twind);
4372 osi_rxWakeup(&call->twind);
4377 rxi_ClearReceiveQueue(register struct rx_call *call)
4379 if (queue_IsNotEmpty(&call->rq)) {
4380 rx_packetReclaims += rxi_FreePackets(0, &call->rq);
4381 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4383 if (call->state == RX_STATE_PRECALL) {
4384 call->flags |= RX_CALL_CLEARED;
4388 /* Send an abort packet for the specified call */
4390 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4391 int istack, int force)
4394 struct clock when, now;
4399 /* Clients should never delay abort messages */
4400 if (rx_IsClientConn(call->conn))
4403 if (call->abortCode != call->error) {
4404 call->abortCode = call->error;
4405 call->abortCount = 0;
4408 if (force || rxi_callAbortThreshhold == 0
4409 || call->abortCount < rxi_callAbortThreshhold) {
4410 if (call->delayedAbortEvent) {
4411 rxevent_Cancel(call->delayedAbortEvent, call,
4412 RX_CALL_REFCOUNT_ABORT);
4414 error = htonl(call->error);
4417 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4418 (char *)&error, sizeof(error), istack);
4419 } else if (!call->delayedAbortEvent) {
4420 clock_GetTime(&now);
4422 clock_Addmsec(&when, rxi_callAbortDelay);
4423 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4424 call->delayedAbortEvent =
4425 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4430 /* Send an abort packet for the specified connection. Packet is an
4431 * optional pointer to a packet that can be used to send the abort.
4432 * Once the number of abort messages reaches the threshhold, an
4433 * event is scheduled to send the abort. Setting the force flag
4434 * overrides sending delayed abort messages.
4436 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4437 * to send the abort packet.
4440 rxi_SendConnectionAbort(register struct rx_connection *conn,
4441 struct rx_packet *packet, int istack, int force)
4444 struct clock when, now;
4449 /* Clients should never delay abort messages */
4450 if (rx_IsClientConn(conn))
4453 if (force || rxi_connAbortThreshhold == 0
4454 || conn->abortCount < rxi_connAbortThreshhold) {
4455 if (conn->delayedAbortEvent) {
4456 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4458 error = htonl(conn->error);
4460 MUTEX_EXIT(&conn->conn_data_lock);
4462 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4463 RX_PACKET_TYPE_ABORT, (char *)&error,
4464 sizeof(error), istack);
4465 MUTEX_ENTER(&conn->conn_data_lock);
4466 } else if (!conn->delayedAbortEvent) {
4467 clock_GetTime(&now);
4469 clock_Addmsec(&when, rxi_connAbortDelay);
4470 conn->delayedAbortEvent =
4471 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4476 /* Associate an error all of the calls owned by a connection. Called
4477 * with error non-zero. This is only for really fatal things, like
4478 * bad authentication responses. The connection itself is set in
4479 * error at this point, so that future packets received will be
4482 rxi_ConnectionError(register struct rx_connection *conn,
4483 register afs_int32 error)
4488 dpf(("rxi_ConnectionError conn %x error %d", conn, error));
4490 MUTEX_ENTER(&conn->conn_data_lock);
4491 if (conn->challengeEvent)
4492 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4493 if (conn->checkReachEvent) {
4494 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4495 conn->checkReachEvent = 0;
4496 conn->flags &= ~RX_CONN_ATTACHWAIT;
4499 MUTEX_EXIT(&conn->conn_data_lock);
4500 for (i = 0; i < RX_MAXCALLS; i++) {
4501 struct rx_call *call = conn->call[i];
4503 MUTEX_ENTER(&call->lock);
4504 rxi_CallError(call, error);
4505 MUTEX_EXIT(&call->lock);
4508 conn->error = error;
4509 rx_MutexIncrement(rx_stats.fatalErrors, rx_stats_mutex);
4514 rxi_CallError(register struct rx_call *call, afs_int32 error)
4516 dpf(("rxi_CallError call %x error %d call->error %d", call, error, call->error));
4518 error = call->error;
4520 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4521 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4522 rxi_ResetCall(call, 0);
4525 rxi_ResetCall(call, 0);
4527 call->error = error;
4528 call->mode = RX_MODE_ERROR;
4531 /* Reset various fields in a call structure, and wakeup waiting
4532 * processes. Some fields aren't changed: state & mode are not
4533 * touched (these must be set by the caller), and bufptr, nLeft, and
4534 * nFree are not reset, since these fields are manipulated by
4535 * unprotected macros, and may only be reset by non-interrupting code.
4538 /* this code requires that call->conn be set properly as a pre-condition. */
4539 #endif /* ADAPT_WINDOW */
4542 rxi_ResetCall(register struct rx_call *call, register int newcall)
4545 register struct rx_peer *peer;
4546 struct rx_packet *packet;
4548 dpf(("rxi_ResetCall(call %x, newcall %d)\n", call, newcall));
4550 /* Notify anyone who is waiting for asynchronous packet arrival */
4551 if (call->arrivalProc) {
4552 (*call->arrivalProc) (call, call->arrivalProcHandle,
4553 call->arrivalProcArg);
4554 call->arrivalProc = (void (*)())0;
4557 if (call->delayedAbortEvent) {
4558 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4559 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4561 rxi_SendCallAbort(call, packet, 0, 1);
4562 rxi_FreePacket(packet);
4567 * Update the peer with the congestion information in this call
4568 * so other calls on this connection can pick up where this call
4569 * left off. If the congestion sequence numbers don't match then
4570 * another call experienced a retransmission.
4572 peer = call->conn->peer;
4573 MUTEX_ENTER(&peer->peer_lock);
4575 if (call->congestSeq == peer->congestSeq) {
4576 peer->cwind = MAX(peer->cwind, call->cwind);
4577 peer->MTU = MAX(peer->MTU, call->MTU);
4578 peer->nDgramPackets =
4579 MAX(peer->nDgramPackets, call->nDgramPackets);
4582 call->abortCode = 0;
4583 call->abortCount = 0;
4585 if (peer->maxDgramPackets > 1) {
4586 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4588 call->MTU = peer->MTU;
4590 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4591 call->ssthresh = rx_maxSendWindow;
4592 call->nDgramPackets = peer->nDgramPackets;
4593 call->congestSeq = peer->congestSeq;
4594 MUTEX_EXIT(&peer->peer_lock);
4596 flags = call->flags;
4597 rxi_ClearReceiveQueue(call);
4598 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4599 if (flags & RX_CALL_TQ_BUSY) {
4600 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4601 call->flags |= (flags & RX_CALL_TQ_WAIT);
4603 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4605 rxi_ClearTransmitQueue(call, 0);
4606 queue_Init(&call->tq);
4607 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4608 dpf(("rcall %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4611 while (call->tqWaiters) {
4612 #ifdef RX_ENABLE_LOCKS
4613 CV_BROADCAST(&call->cv_tq);
4614 #else /* RX_ENABLE_LOCKS */
4615 osi_rxWakeup(&call->tq);
4616 #endif /* RX_ENABLE_LOCKS */
4620 queue_Init(&call->rq);
4622 call->twind = call->conn->twind[call->channel];
4623 call->rwind = call->conn->rwind[call->channel];
4624 call->nSoftAcked = 0;
4625 call->nextCwind = 0;
4628 call->nCwindAcks = 0;
4629 call->nSoftAcks = 0;
4630 call->nHardAcks = 0;
4632 call->tfirst = call->rnext = call->tnext = 1;
4634 call->lastAcked = 0;
4635 call->localStatus = call->remoteStatus = 0;
4637 if (flags & RX_CALL_READER_WAIT) {
4638 #ifdef RX_ENABLE_LOCKS
4639 CV_BROADCAST(&call->cv_rq);
4641 osi_rxWakeup(&call->rq);
4644 if (flags & RX_CALL_WAIT_PACKETS) {
4645 MUTEX_ENTER(&rx_freePktQ_lock);
4646 rxi_PacketsUnWait(); /* XXX */
4647 MUTEX_EXIT(&rx_freePktQ_lock);
4649 #ifdef RX_ENABLE_LOCKS
4650 CV_SIGNAL(&call->cv_twind);
4652 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4653 osi_rxWakeup(&call->twind);
4656 #ifdef RX_ENABLE_LOCKS
4657 /* The following ensures that we don't mess with any queue while some
4658 * other thread might also be doing so. The call_queue_lock field is
4659 * is only modified under the call lock. If the call is in the process
4660 * of being removed from a queue, the call is not locked until the
4661 * the queue lock is dropped and only then is the call_queue_lock field
4662 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4663 * Note that any other routine which removes a call from a queue has to
4664 * obtain the queue lock before examing the queue and removing the call.
4666 if (call->call_queue_lock) {
4667 MUTEX_ENTER(call->call_queue_lock);
4668 if (queue_IsOnQueue(call)) {
4670 if (flags & RX_CALL_WAIT_PROC) {
4671 MUTEX_ENTER(&rx_stats_mutex);
4673 MUTEX_EXIT(&rx_stats_mutex);
4676 MUTEX_EXIT(call->call_queue_lock);
4677 CLEAR_CALL_QUEUE_LOCK(call);
4679 #else /* RX_ENABLE_LOCKS */
4680 if (queue_IsOnQueue(call)) {
4682 if (flags & RX_CALL_WAIT_PROC)
4685 #endif /* RX_ENABLE_LOCKS */
4687 rxi_KeepAliveOff(call);
4688 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4691 /* Send an acknowledge for the indicated packet (seq,serial) of the
4692 * indicated call, for the indicated reason (reason). This
4693 * acknowledge will specifically acknowledge receiving the packet, and
4694 * will also specify which other packets for this call have been
4695 * received. This routine returns the packet that was used to the
4696 * caller. The caller is responsible for freeing it or re-using it.
4697 * This acknowledgement also returns the highest sequence number
4698 * actually read out by the higher level to the sender; the sender
4699 * promises to keep around packets that have not been read by the
4700 * higher level yet (unless, of course, the sender decides to abort
4701 * the call altogether). Any of p, seq, serial, pflags, or reason may
4702 * be set to zero without ill effect. That is, if they are zero, they
4703 * will not convey any information.
4704 * NOW there is a trailer field, after the ack where it will safely be
4705 * ignored by mundanes, which indicates the maximum size packet this
4706 * host can swallow. */
4708 register struct rx_packet *optionalPacket; use to send ack (or null)
4709 int seq; Sequence number of the packet we are acking
4710 int serial; Serial number of the packet
4711 int pflags; Flags field from packet header
4712 int reason; Reason an acknowledge was prompted
4716 rxi_SendAck(register struct rx_call *call,
4717 register struct rx_packet *optionalPacket, int serial, int reason,
4720 struct rx_ackPacket *ap;
4721 register struct rx_packet *rqp;
4722 register struct rx_packet *nxp; /* For queue_Scan */
4723 register struct rx_packet *p;
4726 #ifdef RX_ENABLE_TSFPQ
4727 struct rx_ts_info_t * rx_ts_info;
4731 * Open the receive window once a thread starts reading packets
4733 if (call->rnext > 1) {
4734 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
4737 call->nHardAcks = 0;
4738 call->nSoftAcks = 0;
4739 if (call->rnext > call->lastAcked)
4740 call->lastAcked = call->rnext;
4744 rx_computelen(p, p->length); /* reset length, you never know */
4745 } /* where that's been... */
4746 #ifdef RX_ENABLE_TSFPQ
4748 RX_TS_INFO_GET(rx_ts_info);
4749 if ((p = rx_ts_info->local_special_packet)) {
4750 rx_computelen(p, p->length);
4751 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4752 rx_ts_info->local_special_packet = p;
4753 } else { /* We won't send the ack, but don't panic. */
4754 return optionalPacket;
4758 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4759 /* We won't send the ack, but don't panic. */
4760 return optionalPacket;
4765 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4768 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4769 #ifndef RX_ENABLE_TSFPQ
4770 if (!optionalPacket)
4773 return optionalPacket;
4775 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4776 if (rx_Contiguous(p) < templ) {
4777 #ifndef RX_ENABLE_TSFPQ
4778 if (!optionalPacket)
4781 return optionalPacket;
4786 /* MTUXXX failing to send an ack is very serious. We should */
4787 /* try as hard as possible to send even a partial ack; it's */
4788 /* better than nothing. */
4789 ap = (struct rx_ackPacket *)rx_DataOf(p);
4790 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4791 ap->reason = reason;
4793 /* The skew computation used to be bogus, I think it's better now. */
4794 /* We should start paying attention to skew. XXX */
4795 ap->serial = htonl(serial);
4796 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4798 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4799 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4801 /* No fear of running out of ack packet here because there can only be at most
4802 * one window full of unacknowledged packets. The window size must be constrained
4803 * to be less than the maximum ack size, of course. Also, an ack should always
4804 * fit into a single packet -- it should not ever be fragmented. */
4805 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4806 if (!rqp || !call->rq.next
4807 || (rqp->header.seq > (call->rnext + call->rwind))) {
4808 #ifndef RX_ENABLE_TSFPQ
4809 if (!optionalPacket)
4812 rxi_CallError(call, RX_CALL_DEAD);
4813 return optionalPacket;
4816 while (rqp->header.seq > call->rnext + offset)
4817 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4818 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4820 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4821 #ifndef RX_ENABLE_TSFPQ
4822 if (!optionalPacket)
4825 rxi_CallError(call, RX_CALL_DEAD);
4826 return optionalPacket;
4831 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4833 /* these are new for AFS 3.3 */
4834 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4835 templ = htonl(templ);
4836 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4837 templ = htonl(call->conn->peer->ifMTU);
4838 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4839 sizeof(afs_int32), &templ);
4841 /* new for AFS 3.4 */
4842 templ = htonl(call->rwind);
4843 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4844 sizeof(afs_int32), &templ);
4846 /* new for AFS 3.5 */
4847 templ = htonl(call->conn->peer->ifDgramPackets);
4848 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4849 sizeof(afs_int32), &templ);
4851 p->header.serviceId = call->conn->serviceId;
4852 p->header.cid = (call->conn->cid | call->channel);
4853 p->header.callNumber = *call->callNumber;
4855 p->header.securityIndex = call->conn->securityIndex;
4856 p->header.epoch = call->conn->epoch;
4857 p->header.type = RX_PACKET_TYPE_ACK;
4858 p->header.flags = RX_SLOW_START_OK;
4859 if (reason == RX_ACK_PING) {
4860 p->header.flags |= RX_REQUEST_ACK;
4862 clock_GetTime(&call->pingRequestTime);
4865 if (call->conn->type == RX_CLIENT_CONNECTION)
4866 p->header.flags |= RX_CLIENT_INITIATED;
4870 if (rxdebug_active) {
4874 len = _snprintf(msg, sizeof(msg),
4875 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4876 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4877 ntohl(ap->serial), ntohl(ap->previousPacket),
4878 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
4879 ap->nAcks, ntohs(ap->bufferSpace) );
4883 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
4884 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4888 OutputDebugString(msg);
4890 #else /* AFS_NT40_ENV */
4892 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
4893 ap->reason, ntohl(ap->previousPacket),
4894 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4896 for (offset = 0; offset < ap->nAcks; offset++)
4897 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4902 #endif /* AFS_NT40_ENV */
4905 register int i, nbytes = p->length;
4907 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4908 if (nbytes <= p->wirevec[i].iov_len) {
4909 register int savelen, saven;
4911 savelen = p->wirevec[i].iov_len;
4913 p->wirevec[i].iov_len = nbytes;
4915 rxi_Send(call, p, istack);
4916 p->wirevec[i].iov_len = savelen;
4920 nbytes -= p->wirevec[i].iov_len;
4923 rx_MutexIncrement(rx_stats.ackPacketsSent, rx_stats_mutex);
4924 #ifndef RX_ENABLE_TSFPQ
4925 if (!optionalPacket)
4928 return optionalPacket; /* Return packet for re-use by caller */
4931 /* Send all of the packets in the list in single datagram */
4933 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4934 int istack, int moreFlag, struct clock *now,
4935 struct clock *retryTime, int resending)
4940 struct rx_connection *conn = call->conn;
4941 struct rx_peer *peer = conn->peer;
4943 MUTEX_ENTER(&peer->peer_lock);
4946 peer->reSends += len;
4947 rx_MutexIncrement(rx_stats.dataPacketsSent, rx_stats_mutex);
4948 MUTEX_EXIT(&peer->peer_lock);
4950 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4954 /* Set the packet flags and schedule the resend events */
4955 /* Only request an ack for the last packet in the list */
4956 for (i = 0; i < len; i++) {
4957 list[i]->retryTime = *retryTime;
4958 if (list[i]->header.serial) {
4959 /* Exponentially backoff retry times */
4960 if (list[i]->backoff < MAXBACKOFF) {
4961 /* so it can't stay == 0 */
4962 list[i]->backoff = (list[i]->backoff << 1) + 1;
4965 clock_Addmsec(&(list[i]->retryTime),
4966 ((afs_uint32) list[i]->backoff) << 8);
4969 /* Wait a little extra for the ack on the last packet */
4970 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4971 clock_Addmsec(&(list[i]->retryTime), 400);
4974 /* Record the time sent */
4975 list[i]->timeSent = *now;
4977 /* Ask for an ack on retransmitted packets, on every other packet
4978 * if the peer doesn't support slow start. Ask for an ack on every
4979 * packet until the congestion window reaches the ack rate. */
4980 if (list[i]->header.serial) {
4982 rx_MutexIncrement(rx_stats.dataPacketsReSent, rx_stats_mutex);
4984 /* improved RTO calculation- not Karn */
4985 list[i]->firstSent = *now;
4986 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4987 || (!(call->flags & RX_CALL_SLOW_START_OK)
4988 && (list[i]->header.seq & 1)))) {
4993 MUTEX_ENTER(&peer->peer_lock);
4997 rx_MutexIncrement(rx_stats.dataPacketsSent, rx_stats_mutex);
4998 MUTEX_EXIT(&peer->peer_lock);
5000 /* Tag this packet as not being the last in this group,
5001 * for the receiver's benefit */
5002 if (i < len - 1 || moreFlag) {
5003 list[i]->header.flags |= RX_MORE_PACKETS;
5006 /* Install the new retransmit time for the packet, and
5007 * record the time sent */
5008 list[i]->timeSent = *now;
5012 list[len - 1]->header.flags |= RX_REQUEST_ACK;
5015 /* Since we're about to send a data packet to the peer, it's
5016 * safe to nuke any scheduled end-of-packets ack */
5017 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5019 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5020 MUTEX_EXIT(&call->lock);
5022 rxi_SendPacketList(call, conn, list, len, istack);
5024 rxi_SendPacket(call, conn, list[0], istack);
5026 MUTEX_ENTER(&call->lock);
5027 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5029 /* Update last send time for this call (for keep-alive
5030 * processing), and for the connection (so that we can discover
5031 * idle connections) */
5032 call->lastSendData = conn->lastSendTime = call->lastSendTime = clock_Sec();
5035 /* When sending packets we need to follow these rules:
5036 * 1. Never send more than maxDgramPackets in a jumbogram.
5037 * 2. Never send a packet with more than two iovecs in a jumbogram.
5038 * 3. Never send a retransmitted packet in a jumbogram.
5039 * 4. Never send more than cwind/4 packets in a jumbogram
5040 * We always keep the last list we should have sent so we
5041 * can set the RX_MORE_PACKETS flags correctly.
5044 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5045 int istack, struct clock *now, struct clock *retryTime,
5048 int i, cnt, lastCnt = 0;
5049 struct rx_packet **listP, **lastP = 0;
5050 struct rx_peer *peer = call->conn->peer;
5051 int morePackets = 0;
5053 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
5054 /* Does the current packet force us to flush the current list? */
5056 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5057 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5059 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
5061 /* If the call enters an error state stop sending, or if
5062 * we entered congestion recovery mode, stop sending */
5063 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5071 /* Add the current packet to the list if it hasn't been acked.
5072 * Otherwise adjust the list pointer to skip the current packet. */
5073 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5075 /* Do we need to flush the list? */
5076 if (cnt >= (int)peer->maxDgramPackets
5077 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5078 || list[i]->header.serial
5079 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5081 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5082 retryTime, resending);
5083 /* If the call enters an error state stop sending, or if
5084 * we entered congestion recovery mode, stop sending */
5086 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5091 listP = &list[i + 1];
5096 osi_Panic("rxi_SendList error");
5098 listP = &list[i + 1];
5102 /* Send the whole list when the call is in receive mode, when
5103 * the call is in eof mode, when we are in fast recovery mode,
5104 * and when we have the last packet */
5105 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5106 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5107 || (call->flags & RX_CALL_FAST_RECOVER)) {
5108 /* Check for the case where the current list contains
5109 * an acked packet. Since we always send retransmissions
5110 * in a separate packet, we only need to check the first
5111 * packet in the list */
5112 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5116 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5117 retryTime, resending);
5118 /* If the call enters an error state stop sending, or if
5119 * we entered congestion recovery mode, stop sending */
5120 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5124 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5127 } else if (lastCnt > 0) {
5128 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5133 #ifdef RX_ENABLE_LOCKS
5134 /* Call rxi_Start, below, but with the call lock held. */
5136 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
5137 void *arg1, int istack)
5139 MUTEX_ENTER(&call->lock);
5140 rxi_Start(event, call, arg1, istack);
5141 MUTEX_EXIT(&call->lock);
5143 #endif /* RX_ENABLE_LOCKS */
5145 /* This routine is called when new packets are readied for
5146 * transmission and when retransmission may be necessary, or when the
5147 * transmission window or burst count are favourable. This should be
5148 * better optimized for new packets, the usual case, now that we've
5149 * got rid of queues of send packets. XXXXXXXXXXX */
5151 rxi_Start(struct rxevent *event, register struct rx_call *call,
5152 void *arg1, int istack)
5154 struct rx_packet *p;
5155 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
5156 struct rx_peer *peer = call->conn->peer;
5157 struct clock now, usenow, retryTime;
5161 struct rx_packet **xmitList;
5164 /* If rxi_Start is being called as a result of a resend event,
5165 * then make sure that the event pointer is removed from the call
5166 * structure, since there is no longer a per-call retransmission
5168 if (event && event == call->resendEvent) {
5169 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5170 call->resendEvent = NULL;
5172 if (queue_IsEmpty(&call->tq)) {
5176 /* Timeouts trigger congestion recovery */
5177 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5178 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5179 /* someone else is waiting to start recovery */
5182 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5183 rxi_WaitforTQBusy(call);
5184 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5185 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5186 call->flags |= RX_CALL_FAST_RECOVER;
5187 if (peer->maxDgramPackets > 1) {
5188 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5190 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5192 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5193 call->nDgramPackets = 1;
5195 call->nextCwind = 1;
5198 MUTEX_ENTER(&peer->peer_lock);
5199 peer->MTU = call->MTU;
5200 peer->cwind = call->cwind;
5201 peer->nDgramPackets = 1;
5203 call->congestSeq = peer->congestSeq;
5204 MUTEX_EXIT(&peer->peer_lock);
5205 /* Clear retry times on packets. Otherwise, it's possible for
5206 * some packets in the queue to force resends at rates faster
5207 * than recovery rates.
5209 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5210 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5211 clock_Zero(&p->retryTime);
5216 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5217 rx_MutexIncrement(rx_tq_debug.rxi_start_in_error, rx_stats_mutex);
5222 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5223 /* Get clock to compute the re-transmit time for any packets
5224 * in this burst. Note, if we back off, it's reasonable to
5225 * back off all of the packets in the same manner, even if
5226 * some of them have been retransmitted more times than more
5228 * Do a dance to avoid blocking after setting now. */
5229 clock_Zero(&retryTime);
5230 MUTEX_ENTER(&peer->peer_lock);
5231 clock_Add(&retryTime, &peer->timeout);
5232 MUTEX_EXIT(&peer->peer_lock);
5233 clock_GetTime(&now);
5234 clock_Add(&retryTime, &now);
5236 /* Send (or resend) any packets that need it, subject to
5237 * window restrictions and congestion burst control
5238 * restrictions. Ask for an ack on the last packet sent in
5239 * this burst. For now, we're relying upon the window being
5240 * considerably bigger than the largest number of packets that
5241 * are typically sent at once by one initial call to
5242 * rxi_Start. This is probably bogus (perhaps we should ask
5243 * for an ack when we're half way through the current
5244 * window?). Also, for non file transfer applications, this
5245 * may end up asking for an ack for every packet. Bogus. XXXX
5248 * But check whether we're here recursively, and let the other guy
5251 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5252 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5253 call->flags |= RX_CALL_TQ_BUSY;
5255 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5257 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5258 call->flags &= ~RX_CALL_NEED_START;
5259 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5261 maxXmitPackets = MIN(call->twind, call->cwind);
5262 xmitList = (struct rx_packet **)
5263 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5264 if (xmitList == NULL)
5265 osi_Panic("rxi_Start, failed to allocate xmit list");
5266 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5267 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5268 /* We shouldn't be sending packets if a thread is waiting
5269 * to initiate congestion recovery */
5273 && (call->flags & RX_CALL_FAST_RECOVER)) {
5274 /* Only send one packet during fast recovery */
5277 if ((p->flags & RX_PKTFLAG_FREE)
5278 || (!queue_IsEnd(&call->tq, nxp)
5279 && (nxp->flags & RX_PKTFLAG_FREE))
5280 || (p == (struct rx_packet *)&rx_freePacketQueue)
5281 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5282 osi_Panic("rxi_Start: xmit queue clobbered");
5284 if (p->flags & RX_PKTFLAG_ACKED) {
5285 /* Since we may block, don't trust this */
5286 usenow.sec = usenow.usec = 0;
5287 rx_MutexIncrement(rx_stats.ignoreAckedPacket, rx_stats_mutex);
5288 continue; /* Ignore this packet if it has been acknowledged */
5291 /* Turn off all flags except these ones, which are the same
5292 * on each transmission */
5293 p->header.flags &= RX_PRESET_FLAGS;
5295 if (p->header.seq >=
5296 call->tfirst + MIN((int)call->twind,
5297 (int)(call->nSoftAcked +
5299 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5300 /* Note: if we're waiting for more window space, we can
5301 * still send retransmits; hence we don't return here, but
5302 * break out to schedule a retransmit event */
5303 dpf(("call %d waiting for window",
5304 *(call->callNumber)));
5308 /* Transmit the packet if it needs to be sent. */
5309 if (!clock_Lt(&now, &p->retryTime)) {
5310 if (nXmitPackets == maxXmitPackets) {
5311 rxi_SendXmitList(call, xmitList, nXmitPackets,
5312 istack, &now, &retryTime,
5314 osi_Free(xmitList, maxXmitPackets *
5315 sizeof(struct rx_packet *));
5318 xmitList[nXmitPackets++] = p;
5322 /* xmitList now hold pointers to all of the packets that are
5323 * ready to send. Now we loop to send the packets */
5324 if (nXmitPackets > 0) {
5325 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5326 &now, &retryTime, resending);
5329 maxXmitPackets * sizeof(struct rx_packet *));
5331 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5333 * TQ references no longer protected by this flag; they must remain
5334 * protected by the global lock.
5336 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5337 call->flags &= ~RX_CALL_TQ_BUSY;
5338 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5339 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5340 #ifdef RX_ENABLE_LOCKS
5341 osirx_AssertMine(&call->lock, "rxi_Start start");
5342 CV_BROADCAST(&call->cv_tq);
5343 #else /* RX_ENABLE_LOCKS */
5344 osi_rxWakeup(&call->tq);
5345 #endif /* RX_ENABLE_LOCKS */
5350 /* We went into the error state while sending packets. Now is
5351 * the time to reset the call. This will also inform the using
5352 * process that the call is in an error state.
5354 rx_MutexIncrement(rx_tq_debug.rxi_start_aborted, rx_stats_mutex);
5355 call->flags &= ~RX_CALL_TQ_BUSY;
5356 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5357 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5358 #ifdef RX_ENABLE_LOCKS
5359 osirx_AssertMine(&call->lock, "rxi_Start middle");
5360 CV_BROADCAST(&call->cv_tq);
5361 #else /* RX_ENABLE_LOCKS */
5362 osi_rxWakeup(&call->tq);
5363 #endif /* RX_ENABLE_LOCKS */
5365 rxi_CallError(call, call->error);
5368 #ifdef RX_ENABLE_LOCKS
5369 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5370 register int missing;
5371 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5372 /* Some packets have received acks. If they all have, we can clear
5373 * the transmit queue.
5376 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5377 if (p->header.seq < call->tfirst
5378 && (p->flags & RX_PKTFLAG_ACKED)) {
5385 call->flags |= RX_CALL_TQ_CLEARME;
5387 #endif /* RX_ENABLE_LOCKS */
5388 /* Don't bother doing retransmits if the TQ is cleared. */
5389 if (call->flags & RX_CALL_TQ_CLEARME) {
5390 rxi_ClearTransmitQueue(call, 1);
5392 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5395 /* Always post a resend event, if there is anything in the
5396 * queue, and resend is possible. There should be at least
5397 * one unacknowledged packet in the queue ... otherwise none
5398 * of these packets should be on the queue in the first place.
5400 if (call->resendEvent) {
5401 /* Cancel the existing event and post a new one */
5402 rxevent_Cancel(call->resendEvent, call,
5403 RX_CALL_REFCOUNT_RESEND);
5406 /* The retry time is the retry time on the first unacknowledged
5407 * packet inside the current window */
5409 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5410 /* Don't set timers for packets outside the window */
5411 if (p->header.seq >= call->tfirst + call->twind) {
5415 if (!(p->flags & RX_PKTFLAG_ACKED)
5416 && !clock_IsZero(&p->retryTime)) {
5418 retryTime = p->retryTime;
5423 /* Post a new event to re-run rxi_Start when retries may be needed */
5424 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5425 #ifdef RX_ENABLE_LOCKS
5426 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5428 rxevent_PostNow2(&retryTime, &usenow,
5430 (void *)call, 0, istack);
5431 #else /* RX_ENABLE_LOCKS */
5433 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5434 (void *)call, 0, istack);
5435 #endif /* RX_ENABLE_LOCKS */
5438 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5439 } while (call->flags & RX_CALL_NEED_START);
5441 * TQ references no longer protected by this flag; they must remain
5442 * protected by the global lock.
5444 call->flags &= ~RX_CALL_TQ_BUSY;
5445 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5446 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5447 #ifdef RX_ENABLE_LOCKS
5448 osirx_AssertMine(&call->lock, "rxi_Start end");
5449 CV_BROADCAST(&call->cv_tq);
5450 #else /* RX_ENABLE_LOCKS */
5451 osi_rxWakeup(&call->tq);
5452 #endif /* RX_ENABLE_LOCKS */
5455 call->flags |= RX_CALL_NEED_START;
5457 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5459 if (call->resendEvent) {
5460 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5465 /* Also adjusts the keep alive parameters for the call, to reflect
5466 * that we have just sent a packet (so keep alives aren't sent
5469 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5472 register struct rx_connection *conn = call->conn;
5474 /* Stamp each packet with the user supplied status */
5475 p->header.userStatus = call->localStatus;
5477 /* Allow the security object controlling this call's security to
5478 * make any last-minute changes to the packet */
5479 RXS_SendPacket(conn->securityObject, call, p);
5481 /* Since we're about to send SOME sort of packet to the peer, it's
5482 * safe to nuke any scheduled end-of-packets ack */
5483 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5485 /* Actually send the packet, filling in more connection-specific fields */
5486 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5487 MUTEX_EXIT(&call->lock);
5488 rxi_SendPacket(call, conn, p, istack);
5489 MUTEX_ENTER(&call->lock);
5490 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5492 /* Update last send time for this call (for keep-alive
5493 * processing), and for the connection (so that we can discover
5494 * idle connections) */
5495 conn->lastSendTime = call->lastSendTime = clock_Sec();
5496 /* Don't count keepalives here, so idleness can be tracked. */
5497 if (p->header.type != RX_PACKET_TYPE_ACK)
5498 call->lastSendData = call->lastSendTime;
5502 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5503 * that things are fine. Also called periodically to guarantee that nothing
5504 * falls through the cracks (e.g. (error + dally) connections have keepalive
5505 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5507 * haveCTLock Set if calling from rxi_ReapConnections
5509 #ifdef RX_ENABLE_LOCKS
5511 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5512 #else /* RX_ENABLE_LOCKS */
5514 rxi_CheckCall(register struct rx_call *call)
5515 #endif /* RX_ENABLE_LOCKS */
5517 register struct rx_connection *conn = call->conn;
5519 afs_uint32 deadTime;
5521 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5522 if (call->flags & RX_CALL_TQ_BUSY) {
5523 /* Call is active and will be reset by rxi_Start if it's
5524 * in an error state.
5529 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5531 (((afs_uint32) conn->secondsUntilDead << 10) +
5532 ((afs_uint32) conn->peer->rtt >> 3) +
5533 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5535 /* These are computed to the second (+- 1 second). But that's
5536 * good enough for these values, which should be a significant
5537 * number of seconds. */
5538 if (now > (call->lastReceiveTime + deadTime)) {
5539 if (call->state == RX_STATE_ACTIVE) {
5541 #if defined(KERNEL) && defined(AFS_SUN57_ENV)
5543 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5544 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
5545 ip_stack_t *ipst = ns->netstack_ip;
5547 ire = ire_cache_lookup(call->conn->peer->host
5548 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
5550 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
5552 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5559 if (ire && ire->ire_max_frag > 0)
5560 rxi_SetPeerMtu(call->conn->peer->host, 0, ire->ire_max_frag);
5561 #if defined(GLOBAL_NETSTACKID)
5565 #endif /* ADAPT_PMTU */
5566 rxi_CallError(call, RX_CALL_DEAD);
5569 #ifdef RX_ENABLE_LOCKS
5570 /* Cancel pending events */
5571 rxevent_Cancel(call->delayedAckEvent, call,
5572 RX_CALL_REFCOUNT_DELAY);
5573 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5574 rxevent_Cancel(call->keepAliveEvent, call,
5575 RX_CALL_REFCOUNT_ALIVE);
5576 if (call->refCount == 0) {
5577 rxi_FreeCall(call, haveCTLock);
5581 #else /* RX_ENABLE_LOCKS */
5584 #endif /* RX_ENABLE_LOCKS */
5586 /* Non-active calls are destroyed if they are not responding
5587 * to pings; active calls are simply flagged in error, so the
5588 * attached process can die reasonably gracefully. */
5590 /* see if we have a non-activity timeout */
5591 if (call->startWait && conn->idleDeadTime
5592 && ((call->startWait + conn->idleDeadTime) < now)) {
5593 if (call->state == RX_STATE_ACTIVE) {
5594 rxi_CallError(call, RX_CALL_TIMEOUT);
5598 if (call->lastSendData && conn->idleDeadTime && (conn->idleDeadErr != 0)
5599 && ((call->lastSendData + conn->idleDeadTime) < now)) {
5600 if (call->state == RX_STATE_ACTIVE) {
5601 rxi_CallError(call, conn->idleDeadErr);
5605 /* see if we have a hard timeout */
5606 if (conn->hardDeadTime
5607 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5608 if (call->state == RX_STATE_ACTIVE)
5609 rxi_CallError(call, RX_CALL_TIMEOUT);
5616 /* When a call is in progress, this routine is called occasionally to
5617 * make sure that some traffic has arrived (or been sent to) the peer.
5618 * If nothing has arrived in a reasonable amount of time, the call is
5619 * declared dead; if nothing has been sent for a while, we send a
5620 * keep-alive packet (if we're actually trying to keep the call alive)
5623 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5626 struct rx_connection *conn;
5629 MUTEX_ENTER(&call->lock);
5630 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5631 if (event == call->keepAliveEvent)
5632 call->keepAliveEvent = NULL;
5635 #ifdef RX_ENABLE_LOCKS
5636 if (rxi_CheckCall(call, 0)) {
5637 MUTEX_EXIT(&call->lock);
5640 #else /* RX_ENABLE_LOCKS */
5641 if (rxi_CheckCall(call))
5643 #endif /* RX_ENABLE_LOCKS */
5645 /* Don't try to keep alive dallying calls */
5646 if (call->state == RX_STATE_DALLY) {
5647 MUTEX_EXIT(&call->lock);
5652 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5653 /* Don't try to send keepalives if there is unacknowledged data */
5654 /* the rexmit code should be good enough, this little hack
5655 * doesn't quite work XXX */
5656 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5658 rxi_ScheduleKeepAliveEvent(call);
5659 MUTEX_EXIT(&call->lock);
5664 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5666 if (!call->keepAliveEvent) {
5667 struct clock when, now;
5668 clock_GetTime(&now);
5670 when.sec += call->conn->secondsUntilPing;
5671 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5672 call->keepAliveEvent =
5673 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
5677 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5679 rxi_KeepAliveOn(register struct rx_call *call)
5681 /* Pretend last packet received was received now--i.e. if another
5682 * packet isn't received within the keep alive time, then the call
5683 * will die; Initialize last send time to the current time--even
5684 * if a packet hasn't been sent yet. This will guarantee that a
5685 * keep-alive is sent within the ping time */
5686 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5687 rxi_ScheduleKeepAliveEvent(call);
5690 /* This routine is called to send connection abort messages
5691 * that have been delayed to throttle looping clients. */
5693 rxi_SendDelayedConnAbort(struct rxevent *event,
5694 register struct rx_connection *conn, char *dummy)
5697 struct rx_packet *packet;
5699 MUTEX_ENTER(&conn->conn_data_lock);
5700 conn->delayedAbortEvent = NULL;
5701 error = htonl(conn->error);
5703 MUTEX_EXIT(&conn->conn_data_lock);
5704 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5707 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5708 RX_PACKET_TYPE_ABORT, (char *)&error,
5710 rxi_FreePacket(packet);
5714 /* This routine is called to send call abort messages
5715 * that have been delayed to throttle looping clients. */
5717 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5721 struct rx_packet *packet;
5723 MUTEX_ENTER(&call->lock);
5724 call->delayedAbortEvent = NULL;
5725 error = htonl(call->error);
5727 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5730 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5731 (char *)&error, sizeof(error), 0);
5732 rxi_FreePacket(packet);
5734 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5735 MUTEX_EXIT(&call->lock);
5738 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5739 * seconds) to ask the client to authenticate itself. The routine
5740 * issues a challenge to the client, which is obtained from the
5741 * security object associated with the connection */
5743 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5744 void *arg1, int tries)
5746 conn->challengeEvent = NULL;
5747 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5748 register struct rx_packet *packet;
5749 struct clock when, now;
5752 /* We've failed to authenticate for too long.
5753 * Reset any calls waiting for authentication;
5754 * they are all in RX_STATE_PRECALL.
5758 MUTEX_ENTER(&conn->conn_call_lock);
5759 for (i = 0; i < RX_MAXCALLS; i++) {
5760 struct rx_call *call = conn->call[i];
5762 MUTEX_ENTER(&call->lock);
5763 if (call->state == RX_STATE_PRECALL) {
5764 rxi_CallError(call, RX_CALL_DEAD);
5765 rxi_SendCallAbort(call, NULL, 0, 0);
5767 MUTEX_EXIT(&call->lock);
5770 MUTEX_EXIT(&conn->conn_call_lock);
5774 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5776 /* If there's no packet available, do this later. */
5777 RXS_GetChallenge(conn->securityObject, conn, packet);
5778 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5779 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5780 rxi_FreePacket(packet);
5782 clock_GetTime(&now);
5784 when.sec += RX_CHALLENGE_TIMEOUT;
5785 conn->challengeEvent =
5786 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
5791 /* Call this routine to start requesting the client to authenticate
5792 * itself. This will continue until authentication is established,
5793 * the call times out, or an invalid response is returned. The
5794 * security object associated with the connection is asked to create
5795 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5796 * defined earlier. */
5798 rxi_ChallengeOn(register struct rx_connection *conn)
5800 if (!conn->challengeEvent) {
5801 RXS_CreateChallenge(conn->securityObject, conn);
5802 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5807 /* Compute round trip time of the packet provided, in *rttp.
5810 /* rxi_ComputeRoundTripTime is called with peer locked. */
5811 /* sentp and/or peer may be null */
5813 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5814 register struct clock *sentp,
5815 register struct rx_peer *peer)
5817 struct clock thisRtt, *rttp = &thisRtt;
5819 register int rtt_timeout;
5821 clock_GetTime(rttp);
5823 if (clock_Lt(rttp, sentp)) {
5825 return; /* somebody set the clock back, don't count this time. */
5827 clock_Sub(rttp, sentp);
5828 MUTEX_ENTER(&rx_stats_mutex);
5829 if (clock_Lt(rttp, &rx_stats.minRtt))
5830 rx_stats.minRtt = *rttp;
5831 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5832 if (rttp->sec > 60) {
5833 MUTEX_EXIT(&rx_stats_mutex);
5834 return; /* somebody set the clock ahead */
5836 rx_stats.maxRtt = *rttp;
5838 clock_Add(&rx_stats.totalRtt, rttp);
5839 rx_stats.nRttSamples++;
5840 MUTEX_EXIT(&rx_stats_mutex);
5842 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5844 /* Apply VanJacobson round-trip estimations */
5849 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5850 * srtt is stored as fixed point with 3 bits after the binary
5851 * point (i.e., scaled by 8). The following magic is
5852 * equivalent to the smoothing algorithm in rfc793 with an
5853 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5854 * srtt*8 = srtt*8 + rtt - srtt
5855 * srtt = srtt + rtt/8 - srtt/8
5858 delta = MSEC(rttp) - (peer->rtt >> 3);
5862 * We accumulate a smoothed rtt variance (actually, a smoothed
5863 * mean difference), then set the retransmit timer to smoothed
5864 * rtt + 4 times the smoothed variance (was 2x in van's original
5865 * paper, but 4x works better for me, and apparently for him as
5867 * rttvar is stored as
5868 * fixed point with 2 bits after the binary point (scaled by
5869 * 4). The following is equivalent to rfc793 smoothing with
5870 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5871 * replaces rfc793's wired-in beta.
5872 * dev*4 = dev*4 + (|actual - expected| - dev)
5878 delta -= (peer->rtt_dev >> 2);
5879 peer->rtt_dev += delta;
5881 /* I don't have a stored RTT so I start with this value. Since I'm
5882 * probably just starting a call, and will be pushing more data down
5883 * this, I expect congestion to increase rapidly. So I fudge a
5884 * little, and I set deviance to half the rtt. In practice,
5885 * deviance tends to approach something a little less than
5886 * half the smoothed rtt. */
5887 peer->rtt = (MSEC(rttp) << 3) + 8;
5888 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5890 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5891 * the other of these connections is usually in a user process, and can
5892 * be switched and/or swapped out. So on fast, reliable networks, the
5893 * timeout would otherwise be too short.
5895 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5896 clock_Zero(&(peer->timeout));
5897 clock_Addmsec(&(peer->timeout), rtt_timeout);
5899 dpf(("rxi_ComputeRoundTripTime(rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%0.3d sec)\n", MSEC(rttp), peer->rtt >> 3, peer->rtt_dev >> 2, (peer->timeout.sec), (peer->timeout.usec)));
5903 /* Find all server connections that have not been active for a long time, and
5906 rxi_ReapConnections(void)
5908 struct clock now, when;
5909 clock_GetTime(&now);
5911 /* Find server connection structures that haven't been used for
5912 * greater than rx_idleConnectionTime */
5914 struct rx_connection **conn_ptr, **conn_end;
5915 int i, havecalls = 0;
5916 MUTEX_ENTER(&rx_connHashTable_lock);
5917 for (conn_ptr = &rx_connHashTable[0], conn_end =
5918 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5920 struct rx_connection *conn, *next;
5921 struct rx_call *call;
5925 for (conn = *conn_ptr; conn; conn = next) {
5926 /* XXX -- Shouldn't the connection be locked? */
5929 for (i = 0; i < RX_MAXCALLS; i++) {
5930 call = conn->call[i];
5933 MUTEX_ENTER(&call->lock);
5934 #ifdef RX_ENABLE_LOCKS
5935 result = rxi_CheckCall(call, 1);
5936 #else /* RX_ENABLE_LOCKS */
5937 result = rxi_CheckCall(call);
5938 #endif /* RX_ENABLE_LOCKS */
5939 MUTEX_EXIT(&call->lock);
5941 /* If CheckCall freed the call, it might
5942 * have destroyed the connection as well,
5943 * which screws up the linked lists.
5949 if (conn->type == RX_SERVER_CONNECTION) {
5950 /* This only actually destroys the connection if
5951 * there are no outstanding calls */
5952 MUTEX_ENTER(&conn->conn_data_lock);
5953 if (!havecalls && !conn->refCount
5954 && ((conn->lastSendTime + rx_idleConnectionTime) <
5956 conn->refCount++; /* it will be decr in rx_DestroyConn */
5957 MUTEX_EXIT(&conn->conn_data_lock);
5958 #ifdef RX_ENABLE_LOCKS
5959 rxi_DestroyConnectionNoLock(conn);
5960 #else /* RX_ENABLE_LOCKS */
5961 rxi_DestroyConnection(conn);
5962 #endif /* RX_ENABLE_LOCKS */
5964 #ifdef RX_ENABLE_LOCKS
5966 MUTEX_EXIT(&conn->conn_data_lock);
5968 #endif /* RX_ENABLE_LOCKS */
5972 #ifdef RX_ENABLE_LOCKS
5973 while (rx_connCleanup_list) {
5974 struct rx_connection *conn;
5975 conn = rx_connCleanup_list;
5976 rx_connCleanup_list = rx_connCleanup_list->next;
5977 MUTEX_EXIT(&rx_connHashTable_lock);
5978 rxi_CleanupConnection(conn);
5979 MUTEX_ENTER(&rx_connHashTable_lock);
5981 MUTEX_EXIT(&rx_connHashTable_lock);
5982 #endif /* RX_ENABLE_LOCKS */
5985 /* Find any peer structures that haven't been used (haven't had an
5986 * associated connection) for greater than rx_idlePeerTime */
5988 struct rx_peer **peer_ptr, **peer_end;
5990 MUTEX_ENTER(&rx_rpc_stats);
5991 MUTEX_ENTER(&rx_peerHashTable_lock);
5992 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5993 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5995 struct rx_peer *peer, *next, *prev;
5996 for (prev = peer = *peer_ptr; peer; peer = next) {
5998 code = MUTEX_TRYENTER(&peer->peer_lock);
5999 if ((code) && (peer->refCount == 0)
6000 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6001 rx_interface_stat_p rpc_stat, nrpc_stat;
6003 MUTEX_EXIT(&peer->peer_lock);
6004 MUTEX_DESTROY(&peer->peer_lock);
6006 (&peer->rpcStats, rpc_stat, nrpc_stat,
6007 rx_interface_stat)) {
6008 unsigned int num_funcs;
6011 queue_Remove(&rpc_stat->queue_header);
6012 queue_Remove(&rpc_stat->all_peers);
6013 num_funcs = rpc_stat->stats[0].func_total;
6015 sizeof(rx_interface_stat_t) +
6016 rpc_stat->stats[0].func_total *
6017 sizeof(rx_function_entry_v1_t);
6019 rxi_Free(rpc_stat, space);
6020 rxi_rpc_peer_stat_cnt -= num_funcs;
6023 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6024 if (peer == *peer_ptr) {
6031 MUTEX_EXIT(&peer->peer_lock);
6037 MUTEX_EXIT(&rx_peerHashTable_lock);
6038 MUTEX_EXIT(&rx_rpc_stats);
6041 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
6042 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
6043 * GC, just below. Really, we shouldn't have to keep moving packets from
6044 * one place to another, but instead ought to always know if we can
6045 * afford to hold onto a packet in its particular use. */
6046 MUTEX_ENTER(&rx_freePktQ_lock);
6047 if (rx_waitingForPackets) {
6048 rx_waitingForPackets = 0;
6049 #ifdef RX_ENABLE_LOCKS
6050 CV_BROADCAST(&rx_waitingForPackets_cv);
6052 osi_rxWakeup(&rx_waitingForPackets);
6055 MUTEX_EXIT(&rx_freePktQ_lock);
6058 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
6059 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
6063 /* rxs_Release - This isn't strictly necessary but, since the macro name from
6064 * rx.h is sort of strange this is better. This is called with a security
6065 * object before it is discarded. Each connection using a security object has
6066 * its own refcount to the object so it won't actually be freed until the last
6067 * connection is destroyed.
6069 * This is the only rxs module call. A hold could also be written but no one
6073 rxs_Release(struct rx_securityClass *aobj)
6075 return RXS_Close(aobj);
6079 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
6080 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
6081 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
6082 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
6084 /* Adjust our estimate of the transmission rate to this peer, given
6085 * that the packet p was just acked. We can adjust peer->timeout and
6086 * call->twind. Pragmatically, this is called
6087 * only with packets of maximal length.
6088 * Called with peer and call locked.
6092 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
6093 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
6095 afs_int32 xferSize, xferMs;
6096 register afs_int32 minTime;
6099 /* Count down packets */
6100 if (peer->rateFlag > 0)
6102 /* Do nothing until we're enabled */
6103 if (peer->rateFlag != 0)
6108 /* Count only when the ack seems legitimate */
6109 switch (ackReason) {
6110 case RX_ACK_REQUESTED:
6112 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6116 case RX_ACK_PING_RESPONSE:
6117 if (p) /* want the response to ping-request, not data send */
6119 clock_GetTime(&newTO);
6120 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6121 clock_Sub(&newTO, &call->pingRequestTime);
6122 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6126 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6133 dpf(("CONG peer %lx/%u: sample (%s) size %ld, %ld ms (to %lu.%06lu, rtt %u, ps %u)", ntohl(peer->host), ntohs(peer->port), (ackReason == RX_ACK_REQUESTED ? "dataack" : "pingack"), xferSize, xferMs, peer->timeout.sec, peer->timeout.usec, peer->smRtt, peer->ifMTU));
6135 /* Track only packets that are big enough. */
6136 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6140 /* absorb RTT data (in milliseconds) for these big packets */
6141 if (peer->smRtt == 0) {
6142 peer->smRtt = xferMs;
6144 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6149 if (peer->countDown) {
6153 peer->countDown = 10; /* recalculate only every so often */
6155 /* In practice, we can measure only the RTT for full packets,
6156 * because of the way Rx acks the data that it receives. (If it's
6157 * smaller than a full packet, it often gets implicitly acked
6158 * either by the call response (from a server) or by the next call
6159 * (from a client), and either case confuses transmission times
6160 * with processing times.) Therefore, replace the above
6161 * more-sophisticated processing with a simpler version, where the
6162 * smoothed RTT is kept for full-size packets, and the time to
6163 * transmit a windowful of full-size packets is simply RTT *
6164 * windowSize. Again, we take two steps:
6165 - ensure the timeout is large enough for a single packet's RTT;
6166 - ensure that the window is small enough to fit in the desired timeout.*/
6168 /* First, the timeout check. */
6169 minTime = peer->smRtt;
6170 /* Get a reasonable estimate for a timeout period */
6172 newTO.sec = minTime / 1000;
6173 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6175 /* Increase the timeout period so that we can always do at least
6176 * one packet exchange */
6177 if (clock_Gt(&newTO, &peer->timeout)) {
6179 dpf(("CONG peer %lx/%u: timeout %lu.%06lu ==> %lu.%06lu (rtt %u, ps %u)", ntohl(peer->host), ntohs(peer->port), peer->timeout.sec, peer->timeout.usec, newTO.sec, newTO.usec, peer->smRtt, peer->packetSize));
6181 peer->timeout = newTO;
6184 /* Now, get an estimate for the transmit window size. */
6185 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6186 /* Now, convert to the number of full packets that could fit in a
6187 * reasonable fraction of that interval */
6188 minTime /= (peer->smRtt << 1);
6189 xferSize = minTime; /* (make a copy) */
6191 /* Now clamp the size to reasonable bounds. */
6194 else if (minTime > rx_Window)
6195 minTime = rx_Window;
6196 /* if (minTime != peer->maxWindow) {
6197 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6198 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6199 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6201 peer->maxWindow = minTime;
6202 elide... call->twind = minTime;
6206 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6207 * Discern this by calculating the timeout necessary for rx_Window
6209 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6210 /* calculate estimate for transmission interval in milliseconds */
6211 minTime = rx_Window * peer->smRtt;
6212 if (minTime < 1000) {
6213 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
6214 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6215 peer->timeout.usec, peer->smRtt, peer->packetSize));
6217 newTO.sec = 0; /* cut back on timeout by half a second */
6218 newTO.usec = 500000;
6219 clock_Sub(&peer->timeout, &newTO);
6224 } /* end of rxi_ComputeRate */
6225 #endif /* ADAPT_WINDOW */
6233 #define TRACE_OPTION_DEBUGLOG 4
6241 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6242 0, KEY_QUERY_VALUE, &parmKey);
6243 if (code != ERROR_SUCCESS)
6246 dummyLen = sizeof(TraceOption);
6247 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6248 (BYTE *) &TraceOption, &dummyLen);
6249 if (code == ERROR_SUCCESS) {
6250 rxdebug_active = (TraceOption & TRACE_OPTION_DEBUGLOG) ? 1 : 0;
6252 RegCloseKey (parmKey);
6253 #endif /* AFS_NT40_ENV */
6258 rx_DebugOnOff(int on)
6260 rxdebug_active = on;
6262 #endif /* AFS_NT40_ENV */
6265 /* Don't call this debugging routine directly; use dpf */
6267 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
6268 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
6276 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6279 len = _snprintf(msg, sizeof(msg)-2,
6280 tformat, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10,
6281 a11, a12, a13, a14, a15);
6283 if (msg[len-1] != '\n') {
6287 OutputDebugString(msg);
6292 clock_GetTime(&now);
6293 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
6294 (unsigned int)now.usec / 1000);
6295 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
6302 * This function is used to process the rx_stats structure that is local
6303 * to a process as well as an rx_stats structure received from a remote
6304 * process (via rxdebug). Therefore, it needs to do minimal version
6308 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
6309 afs_int32 freePackets, char version)
6313 if (size != sizeof(struct rx_stats)) {
6315 "Unexpected size of stats structure: was %d, expected %d\n",
6316 size, sizeof(struct rx_stats));
6319 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6322 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6323 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6324 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6325 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6326 s->specialPktAllocFailures);
6328 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6329 s->receivePktAllocFailures, s->sendPktAllocFailures,
6330 s->specialPktAllocFailures);
6334 " greedy %d, " "bogusReads %d (last from host %x), "
6335 "noPackets %d, " "noBuffers %d, " "selects %d, "
6336 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6337 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6338 s->selects, s->sendSelects);
6340 fprintf(file, " packets read: ");
6341 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6342 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6344 fprintf(file, "\n");
6347 " other read counters: data %d, " "ack %d, " "dup %d "
6348 "spurious %d " "dally %d\n", s->dataPacketsRead,
6349 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6350 s->ignorePacketDally);
6352 fprintf(file, " packets sent: ");
6353 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6354 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6356 fprintf(file, "\n");
6359 " other send counters: ack %d, " "data %d (not resends), "
6360 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6361 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6362 s->dataPacketsPushed, s->ignoreAckedPacket);
6365 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6366 s->netSendFailures, (int)s->fatalErrors);
6368 if (s->nRttSamples) {
6369 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6370 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6372 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6373 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6377 " %d server connections, " "%d client connections, "
6378 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6379 s->nServerConns, s->nClientConns, s->nPeerStructs,
6380 s->nCallStructs, s->nFreeCallStructs);
6382 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6383 fprintf(file, " %d clock updates\n", clock_nUpdates);
6388 /* for backward compatibility */
6390 rx_PrintStats(FILE * file)
6392 MUTEX_ENTER(&rx_stats_mutex);
6393 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6395 MUTEX_EXIT(&rx_stats_mutex);
6399 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6401 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6402 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6403 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6406 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6407 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6408 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6411 " Packet size %d, " "max in packet skew %d, "
6412 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6413 (int)peer->outPacketSkew);
6416 #ifdef AFS_PTHREAD_ENV
6418 * This mutex protects the following static variables:
6422 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6423 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6425 #define LOCK_RX_DEBUG
6426 #define UNLOCK_RX_DEBUG
6427 #endif /* AFS_PTHREAD_ENV */
6430 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6431 u_char type, void *inputData, size_t inputLength,
6432 void *outputData, size_t outputLength)
6434 static afs_int32 counter = 100;
6435 time_t waitTime, waitCount, startTime;
6436 struct rx_header theader;
6438 register afs_int32 code;
6439 struct timeval tv_now, tv_wake, tv_delta;
6440 struct sockaddr_in taddr, faddr;
6445 startTime = time(0);
6451 tp = &tbuffer[sizeof(struct rx_header)];
6452 taddr.sin_family = AF_INET;
6453 taddr.sin_port = remotePort;
6454 taddr.sin_addr.s_addr = remoteAddr;
6455 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6456 taddr.sin_len = sizeof(struct sockaddr_in);
6459 memset(&theader, 0, sizeof(theader));
6460 theader.epoch = htonl(999);
6462 theader.callNumber = htonl(counter);
6465 theader.type = type;
6466 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6467 theader.serviceId = 0;
6469 memcpy(tbuffer, &theader, sizeof(theader));
6470 memcpy(tp, inputData, inputLength);
6472 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6473 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6475 /* see if there's a packet available */
6476 gettimeofday(&tv_wake,0);
6477 tv_wake.tv_sec += waitTime;
6480 FD_SET(socket, &imask);
6481 tv_delta.tv_sec = tv_wake.tv_sec;
6482 tv_delta.tv_usec = tv_wake.tv_usec;
6483 gettimeofday(&tv_now, 0);
6485 if (tv_delta.tv_usec < tv_now.tv_usec) {
6487 tv_delta.tv_usec += 1000000;
6490 tv_delta.tv_usec -= tv_now.tv_usec;
6492 if (tv_delta.tv_sec < tv_now.tv_sec) {
6496 tv_delta.tv_sec -= tv_now.tv_sec;
6498 code = select(socket + 1, &imask, 0, 0, &tv_delta);
6499 if (code == 1 && FD_ISSET(socket, &imask)) {
6500 /* now receive a packet */
6501 faddrLen = sizeof(struct sockaddr_in);
6503 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6504 (struct sockaddr *)&faddr, &faddrLen);
6507 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6508 if (counter == ntohl(theader.callNumber))
6516 /* see if we've timed out */
6524 code -= sizeof(struct rx_header);
6525 if (code > outputLength)
6526 code = outputLength;
6527 memcpy(outputData, tp, code);
6532 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6533 afs_uint16 remotePort, struct rx_debugStats * stat,
6534 afs_uint32 * supportedValues)
6536 struct rx_debugIn in;
6539 *supportedValues = 0;
6540 in.type = htonl(RX_DEBUGI_GETSTATS);
6543 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6544 &in, sizeof(in), stat, sizeof(*stat));
6547 * If the call was successful, fixup the version and indicate
6548 * what contents of the stat structure are valid.
6549 * Also do net to host conversion of fields here.
6553 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6554 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6556 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6557 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6559 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6560 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6562 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6563 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6565 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6566 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6568 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6569 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6571 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6572 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6574 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6575 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6578 stat->nFreePackets = ntohl(stat->nFreePackets);
6579 stat->packetReclaims = ntohl(stat->packetReclaims);
6580 stat->callsExecuted = ntohl(stat->callsExecuted);
6581 stat->nWaiting = ntohl(stat->nWaiting);
6582 stat->idleThreads = ntohl(stat->idleThreads);
6589 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6590 afs_uint16 remotePort, struct rx_stats * stat,
6591 afs_uint32 * supportedValues)
6593 struct rx_debugIn in;
6594 afs_int32 *lp = (afs_int32 *) stat;
6599 * supportedValues is currently unused, but added to allow future
6600 * versioning of this function.
6603 *supportedValues = 0;
6604 in.type = htonl(RX_DEBUGI_RXSTATS);
6606 memset(stat, 0, sizeof(*stat));
6608 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6609 &in, sizeof(in), stat, sizeof(*stat));
6614 * Do net to host conversion here
6617 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6626 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6627 afs_uint16 remotePort, size_t version_length,
6631 return MakeDebugCall(socket, remoteAddr, remotePort,
6632 RX_PACKET_TYPE_VERSION, a, 1, version,
6637 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6638 afs_uint16 remotePort, afs_int32 * nextConnection,
6639 int allConnections, afs_uint32 debugSupportedValues,
6640 struct rx_debugConn * conn,
6641 afs_uint32 * supportedValues)
6643 struct rx_debugIn in;
6648 * supportedValues is currently unused, but added to allow future
6649 * versioning of this function.
6652 *supportedValues = 0;
6653 if (allConnections) {
6654 in.type = htonl(RX_DEBUGI_GETALLCONN);
6656 in.type = htonl(RX_DEBUGI_GETCONN);
6658 in.index = htonl(*nextConnection);
6659 memset(conn, 0, sizeof(*conn));
6661 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6662 &in, sizeof(in), conn, sizeof(*conn));
6665 *nextConnection += 1;
6668 * Convert old connection format to new structure.
6671 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6672 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6673 #define MOVEvL(a) (conn->a = vL->a)
6675 /* any old or unrecognized version... */
6676 for (i = 0; i < RX_MAXCALLS; i++) {
6677 MOVEvL(callState[i]);
6678 MOVEvL(callMode[i]);
6679 MOVEvL(callFlags[i]);
6680 MOVEvL(callOther[i]);
6682 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6683 MOVEvL(secStats.type);
6684 MOVEvL(secStats.level);
6685 MOVEvL(secStats.flags);
6686 MOVEvL(secStats.expires);
6687 MOVEvL(secStats.packetsReceived);
6688 MOVEvL(secStats.packetsSent);
6689 MOVEvL(secStats.bytesReceived);
6690 MOVEvL(secStats.bytesSent);
6695 * Do net to host conversion here
6697 * I don't convert host or port since we are most likely
6698 * going to want these in NBO.
6700 conn->cid = ntohl(conn->cid);
6701 conn->serial = ntohl(conn->serial);
6702 for (i = 0; i < RX_MAXCALLS; i++) {
6703 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6705 conn->error = ntohl(conn->error);
6706 conn->secStats.flags = ntohl(conn->secStats.flags);
6707 conn->secStats.expires = ntohl(conn->secStats.expires);
6708 conn->secStats.packetsReceived =
6709 ntohl(conn->secStats.packetsReceived);
6710 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6711 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6712 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6713 conn->epoch = ntohl(conn->epoch);
6714 conn->natMTU = ntohl(conn->natMTU);
6721 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6722 afs_uint16 remotePort, afs_int32 * nextPeer,
6723 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6724 afs_uint32 * supportedValues)
6726 struct rx_debugIn in;
6730 * supportedValues is currently unused, but added to allow future
6731 * versioning of this function.
6734 *supportedValues = 0;
6735 in.type = htonl(RX_DEBUGI_GETPEER);
6736 in.index = htonl(*nextPeer);
6737 memset(peer, 0, sizeof(*peer));
6739 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6740 &in, sizeof(in), peer, sizeof(*peer));
6746 * Do net to host conversion here
6748 * I don't convert host or port since we are most likely
6749 * going to want these in NBO.
6751 peer->ifMTU = ntohs(peer->ifMTU);
6752 peer->idleWhen = ntohl(peer->idleWhen);
6753 peer->refCount = ntohs(peer->refCount);
6754 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6755 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6756 peer->rtt = ntohl(peer->rtt);
6757 peer->rtt_dev = ntohl(peer->rtt_dev);
6758 peer->timeout.sec = ntohl(peer->timeout.sec);
6759 peer->timeout.usec = ntohl(peer->timeout.usec);
6760 peer->nSent = ntohl(peer->nSent);
6761 peer->reSends = ntohl(peer->reSends);
6762 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6763 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6764 peer->rateFlag = ntohl(peer->rateFlag);
6765 peer->natMTU = ntohs(peer->natMTU);
6766 peer->maxMTU = ntohs(peer->maxMTU);
6767 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6768 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6769 peer->MTU = ntohs(peer->MTU);
6770 peer->cwind = ntohs(peer->cwind);
6771 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6772 peer->congestSeq = ntohs(peer->congestSeq);
6773 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6774 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6775 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6776 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6781 #endif /* RXDEBUG */
6786 struct rx_serverQueueEntry *np;
6789 register struct rx_call *call;
6790 register struct rx_serverQueueEntry *sq;
6794 if (rxinit_status == 1) {
6796 return; /* Already shutdown. */
6800 #ifndef AFS_PTHREAD_ENV
6801 FD_ZERO(&rx_selectMask);
6802 #endif /* AFS_PTHREAD_ENV */
6803 rxi_dataQuota = RX_MAX_QUOTA;
6804 #ifndef AFS_PTHREAD_ENV
6806 #endif /* AFS_PTHREAD_ENV */
6809 #ifndef AFS_PTHREAD_ENV
6810 #ifndef AFS_USE_GETTIMEOFDAY
6812 #endif /* AFS_USE_GETTIMEOFDAY */
6813 #endif /* AFS_PTHREAD_ENV */
6815 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6816 call = queue_First(&rx_freeCallQueue, rx_call);
6818 rxi_Free(call, sizeof(struct rx_call));
6821 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6822 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6828 struct rx_peer **peer_ptr, **peer_end;
6829 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6830 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6832 struct rx_peer *peer, *next;
6833 for (peer = *peer_ptr; peer; peer = next) {
6834 rx_interface_stat_p rpc_stat, nrpc_stat;
6837 (&peer->rpcStats, rpc_stat, nrpc_stat,
6838 rx_interface_stat)) {
6839 unsigned int num_funcs;
6842 queue_Remove(&rpc_stat->queue_header);
6843 queue_Remove(&rpc_stat->all_peers);
6844 num_funcs = rpc_stat->stats[0].func_total;
6846 sizeof(rx_interface_stat_t) +
6847 rpc_stat->stats[0].func_total *
6848 sizeof(rx_function_entry_v1_t);
6850 rxi_Free(rpc_stat, space);
6851 MUTEX_ENTER(&rx_rpc_stats);
6852 rxi_rpc_peer_stat_cnt -= num_funcs;
6853 MUTEX_EXIT(&rx_rpc_stats);
6857 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6861 for (i = 0; i < RX_MAX_SERVICES; i++) {
6863 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6865 for (i = 0; i < rx_hashTableSize; i++) {
6866 register struct rx_connection *tc, *ntc;
6867 MUTEX_ENTER(&rx_connHashTable_lock);
6868 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6870 for (j = 0; j < RX_MAXCALLS; j++) {
6872 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6875 rxi_Free(tc, sizeof(*tc));
6877 MUTEX_EXIT(&rx_connHashTable_lock);
6880 MUTEX_ENTER(&freeSQEList_lock);
6882 while ((np = rx_FreeSQEList)) {
6883 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6884 MUTEX_DESTROY(&np->lock);
6885 rxi_Free(np, sizeof(*np));
6888 MUTEX_EXIT(&freeSQEList_lock);
6889 MUTEX_DESTROY(&freeSQEList_lock);
6890 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6891 MUTEX_DESTROY(&rx_connHashTable_lock);
6892 MUTEX_DESTROY(&rx_peerHashTable_lock);
6893 MUTEX_DESTROY(&rx_serverPool_lock);
6895 osi_Free(rx_connHashTable,
6896 rx_hashTableSize * sizeof(struct rx_connection *));
6897 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6899 UNPIN(rx_connHashTable,
6900 rx_hashTableSize * sizeof(struct rx_connection *));
6901 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6903 rxi_FreeAllPackets();
6905 MUTEX_ENTER(&rx_stats_mutex);
6906 rxi_dataQuota = RX_MAX_QUOTA;
6907 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6908 MUTEX_EXIT(&rx_stats_mutex);
6914 #ifdef RX_ENABLE_LOCKS
6916 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6918 if (!MUTEX_ISMINE(lockaddr))
6919 osi_Panic("Lock not held: %s", msg);
6921 #endif /* RX_ENABLE_LOCKS */
6926 * Routines to implement connection specific data.
6930 rx_KeyCreate(rx_destructor_t rtn)
6933 MUTEX_ENTER(&rxi_keyCreate_lock);
6934 key = rxi_keyCreate_counter++;
6935 rxi_keyCreate_destructor = (rx_destructor_t *)
6936 realloc((void *)rxi_keyCreate_destructor,
6937 (key + 1) * sizeof(rx_destructor_t));
6938 rxi_keyCreate_destructor[key] = rtn;
6939 MUTEX_EXIT(&rxi_keyCreate_lock);
6944 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6947 MUTEX_ENTER(&conn->conn_data_lock);
6948 if (!conn->specific) {
6949 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6950 for (i = 0; i < key; i++)
6951 conn->specific[i] = NULL;
6952 conn->nSpecific = key + 1;
6953 conn->specific[key] = ptr;
6954 } else if (key >= conn->nSpecific) {
6955 conn->specific = (void **)
6956 realloc(conn->specific, (key + 1) * sizeof(void *));
6957 for (i = conn->nSpecific; i < key; i++)
6958 conn->specific[i] = NULL;
6959 conn->nSpecific = key + 1;
6960 conn->specific[key] = ptr;
6962 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6963 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6964 conn->specific[key] = ptr;
6966 MUTEX_EXIT(&conn->conn_data_lock);
6970 rx_GetSpecific(struct rx_connection *conn, int key)
6973 MUTEX_ENTER(&conn->conn_data_lock);
6974 if (key >= conn->nSpecific)
6977 ptr = conn->specific[key];
6978 MUTEX_EXIT(&conn->conn_data_lock);
6982 #endif /* !KERNEL */
6985 * processStats is a queue used to store the statistics for the local
6986 * process. Its contents are similar to the contents of the rpcStats
6987 * queue on a rx_peer structure, but the actual data stored within
6988 * this queue contains totals across the lifetime of the process (assuming
6989 * the stats have not been reset) - unlike the per peer structures
6990 * which can come and go based upon the peer lifetime.
6993 static struct rx_queue processStats = { &processStats, &processStats };
6996 * peerStats is a queue used to store the statistics for all peer structs.
6997 * Its contents are the union of all the peer rpcStats queues.
7000 static struct rx_queue peerStats = { &peerStats, &peerStats };
7003 * rxi_monitor_processStats is used to turn process wide stat collection
7007 static int rxi_monitor_processStats = 0;
7010 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
7013 static int rxi_monitor_peerStats = 0;
7016 * rxi_AddRpcStat - given all of the information for a particular rpc
7017 * call, create (if needed) and update the stat totals for the rpc.
7021 * IN stats - the queue of stats that will be updated with the new value
7023 * IN rxInterface - a unique number that identifies the rpc interface
7025 * IN currentFunc - the index of the function being invoked
7027 * IN totalFunc - the total number of functions in this interface
7029 * IN queueTime - the amount of time this function waited for a thread
7031 * IN execTime - the amount of time this function invocation took to execute
7033 * IN bytesSent - the number bytes sent by this invocation
7035 * IN bytesRcvd - the number bytes received by this invocation
7037 * IN isServer - if true, this invocation was made to a server
7039 * IN remoteHost - the ip address of the remote host
7041 * IN remotePort - the port of the remote host
7043 * IN addToPeerList - if != 0, add newly created stat to the global peer list
7045 * INOUT counter - if a new stats structure is allocated, the counter will
7046 * be updated with the new number of allocated stat structures
7054 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
7055 afs_uint32 currentFunc, afs_uint32 totalFunc,
7056 struct clock *queueTime, struct clock *execTime,
7057 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
7058 afs_uint32 remoteHost, afs_uint32 remotePort,
7059 int addToPeerList, unsigned int *counter)
7062 rx_interface_stat_p rpc_stat, nrpc_stat;
7065 * See if there's already a structure for this interface
7068 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7069 if ((rpc_stat->stats[0].interfaceId == rxInterface)
7070 && (rpc_stat->stats[0].remote_is_server == isServer))
7075 * Didn't find a match so allocate a new structure and add it to the
7079 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
7080 || (rpc_stat->stats[0].interfaceId != rxInterface)
7081 || (rpc_stat->stats[0].remote_is_server != isServer)) {
7086 sizeof(rx_interface_stat_t) +
7087 totalFunc * sizeof(rx_function_entry_v1_t);
7089 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
7090 if (rpc_stat == NULL) {
7094 *counter += totalFunc;
7095 for (i = 0; i < totalFunc; i++) {
7096 rpc_stat->stats[i].remote_peer = remoteHost;
7097 rpc_stat->stats[i].remote_port = remotePort;
7098 rpc_stat->stats[i].remote_is_server = isServer;
7099 rpc_stat->stats[i].interfaceId = rxInterface;
7100 rpc_stat->stats[i].func_total = totalFunc;
7101 rpc_stat->stats[i].func_index = i;
7102 hzero(rpc_stat->stats[i].invocations);
7103 hzero(rpc_stat->stats[i].bytes_sent);
7104 hzero(rpc_stat->stats[i].bytes_rcvd);
7105 rpc_stat->stats[i].queue_time_sum.sec = 0;
7106 rpc_stat->stats[i].queue_time_sum.usec = 0;
7107 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7108 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7109 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7110 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7111 rpc_stat->stats[i].queue_time_max.sec = 0;
7112 rpc_stat->stats[i].queue_time_max.usec = 0;
7113 rpc_stat->stats[i].execution_time_sum.sec = 0;
7114 rpc_stat->stats[i].execution_time_sum.usec = 0;
7115 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7116 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7117 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7118 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7119 rpc_stat->stats[i].execution_time_max.sec = 0;
7120 rpc_stat->stats[i].execution_time_max.usec = 0;
7122 queue_Prepend(stats, rpc_stat);
7123 if (addToPeerList) {
7124 queue_Prepend(&peerStats, &rpc_stat->all_peers);
7129 * Increment the stats for this function
7132 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
7133 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
7134 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
7135 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
7136 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7137 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7138 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7140 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7141 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7143 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7144 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7146 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7147 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7149 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7150 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7158 * rx_IncrementTimeAndCount - increment the times and count for a particular
7163 * IN peer - the peer who invoked the rpc
7165 * IN rxInterface - a unique number that identifies the rpc interface
7167 * IN currentFunc - the index of the function being invoked
7169 * IN totalFunc - the total number of functions in this interface
7171 * IN queueTime - the amount of time this function waited for a thread
7173 * IN execTime - the amount of time this function invocation took to execute
7175 * IN bytesSent - the number bytes sent by this invocation
7177 * IN bytesRcvd - the number bytes received by this invocation
7179 * IN isServer - if true, this invocation was made to a server
7187 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7188 afs_uint32 currentFunc, afs_uint32 totalFunc,
7189 struct clock *queueTime, struct clock *execTime,
7190 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7194 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
7197 MUTEX_ENTER(&rx_rpc_stats);
7198 MUTEX_ENTER(&peer->peer_lock);
7200 if (rxi_monitor_peerStats) {
7201 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7202 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7203 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7206 if (rxi_monitor_processStats) {
7207 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7208 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7209 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7212 MUTEX_EXIT(&peer->peer_lock);
7213 MUTEX_EXIT(&rx_rpc_stats);
7218 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7222 * IN callerVersion - the rpc stat version of the caller.
7224 * IN count - the number of entries to marshall.
7226 * IN stats - pointer to stats to be marshalled.
7228 * OUT ptr - Where to store the marshalled data.
7235 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7236 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7242 * We only support the first version
7244 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7245 *(ptr++) = stats->remote_peer;
7246 *(ptr++) = stats->remote_port;
7247 *(ptr++) = stats->remote_is_server;
7248 *(ptr++) = stats->interfaceId;
7249 *(ptr++) = stats->func_total;
7250 *(ptr++) = stats->func_index;
7251 *(ptr++) = hgethi(stats->invocations);
7252 *(ptr++) = hgetlo(stats->invocations);
7253 *(ptr++) = hgethi(stats->bytes_sent);
7254 *(ptr++) = hgetlo(stats->bytes_sent);
7255 *(ptr++) = hgethi(stats->bytes_rcvd);
7256 *(ptr++) = hgetlo(stats->bytes_rcvd);
7257 *(ptr++) = stats->queue_time_sum.sec;
7258 *(ptr++) = stats->queue_time_sum.usec;
7259 *(ptr++) = stats->queue_time_sum_sqr.sec;
7260 *(ptr++) = stats->queue_time_sum_sqr.usec;
7261 *(ptr++) = stats->queue_time_min.sec;
7262 *(ptr++) = stats->queue_time_min.usec;
7263 *(ptr++) = stats->queue_time_max.sec;
7264 *(ptr++) = stats->queue_time_max.usec;
7265 *(ptr++) = stats->execution_time_sum.sec;
7266 *(ptr++) = stats->execution_time_sum.usec;
7267 *(ptr++) = stats->execution_time_sum_sqr.sec;
7268 *(ptr++) = stats->execution_time_sum_sqr.usec;
7269 *(ptr++) = stats->execution_time_min.sec;
7270 *(ptr++) = stats->execution_time_min.usec;
7271 *(ptr++) = stats->execution_time_max.sec;
7272 *(ptr++) = stats->execution_time_max.usec;
7278 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7283 * IN callerVersion - the rpc stat version of the caller
7285 * OUT myVersion - the rpc stat version of this function
7287 * OUT clock_sec - local time seconds
7289 * OUT clock_usec - local time microseconds
7291 * OUT allocSize - the number of bytes allocated to contain stats
7293 * OUT statCount - the number stats retrieved from this process.
7295 * OUT stats - the actual stats retrieved from this process.
7299 * Returns void. If successful, stats will != NULL.
7303 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7304 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7305 size_t * allocSize, afs_uint32 * statCount,
7306 afs_uint32 ** stats)
7316 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7319 * Check to see if stats are enabled
7322 MUTEX_ENTER(&rx_rpc_stats);
7323 if (!rxi_monitor_processStats) {
7324 MUTEX_EXIT(&rx_rpc_stats);
7328 clock_GetTime(&now);
7329 *clock_sec = now.sec;
7330 *clock_usec = now.usec;
7333 * Allocate the space based upon the caller version
7335 * If the client is at an older version than we are,
7336 * we return the statistic data in the older data format, but
7337 * we still return our version number so the client knows we
7338 * are maintaining more data than it can retrieve.
7341 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7342 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7343 *statCount = rxi_rpc_process_stat_cnt;
7346 * This can't happen yet, but in the future version changes
7347 * can be handled by adding additional code here
7351 if (space > (size_t) 0) {
7353 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7356 rx_interface_stat_p rpc_stat, nrpc_stat;
7360 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7362 * Copy the data based upon the caller version
7364 rx_MarshallProcessRPCStats(callerVersion,
7365 rpc_stat->stats[0].func_total,
7366 rpc_stat->stats, &ptr);
7372 MUTEX_EXIT(&rx_rpc_stats);
7377 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7381 * IN callerVersion - the rpc stat version of the caller
7383 * OUT myVersion - the rpc stat version of this function
7385 * OUT clock_sec - local time seconds
7387 * OUT clock_usec - local time microseconds
7389 * OUT allocSize - the number of bytes allocated to contain stats
7391 * OUT statCount - the number of stats retrieved from the individual
7394 * OUT stats - the actual stats retrieved from the individual peer structures.
7398 * Returns void. If successful, stats will != NULL.
7402 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7403 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7404 size_t * allocSize, afs_uint32 * statCount,
7405 afs_uint32 ** stats)
7415 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7418 * Check to see if stats are enabled
7421 MUTEX_ENTER(&rx_rpc_stats);
7422 if (!rxi_monitor_peerStats) {
7423 MUTEX_EXIT(&rx_rpc_stats);
7427 clock_GetTime(&now);
7428 *clock_sec = now.sec;
7429 *clock_usec = now.usec;
7432 * Allocate the space based upon the caller version
7434 * If the client is at an older version than we are,
7435 * we return the statistic data in the older data format, but
7436 * we still return our version number so the client knows we
7437 * are maintaining more data than it can retrieve.
7440 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7441 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7442 *statCount = rxi_rpc_peer_stat_cnt;
7445 * This can't happen yet, but in the future version changes
7446 * can be handled by adding additional code here
7450 if (space > (size_t) 0) {
7452 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7455 rx_interface_stat_p rpc_stat, nrpc_stat;
7459 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7461 * We have to fix the offset of rpc_stat since we are
7462 * keeping this structure on two rx_queues. The rx_queue
7463 * package assumes that the rx_queue member is the first
7464 * member of the structure. That is, rx_queue assumes that
7465 * any one item is only on one queue at a time. We are
7466 * breaking that assumption and so we have to do a little
7467 * math to fix our pointers.
7470 fix_offset = (char *)rpc_stat;
7471 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7472 rpc_stat = (rx_interface_stat_p) fix_offset;
7475 * Copy the data based upon the caller version
7477 rx_MarshallProcessRPCStats(callerVersion,
7478 rpc_stat->stats[0].func_total,
7479 rpc_stat->stats, &ptr);
7485 MUTEX_EXIT(&rx_rpc_stats);
7490 * rx_FreeRPCStats - free memory allocated by
7491 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7495 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7496 * rx_RetrievePeerRPCStats
7498 * IN allocSize - the number of bytes in stats.
7506 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7508 rxi_Free(stats, allocSize);
7512 * rx_queryProcessRPCStats - see if process rpc stat collection is
7513 * currently enabled.
7519 * Returns 0 if stats are not enabled != 0 otherwise
7523 rx_queryProcessRPCStats(void)
7526 MUTEX_ENTER(&rx_rpc_stats);
7527 rc = rxi_monitor_processStats;
7528 MUTEX_EXIT(&rx_rpc_stats);
7533 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7539 * Returns 0 if stats are not enabled != 0 otherwise
7543 rx_queryPeerRPCStats(void)
7546 MUTEX_ENTER(&rx_rpc_stats);
7547 rc = rxi_monitor_peerStats;
7548 MUTEX_EXIT(&rx_rpc_stats);
7553 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7563 rx_enableProcessRPCStats(void)
7565 MUTEX_ENTER(&rx_rpc_stats);
7566 rx_enable_stats = 1;
7567 rxi_monitor_processStats = 1;
7568 MUTEX_EXIT(&rx_rpc_stats);
7572 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7582 rx_enablePeerRPCStats(void)
7584 MUTEX_ENTER(&rx_rpc_stats);
7585 rx_enable_stats = 1;
7586 rxi_monitor_peerStats = 1;
7587 MUTEX_EXIT(&rx_rpc_stats);
7591 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7601 rx_disableProcessRPCStats(void)
7603 rx_interface_stat_p rpc_stat, nrpc_stat;
7606 MUTEX_ENTER(&rx_rpc_stats);
7609 * Turn off process statistics and if peer stats is also off, turn
7613 rxi_monitor_processStats = 0;
7614 if (rxi_monitor_peerStats == 0) {
7615 rx_enable_stats = 0;
7618 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7619 unsigned int num_funcs = 0;
7622 queue_Remove(rpc_stat);
7623 num_funcs = rpc_stat->stats[0].func_total;
7625 sizeof(rx_interface_stat_t) +
7626 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7628 rxi_Free(rpc_stat, space);
7629 rxi_rpc_process_stat_cnt -= num_funcs;
7631 MUTEX_EXIT(&rx_rpc_stats);
7635 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7645 rx_disablePeerRPCStats(void)
7647 struct rx_peer **peer_ptr, **peer_end;
7650 MUTEX_ENTER(&rx_rpc_stats);
7653 * Turn off peer statistics and if process stats is also off, turn
7657 rxi_monitor_peerStats = 0;
7658 if (rxi_monitor_processStats == 0) {
7659 rx_enable_stats = 0;
7662 MUTEX_ENTER(&rx_peerHashTable_lock);
7663 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7664 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7666 struct rx_peer *peer, *next, *prev;
7667 for (prev = peer = *peer_ptr; peer; peer = next) {
7669 code = MUTEX_TRYENTER(&peer->peer_lock);
7671 rx_interface_stat_p rpc_stat, nrpc_stat;
7674 (&peer->rpcStats, rpc_stat, nrpc_stat,
7675 rx_interface_stat)) {
7676 unsigned int num_funcs = 0;
7679 queue_Remove(&rpc_stat->queue_header);
7680 queue_Remove(&rpc_stat->all_peers);
7681 num_funcs = rpc_stat->stats[0].func_total;
7683 sizeof(rx_interface_stat_t) +
7684 rpc_stat->stats[0].func_total *
7685 sizeof(rx_function_entry_v1_t);
7687 rxi_Free(rpc_stat, space);
7688 rxi_rpc_peer_stat_cnt -= num_funcs;
7690 MUTEX_EXIT(&peer->peer_lock);
7691 if (prev == *peer_ptr) {
7701 MUTEX_EXIT(&rx_peerHashTable_lock);
7702 MUTEX_EXIT(&rx_rpc_stats);
7706 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7711 * IN clearFlag - flag indicating which stats to clear
7719 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7721 rx_interface_stat_p rpc_stat, nrpc_stat;
7723 MUTEX_ENTER(&rx_rpc_stats);
7725 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7726 unsigned int num_funcs = 0, i;
7727 num_funcs = rpc_stat->stats[0].func_total;
7728 for (i = 0; i < num_funcs; i++) {
7729 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7730 hzero(rpc_stat->stats[i].invocations);
7732 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7733 hzero(rpc_stat->stats[i].bytes_sent);
7735 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7736 hzero(rpc_stat->stats[i].bytes_rcvd);
7738 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7739 rpc_stat->stats[i].queue_time_sum.sec = 0;
7740 rpc_stat->stats[i].queue_time_sum.usec = 0;
7742 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7743 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7744 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7746 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7747 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7748 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7750 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7751 rpc_stat->stats[i].queue_time_max.sec = 0;
7752 rpc_stat->stats[i].queue_time_max.usec = 0;
7754 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7755 rpc_stat->stats[i].execution_time_sum.sec = 0;
7756 rpc_stat->stats[i].execution_time_sum.usec = 0;
7758 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7759 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7760 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7762 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7763 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7764 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7766 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7767 rpc_stat->stats[i].execution_time_max.sec = 0;
7768 rpc_stat->stats[i].execution_time_max.usec = 0;
7773 MUTEX_EXIT(&rx_rpc_stats);
7777 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7782 * IN clearFlag - flag indicating which stats to clear
7790 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7792 rx_interface_stat_p rpc_stat, nrpc_stat;
7794 MUTEX_ENTER(&rx_rpc_stats);
7796 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7797 unsigned int num_funcs = 0, i;
7800 * We have to fix the offset of rpc_stat since we are
7801 * keeping this structure on two rx_queues. The rx_queue
7802 * package assumes that the rx_queue member is the first
7803 * member of the structure. That is, rx_queue assumes that
7804 * any one item is only on one queue at a time. We are
7805 * breaking that assumption and so we have to do a little
7806 * math to fix our pointers.
7809 fix_offset = (char *)rpc_stat;
7810 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7811 rpc_stat = (rx_interface_stat_p) fix_offset;
7813 num_funcs = rpc_stat->stats[0].func_total;
7814 for (i = 0; i < num_funcs; i++) {
7815 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7816 hzero(rpc_stat->stats[i].invocations);
7818 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7819 hzero(rpc_stat->stats[i].bytes_sent);
7821 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7822 hzero(rpc_stat->stats[i].bytes_rcvd);
7824 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7825 rpc_stat->stats[i].queue_time_sum.sec = 0;
7826 rpc_stat->stats[i].queue_time_sum.usec = 0;
7828 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7829 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7830 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7832 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7833 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7834 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7836 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7837 rpc_stat->stats[i].queue_time_max.sec = 0;
7838 rpc_stat->stats[i].queue_time_max.usec = 0;
7840 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7841 rpc_stat->stats[i].execution_time_sum.sec = 0;
7842 rpc_stat->stats[i].execution_time_sum.usec = 0;
7844 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7845 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7846 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7848 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7849 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7850 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7852 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7853 rpc_stat->stats[i].execution_time_max.sec = 0;
7854 rpc_stat->stats[i].execution_time_max.usec = 0;
7859 MUTEX_EXIT(&rx_rpc_stats);
7863 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7864 * is authorized to enable/disable/clear RX statistics.
7866 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7869 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7871 rxi_rxstat_userok = proc;
7875 rx_RxStatUserOk(struct rx_call *call)
7877 if (!rxi_rxstat_userok)
7879 return rxi_rxstat_userok(call);
7884 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
7885 * function in the MSVC runtime DLL (msvcrt.dll).
7887 * Note: the system serializes calls to this function.
7890 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
7891 DWORD reason, /* reason function is being called */
7892 LPVOID reserved) /* reserved for future use */
7895 case DLL_PROCESS_ATTACH:
7896 /* library is being attached to a process */
7900 case DLL_PROCESS_DETACH: