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
361 static int rxinit_status = 1;
362 #ifdef AFS_PTHREAD_ENV
364 * This mutex protects the following global variables:
368 #define LOCK_RX_INIT assert(pthread_mutex_lock(&rx_init_mutex)==0)
369 #define UNLOCK_RX_INIT assert(pthread_mutex_unlock(&rx_init_mutex)==0)
372 #define UNLOCK_RX_INIT
376 rx_InitHost(u_int host, u_int port)
383 char *htable, *ptable;
386 #if defined(AFS_DJGPP_ENV) && !defined(DEBUG)
387 __djgpp_set_quiet_socket(1);
394 if (rxinit_status == 0) {
395 tmp_status = rxinit_status;
397 return tmp_status; /* Already started; return previous error code. */
403 if (afs_winsockInit() < 0)
409 * Initialize anything necessary to provide a non-premptive threading
412 rxi_InitializeThreadSupport();
415 /* Allocate and initialize a socket for client and perhaps server
418 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
419 if (rx_socket == OSI_NULLSOCKET) {
423 #ifdef RX_ENABLE_LOCKS
426 #endif /* RX_LOCKS_DB */
427 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
428 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
429 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
430 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
431 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
433 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
435 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
437 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
439 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
441 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
443 #if defined(KERNEL) && defined(AFS_HPUX110_ENV)
445 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
446 #endif /* KERNEL && AFS_HPUX110_ENV */
447 #endif /* RX_ENABLE_LOCKS */
450 rx_connDeadTime = 12;
451 rx_tranquil = 0; /* reset flag */
452 memset((char *)&rx_stats, 0, sizeof(struct rx_stats));
454 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
455 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
456 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
457 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
458 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
459 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
461 /* Malloc up a bunch of packets & buffers */
463 queue_Init(&rx_freePacketQueue);
464 rxi_NeedMorePackets = FALSE;
465 #ifdef RX_ENABLE_TSFPQ
466 rx_nPackets = 0; /* in TSFPQ version, rx_nPackets is managed by rxi_MorePackets* */
467 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
468 #else /* RX_ENABLE_TSFPQ */
469 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
470 rxi_MorePackets(rx_nPackets);
471 #endif /* RX_ENABLE_TSFPQ */
478 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
479 tv.tv_sec = clock_now.sec;
480 tv.tv_usec = clock_now.usec;
481 srand((unsigned int)tv.tv_usec);
488 #if defined(KERNEL) && !defined(UKERNEL)
489 /* Really, this should never happen in a real kernel */
492 struct sockaddr_in addr;
493 int addrlen = sizeof(addr);
494 if (getsockname((int)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
498 rx_port = addr.sin_port;
501 rx_stats.minRtt.sec = 9999999;
503 rx_SetEpoch(tv.tv_sec | 0x80000000);
505 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
506 * will provide a randomer value. */
508 MUTEX_ENTER(&rx_stats_mutex);
509 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
510 MUTEX_EXIT(&rx_stats_mutex);
511 /* *Slightly* random start time for the cid. This is just to help
512 * out with the hashing function at the peer */
513 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
514 rx_connHashTable = (struct rx_connection **)htable;
515 rx_peerHashTable = (struct rx_peer **)ptable;
517 rx_lastAckDelay.sec = 0;
518 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
519 rx_hardAckDelay.sec = 0;
520 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
521 rx_softAckDelay.sec = 0;
522 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
524 rxevent_Init(20, rxi_ReScheduleEvents);
526 /* Initialize various global queues */
527 queue_Init(&rx_idleServerQueue);
528 queue_Init(&rx_incomingCallQueue);
529 queue_Init(&rx_freeCallQueue);
531 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
532 /* Initialize our list of usable IP addresses. */
536 /* Start listener process (exact function is dependent on the
537 * implementation environment--kernel or user space) */
541 tmp_status = rxinit_status = 0;
549 return rx_InitHost(htonl(INADDR_ANY), port);
552 /* called with unincremented nRequestsRunning to see if it is OK to start
553 * a new thread in this service. Could be "no" for two reasons: over the
554 * max quota, or would prevent others from reaching their min quota.
556 #ifdef RX_ENABLE_LOCKS
557 /* This verion of QuotaOK reserves quota if it's ok while the
558 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
561 QuotaOK(register struct rx_service *aservice)
563 /* check if over max quota */
564 if (aservice->nRequestsRunning >= aservice->maxProcs) {
568 /* under min quota, we're OK */
569 /* otherwise, can use only if there are enough to allow everyone
570 * to go to their min quota after this guy starts.
572 MUTEX_ENTER(&rx_stats_mutex);
573 if ((aservice->nRequestsRunning < aservice->minProcs)
574 || (rxi_availProcs > rxi_minDeficit)) {
575 aservice->nRequestsRunning++;
576 /* just started call in minProcs pool, need fewer to maintain
578 if (aservice->nRequestsRunning <= aservice->minProcs)
581 MUTEX_EXIT(&rx_stats_mutex);
584 MUTEX_EXIT(&rx_stats_mutex);
590 ReturnToServerPool(register struct rx_service *aservice)
592 aservice->nRequestsRunning--;
593 MUTEX_ENTER(&rx_stats_mutex);
594 if (aservice->nRequestsRunning < aservice->minProcs)
597 MUTEX_EXIT(&rx_stats_mutex);
600 #else /* RX_ENABLE_LOCKS */
602 QuotaOK(register struct rx_service *aservice)
605 /* under min quota, we're OK */
606 if (aservice->nRequestsRunning < aservice->minProcs)
609 /* check if over max quota */
610 if (aservice->nRequestsRunning >= aservice->maxProcs)
613 /* otherwise, can use only if there are enough to allow everyone
614 * to go to their min quota after this guy starts.
616 if (rxi_availProcs > rxi_minDeficit)
620 #endif /* RX_ENABLE_LOCKS */
623 /* Called by rx_StartServer to start up lwp's to service calls.
624 NExistingProcs gives the number of procs already existing, and which
625 therefore needn't be created. */
627 rxi_StartServerProcs(int nExistingProcs)
629 register struct rx_service *service;
634 /* For each service, reserve N processes, where N is the "minimum"
635 * number of processes that MUST be able to execute a request in parallel,
636 * at any time, for that process. Also compute the maximum difference
637 * between any service's maximum number of processes that can run
638 * (i.e. the maximum number that ever will be run, and a guarantee
639 * that this number will run if other services aren't running), and its
640 * minimum number. The result is the extra number of processes that
641 * we need in order to provide the latter guarantee */
642 for (i = 0; i < RX_MAX_SERVICES; i++) {
644 service = rx_services[i];
645 if (service == (struct rx_service *)0)
647 nProcs += service->minProcs;
648 diff = service->maxProcs - service->minProcs;
652 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
653 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
654 for (i = 0; i < nProcs; i++) {
655 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
661 /* This routine is only required on Windows */
663 rx_StartClientThread(void)
665 #ifdef AFS_PTHREAD_ENV
667 pid = pthread_self();
668 #endif /* AFS_PTHREAD_ENV */
670 #endif /* AFS_NT40_ENV */
672 /* This routine must be called if any services are exported. If the
673 * donateMe flag is set, the calling process is donated to the server
676 rx_StartServer(int donateMe)
678 register struct rx_service *service;
684 /* Start server processes, if necessary (exact function is dependent
685 * on the implementation environment--kernel or user space). DonateMe
686 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
687 * case, one less new proc will be created rx_StartServerProcs.
689 rxi_StartServerProcs(donateMe);
691 /* count up the # of threads in minProcs, and add set the min deficit to
692 * be that value, too.
694 for (i = 0; i < RX_MAX_SERVICES; i++) {
695 service = rx_services[i];
696 if (service == (struct rx_service *)0)
698 MUTEX_ENTER(&rx_stats_mutex);
699 rxi_totalMin += service->minProcs;
700 /* below works even if a thread is running, since minDeficit would
701 * still have been decremented and later re-incremented.
703 rxi_minDeficit += service->minProcs;
704 MUTEX_EXIT(&rx_stats_mutex);
707 /* Turn on reaping of idle server connections */
708 rxi_ReapConnections();
717 #ifdef AFS_PTHREAD_ENV
719 pid = (pid_t) pthread_self();
720 #else /* AFS_PTHREAD_ENV */
722 LWP_CurrentProcess(&pid);
723 #endif /* AFS_PTHREAD_ENV */
725 sprintf(name, "srv_%d", ++nProcs);
727 (*registerProgram) (pid, name);
729 #endif /* AFS_NT40_ENV */
730 rx_ServerProc(NULL); /* Never returns */
732 #ifdef RX_ENABLE_TSFPQ
733 /* no use leaving packets around in this thread's local queue if
734 * it isn't getting donated to the server thread pool.
736 rxi_FlushLocalPacketsTSFPQ();
737 #endif /* RX_ENABLE_TSFPQ */
741 /* Create a new client connection to the specified service, using the
742 * specified security object to implement the security model for this
744 struct rx_connection *
745 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
746 register struct rx_securityClass *securityObject,
747 int serviceSecurityIndex)
750 afs_int32 cid, cix, nclones;
751 register struct rx_connection *conn, *tconn, *ptconn;
756 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", ntohl(shost), ntohs(sport), sservice, securityObject, serviceSecurityIndex));
759 nclones = rx_max_clones_per_connection;
761 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
762 * the case of kmem_alloc? */
765 MUTEX_ENTER(&rx_connHashTable_lock);
767 /* send in the clones */
768 for(cix = 0; cix <= nclones; ++cix) {
771 tconn = rxi_AllocConnection();
772 tconn->type = RX_CLIENT_CONNECTION;
773 tconn->epoch = rx_epoch;
774 tconn->peer = rxi_FindPeer(shost, sport, 0, 1);
775 tconn->serviceId = sservice;
776 tconn->securityObject = securityObject;
777 tconn->securityData = (void *) 0;
778 tconn->securityIndex = serviceSecurityIndex;
779 tconn->ackRate = RX_FAST_ACK_RATE;
780 tconn->nSpecific = 0;
781 tconn->specific = NULL;
782 tconn->challengeEvent = NULL;
783 tconn->delayedAbortEvent = NULL;
784 tconn->abortCount = 0;
786 for (i = 0; i < RX_MAXCALLS; i++) {
787 tconn->twind[i] = rx_initSendWindow;
788 tconn->rwind[i] = rx_initReceiveWindow;
791 tconn->next_clone = 0;
792 tconn->nclones = nclones;
793 rx_SetConnDeadTime(tconn, rx_connDeadTime);
798 tconn->flags |= RX_CLONED_CONNECTION;
799 tconn->parent = conn;
800 ptconn->next_clone = tconn;
803 /* generic connection setup */
804 #ifdef RX_ENABLE_LOCKS
805 MUTEX_INIT(&tconn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
806 MUTEX_INIT(&tconn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
807 CV_INIT(&tconn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
809 cid = (rx_nextCid += RX_MAXCALLS);
811 RXS_NewConnection(securityObject, tconn);
813 CONN_HASH(shost, sport, tconn->cid, tconn->epoch,
814 RX_CLIENT_CONNECTION);
815 tconn->refCount++; /* no lock required since only this thread knows */
816 tconn->next = rx_connHashTable[hashindex];
817 rx_connHashTable[hashindex] = tconn;
818 rx_MutexIncrement(rx_stats.nClientConns, rx_stats_mutex);
821 MUTEX_EXIT(&rx_connHashTable_lock);
827 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
829 /* The idea is to set the dead time to a value that allows several
830 * keepalives to be dropped without timing out the connection. */
831 struct rx_connection *tconn;
834 tconn->secondsUntilDead = MAX(seconds, 6);
835 tconn->secondsUntilPing = tconn->secondsUntilDead / 6;
836 } while(tconn->next_clone && (tconn = tconn->next_clone));
839 int rxi_lowPeerRefCount = 0;
840 int rxi_lowConnRefCount = 0;
843 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
844 * NOTE: must not be called with rx_connHashTable_lock held.
847 rxi_CleanupConnection(struct rx_connection *conn)
849 /* Notify the service exporter, if requested, that this connection
850 * is being destroyed */
851 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
852 (*conn->service->destroyConnProc) (conn);
854 /* Notify the security module that this connection is being destroyed */
855 RXS_DestroyConnection(conn->securityObject, conn);
857 /* If this is the last connection using the rx_peer struct, set its
858 * idle time to now. rxi_ReapConnections will reap it if it's still
859 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
861 MUTEX_ENTER(&rx_peerHashTable_lock);
862 if (conn->peer->refCount < 2) {
863 conn->peer->idleWhen = clock_Sec();
864 if (conn->peer->refCount < 1) {
865 conn->peer->refCount = 1;
866 MUTEX_ENTER(&rx_stats_mutex);
867 rxi_lowPeerRefCount++;
868 MUTEX_EXIT(&rx_stats_mutex);
871 conn->peer->refCount--;
872 MUTEX_EXIT(&rx_peerHashTable_lock);
874 if (conn->type == RX_SERVER_CONNECTION)
875 rx_MutexDecrement(rx_stats.nServerConns, rx_stats_mutex);
877 rx_MutexDecrement(rx_stats.nClientConns, rx_stats_mutex);
879 if (conn->specific) {
881 for (i = 0; i < conn->nSpecific; i++) {
882 if (conn->specific[i] && rxi_keyCreate_destructor[i])
883 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
884 conn->specific[i] = NULL;
886 free(conn->specific);
888 conn->specific = NULL;
892 MUTEX_DESTROY(&conn->conn_call_lock);
893 MUTEX_DESTROY(&conn->conn_data_lock);
894 CV_DESTROY(&conn->conn_call_cv);
896 rxi_FreeConnection(conn);
899 /* Destroy the specified connection */
901 rxi_DestroyConnection(register struct rx_connection *conn)
903 register struct rx_connection *tconn, *dtconn;
905 MUTEX_ENTER(&rx_connHashTable_lock);
907 if(!(conn->flags & RX_CLONED_CONNECTION)) {
908 tconn = conn->next_clone;
909 conn->next_clone = 0; /* once */
913 tconn = tconn->next_clone;
914 rxi_DestroyConnectionNoLock(dtconn);
916 if (dtconn == rx_connCleanup_list) {
917 rx_connCleanup_list = rx_connCleanup_list->next;
918 MUTEX_EXIT(&rx_connHashTable_lock);
919 /* rxi_CleanupConnection will free tconn */
920 rxi_CleanupConnection(dtconn);
921 MUTEX_ENTER(&rx_connHashTable_lock);
928 rxi_DestroyConnectionNoLock(conn);
929 /* conn should be at the head of the cleanup list */
930 if (conn == rx_connCleanup_list) {
931 rx_connCleanup_list = rx_connCleanup_list->next;
932 MUTEX_EXIT(&rx_connHashTable_lock);
933 rxi_CleanupConnection(conn);
935 #ifdef RX_ENABLE_LOCKS
937 MUTEX_EXIT(&rx_connHashTable_lock);
939 #endif /* RX_ENABLE_LOCKS */
943 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
945 register struct rx_connection **conn_ptr;
946 register int havecalls = 0;
947 struct rx_packet *packet;
954 MUTEX_ENTER(&conn->conn_data_lock);
955 if (conn->refCount > 0)
958 MUTEX_ENTER(&rx_stats_mutex);
959 rxi_lowConnRefCount++;
960 MUTEX_EXIT(&rx_stats_mutex);
963 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
964 /* Busy; wait till the last guy before proceeding */
965 MUTEX_EXIT(&conn->conn_data_lock);
970 /* If the client previously called rx_NewCall, but it is still
971 * waiting, treat this as a running call, and wait to destroy the
972 * connection later when the call completes. */
973 if ((conn->type == RX_CLIENT_CONNECTION)
974 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
975 conn->flags |= RX_CONN_DESTROY_ME;
976 MUTEX_EXIT(&conn->conn_data_lock);
980 MUTEX_EXIT(&conn->conn_data_lock);
982 /* Check for extant references to this connection */
983 for (i = 0; i < RX_MAXCALLS; i++) {
984 register struct rx_call *call = conn->call[i];
987 if (conn->type == RX_CLIENT_CONNECTION) {
988 MUTEX_ENTER(&call->lock);
989 if (call->delayedAckEvent) {
990 /* Push the final acknowledgment out now--there
991 * won't be a subsequent call to acknowledge the
992 * last reply packets */
993 rxevent_Cancel(call->delayedAckEvent, call,
994 RX_CALL_REFCOUNT_DELAY);
995 if (call->state == RX_STATE_PRECALL
996 || call->state == RX_STATE_ACTIVE) {
997 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
999 rxi_AckAll(NULL, call, 0);
1002 MUTEX_EXIT(&call->lock);
1006 #ifdef RX_ENABLE_LOCKS
1008 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1009 MUTEX_EXIT(&conn->conn_data_lock);
1011 /* Someone is accessing a packet right now. */
1015 #endif /* RX_ENABLE_LOCKS */
1018 /* Don't destroy the connection if there are any call
1019 * structures still in use */
1020 MUTEX_ENTER(&conn->conn_data_lock);
1021 conn->flags |= RX_CONN_DESTROY_ME;
1022 MUTEX_EXIT(&conn->conn_data_lock);
1027 if (conn->delayedAbortEvent) {
1028 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
1029 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1031 MUTEX_ENTER(&conn->conn_data_lock);
1032 rxi_SendConnectionAbort(conn, packet, 0, 1);
1033 MUTEX_EXIT(&conn->conn_data_lock);
1034 rxi_FreePacket(packet);
1038 /* Remove from connection hash table before proceeding */
1040 &rx_connHashTable[CONN_HASH
1041 (peer->host, peer->port, conn->cid, conn->epoch,
1043 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1044 if (*conn_ptr == conn) {
1045 *conn_ptr = conn->next;
1049 /* if the conn that we are destroying was the last connection, then we
1050 * clear rxLastConn as well */
1051 if (rxLastConn == conn)
1054 /* Make sure the connection is completely reset before deleting it. */
1055 /* get rid of pending events that could zap us later */
1056 if (conn->challengeEvent)
1057 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1058 if (conn->checkReachEvent)
1059 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1061 /* Add the connection to the list of destroyed connections that
1062 * need to be cleaned up. This is necessary to avoid deadlocks
1063 * in the routines we call to inform others that this connection is
1064 * being destroyed. */
1065 conn->next = rx_connCleanup_list;
1066 rx_connCleanup_list = conn;
1069 /* Externally available version */
1071 rx_DestroyConnection(register struct rx_connection *conn)
1076 rxi_DestroyConnection(conn);
1081 rx_GetConnection(register struct rx_connection *conn)
1086 MUTEX_ENTER(&conn->conn_data_lock);
1088 MUTEX_EXIT(&conn->conn_data_lock);
1092 /* Wait for the transmit queue to no longer be busy.
1093 * requires the call->lock to be held */
1094 static void rxi_WaitforTQBusy(struct rx_call *call) {
1095 while (call->flags & RX_CALL_TQ_BUSY) {
1096 call->flags |= RX_CALL_TQ_WAIT;
1098 #ifdef RX_ENABLE_LOCKS
1099 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1100 CV_WAIT(&call->cv_tq, &call->lock);
1101 #else /* RX_ENABLE_LOCKS */
1102 osi_rxSleep(&call->tq);
1103 #endif /* RX_ENABLE_LOCKS */
1105 if (call->tqWaiters == 0) {
1106 call->flags &= ~RX_CALL_TQ_WAIT;
1110 /* Start a new rx remote procedure call, on the specified connection.
1111 * If wait is set to 1, wait for a free call channel; otherwise return
1112 * 0. Maxtime gives the maximum number of seconds this call may take,
1113 * after rx_NewCall returns. After this time interval, a call to any
1114 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1115 * For fine grain locking, we hold the conn_call_lock in order to
1116 * to ensure that we don't get signalle after we found a call in an active
1117 * state and before we go to sleep.
1120 rx_NewCall(register struct rx_connection *conn)
1123 register struct rx_call *call;
1124 register struct rx_connection *tconn;
1125 struct clock queueTime;
1129 dpf(("rx_NewCall(conn %x)\n", conn));
1132 clock_GetTime(&queueTime);
1133 MUTEX_ENTER(&conn->conn_call_lock);
1136 * Check if there are others waiting for a new call.
1137 * If so, let them go first to avoid starving them.
1138 * This is a fairly simple scheme, and might not be
1139 * a complete solution for large numbers of waiters.
1141 * makeCallWaiters keeps track of the number of
1142 * threads waiting to make calls and the
1143 * RX_CONN_MAKECALL_WAITING flag bit is used to
1144 * indicate that there are indeed calls waiting.
1145 * The flag is set when the waiter is incremented.
1146 * It is only cleared in rx_EndCall when
1147 * makeCallWaiters is 0. This prevents us from
1148 * accidently destroying the connection while it
1149 * is potentially about to be used.
1151 MUTEX_ENTER(&conn->conn_data_lock);
1152 if (conn->makeCallWaiters) {
1153 conn->flags |= RX_CONN_MAKECALL_WAITING;
1154 conn->makeCallWaiters++;
1155 MUTEX_EXIT(&conn->conn_data_lock);
1157 #ifdef RX_ENABLE_LOCKS
1158 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1162 MUTEX_ENTER(&conn->conn_data_lock);
1163 conn->makeCallWaiters--;
1165 MUTEX_EXIT(&conn->conn_data_lock);
1167 /* search for next free call on this connection or
1168 * its clones, if any */
1172 for (i = 0; i < RX_MAXCALLS; i++) {
1173 call = tconn->call[i];
1175 MUTEX_ENTER(&call->lock);
1176 if (call->state == RX_STATE_DALLY) {
1177 rxi_ResetCall(call, 0);
1178 (*call->callNumber)++;
1181 MUTEX_EXIT(&call->lock);
1183 call = rxi_NewCall(tconn, i);
1186 } /* for i < RX_MAXCALLS */
1187 } while (tconn->next_clone && (tconn = tconn->next_clone));
1191 if (i < RX_MAXCALLS) {
1195 /* to be here, all available calls for this connection (and all
1196 * its clones) must be in use */
1198 MUTEX_ENTER(&conn->conn_data_lock);
1199 conn->flags |= RX_CONN_MAKECALL_WAITING;
1200 conn->makeCallWaiters++;
1201 MUTEX_EXIT(&conn->conn_data_lock);
1203 #ifdef RX_ENABLE_LOCKS
1204 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1208 MUTEX_ENTER(&conn->conn_data_lock);
1209 conn->makeCallWaiters--;
1210 MUTEX_EXIT(&conn->conn_data_lock);
1213 * Wake up anyone else who might be giving us a chance to
1214 * run (see code above that avoids resource starvation).
1216 #ifdef RX_ENABLE_LOCKS
1217 CV_BROADCAST(&conn->conn_call_cv);
1222 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1224 /* Client is initially in send mode */
1225 call->state = RX_STATE_ACTIVE;
1226 call->error = conn->error;
1228 call->mode = RX_MODE_ERROR;
1230 call->mode = RX_MODE_SENDING;
1232 /* remember start time for call in case we have hard dead time limit */
1233 call->queueTime = queueTime;
1234 clock_GetTime(&call->startTime);
1235 hzero(call->bytesSent);
1236 hzero(call->bytesRcvd);
1238 /* Turn on busy protocol. */
1239 rxi_KeepAliveOn(call);
1241 MUTEX_EXIT(&call->lock);
1242 MUTEX_EXIT(&conn->conn_call_lock);
1245 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1246 /* Now, if TQ wasn't cleared earlier, do it now. */
1247 MUTEX_ENTER(&call->lock);
1248 rxi_WaitforTQBusy(call);
1249 if (call->flags & RX_CALL_TQ_CLEARME) {
1250 rxi_ClearTransmitQueue(call, 0);
1251 queue_Init(&call->tq);
1253 MUTEX_EXIT(&call->lock);
1254 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1256 dpf(("rx_NewCall(call %x)\n", call));
1261 rxi_HasActiveCalls(register struct rx_connection *aconn)
1264 register struct rx_call *tcall;
1268 for (i = 0; i < RX_MAXCALLS; i++) {
1269 if ((tcall = aconn->call[i])) {
1270 if ((tcall->state == RX_STATE_ACTIVE)
1271 || (tcall->state == RX_STATE_PRECALL)) {
1282 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1283 register afs_int32 * aint32s)
1286 register struct rx_call *tcall;
1290 for (i = 0; i < RX_MAXCALLS; i++) {
1291 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1292 aint32s[i] = aconn->callNumber[i] + 1;
1294 aint32s[i] = aconn->callNumber[i];
1301 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1302 register afs_int32 * aint32s)
1305 register struct rx_call *tcall;
1309 for (i = 0; i < RX_MAXCALLS; i++) {
1310 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1311 aconn->callNumber[i] = aint32s[i] - 1;
1313 aconn->callNumber[i] = aint32s[i];
1319 /* Advertise a new service. A service is named locally by a UDP port
1320 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1323 char *serviceName; Name for identification purposes (e.g. the
1324 service name might be used for probing for
1327 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1328 char *serviceName, struct rx_securityClass **securityObjects,
1329 int nSecurityObjects,
1330 afs_int32(*serviceProc) (struct rx_call * acall))
1332 osi_socket socket = OSI_NULLSOCKET;
1333 register struct rx_service *tservice;
1339 if (serviceId == 0) {
1341 "rx_NewService: service id for service %s is not non-zero.\n",
1348 "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",
1356 tservice = rxi_AllocService();
1358 for (i = 0; i < RX_MAX_SERVICES; i++) {
1359 register struct rx_service *service = rx_services[i];
1361 if (port == service->servicePort && host == service->serviceHost) {
1362 if (service->serviceId == serviceId) {
1363 /* The identical service has already been
1364 * installed; if the caller was intending to
1365 * change the security classes used by this
1366 * service, he/she loses. */
1368 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1369 serviceName, serviceId, service->serviceName);
1371 rxi_FreeService(tservice);
1374 /* Different service, same port: re-use the socket
1375 * which is bound to the same port */
1376 socket = service->socket;
1379 if (socket == OSI_NULLSOCKET) {
1380 /* If we don't already have a socket (from another
1381 * service on same port) get a new one */
1382 socket = rxi_GetHostUDPSocket(host, port);
1383 if (socket == OSI_NULLSOCKET) {
1385 rxi_FreeService(tservice);
1390 service->socket = socket;
1391 service->serviceHost = host;
1392 service->servicePort = port;
1393 service->serviceId = serviceId;
1394 service->serviceName = serviceName;
1395 service->nSecurityObjects = nSecurityObjects;
1396 service->securityObjects = securityObjects;
1397 service->minProcs = 0;
1398 service->maxProcs = 1;
1399 service->idleDeadTime = 60;
1400 service->idleDeadErr = 0;
1401 service->connDeadTime = rx_connDeadTime;
1402 service->executeRequestProc = serviceProc;
1403 service->checkReach = 0;
1404 rx_services[i] = service; /* not visible until now */
1410 rxi_FreeService(tservice);
1411 (osi_Msg "rx_NewService: cannot support > %d services\n",
1416 /* Set configuration options for all of a service's security objects */
1419 rx_SetSecurityConfiguration(struct rx_service *service,
1420 rx_securityConfigVariables type,
1424 for (i = 0; i<service->nSecurityObjects; i++) {
1425 if (service->securityObjects[i]) {
1426 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1434 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1435 struct rx_securityClass **securityObjects, int nSecurityObjects,
1436 afs_int32(*serviceProc) (struct rx_call * acall))
1438 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1441 /* Generic request processing loop. This routine should be called
1442 * by the implementation dependent rx_ServerProc. If socketp is
1443 * non-null, it will be set to the file descriptor that this thread
1444 * is now listening on. If socketp is null, this routine will never
1447 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1449 register struct rx_call *call;
1450 register afs_int32 code;
1451 register struct rx_service *tservice = NULL;
1458 call = rx_GetCall(threadID, tservice, socketp);
1459 if (socketp && *socketp != OSI_NULLSOCKET) {
1460 /* We are now a listener thread */
1465 /* if server is restarting( typically smooth shutdown) then do not
1466 * allow any new calls.
1469 if (rx_tranquil && (call != NULL)) {
1473 MUTEX_ENTER(&call->lock);
1475 rxi_CallError(call, RX_RESTARTING);
1476 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1478 MUTEX_EXIT(&call->lock);
1482 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1483 #ifdef RX_ENABLE_LOCKS
1485 #endif /* RX_ENABLE_LOCKS */
1486 afs_termState = AFSOP_STOP_AFS;
1487 afs_osi_Wakeup(&afs_termState);
1488 #ifdef RX_ENABLE_LOCKS
1490 #endif /* RX_ENABLE_LOCKS */
1495 tservice = call->conn->service;
1497 if (tservice->beforeProc)
1498 (*tservice->beforeProc) (call);
1500 code = call->conn->service->executeRequestProc(call);
1502 if (tservice->afterProc)
1503 (*tservice->afterProc) (call, code);
1505 rx_EndCall(call, code);
1506 MUTEX_ENTER(&rx_stats_mutex);
1508 MUTEX_EXIT(&rx_stats_mutex);
1514 rx_WakeupServerProcs(void)
1516 struct rx_serverQueueEntry *np, *tqp;
1520 MUTEX_ENTER(&rx_serverPool_lock);
1522 #ifdef RX_ENABLE_LOCKS
1523 if (rx_waitForPacket)
1524 CV_BROADCAST(&rx_waitForPacket->cv);
1525 #else /* RX_ENABLE_LOCKS */
1526 if (rx_waitForPacket)
1527 osi_rxWakeup(rx_waitForPacket);
1528 #endif /* RX_ENABLE_LOCKS */
1529 MUTEX_ENTER(&freeSQEList_lock);
1530 for (np = rx_FreeSQEList; np; np = tqp) {
1531 tqp = *(struct rx_serverQueueEntry **)np;
1532 #ifdef RX_ENABLE_LOCKS
1533 CV_BROADCAST(&np->cv);
1534 #else /* RX_ENABLE_LOCKS */
1536 #endif /* RX_ENABLE_LOCKS */
1538 MUTEX_EXIT(&freeSQEList_lock);
1539 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1540 #ifdef RX_ENABLE_LOCKS
1541 CV_BROADCAST(&np->cv);
1542 #else /* RX_ENABLE_LOCKS */
1544 #endif /* RX_ENABLE_LOCKS */
1546 MUTEX_EXIT(&rx_serverPool_lock);
1551 * One thing that seems to happen is that all the server threads get
1552 * tied up on some empty or slow call, and then a whole bunch of calls
1553 * arrive at once, using up the packet pool, so now there are more
1554 * empty calls. The most critical resources here are server threads
1555 * and the free packet pool. The "doreclaim" code seems to help in
1556 * general. I think that eventually we arrive in this state: there
1557 * are lots of pending calls which do have all their packets present,
1558 * so they won't be reclaimed, are multi-packet calls, so they won't
1559 * be scheduled until later, and thus are tying up most of the free
1560 * packet pool for a very long time.
1562 * 1. schedule multi-packet calls if all the packets are present.
1563 * Probably CPU-bound operation, useful to return packets to pool.
1564 * Do what if there is a full window, but the last packet isn't here?
1565 * 3. preserve one thread which *only* runs "best" calls, otherwise
1566 * it sleeps and waits for that type of call.
1567 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1568 * the current dataquota business is badly broken. The quota isn't adjusted
1569 * to reflect how many packets are presently queued for a running call.
1570 * So, when we schedule a queued call with a full window of packets queued
1571 * up for it, that *should* free up a window full of packets for other 2d-class
1572 * calls to be able to use from the packet pool. But it doesn't.
1574 * NB. Most of the time, this code doesn't run -- since idle server threads
1575 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1576 * as a new call arrives.
1578 /* Sleep until a call arrives. Returns a pointer to the call, ready
1579 * for an rx_Read. */
1580 #ifdef RX_ENABLE_LOCKS
1582 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1584 struct rx_serverQueueEntry *sq;
1585 register struct rx_call *call = (struct rx_call *)0;
1586 struct rx_service *service = NULL;
1589 MUTEX_ENTER(&freeSQEList_lock);
1591 if ((sq = rx_FreeSQEList)) {
1592 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1593 MUTEX_EXIT(&freeSQEList_lock);
1594 } else { /* otherwise allocate a new one and return that */
1595 MUTEX_EXIT(&freeSQEList_lock);
1596 sq = (struct rx_serverQueueEntry *)
1597 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1598 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1599 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1602 MUTEX_ENTER(&rx_serverPool_lock);
1603 if (cur_service != NULL) {
1604 ReturnToServerPool(cur_service);
1607 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1608 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1610 /* Scan for eligible incoming calls. A call is not eligible
1611 * if the maximum number of calls for its service type are
1612 * already executing */
1613 /* One thread will process calls FCFS (to prevent starvation),
1614 * while the other threads may run ahead looking for calls which
1615 * have all their input data available immediately. This helps
1616 * keep threads from blocking, waiting for data from the client. */
1617 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1618 service = tcall->conn->service;
1619 if (!QuotaOK(service)) {
1622 if (tno == rxi_fcfs_thread_num
1623 || !tcall->queue_item_header.next) {
1624 /* If we're the fcfs thread , then we'll just use
1625 * this call. If we haven't been able to find an optimal
1626 * choice, and we're at the end of the list, then use a
1627 * 2d choice if one has been identified. Otherwise... */
1628 call = (choice2 ? choice2 : tcall);
1629 service = call->conn->service;
1630 } else if (!queue_IsEmpty(&tcall->rq)) {
1631 struct rx_packet *rp;
1632 rp = queue_First(&tcall->rq, rx_packet);
1633 if (rp->header.seq == 1) {
1635 || (rp->header.flags & RX_LAST_PACKET)) {
1637 } else if (rxi_2dchoice && !choice2
1638 && !(tcall->flags & RX_CALL_CLEARED)
1639 && (tcall->rprev > rxi_HardAckRate)) {
1648 ReturnToServerPool(service);
1655 MUTEX_EXIT(&rx_serverPool_lock);
1656 MUTEX_ENTER(&call->lock);
1658 if (call->flags & RX_CALL_WAIT_PROC) {
1659 call->flags &= ~RX_CALL_WAIT_PROC;
1660 MUTEX_ENTER(&rx_stats_mutex);
1662 MUTEX_EXIT(&rx_stats_mutex);
1665 if (call->state != RX_STATE_PRECALL || call->error) {
1666 MUTEX_EXIT(&call->lock);
1667 MUTEX_ENTER(&rx_serverPool_lock);
1668 ReturnToServerPool(service);
1673 if (queue_IsEmpty(&call->rq)
1674 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1675 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1677 CLEAR_CALL_QUEUE_LOCK(call);
1680 /* If there are no eligible incoming calls, add this process
1681 * to the idle server queue, to wait for one */
1685 *socketp = OSI_NULLSOCKET;
1687 sq->socketp = socketp;
1688 queue_Append(&rx_idleServerQueue, sq);
1689 #ifndef AFS_AIX41_ENV
1690 rx_waitForPacket = sq;
1692 rx_waitingForPacket = sq;
1693 #endif /* AFS_AIX41_ENV */
1695 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1697 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1698 MUTEX_EXIT(&rx_serverPool_lock);
1699 return (struct rx_call *)0;
1702 } while (!(call = sq->newcall)
1703 && !(socketp && *socketp != OSI_NULLSOCKET));
1704 MUTEX_EXIT(&rx_serverPool_lock);
1706 MUTEX_ENTER(&call->lock);
1712 MUTEX_ENTER(&freeSQEList_lock);
1713 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1714 rx_FreeSQEList = sq;
1715 MUTEX_EXIT(&freeSQEList_lock);
1718 clock_GetTime(&call->startTime);
1719 call->state = RX_STATE_ACTIVE;
1720 call->mode = RX_MODE_RECEIVING;
1721 #ifdef RX_KERNEL_TRACE
1722 if (ICL_SETACTIVE(afs_iclSetp)) {
1723 int glockOwner = ISAFS_GLOCK();
1726 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1727 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1734 rxi_calltrace(RX_CALL_START, call);
1735 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1736 call->conn->service->servicePort, call->conn->service->serviceId,
1739 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1740 MUTEX_EXIT(&call->lock);
1742 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1747 #else /* RX_ENABLE_LOCKS */
1749 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1751 struct rx_serverQueueEntry *sq;
1752 register struct rx_call *call = (struct rx_call *)0, *choice2;
1753 struct rx_service *service = NULL;
1757 MUTEX_ENTER(&freeSQEList_lock);
1759 if ((sq = rx_FreeSQEList)) {
1760 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1761 MUTEX_EXIT(&freeSQEList_lock);
1762 } else { /* otherwise allocate a new one and return that */
1763 MUTEX_EXIT(&freeSQEList_lock);
1764 sq = (struct rx_serverQueueEntry *)
1765 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1766 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1767 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1769 MUTEX_ENTER(&sq->lock);
1771 if (cur_service != NULL) {
1772 cur_service->nRequestsRunning--;
1773 if (cur_service->nRequestsRunning < cur_service->minProcs)
1777 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1778 register struct rx_call *tcall, *ncall;
1779 /* Scan for eligible incoming calls. A call is not eligible
1780 * if the maximum number of calls for its service type are
1781 * already executing */
1782 /* One thread will process calls FCFS (to prevent starvation),
1783 * while the other threads may run ahead looking for calls which
1784 * have all their input data available immediately. This helps
1785 * keep threads from blocking, waiting for data from the client. */
1786 choice2 = (struct rx_call *)0;
1787 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1788 service = tcall->conn->service;
1789 if (QuotaOK(service)) {
1790 if (tno == rxi_fcfs_thread_num
1791 || !tcall->queue_item_header.next) {
1792 /* If we're the fcfs thread, then we'll just use
1793 * this call. If we haven't been able to find an optimal
1794 * choice, and we're at the end of the list, then use a
1795 * 2d choice if one has been identified. Otherwise... */
1796 call = (choice2 ? choice2 : tcall);
1797 service = call->conn->service;
1798 } else if (!queue_IsEmpty(&tcall->rq)) {
1799 struct rx_packet *rp;
1800 rp = queue_First(&tcall->rq, rx_packet);
1801 if (rp->header.seq == 1
1803 || (rp->header.flags & RX_LAST_PACKET))) {
1805 } else if (rxi_2dchoice && !choice2
1806 && !(tcall->flags & RX_CALL_CLEARED)
1807 && (tcall->rprev > rxi_HardAckRate)) {
1820 /* we can't schedule a call if there's no data!!! */
1821 /* send an ack if there's no data, if we're missing the
1822 * first packet, or we're missing something between first
1823 * and last -- there's a "hole" in the incoming data. */
1824 if (queue_IsEmpty(&call->rq)
1825 || queue_First(&call->rq, rx_packet)->header.seq != 1
1826 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1827 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1829 call->flags &= (~RX_CALL_WAIT_PROC);
1830 service->nRequestsRunning++;
1831 /* just started call in minProcs pool, need fewer to maintain
1833 if (service->nRequestsRunning <= service->minProcs)
1837 /* MUTEX_EXIT(&call->lock); */
1839 /* If there are no eligible incoming calls, add this process
1840 * to the idle server queue, to wait for one */
1843 *socketp = OSI_NULLSOCKET;
1845 sq->socketp = socketp;
1846 queue_Append(&rx_idleServerQueue, sq);
1850 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1852 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1853 return (struct rx_call *)0;
1856 } while (!(call = sq->newcall)
1857 && !(socketp && *socketp != OSI_NULLSOCKET));
1859 MUTEX_EXIT(&sq->lock);
1861 MUTEX_ENTER(&freeSQEList_lock);
1862 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1863 rx_FreeSQEList = sq;
1864 MUTEX_EXIT(&freeSQEList_lock);
1867 clock_GetTime(&call->startTime);
1868 call->state = RX_STATE_ACTIVE;
1869 call->mode = RX_MODE_RECEIVING;
1870 #ifdef RX_KERNEL_TRACE
1871 if (ICL_SETACTIVE(afs_iclSetp)) {
1872 int glockOwner = ISAFS_GLOCK();
1875 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1876 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1883 rxi_calltrace(RX_CALL_START, call);
1884 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1885 call->conn->service->servicePort, call->conn->service->serviceId,
1888 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1895 #endif /* RX_ENABLE_LOCKS */
1899 /* Establish a procedure to be called when a packet arrives for a
1900 * call. This routine will be called at most once after each call,
1901 * and will also be called if there is an error condition on the or
1902 * the call is complete. Used by multi rx to build a selection
1903 * function which determines which of several calls is likely to be a
1904 * good one to read from.
1905 * NOTE: the way this is currently implemented it is probably only a
1906 * good idea to (1) use it immediately after a newcall (clients only)
1907 * and (2) only use it once. Other uses currently void your warranty
1910 rx_SetArrivalProc(register struct rx_call *call,
1911 register void (*proc) (register struct rx_call * call,
1913 register int index),
1914 register void * handle, register int arg)
1916 call->arrivalProc = proc;
1917 call->arrivalProcHandle = handle;
1918 call->arrivalProcArg = arg;
1921 /* Call is finished (possibly prematurely). Return rc to the peer, if
1922 * appropriate, and return the final error code from the conversation
1926 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1928 register struct rx_connection *conn = call->conn;
1929 register struct rx_service *service;
1935 dpf(("rx_EndCall(call %x rc %d error %d abortCode %d)\n", call, rc, call->error, call->abortCode));
1938 MUTEX_ENTER(&call->lock);
1940 if (rc == 0 && call->error == 0) {
1941 call->abortCode = 0;
1942 call->abortCount = 0;
1945 call->arrivalProc = (void (*)())0;
1946 if (rc && call->error == 0) {
1947 rxi_CallError(call, rc);
1948 /* Send an abort message to the peer if this error code has
1949 * only just been set. If it was set previously, assume the
1950 * peer has already been sent the error code or will request it
1952 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1954 if (conn->type == RX_SERVER_CONNECTION) {
1955 /* Make sure reply or at least dummy reply is sent */
1956 if (call->mode == RX_MODE_RECEIVING) {
1957 rxi_WriteProc(call, 0, 0);
1959 if (call->mode == RX_MODE_SENDING) {
1960 rxi_FlushWrite(call);
1962 service = conn->service;
1963 rxi_calltrace(RX_CALL_END, call);
1964 /* Call goes to hold state until reply packets are acknowledged */
1965 if (call->tfirst + call->nSoftAcked < call->tnext) {
1966 call->state = RX_STATE_HOLD;
1968 call->state = RX_STATE_DALLY;
1969 rxi_ClearTransmitQueue(call, 0);
1970 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1971 rxevent_Cancel(call->keepAliveEvent, call,
1972 RX_CALL_REFCOUNT_ALIVE);
1974 } else { /* Client connection */
1976 /* Make sure server receives input packets, in the case where
1977 * no reply arguments are expected */
1978 if ((call->mode == RX_MODE_SENDING)
1979 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1980 (void)rxi_ReadProc(call, &dummy, 1);
1983 /* If we had an outstanding delayed ack, be nice to the server
1984 * and force-send it now.
1986 if (call->delayedAckEvent) {
1987 rxevent_Cancel(call->delayedAckEvent, call,
1988 RX_CALL_REFCOUNT_DELAY);
1989 call->delayedAckEvent = NULL;
1990 rxi_SendDelayedAck(NULL, call, NULL);
1993 /* We need to release the call lock since it's lower than the
1994 * conn_call_lock and we don't want to hold the conn_call_lock
1995 * over the rx_ReadProc call. The conn_call_lock needs to be held
1996 * here for the case where rx_NewCall is perusing the calls on
1997 * the connection structure. We don't want to signal until
1998 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1999 * have checked this call, found it active and by the time it
2000 * goes to sleep, will have missed the signal.
2002 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
2003 * there are threads waiting to use the conn object.
2005 MUTEX_EXIT(&call->lock);
2006 MUTEX_ENTER(&conn->conn_call_lock);
2007 MUTEX_ENTER(&call->lock);
2008 MUTEX_ENTER(&conn->conn_data_lock);
2009 conn->flags |= RX_CONN_BUSY;
2010 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2011 if (conn->makeCallWaiters == 0)
2012 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
2013 MUTEX_EXIT(&conn->conn_data_lock);
2014 #ifdef RX_ENABLE_LOCKS
2015 CV_BROADCAST(&conn->conn_call_cv);
2020 #ifdef RX_ENABLE_LOCKS
2022 MUTEX_EXIT(&conn->conn_data_lock);
2024 #endif /* RX_ENABLE_LOCKS */
2025 call->state = RX_STATE_DALLY;
2027 error = call->error;
2029 /* currentPacket, nLeft, and NFree must be zeroed here, because
2030 * ResetCall cannot: ResetCall may be called at splnet(), in the
2031 * kernel version, and may interrupt the macros rx_Read or
2032 * rx_Write, which run at normal priority for efficiency. */
2033 if (call->currentPacket) {
2034 queue_Prepend(&call->iovq, call->currentPacket);
2035 call->currentPacket = (struct rx_packet *)0;
2038 call->nLeft = call->nFree = call->curlen = 0;
2040 /* Free any packets from the last call to ReadvProc/WritevProc */
2041 rxi_FreePackets(0, &call->iovq);
2043 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2044 MUTEX_EXIT(&call->lock);
2045 if (conn->type == RX_CLIENT_CONNECTION) {
2046 MUTEX_EXIT(&conn->conn_call_lock);
2047 conn->flags &= ~RX_CONN_BUSY;
2051 * Map errors to the local host's errno.h format.
2053 error = ntoh_syserr_conv(error);
2057 #if !defined(KERNEL)
2059 /* Call this routine when shutting down a server or client (especially
2060 * clients). This will allow Rx to gracefully garbage collect server
2061 * connections, and reduce the number of retries that a server might
2062 * make to a dead client.
2063 * This is not quite right, since some calls may still be ongoing and
2064 * we can't lock them to destroy them. */
2068 register struct rx_connection **conn_ptr, **conn_end;
2072 if (rxinit_status == 1) {
2074 return; /* Already shutdown. */
2076 rxi_DeleteCachedConnections();
2077 if (rx_connHashTable) {
2078 MUTEX_ENTER(&rx_connHashTable_lock);
2079 for (conn_ptr = &rx_connHashTable[0], conn_end =
2080 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2082 struct rx_connection *conn, *next;
2083 for (conn = *conn_ptr; conn; conn = next) {
2085 if (conn->type == RX_CLIENT_CONNECTION) {
2086 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2088 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2089 #ifdef RX_ENABLE_LOCKS
2090 rxi_DestroyConnectionNoLock(conn);
2091 #else /* RX_ENABLE_LOCKS */
2092 rxi_DestroyConnection(conn);
2093 #endif /* RX_ENABLE_LOCKS */
2097 #ifdef RX_ENABLE_LOCKS
2098 while (rx_connCleanup_list) {
2099 struct rx_connection *conn;
2100 conn = rx_connCleanup_list;
2101 rx_connCleanup_list = rx_connCleanup_list->next;
2102 MUTEX_EXIT(&rx_connHashTable_lock);
2103 rxi_CleanupConnection(conn);
2104 MUTEX_ENTER(&rx_connHashTable_lock);
2106 MUTEX_EXIT(&rx_connHashTable_lock);
2107 #endif /* RX_ENABLE_LOCKS */
2112 afs_winsockCleanup();
2120 /* if we wakeup packet waiter too often, can get in loop with two
2121 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2123 rxi_PacketsUnWait(void)
2125 if (!rx_waitingForPackets) {
2129 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2130 return; /* still over quota */
2133 rx_waitingForPackets = 0;
2134 #ifdef RX_ENABLE_LOCKS
2135 CV_BROADCAST(&rx_waitingForPackets_cv);
2137 osi_rxWakeup(&rx_waitingForPackets);
2143 /* ------------------Internal interfaces------------------------- */
2145 /* Return this process's service structure for the
2146 * specified socket and service */
2148 rxi_FindService(register osi_socket socket, register u_short serviceId)
2150 register struct rx_service **sp;
2151 for (sp = &rx_services[0]; *sp; sp++) {
2152 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2158 /* Allocate a call structure, for the indicated channel of the
2159 * supplied connection. The mode and state of the call must be set by
2160 * the caller. Returns the call with mutex locked. */
2162 rxi_NewCall(register struct rx_connection *conn, register int channel)
2164 register struct rx_call *call;
2165 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2166 register struct rx_call *cp; /* Call pointer temp */
2167 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2168 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2170 dpf(("rxi_NewCall(conn %x, channel %d)\n", conn, channel));
2172 /* Grab an existing call structure, or allocate a new one.
2173 * Existing call structures are assumed to have been left reset by
2175 MUTEX_ENTER(&rx_freeCallQueue_lock);
2177 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2179 * EXCEPT that the TQ might not yet be cleared out.
2180 * Skip over those with in-use TQs.
2183 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2184 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2190 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2191 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2192 call = queue_First(&rx_freeCallQueue, rx_call);
2193 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2195 rx_MutexDecrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2196 MUTEX_EXIT(&rx_freeCallQueue_lock);
2197 MUTEX_ENTER(&call->lock);
2198 CLEAR_CALL_QUEUE_LOCK(call);
2199 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2200 /* Now, if TQ wasn't cleared earlier, do it now. */
2201 if (call->flags & RX_CALL_TQ_CLEARME) {
2202 rxi_ClearTransmitQueue(call, 0);
2203 queue_Init(&call->tq);
2205 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2206 /* Bind the call to its connection structure */
2208 rxi_ResetCall(call, 1);
2210 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2212 MUTEX_EXIT(&rx_freeCallQueue_lock);
2213 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2214 MUTEX_ENTER(&call->lock);
2215 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2216 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2217 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2219 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2220 /* Initialize once-only items */
2221 queue_Init(&call->tq);
2222 queue_Init(&call->rq);
2223 queue_Init(&call->iovq);
2224 /* Bind the call to its connection structure (prereq for reset) */
2226 rxi_ResetCall(call, 1);
2228 call->channel = channel;
2229 call->callNumber = &conn->callNumber[channel];
2230 call->rwind = conn->rwind[channel];
2231 call->twind = conn->twind[channel];
2232 /* Note that the next expected call number is retained (in
2233 * conn->callNumber[i]), even if we reallocate the call structure
2235 conn->call[channel] = call;
2236 /* if the channel's never been used (== 0), we should start at 1, otherwise
2237 * the call number is valid from the last time this channel was used */
2238 if (*call->callNumber == 0)
2239 *call->callNumber = 1;
2244 /* A call has been inactive long enough that so we can throw away
2245 * state, including the call structure, which is placed on the call
2247 * Call is locked upon entry.
2248 * haveCTLock set if called from rxi_ReapConnections
2250 #ifdef RX_ENABLE_LOCKS
2252 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2253 #else /* RX_ENABLE_LOCKS */
2255 rxi_FreeCall(register struct rx_call *call)
2256 #endif /* RX_ENABLE_LOCKS */
2258 register int channel = call->channel;
2259 register struct rx_connection *conn = call->conn;
2262 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2263 (*call->callNumber)++;
2264 rxi_ResetCall(call, 0);
2265 call->conn->call[channel] = (struct rx_call *)0;
2267 MUTEX_ENTER(&rx_freeCallQueue_lock);
2268 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2269 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2270 /* A call may be free even though its transmit queue is still in use.
2271 * Since we search the call list from head to tail, put busy calls at
2272 * the head of the list, and idle calls at the tail.
2274 if (call->flags & RX_CALL_TQ_BUSY)
2275 queue_Prepend(&rx_freeCallQueue, call);
2277 queue_Append(&rx_freeCallQueue, call);
2278 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2279 queue_Append(&rx_freeCallQueue, call);
2280 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2281 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2282 MUTEX_EXIT(&rx_freeCallQueue_lock);
2284 /* Destroy the connection if it was previously slated for
2285 * destruction, i.e. the Rx client code previously called
2286 * rx_DestroyConnection (client connections), or
2287 * rxi_ReapConnections called the same routine (server
2288 * connections). Only do this, however, if there are no
2289 * outstanding calls. Note that for fine grain locking, there appears
2290 * to be a deadlock in that rxi_FreeCall has a call locked and
2291 * DestroyConnectionNoLock locks each call in the conn. But note a
2292 * few lines up where we have removed this call from the conn.
2293 * If someone else destroys a connection, they either have no
2294 * call lock held or are going through this section of code.
2296 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2297 MUTEX_ENTER(&conn->conn_data_lock);
2299 MUTEX_EXIT(&conn->conn_data_lock);
2300 #ifdef RX_ENABLE_LOCKS
2302 rxi_DestroyConnectionNoLock(conn);
2304 rxi_DestroyConnection(conn);
2305 #else /* RX_ENABLE_LOCKS */
2306 rxi_DestroyConnection(conn);
2307 #endif /* RX_ENABLE_LOCKS */
2311 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2313 rxi_Alloc(register size_t size)
2317 rx_MutexAdd1Increment2(rxi_Allocsize, (afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2318 p = (char *)osi_Alloc(size);
2321 osi_Panic("rxi_Alloc error");
2327 rxi_Free(void *addr, register size_t size)
2329 rx_MutexAdd1Decrement2(rxi_Allocsize, -(afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2330 osi_Free(addr, size);
2334 rxi_SetPeerMtu(register afs_uint32 host, register afs_uint32 port, int mtu)
2336 struct rx_peer **peer_ptr, **peer_end;
2339 MUTEX_ENTER(&rx_peerHashTable_lock);
2341 for (peer_ptr = &rx_peerHashTable[0], peer_end =
2342 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
2344 struct rx_peer *peer, *next;
2345 for (peer = *peer_ptr; peer; peer = next) {
2347 if (host == peer->host) {
2348 MUTEX_ENTER(&peer->peer_lock);
2349 peer->ifMTU=MIN(mtu, peer->ifMTU);
2350 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2351 MUTEX_EXIT(&peer->peer_lock);
2356 struct rx_peer *peer, *next;
2357 hashIndex = PEER_HASH(host, port);
2358 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2359 if ((peer->host == host) && (peer->port == port)) {
2360 MUTEX_ENTER(&peer->peer_lock);
2361 peer->ifMTU=MIN(mtu, peer->ifMTU);
2362 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2363 MUTEX_EXIT(&peer->peer_lock);
2367 MUTEX_EXIT(&rx_peerHashTable_lock);
2370 /* Find the peer process represented by the supplied (host,port)
2371 * combination. If there is no appropriate active peer structure, a
2372 * new one will be allocated and initialized
2373 * The origPeer, if set, is a pointer to a peer structure on which the
2374 * refcount will be be decremented. This is used to replace the peer
2375 * structure hanging off a connection structure */
2377 rxi_FindPeer(register afs_uint32 host, register u_short port,
2378 struct rx_peer *origPeer, int create)
2380 register struct rx_peer *pp;
2382 hashIndex = PEER_HASH(host, port);
2383 MUTEX_ENTER(&rx_peerHashTable_lock);
2384 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2385 if ((pp->host == host) && (pp->port == port))
2390 pp = rxi_AllocPeer(); /* This bzero's *pp */
2391 pp->host = host; /* set here or in InitPeerParams is zero */
2393 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2394 queue_Init(&pp->congestionQueue);
2395 queue_Init(&pp->rpcStats);
2396 pp->next = rx_peerHashTable[hashIndex];
2397 rx_peerHashTable[hashIndex] = pp;
2398 rxi_InitPeerParams(pp);
2399 rx_MutexIncrement(rx_stats.nPeerStructs, rx_stats_mutex);
2406 origPeer->refCount--;
2407 MUTEX_EXIT(&rx_peerHashTable_lock);
2412 /* Find the connection at (host, port) started at epoch, and with the
2413 * given connection id. Creates the server connection if necessary.
2414 * The type specifies whether a client connection or a server
2415 * connection is desired. In both cases, (host, port) specify the
2416 * peer's (host, pair) pair. Client connections are not made
2417 * automatically by this routine. The parameter socket gives the
2418 * socket descriptor on which the packet was received. This is used,
2419 * in the case of server connections, to check that *new* connections
2420 * come via a valid (port, serviceId). Finally, the securityIndex
2421 * parameter must match the existing index for the connection. If a
2422 * server connection is created, it will be created using the supplied
2423 * index, if the index is valid for this service */
2424 struct rx_connection *
2425 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2426 register u_short port, u_short serviceId, afs_uint32 cid,
2427 afs_uint32 epoch, int type, u_int securityIndex)
2429 int hashindex, flag, i;
2430 register struct rx_connection *conn;
2431 hashindex = CONN_HASH(host, port, cid, epoch, type);
2432 MUTEX_ENTER(&rx_connHashTable_lock);
2433 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2434 rx_connHashTable[hashindex],
2437 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2438 && (epoch == conn->epoch)) {
2439 register struct rx_peer *pp = conn->peer;
2440 if (securityIndex != conn->securityIndex) {
2441 /* this isn't supposed to happen, but someone could forge a packet
2442 * like this, and there seems to be some CM bug that makes this
2443 * happen from time to time -- in which case, the fileserver
2445 MUTEX_EXIT(&rx_connHashTable_lock);
2446 return (struct rx_connection *)0;
2448 if (pp->host == host && pp->port == port)
2450 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2452 /* So what happens when it's a callback connection? */
2453 if ( /*type == RX_CLIENT_CONNECTION && */
2454 (conn->epoch & 0x80000000))
2458 /* the connection rxLastConn that was used the last time is not the
2459 ** one we are looking for now. Hence, start searching in the hash */
2461 conn = rx_connHashTable[hashindex];
2466 struct rx_service *service;
2467 if (type == RX_CLIENT_CONNECTION) {
2468 MUTEX_EXIT(&rx_connHashTable_lock);
2469 return (struct rx_connection *)0;
2471 service = rxi_FindService(socket, serviceId);
2472 if (!service || (securityIndex >= service->nSecurityObjects)
2473 || (service->securityObjects[securityIndex] == 0)) {
2474 MUTEX_EXIT(&rx_connHashTable_lock);
2475 return (struct rx_connection *)0;
2477 conn = rxi_AllocConnection(); /* This bzero's the connection */
2478 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2479 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2480 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2481 conn->next = rx_connHashTable[hashindex];
2482 rx_connHashTable[hashindex] = conn;
2483 conn->peer = rxi_FindPeer(host, port, 0, 1);
2484 conn->type = RX_SERVER_CONNECTION;
2485 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2486 conn->epoch = epoch;
2487 conn->cid = cid & RX_CIDMASK;
2488 /* conn->serial = conn->lastSerial = 0; */
2489 /* conn->timeout = 0; */
2490 conn->ackRate = RX_FAST_ACK_RATE;
2491 conn->service = service;
2492 conn->serviceId = serviceId;
2493 conn->securityIndex = securityIndex;
2494 conn->securityObject = service->securityObjects[securityIndex];
2495 conn->nSpecific = 0;
2496 conn->specific = NULL;
2497 rx_SetConnDeadTime(conn, service->connDeadTime);
2498 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2499 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2500 for (i = 0; i < RX_MAXCALLS; i++) {
2501 conn->twind[i] = rx_initSendWindow;
2502 conn->rwind[i] = rx_initReceiveWindow;
2504 /* Notify security object of the new connection */
2505 RXS_NewConnection(conn->securityObject, conn);
2506 /* XXXX Connection timeout? */
2507 if (service->newConnProc)
2508 (*service->newConnProc) (conn);
2509 rx_MutexIncrement(rx_stats.nServerConns, rx_stats_mutex);
2512 MUTEX_ENTER(&conn->conn_data_lock);
2514 MUTEX_EXIT(&conn->conn_data_lock);
2516 rxLastConn = conn; /* store this connection as the last conn used */
2517 MUTEX_EXIT(&rx_connHashTable_lock);
2521 /* There are two packet tracing routines available for testing and monitoring
2522 * Rx. One is called just after every packet is received and the other is
2523 * called just before every packet is sent. Received packets, have had their
2524 * headers decoded, and packets to be sent have not yet had their headers
2525 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2526 * containing the network address. Both can be modified. The return value, if
2527 * non-zero, indicates that the packet should be dropped. */
2529 int (*rx_justReceived) () = 0;
2530 int (*rx_almostSent) () = 0;
2532 /* A packet has been received off the interface. Np is the packet, socket is
2533 * the socket number it was received from (useful in determining which service
2534 * this packet corresponds to), and (host, port) reflect the host,port of the
2535 * sender. This call returns the packet to the caller if it is finished with
2536 * it, rather than de-allocating it, just as a small performance hack */
2539 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2540 afs_uint32 host, u_short port, int *tnop,
2541 struct rx_call **newcallp)
2543 register struct rx_call *call;
2544 register struct rx_connection *conn;
2546 afs_uint32 currentCallNumber;
2552 struct rx_packet *tnp;
2555 /* We don't print out the packet until now because (1) the time may not be
2556 * accurate enough until now in the lwp implementation (rx_Listener only gets
2557 * the time after the packet is read) and (2) from a protocol point of view,
2558 * this is the first time the packet has been seen */
2559 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2560 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2561 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2562 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2563 np->header.epoch, np->header.cid, np->header.callNumber,
2564 np->header.seq, np->header.flags, np));
2567 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2568 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2571 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2572 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2575 /* If an input tracer function is defined, call it with the packet and
2576 * network address. Note this function may modify its arguments. */
2577 if (rx_justReceived) {
2578 struct sockaddr_in addr;
2580 addr.sin_family = AF_INET;
2581 addr.sin_port = port;
2582 addr.sin_addr.s_addr = host;
2583 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2584 addr.sin_len = sizeof(addr);
2585 #endif /* AFS_OSF_ENV */
2586 drop = (*rx_justReceived) (np, &addr);
2587 /* drop packet if return value is non-zero */
2590 port = addr.sin_port; /* in case fcn changed addr */
2591 host = addr.sin_addr.s_addr;
2595 /* If packet was not sent by the client, then *we* must be the client */
2596 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2597 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2599 /* Find the connection (or fabricate one, if we're the server & if
2600 * necessary) associated with this packet */
2602 rxi_FindConnection(socket, host, port, np->header.serviceId,
2603 np->header.cid, np->header.epoch, type,
2604 np->header.securityIndex);
2607 /* If no connection found or fabricated, just ignore the packet.
2608 * (An argument could be made for sending an abort packet for
2613 MUTEX_ENTER(&conn->conn_data_lock);
2614 if (conn->maxSerial < np->header.serial)
2615 conn->maxSerial = np->header.serial;
2616 MUTEX_EXIT(&conn->conn_data_lock);
2618 /* If the connection is in an error state, send an abort packet and ignore
2619 * the incoming packet */
2621 /* Don't respond to an abort packet--we don't want loops! */
2622 MUTEX_ENTER(&conn->conn_data_lock);
2623 if (np->header.type != RX_PACKET_TYPE_ABORT)
2624 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2626 MUTEX_EXIT(&conn->conn_data_lock);
2630 /* Check for connection-only requests (i.e. not call specific). */
2631 if (np->header.callNumber == 0) {
2632 switch (np->header.type) {
2633 case RX_PACKET_TYPE_ABORT: {
2634 /* What if the supplied error is zero? */
2635 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2636 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2637 rxi_ConnectionError(conn, errcode);
2638 MUTEX_ENTER(&conn->conn_data_lock);
2640 MUTEX_EXIT(&conn->conn_data_lock);
2643 case RX_PACKET_TYPE_CHALLENGE:
2644 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2645 MUTEX_ENTER(&conn->conn_data_lock);
2647 MUTEX_EXIT(&conn->conn_data_lock);
2649 case RX_PACKET_TYPE_RESPONSE:
2650 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2651 MUTEX_ENTER(&conn->conn_data_lock);
2653 MUTEX_EXIT(&conn->conn_data_lock);
2655 case RX_PACKET_TYPE_PARAMS:
2656 case RX_PACKET_TYPE_PARAMS + 1:
2657 case RX_PACKET_TYPE_PARAMS + 2:
2658 /* ignore these packet types for now */
2659 MUTEX_ENTER(&conn->conn_data_lock);
2661 MUTEX_EXIT(&conn->conn_data_lock);
2666 /* Should not reach here, unless the peer is broken: send an
2668 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2669 MUTEX_ENTER(&conn->conn_data_lock);
2670 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2672 MUTEX_EXIT(&conn->conn_data_lock);
2677 channel = np->header.cid & RX_CHANNELMASK;
2678 call = conn->call[channel];
2679 #ifdef RX_ENABLE_LOCKS
2681 MUTEX_ENTER(&call->lock);
2682 /* Test to see if call struct is still attached to conn. */
2683 if (call != conn->call[channel]) {
2685 MUTEX_EXIT(&call->lock);
2686 if (type == RX_SERVER_CONNECTION) {
2687 call = conn->call[channel];
2688 /* If we started with no call attached and there is one now,
2689 * another thread is also running this routine and has gotten
2690 * the connection channel. We should drop this packet in the tests
2691 * below. If there was a call on this connection and it's now
2692 * gone, then we'll be making a new call below.
2693 * If there was previously a call and it's now different then
2694 * the old call was freed and another thread running this routine
2695 * has created a call on this channel. One of these two threads
2696 * has a packet for the old call and the code below handles those
2700 MUTEX_ENTER(&call->lock);
2702 /* This packet can't be for this call. If the new call address is
2703 * 0 then no call is running on this channel. If there is a call
2704 * then, since this is a client connection we're getting data for
2705 * it must be for the previous call.
2707 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2708 MUTEX_ENTER(&conn->conn_data_lock);
2710 MUTEX_EXIT(&conn->conn_data_lock);
2715 currentCallNumber = conn->callNumber[channel];
2717 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2718 if (np->header.callNumber < currentCallNumber) {
2719 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2720 #ifdef RX_ENABLE_LOCKS
2722 MUTEX_EXIT(&call->lock);
2724 MUTEX_ENTER(&conn->conn_data_lock);
2726 MUTEX_EXIT(&conn->conn_data_lock);
2730 MUTEX_ENTER(&conn->conn_call_lock);
2731 call = rxi_NewCall(conn, channel);
2732 MUTEX_EXIT(&conn->conn_call_lock);
2733 *call->callNumber = np->header.callNumber;
2734 if (np->header.callNumber == 0)
2735 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));
2737 call->state = RX_STATE_PRECALL;
2738 clock_GetTime(&call->queueTime);
2739 hzero(call->bytesSent);
2740 hzero(call->bytesRcvd);
2742 * If the number of queued calls exceeds the overload
2743 * threshold then abort this call.
2745 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2746 struct rx_packet *tp;
2748 rxi_CallError(call, rx_BusyError);
2749 tp = rxi_SendCallAbort(call, np, 1, 0);
2750 MUTEX_EXIT(&call->lock);
2751 MUTEX_ENTER(&conn->conn_data_lock);
2753 MUTEX_EXIT(&conn->conn_data_lock);
2754 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2757 rxi_KeepAliveOn(call);
2758 } else if (np->header.callNumber != currentCallNumber) {
2759 /* Wait until the transmit queue is idle before deciding
2760 * whether to reset the current call. Chances are that the
2761 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2764 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2765 while ((call->state == RX_STATE_ACTIVE)
2766 && (call->flags & RX_CALL_TQ_BUSY)) {
2767 call->flags |= RX_CALL_TQ_WAIT;
2769 #ifdef RX_ENABLE_LOCKS
2770 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2771 CV_WAIT(&call->cv_tq, &call->lock);
2772 #else /* RX_ENABLE_LOCKS */
2773 osi_rxSleep(&call->tq);
2774 #endif /* RX_ENABLE_LOCKS */
2776 if (call->tqWaiters == 0)
2777 call->flags &= ~RX_CALL_TQ_WAIT;
2779 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2780 /* If the new call cannot be taken right now send a busy and set
2781 * the error condition in this call, so that it terminates as
2782 * quickly as possible */
2783 if (call->state == RX_STATE_ACTIVE) {
2784 struct rx_packet *tp;
2786 rxi_CallError(call, RX_CALL_DEAD);
2787 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2789 MUTEX_EXIT(&call->lock);
2790 MUTEX_ENTER(&conn->conn_data_lock);
2792 MUTEX_EXIT(&conn->conn_data_lock);
2795 rxi_ResetCall(call, 0);
2796 *call->callNumber = np->header.callNumber;
2797 if (np->header.callNumber == 0)
2798 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));
2800 call->state = RX_STATE_PRECALL;
2801 clock_GetTime(&call->queueTime);
2802 hzero(call->bytesSent);
2803 hzero(call->bytesRcvd);
2805 * If the number of queued calls exceeds the overload
2806 * threshold then abort this call.
2808 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2809 struct rx_packet *tp;
2811 rxi_CallError(call, rx_BusyError);
2812 tp = rxi_SendCallAbort(call, np, 1, 0);
2813 MUTEX_EXIT(&call->lock);
2814 MUTEX_ENTER(&conn->conn_data_lock);
2816 MUTEX_EXIT(&conn->conn_data_lock);
2817 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2820 rxi_KeepAliveOn(call);
2822 /* Continuing call; do nothing here. */
2824 } else { /* we're the client */
2825 /* Ignore all incoming acknowledgements for calls in DALLY state */
2826 if (call && (call->state == RX_STATE_DALLY)
2827 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2828 rx_MutexIncrement(rx_stats.ignorePacketDally, rx_stats_mutex);
2829 #ifdef RX_ENABLE_LOCKS
2831 MUTEX_EXIT(&call->lock);
2834 MUTEX_ENTER(&conn->conn_data_lock);
2836 MUTEX_EXIT(&conn->conn_data_lock);
2840 /* Ignore anything that's not relevant to the current call. If there
2841 * isn't a current call, then no packet is relevant. */
2842 if (!call || (np->header.callNumber != currentCallNumber)) {
2843 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2844 #ifdef RX_ENABLE_LOCKS
2846 MUTEX_EXIT(&call->lock);
2849 MUTEX_ENTER(&conn->conn_data_lock);
2851 MUTEX_EXIT(&conn->conn_data_lock);
2854 /* If the service security object index stamped in the packet does not
2855 * match the connection's security index, ignore the packet */
2856 if (np->header.securityIndex != conn->securityIndex) {
2857 #ifdef RX_ENABLE_LOCKS
2858 MUTEX_EXIT(&call->lock);
2860 MUTEX_ENTER(&conn->conn_data_lock);
2862 MUTEX_EXIT(&conn->conn_data_lock);
2866 /* If we're receiving the response, then all transmit packets are
2867 * implicitly acknowledged. Get rid of them. */
2868 if (np->header.type == RX_PACKET_TYPE_DATA) {
2869 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2870 /* XXX Hack. Because we must release the global rx lock when
2871 * sending packets (osi_NetSend) we drop all acks while we're
2872 * traversing the tq in rxi_Start sending packets out because
2873 * packets may move to the freePacketQueue as result of being here!
2874 * So we drop these packets until we're safely out of the
2875 * traversing. Really ugly!
2876 * For fine grain RX locking, we set the acked field in the
2877 * packets and let rxi_Start remove them from the transmit queue.
2879 if (call->flags & RX_CALL_TQ_BUSY) {
2880 #ifdef RX_ENABLE_LOCKS
2881 rxi_SetAcksInTransmitQueue(call);
2884 return np; /* xmitting; drop packet */
2887 rxi_ClearTransmitQueue(call, 0);
2889 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2890 rxi_ClearTransmitQueue(call, 0);
2891 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2893 if (np->header.type == RX_PACKET_TYPE_ACK) {
2894 /* now check to see if this is an ack packet acknowledging that the
2895 * server actually *lost* some hard-acked data. If this happens we
2896 * ignore this packet, as it may indicate that the server restarted in
2897 * the middle of a call. It is also possible that this is an old ack
2898 * packet. We don't abort the connection in this case, because this
2899 * *might* just be an old ack packet. The right way to detect a server
2900 * restart in the midst of a call is to notice that the server epoch
2902 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2903 * XXX unacknowledged. I think that this is off-by-one, but
2904 * XXX I don't dare change it just yet, since it will
2905 * XXX interact badly with the server-restart detection
2906 * XXX code in receiveackpacket. */
2907 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2908 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2909 MUTEX_EXIT(&call->lock);
2910 MUTEX_ENTER(&conn->conn_data_lock);
2912 MUTEX_EXIT(&conn->conn_data_lock);
2916 } /* else not a data packet */
2919 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2920 /* Set remote user defined status from packet */
2921 call->remoteStatus = np->header.userStatus;
2923 /* Note the gap between the expected next packet and the actual
2924 * packet that arrived, when the new packet has a smaller serial number
2925 * than expected. Rioses frequently reorder packets all by themselves,
2926 * so this will be quite important with very large window sizes.
2927 * Skew is checked against 0 here to avoid any dependence on the type of
2928 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2930 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2931 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2932 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2934 MUTEX_ENTER(&conn->conn_data_lock);
2935 skew = conn->lastSerial - np->header.serial;
2936 conn->lastSerial = np->header.serial;
2937 MUTEX_EXIT(&conn->conn_data_lock);
2939 register struct rx_peer *peer;
2941 if (skew > peer->inPacketSkew) {
2942 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2944 peer->inPacketSkew = skew;
2948 /* Now do packet type-specific processing */
2949 switch (np->header.type) {
2950 case RX_PACKET_TYPE_DATA:
2951 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2954 case RX_PACKET_TYPE_ACK:
2955 /* Respond immediately to ack packets requesting acknowledgement
2957 if (np->header.flags & RX_REQUEST_ACK) {
2959 (void)rxi_SendCallAbort(call, 0, 1, 0);
2961 (void)rxi_SendAck(call, 0, np->header.serial,
2962 RX_ACK_PING_RESPONSE, 1);
2964 np = rxi_ReceiveAckPacket(call, np, 1);
2966 case RX_PACKET_TYPE_ABORT: {
2967 /* An abort packet: reset the call, passing the error up to the user. */
2968 /* What if error is zero? */
2969 /* What if the error is -1? the application will treat it as a timeout. */
2970 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
2971 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
2972 rxi_CallError(call, errdata);
2973 MUTEX_EXIT(&call->lock);
2974 MUTEX_ENTER(&conn->conn_data_lock);
2976 MUTEX_EXIT(&conn->conn_data_lock);
2977 return np; /* xmitting; drop packet */
2979 case RX_PACKET_TYPE_BUSY:
2982 case RX_PACKET_TYPE_ACKALL:
2983 /* All packets acknowledged, so we can drop all packets previously
2984 * readied for sending */
2985 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2986 /* XXX Hack. We because we can't release the global rx lock when
2987 * sending packets (osi_NetSend) we drop all ack pkts while we're
2988 * traversing the tq in rxi_Start sending packets out because
2989 * packets may move to the freePacketQueue as result of being
2990 * here! So we drop these packets until we're safely out of the
2991 * traversing. Really ugly!
2992 * For fine grain RX locking, we set the acked field in the packets
2993 * and let rxi_Start remove the packets from the transmit queue.
2995 if (call->flags & RX_CALL_TQ_BUSY) {
2996 #ifdef RX_ENABLE_LOCKS
2997 rxi_SetAcksInTransmitQueue(call);
2999 #else /* RX_ENABLE_LOCKS */
3000 MUTEX_EXIT(&call->lock);
3001 MUTEX_ENTER(&conn->conn_data_lock);
3003 MUTEX_EXIT(&conn->conn_data_lock);
3004 return np; /* xmitting; drop packet */
3005 #endif /* RX_ENABLE_LOCKS */
3007 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3008 rxi_ClearTransmitQueue(call, 0);
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);
3923 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
3924 maxDgramPackets = MIN(maxDgramPackets, tSize);
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 } else if (!queue_IsEmpty(&call->tq)) {
4058 rxi_Start(0, call, 0, istack);
4063 /* Received a response to a challenge packet */
4065 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
4066 register struct rx_packet *np, int istack)
4070 /* Ignore the packet if we're the client */
4071 if (conn->type == RX_CLIENT_CONNECTION)
4074 /* If already authenticated, ignore the packet (it's probably a retry) */
4075 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4078 /* Otherwise, have the security object evaluate the response packet */
4079 error = RXS_CheckResponse(conn->securityObject, conn, np);
4081 /* If the response is invalid, reset the connection, sending
4082 * an abort to the peer */
4086 rxi_ConnectionError(conn, error);
4087 MUTEX_ENTER(&conn->conn_data_lock);
4088 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4089 MUTEX_EXIT(&conn->conn_data_lock);
4092 /* If the response is valid, any calls waiting to attach
4093 * servers can now do so */
4096 for (i = 0; i < RX_MAXCALLS; i++) {
4097 struct rx_call *call = conn->call[i];
4099 MUTEX_ENTER(&call->lock);
4100 if (call->state == RX_STATE_PRECALL)
4101 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4102 /* tnop can be null if newcallp is null */
4103 MUTEX_EXIT(&call->lock);
4107 /* Update the peer reachability information, just in case
4108 * some calls went into attach-wait while we were waiting
4109 * for authentication..
4111 rxi_UpdatePeerReach(conn, NULL);
4116 /* A client has received an authentication challenge: the security
4117 * object is asked to cough up a respectable response packet to send
4118 * back to the server. The server is responsible for retrying the
4119 * challenge if it fails to get a response. */
4122 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
4123 register struct rx_packet *np, int istack)
4127 /* Ignore the challenge if we're the server */
4128 if (conn->type == RX_SERVER_CONNECTION)
4131 /* Ignore the challenge if the connection is otherwise idle; someone's
4132 * trying to use us as an oracle. */
4133 if (!rxi_HasActiveCalls(conn))
4136 /* Send the security object the challenge packet. It is expected to fill
4137 * in the response. */
4138 error = RXS_GetResponse(conn->securityObject, conn, np);
4140 /* If the security object is unable to return a valid response, reset the
4141 * connection and send an abort to the peer. Otherwise send the response
4142 * packet to the peer connection. */
4144 rxi_ConnectionError(conn, error);
4145 MUTEX_ENTER(&conn->conn_data_lock);
4146 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4147 MUTEX_EXIT(&conn->conn_data_lock);
4149 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4150 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4156 /* Find an available server process to service the current request in
4157 * the given call structure. If one isn't available, queue up this
4158 * call so it eventually gets one */
4160 rxi_AttachServerProc(register struct rx_call *call,
4161 register osi_socket socket, register int *tnop,
4162 register struct rx_call **newcallp)
4164 register struct rx_serverQueueEntry *sq;
4165 register struct rx_service *service = call->conn->service;
4166 register int haveQuota = 0;
4168 /* May already be attached */
4169 if (call->state == RX_STATE_ACTIVE)
4172 MUTEX_ENTER(&rx_serverPool_lock);
4174 haveQuota = QuotaOK(service);
4175 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4176 /* If there are no processes available to service this call,
4177 * put the call on the incoming call queue (unless it's
4178 * already on the queue).
4180 #ifdef RX_ENABLE_LOCKS
4182 ReturnToServerPool(service);
4183 #endif /* RX_ENABLE_LOCKS */
4185 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4186 call->flags |= RX_CALL_WAIT_PROC;
4187 MUTEX_ENTER(&rx_stats_mutex);
4190 MUTEX_EXIT(&rx_stats_mutex);
4191 rxi_calltrace(RX_CALL_ARRIVAL, call);
4192 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4193 queue_Append(&rx_incomingCallQueue, call);
4196 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4198 /* If hot threads are enabled, and both newcallp and sq->socketp
4199 * are non-null, then this thread will process the call, and the
4200 * idle server thread will start listening on this threads socket.
4203 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4206 *sq->socketp = socket;
4207 clock_GetTime(&call->startTime);
4208 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4212 if (call->flags & RX_CALL_WAIT_PROC) {
4213 /* Conservative: I don't think this should happen */
4214 call->flags &= ~RX_CALL_WAIT_PROC;
4215 if (queue_IsOnQueue(call)) {
4217 MUTEX_ENTER(&rx_stats_mutex);
4219 MUTEX_EXIT(&rx_stats_mutex);
4222 call->state = RX_STATE_ACTIVE;
4223 call->mode = RX_MODE_RECEIVING;
4224 #ifdef RX_KERNEL_TRACE
4226 int glockOwner = ISAFS_GLOCK();
4229 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4230 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4236 if (call->flags & RX_CALL_CLEARED) {
4237 /* send an ack now to start the packet flow up again */
4238 call->flags &= ~RX_CALL_CLEARED;
4239 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4241 #ifdef RX_ENABLE_LOCKS
4244 service->nRequestsRunning++;
4245 if (service->nRequestsRunning <= service->minProcs)
4251 MUTEX_EXIT(&rx_serverPool_lock);
4254 /* Delay the sending of an acknowledge event for a short while, while
4255 * a new call is being prepared (in the case of a client) or a reply
4256 * is being prepared (in the case of a server). Rather than sending
4257 * an ack packet, an ACKALL packet is sent. */
4259 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4261 #ifdef RX_ENABLE_LOCKS
4263 MUTEX_ENTER(&call->lock);
4264 call->delayedAckEvent = NULL;
4265 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4267 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4268 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4270 MUTEX_EXIT(&call->lock);
4271 #else /* RX_ENABLE_LOCKS */
4273 call->delayedAckEvent = NULL;
4274 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4275 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4276 #endif /* RX_ENABLE_LOCKS */
4280 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4283 #ifdef RX_ENABLE_LOCKS
4285 MUTEX_ENTER(&call->lock);
4286 if (event == call->delayedAckEvent)
4287 call->delayedAckEvent = NULL;
4288 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4290 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4292 MUTEX_EXIT(&call->lock);
4293 #else /* RX_ENABLE_LOCKS */
4295 call->delayedAckEvent = NULL;
4296 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4297 #endif /* RX_ENABLE_LOCKS */
4301 #ifdef RX_ENABLE_LOCKS
4302 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4303 * clearing them out.
4306 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4308 register struct rx_packet *p, *tp;
4311 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4312 p->flags |= RX_PKTFLAG_ACKED;
4316 call->flags |= RX_CALL_TQ_CLEARME;
4317 call->flags |= RX_CALL_TQ_SOME_ACKED;
4320 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4321 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
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 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4363 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4364 call->nSoftAcked = 0;
4366 if (call->flags & RX_CALL_FAST_RECOVER) {
4367 call->flags &= ~RX_CALL_FAST_RECOVER;
4368 call->cwind = call->nextCwind;
4370 #ifdef RX_ENABLE_LOCKS
4371 CV_SIGNAL(&call->cv_twind);
4373 osi_rxWakeup(&call->twind);
4378 rxi_ClearReceiveQueue(register struct rx_call *call)
4380 if (queue_IsNotEmpty(&call->rq)) {
4381 rx_packetReclaims += rxi_FreePackets(0, &call->rq);
4382 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4384 if (call->state == RX_STATE_PRECALL) {
4385 call->flags |= RX_CALL_CLEARED;
4389 /* Send an abort packet for the specified call */
4391 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4392 int istack, int force)
4395 struct clock when, now;
4400 /* Clients should never delay abort messages */
4401 if (rx_IsClientConn(call->conn))
4404 if (call->abortCode != call->error) {
4405 call->abortCode = call->error;
4406 call->abortCount = 0;
4409 if (force || rxi_callAbortThreshhold == 0
4410 || call->abortCount < rxi_callAbortThreshhold) {
4411 if (call->delayedAbortEvent) {
4412 rxevent_Cancel(call->delayedAbortEvent, call,
4413 RX_CALL_REFCOUNT_ABORT);
4415 error = htonl(call->error);
4418 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4419 (char *)&error, sizeof(error), istack);
4420 } else if (!call->delayedAbortEvent) {
4421 clock_GetTime(&now);
4423 clock_Addmsec(&when, rxi_callAbortDelay);
4424 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4425 call->delayedAbortEvent =
4426 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4431 /* Send an abort packet for the specified connection. Packet is an
4432 * optional pointer to a packet that can be used to send the abort.
4433 * Once the number of abort messages reaches the threshhold, an
4434 * event is scheduled to send the abort. Setting the force flag
4435 * overrides sending delayed abort messages.
4437 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4438 * to send the abort packet.
4441 rxi_SendConnectionAbort(register struct rx_connection *conn,
4442 struct rx_packet *packet, int istack, int force)
4445 struct clock when, now;
4450 /* Clients should never delay abort messages */
4451 if (rx_IsClientConn(conn))
4454 if (force || rxi_connAbortThreshhold == 0
4455 || conn->abortCount < rxi_connAbortThreshhold) {
4456 if (conn->delayedAbortEvent) {
4457 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4459 error = htonl(conn->error);
4461 MUTEX_EXIT(&conn->conn_data_lock);
4463 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4464 RX_PACKET_TYPE_ABORT, (char *)&error,
4465 sizeof(error), istack);
4466 MUTEX_ENTER(&conn->conn_data_lock);
4467 } else if (!conn->delayedAbortEvent) {
4468 clock_GetTime(&now);
4470 clock_Addmsec(&when, rxi_connAbortDelay);
4471 conn->delayedAbortEvent =
4472 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4477 /* Associate an error all of the calls owned by a connection. Called
4478 * with error non-zero. This is only for really fatal things, like
4479 * bad authentication responses. The connection itself is set in
4480 * error at this point, so that future packets received will be
4483 rxi_ConnectionError(register struct rx_connection *conn,
4484 register afs_int32 error)
4489 dpf(("rxi_ConnectionError conn %x error %d", conn, error));
4491 MUTEX_ENTER(&conn->conn_data_lock);
4492 if (conn->challengeEvent)
4493 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4494 if (conn->checkReachEvent) {
4495 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4496 conn->checkReachEvent = 0;
4497 conn->flags &= ~RX_CONN_ATTACHWAIT;
4500 MUTEX_EXIT(&conn->conn_data_lock);
4501 for (i = 0; i < RX_MAXCALLS; i++) {
4502 struct rx_call *call = conn->call[i];
4504 MUTEX_ENTER(&call->lock);
4505 rxi_CallError(call, error);
4506 MUTEX_EXIT(&call->lock);
4509 conn->error = error;
4510 rx_MutexIncrement(rx_stats.fatalErrors, rx_stats_mutex);
4515 rxi_CallError(register struct rx_call *call, afs_int32 error)
4517 dpf(("rxi_CallError call %x error %d call->error %d", call, error, call->error));
4519 error = call->error;
4521 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4522 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4523 rxi_ResetCall(call, 0);
4526 rxi_ResetCall(call, 0);
4528 call->error = error;
4529 call->mode = RX_MODE_ERROR;
4532 /* Reset various fields in a call structure, and wakeup waiting
4533 * processes. Some fields aren't changed: state & mode are not
4534 * touched (these must be set by the caller), and bufptr, nLeft, and
4535 * nFree are not reset, since these fields are manipulated by
4536 * unprotected macros, and may only be reset by non-interrupting code.
4539 /* this code requires that call->conn be set properly as a pre-condition. */
4540 #endif /* ADAPT_WINDOW */
4543 rxi_ResetCall(register struct rx_call *call, register int newcall)
4546 register struct rx_peer *peer;
4547 struct rx_packet *packet;
4549 dpf(("rxi_ResetCall(call %x, newcall %d)\n", call, newcall));
4551 /* Notify anyone who is waiting for asynchronous packet arrival */
4552 if (call->arrivalProc) {
4553 (*call->arrivalProc) (call, call->arrivalProcHandle,
4554 call->arrivalProcArg);
4555 call->arrivalProc = (void (*)())0;
4558 if (call->delayedAbortEvent) {
4559 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4560 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4562 rxi_SendCallAbort(call, packet, 0, 1);
4563 rxi_FreePacket(packet);
4568 * Update the peer with the congestion information in this call
4569 * so other calls on this connection can pick up where this call
4570 * left off. If the congestion sequence numbers don't match then
4571 * another call experienced a retransmission.
4573 peer = call->conn->peer;
4574 MUTEX_ENTER(&peer->peer_lock);
4576 if (call->congestSeq == peer->congestSeq) {
4577 peer->cwind = MAX(peer->cwind, call->cwind);
4578 peer->MTU = MAX(peer->MTU, call->MTU);
4579 peer->nDgramPackets =
4580 MAX(peer->nDgramPackets, call->nDgramPackets);
4583 call->abortCode = 0;
4584 call->abortCount = 0;
4586 if (peer->maxDgramPackets > 1) {
4587 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4589 call->MTU = peer->MTU;
4591 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4592 call->ssthresh = rx_maxSendWindow;
4593 call->nDgramPackets = peer->nDgramPackets;
4594 call->congestSeq = peer->congestSeq;
4595 MUTEX_EXIT(&peer->peer_lock);
4597 flags = call->flags;
4598 rxi_ClearReceiveQueue(call);
4599 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4600 if (flags & RX_CALL_TQ_BUSY) {
4601 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4602 call->flags |= (flags & RX_CALL_TQ_WAIT);
4604 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4606 rxi_ClearTransmitQueue(call, 0);
4607 queue_Init(&call->tq);
4608 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4609 dpf(("rcall %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4612 while (call->tqWaiters) {
4613 #ifdef RX_ENABLE_LOCKS
4614 CV_BROADCAST(&call->cv_tq);
4615 #else /* RX_ENABLE_LOCKS */
4616 osi_rxWakeup(&call->tq);
4617 #endif /* RX_ENABLE_LOCKS */
4621 queue_Init(&call->rq);
4623 call->twind = call->conn->twind[call->channel];
4624 call->rwind = call->conn->rwind[call->channel];
4625 call->nSoftAcked = 0;
4626 call->nextCwind = 0;
4629 call->nCwindAcks = 0;
4630 call->nSoftAcks = 0;
4631 call->nHardAcks = 0;
4633 call->tfirst = call->rnext = call->tnext = 1;
4635 call->lastAcked = 0;
4636 call->localStatus = call->remoteStatus = 0;
4638 if (flags & RX_CALL_READER_WAIT) {
4639 #ifdef RX_ENABLE_LOCKS
4640 CV_BROADCAST(&call->cv_rq);
4642 osi_rxWakeup(&call->rq);
4645 if (flags & RX_CALL_WAIT_PACKETS) {
4646 MUTEX_ENTER(&rx_freePktQ_lock);
4647 rxi_PacketsUnWait(); /* XXX */
4648 MUTEX_EXIT(&rx_freePktQ_lock);
4650 #ifdef RX_ENABLE_LOCKS
4651 CV_SIGNAL(&call->cv_twind);
4653 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4654 osi_rxWakeup(&call->twind);
4657 #ifdef RX_ENABLE_LOCKS
4658 /* The following ensures that we don't mess with any queue while some
4659 * other thread might also be doing so. The call_queue_lock field is
4660 * is only modified under the call lock. If the call is in the process
4661 * of being removed from a queue, the call is not locked until the
4662 * the queue lock is dropped and only then is the call_queue_lock field
4663 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4664 * Note that any other routine which removes a call from a queue has to
4665 * obtain the queue lock before examing the queue and removing the call.
4667 if (call->call_queue_lock) {
4668 MUTEX_ENTER(call->call_queue_lock);
4669 if (queue_IsOnQueue(call)) {
4671 if (flags & RX_CALL_WAIT_PROC) {
4672 MUTEX_ENTER(&rx_stats_mutex);
4674 MUTEX_EXIT(&rx_stats_mutex);
4677 MUTEX_EXIT(call->call_queue_lock);
4678 CLEAR_CALL_QUEUE_LOCK(call);
4680 #else /* RX_ENABLE_LOCKS */
4681 if (queue_IsOnQueue(call)) {
4683 if (flags & RX_CALL_WAIT_PROC)
4686 #endif /* RX_ENABLE_LOCKS */
4688 rxi_KeepAliveOff(call);
4689 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4692 /* Send an acknowledge for the indicated packet (seq,serial) of the
4693 * indicated call, for the indicated reason (reason). This
4694 * acknowledge will specifically acknowledge receiving the packet, and
4695 * will also specify which other packets for this call have been
4696 * received. This routine returns the packet that was used to the
4697 * caller. The caller is responsible for freeing it or re-using it.
4698 * This acknowledgement also returns the highest sequence number
4699 * actually read out by the higher level to the sender; the sender
4700 * promises to keep around packets that have not been read by the
4701 * higher level yet (unless, of course, the sender decides to abort
4702 * the call altogether). Any of p, seq, serial, pflags, or reason may
4703 * be set to zero without ill effect. That is, if they are zero, they
4704 * will not convey any information.
4705 * NOW there is a trailer field, after the ack where it will safely be
4706 * ignored by mundanes, which indicates the maximum size packet this
4707 * host can swallow. */
4709 register struct rx_packet *optionalPacket; use to send ack (or null)
4710 int seq; Sequence number of the packet we are acking
4711 int serial; Serial number of the packet
4712 int pflags; Flags field from packet header
4713 int reason; Reason an acknowledge was prompted
4717 rxi_SendAck(register struct rx_call *call,
4718 register struct rx_packet *optionalPacket, int serial, int reason,
4721 struct rx_ackPacket *ap;
4722 register struct rx_packet *rqp;
4723 register struct rx_packet *nxp; /* For queue_Scan */
4724 register struct rx_packet *p;
4727 #ifdef RX_ENABLE_TSFPQ
4728 struct rx_ts_info_t * rx_ts_info;
4732 * Open the receive window once a thread starts reading packets
4734 if (call->rnext > 1) {
4735 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
4738 call->nHardAcks = 0;
4739 call->nSoftAcks = 0;
4740 if (call->rnext > call->lastAcked)
4741 call->lastAcked = call->rnext;
4745 rx_computelen(p, p->length); /* reset length, you never know */
4746 } /* where that's been... */
4747 #ifdef RX_ENABLE_TSFPQ
4749 RX_TS_INFO_GET(rx_ts_info);
4750 if ((p = rx_ts_info->local_special_packet)) {
4751 rx_computelen(p, p->length);
4752 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4753 rx_ts_info->local_special_packet = p;
4754 } else { /* We won't send the ack, but don't panic. */
4755 return optionalPacket;
4759 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4760 /* We won't send the ack, but don't panic. */
4761 return optionalPacket;
4766 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4769 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4770 #ifndef RX_ENABLE_TSFPQ
4771 if (!optionalPacket)
4774 return optionalPacket;
4776 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4777 if (rx_Contiguous(p) < templ) {
4778 #ifndef RX_ENABLE_TSFPQ
4779 if (!optionalPacket)
4782 return optionalPacket;
4787 /* MTUXXX failing to send an ack is very serious. We should */
4788 /* try as hard as possible to send even a partial ack; it's */
4789 /* better than nothing. */
4790 ap = (struct rx_ackPacket *)rx_DataOf(p);
4791 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4792 ap->reason = reason;
4794 /* The skew computation used to be bogus, I think it's better now. */
4795 /* We should start paying attention to skew. XXX */
4796 ap->serial = htonl(serial);
4797 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4799 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4800 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4802 /* No fear of running out of ack packet here because there can only be at most
4803 * one window full of unacknowledged packets. The window size must be constrained
4804 * to be less than the maximum ack size, of course. Also, an ack should always
4805 * fit into a single packet -- it should not ever be fragmented. */
4806 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4807 if (!rqp || !call->rq.next
4808 || (rqp->header.seq > (call->rnext + call->rwind))) {
4809 #ifndef RX_ENABLE_TSFPQ
4810 if (!optionalPacket)
4813 rxi_CallError(call, RX_CALL_DEAD);
4814 return optionalPacket;
4817 while (rqp->header.seq > call->rnext + offset)
4818 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4819 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4821 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4822 #ifndef RX_ENABLE_TSFPQ
4823 if (!optionalPacket)
4826 rxi_CallError(call, RX_CALL_DEAD);
4827 return optionalPacket;
4832 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4834 /* these are new for AFS 3.3 */
4835 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4836 templ = htonl(templ);
4837 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4838 templ = htonl(call->conn->peer->ifMTU);
4839 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4840 sizeof(afs_int32), &templ);
4842 /* new for AFS 3.4 */
4843 templ = htonl(call->rwind);
4844 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4845 sizeof(afs_int32), &templ);
4847 /* new for AFS 3.5 */
4848 templ = htonl(call->conn->peer->ifDgramPackets);
4849 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4850 sizeof(afs_int32), &templ);
4852 p->header.serviceId = call->conn->serviceId;
4853 p->header.cid = (call->conn->cid | call->channel);
4854 p->header.callNumber = *call->callNumber;
4856 p->header.securityIndex = call->conn->securityIndex;
4857 p->header.epoch = call->conn->epoch;
4858 p->header.type = RX_PACKET_TYPE_ACK;
4859 p->header.flags = RX_SLOW_START_OK;
4860 if (reason == RX_ACK_PING) {
4861 p->header.flags |= RX_REQUEST_ACK;
4863 clock_GetTime(&call->pingRequestTime);
4866 if (call->conn->type == RX_CLIENT_CONNECTION)
4867 p->header.flags |= RX_CLIENT_INITIATED;
4871 if (rxdebug_active) {
4875 len = _snprintf(msg, sizeof(msg),
4876 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4877 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4878 ntohl(ap->serial), ntohl(ap->previousPacket),
4879 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
4880 ap->nAcks, ntohs(ap->bufferSpace) );
4884 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
4885 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4889 OutputDebugString(msg);
4891 #else /* AFS_NT40_ENV */
4893 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
4894 ap->reason, ntohl(ap->previousPacket),
4895 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4897 for (offset = 0; offset < ap->nAcks; offset++)
4898 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4903 #endif /* AFS_NT40_ENV */
4906 register int i, nbytes = p->length;
4908 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4909 if (nbytes <= p->wirevec[i].iov_len) {
4910 register int savelen, saven;
4912 savelen = p->wirevec[i].iov_len;
4914 p->wirevec[i].iov_len = nbytes;
4916 rxi_Send(call, p, istack);
4917 p->wirevec[i].iov_len = savelen;
4921 nbytes -= p->wirevec[i].iov_len;
4924 rx_MutexIncrement(rx_stats.ackPacketsSent, rx_stats_mutex);
4925 #ifndef RX_ENABLE_TSFPQ
4926 if (!optionalPacket)
4929 return optionalPacket; /* Return packet for re-use by caller */
4932 /* Send all of the packets in the list in single datagram */
4934 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4935 int istack, int moreFlag, struct clock *now,
4936 struct clock *retryTime, int resending)
4941 struct rx_connection *conn = call->conn;
4942 struct rx_peer *peer = conn->peer;
4944 MUTEX_ENTER(&peer->peer_lock);
4947 peer->reSends += len;
4948 rx_MutexIncrement(rx_stats.dataPacketsSent, rx_stats_mutex);
4949 MUTEX_EXIT(&peer->peer_lock);
4951 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4955 /* Set the packet flags and schedule the resend events */
4956 /* Only request an ack for the last packet in the list */
4957 for (i = 0; i < len; i++) {
4958 list[i]->retryTime = *retryTime;
4959 if (list[i]->header.serial) {
4960 /* Exponentially backoff retry times */
4961 if (list[i]->backoff < MAXBACKOFF) {
4962 /* so it can't stay == 0 */
4963 list[i]->backoff = (list[i]->backoff << 1) + 1;
4966 clock_Addmsec(&(list[i]->retryTime),
4967 ((afs_uint32) list[i]->backoff) << 8);
4970 /* Wait a little extra for the ack on the last packet */
4971 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4972 clock_Addmsec(&(list[i]->retryTime), 400);
4975 /* Record the time sent */
4976 list[i]->timeSent = *now;
4978 /* Ask for an ack on retransmitted packets, on every other packet
4979 * if the peer doesn't support slow start. Ask for an ack on every
4980 * packet until the congestion window reaches the ack rate. */
4981 if (list[i]->header.serial) {
4983 rx_MutexIncrement(rx_stats.dataPacketsReSent, rx_stats_mutex);
4985 /* improved RTO calculation- not Karn */
4986 list[i]->firstSent = *now;
4987 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4988 || (!(call->flags & RX_CALL_SLOW_START_OK)
4989 && (list[i]->header.seq & 1)))) {
4994 MUTEX_ENTER(&peer->peer_lock);
4998 rx_MutexIncrement(rx_stats.dataPacketsSent, rx_stats_mutex);
4999 MUTEX_EXIT(&peer->peer_lock);
5001 /* Tag this packet as not being the last in this group,
5002 * for the receiver's benefit */
5003 if (i < len - 1 || moreFlag) {
5004 list[i]->header.flags |= RX_MORE_PACKETS;
5007 /* Install the new retransmit time for the packet, and
5008 * record the time sent */
5009 list[i]->timeSent = *now;
5013 list[len - 1]->header.flags |= RX_REQUEST_ACK;
5016 /* Since we're about to send a data packet to the peer, it's
5017 * safe to nuke any scheduled end-of-packets ack */
5018 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5020 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5021 MUTEX_EXIT(&call->lock);
5023 rxi_SendPacketList(call, conn, list, len, istack);
5025 rxi_SendPacket(call, conn, list[0], istack);
5027 MUTEX_ENTER(&call->lock);
5028 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5030 /* Update last send time for this call (for keep-alive
5031 * processing), and for the connection (so that we can discover
5032 * idle connections) */
5033 call->lastSendData = conn->lastSendTime = call->lastSendTime = clock_Sec();
5036 /* When sending packets we need to follow these rules:
5037 * 1. Never send more than maxDgramPackets in a jumbogram.
5038 * 2. Never send a packet with more than two iovecs in a jumbogram.
5039 * 3. Never send a retransmitted packet in a jumbogram.
5040 * 4. Never send more than cwind/4 packets in a jumbogram
5041 * We always keep the last list we should have sent so we
5042 * can set the RX_MORE_PACKETS flags correctly.
5045 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5046 int istack, struct clock *now, struct clock *retryTime,
5049 int i, cnt, lastCnt = 0;
5050 struct rx_packet **listP, **lastP = 0;
5051 struct rx_peer *peer = call->conn->peer;
5052 int morePackets = 0;
5054 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
5055 /* Does the current packet force us to flush the current list? */
5057 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5058 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5060 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
5062 /* If the call enters an error state stop sending, or if
5063 * we entered congestion recovery mode, stop sending */
5064 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5072 /* Add the current packet to the list if it hasn't been acked.
5073 * Otherwise adjust the list pointer to skip the current packet. */
5074 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5076 /* Do we need to flush the list? */
5077 if (cnt >= (int)peer->maxDgramPackets
5078 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5079 || list[i]->header.serial
5080 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5082 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5083 retryTime, resending);
5084 /* If the call enters an error state stop sending, or if
5085 * we entered congestion recovery mode, stop sending */
5087 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5092 listP = &list[i + 1];
5097 osi_Panic("rxi_SendList error");
5099 listP = &list[i + 1];
5103 /* Send the whole list when the call is in receive mode, when
5104 * the call is in eof mode, when we are in fast recovery mode,
5105 * and when we have the last packet */
5106 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5107 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5108 || (call->flags & RX_CALL_FAST_RECOVER)) {
5109 /* Check for the case where the current list contains
5110 * an acked packet. Since we always send retransmissions
5111 * in a separate packet, we only need to check the first
5112 * packet in the list */
5113 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5117 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5118 retryTime, resending);
5119 /* If the call enters an error state stop sending, or if
5120 * we entered congestion recovery mode, stop sending */
5121 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5125 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5128 } else if (lastCnt > 0) {
5129 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5134 #ifdef RX_ENABLE_LOCKS
5135 /* Call rxi_Start, below, but with the call lock held. */
5137 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
5138 void *arg1, int istack)
5140 MUTEX_ENTER(&call->lock);
5141 rxi_Start(event, call, arg1, istack);
5142 MUTEX_EXIT(&call->lock);
5144 #endif /* RX_ENABLE_LOCKS */
5146 /* This routine is called when new packets are readied for
5147 * transmission and when retransmission may be necessary, or when the
5148 * transmission window or burst count are favourable. This should be
5149 * better optimized for new packets, the usual case, now that we've
5150 * got rid of queues of send packets. XXXXXXXXXXX */
5152 rxi_Start(struct rxevent *event, register struct rx_call *call,
5153 void *arg1, int istack)
5155 struct rx_packet *p;
5156 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
5157 struct rx_peer *peer = call->conn->peer;
5158 struct clock now, usenow, retryTime;
5162 struct rx_packet **xmitList;
5165 /* If rxi_Start is being called as a result of a resend event,
5166 * then make sure that the event pointer is removed from the call
5167 * structure, since there is no longer a per-call retransmission
5169 if (event && event == call->resendEvent) {
5170 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5171 call->resendEvent = NULL;
5173 if (queue_IsEmpty(&call->tq)) {
5177 /* Timeouts trigger congestion recovery */
5178 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5179 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5180 /* someone else is waiting to start recovery */
5183 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5184 rxi_WaitforTQBusy(call);
5185 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5186 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5187 call->flags |= RX_CALL_FAST_RECOVER;
5188 if (peer->maxDgramPackets > 1) {
5189 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5191 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5193 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5194 call->nDgramPackets = 1;
5196 call->nextCwind = 1;
5199 MUTEX_ENTER(&peer->peer_lock);
5200 peer->MTU = call->MTU;
5201 peer->cwind = call->cwind;
5202 peer->nDgramPackets = 1;
5204 call->congestSeq = peer->congestSeq;
5205 MUTEX_EXIT(&peer->peer_lock);
5206 /* Clear retry times on packets. Otherwise, it's possible for
5207 * some packets in the queue to force resends at rates faster
5208 * than recovery rates.
5210 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5211 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5212 clock_Zero(&p->retryTime);
5217 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5218 rx_MutexIncrement(rx_tq_debug.rxi_start_in_error, rx_stats_mutex);
5223 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5224 /* Get clock to compute the re-transmit time for any packets
5225 * in this burst. Note, if we back off, it's reasonable to
5226 * back off all of the packets in the same manner, even if
5227 * some of them have been retransmitted more times than more
5229 * Do a dance to avoid blocking after setting now. */
5230 clock_Zero(&retryTime);
5231 MUTEX_ENTER(&peer->peer_lock);
5232 clock_Add(&retryTime, &peer->timeout);
5233 MUTEX_EXIT(&peer->peer_lock);
5234 clock_GetTime(&now);
5235 clock_Add(&retryTime, &now);
5237 /* Send (or resend) any packets that need it, subject to
5238 * window restrictions and congestion burst control
5239 * restrictions. Ask for an ack on the last packet sent in
5240 * this burst. For now, we're relying upon the window being
5241 * considerably bigger than the largest number of packets that
5242 * are typically sent at once by one initial call to
5243 * rxi_Start. This is probably bogus (perhaps we should ask
5244 * for an ack when we're half way through the current
5245 * window?). Also, for non file transfer applications, this
5246 * may end up asking for an ack for every packet. Bogus. XXXX
5249 * But check whether we're here recursively, and let the other guy
5252 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5253 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5254 call->flags |= RX_CALL_TQ_BUSY;
5256 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5258 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5259 call->flags &= ~RX_CALL_NEED_START;
5260 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5262 maxXmitPackets = MIN(call->twind, call->cwind);
5263 xmitList = (struct rx_packet **)
5264 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5265 if (xmitList == NULL)
5266 osi_Panic("rxi_Start, failed to allocate xmit list");
5267 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5268 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5269 /* We shouldn't be sending packets if a thread is waiting
5270 * to initiate congestion recovery */
5274 && (call->flags & RX_CALL_FAST_RECOVER)) {
5275 /* Only send one packet during fast recovery */
5278 if ((p->flags & RX_PKTFLAG_FREE)
5279 || (!queue_IsEnd(&call->tq, nxp)
5280 && (nxp->flags & RX_PKTFLAG_FREE))
5281 || (p == (struct rx_packet *)&rx_freePacketQueue)
5282 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5283 osi_Panic("rxi_Start: xmit queue clobbered");
5285 if (p->flags & RX_PKTFLAG_ACKED) {
5286 /* Since we may block, don't trust this */
5287 usenow.sec = usenow.usec = 0;
5288 rx_MutexIncrement(rx_stats.ignoreAckedPacket, rx_stats_mutex);
5289 continue; /* Ignore this packet if it has been acknowledged */
5292 /* Turn off all flags except these ones, which are the same
5293 * on each transmission */
5294 p->header.flags &= RX_PRESET_FLAGS;
5296 if (p->header.seq >=
5297 call->tfirst + MIN((int)call->twind,
5298 (int)(call->nSoftAcked +
5300 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5301 /* Note: if we're waiting for more window space, we can
5302 * still send retransmits; hence we don't return here, but
5303 * break out to schedule a retransmit event */
5304 dpf(("call %d waiting for window",
5305 *(call->callNumber)));
5309 /* Transmit the packet if it needs to be sent. */
5310 if (!clock_Lt(&now, &p->retryTime)) {
5311 if (nXmitPackets == maxXmitPackets) {
5312 rxi_SendXmitList(call, xmitList, nXmitPackets,
5313 istack, &now, &retryTime,
5315 osi_Free(xmitList, maxXmitPackets *
5316 sizeof(struct rx_packet *));
5319 xmitList[nXmitPackets++] = p;
5323 /* xmitList now hold pointers to all of the packets that are
5324 * ready to send. Now we loop to send the packets */
5325 if (nXmitPackets > 0) {
5326 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5327 &now, &retryTime, resending);
5330 maxXmitPackets * sizeof(struct rx_packet *));
5332 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5334 * TQ references no longer protected by this flag; they must remain
5335 * protected by the global lock.
5337 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5338 call->flags &= ~RX_CALL_TQ_BUSY;
5339 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5340 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5341 #ifdef RX_ENABLE_LOCKS
5342 osirx_AssertMine(&call->lock, "rxi_Start start");
5343 CV_BROADCAST(&call->cv_tq);
5344 #else /* RX_ENABLE_LOCKS */
5345 osi_rxWakeup(&call->tq);
5346 #endif /* RX_ENABLE_LOCKS */
5351 /* We went into the error state while sending packets. Now is
5352 * the time to reset the call. This will also inform the using
5353 * process that the call is in an error state.
5355 rx_MutexIncrement(rx_tq_debug.rxi_start_aborted, rx_stats_mutex);
5356 call->flags &= ~RX_CALL_TQ_BUSY;
5357 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5358 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5359 #ifdef RX_ENABLE_LOCKS
5360 osirx_AssertMine(&call->lock, "rxi_Start middle");
5361 CV_BROADCAST(&call->cv_tq);
5362 #else /* RX_ENABLE_LOCKS */
5363 osi_rxWakeup(&call->tq);
5364 #endif /* RX_ENABLE_LOCKS */
5366 rxi_CallError(call, call->error);
5369 #ifdef RX_ENABLE_LOCKS
5370 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5371 register int missing;
5372 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5373 /* Some packets have received acks. If they all have, we can clear
5374 * the transmit queue.
5377 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5378 if (p->header.seq < call->tfirst
5379 && (p->flags & RX_PKTFLAG_ACKED)) {
5386 call->flags |= RX_CALL_TQ_CLEARME;
5388 #endif /* RX_ENABLE_LOCKS */
5389 /* Don't bother doing retransmits if the TQ is cleared. */
5390 if (call->flags & RX_CALL_TQ_CLEARME) {
5391 rxi_ClearTransmitQueue(call, 1);
5393 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5396 /* Always post a resend event, if there is anything in the
5397 * queue, and resend is possible. There should be at least
5398 * one unacknowledged packet in the queue ... otherwise none
5399 * of these packets should be on the queue in the first place.
5401 if (call->resendEvent) {
5402 /* Cancel the existing event and post a new one */
5403 rxevent_Cancel(call->resendEvent, call,
5404 RX_CALL_REFCOUNT_RESEND);
5407 /* The retry time is the retry time on the first unacknowledged
5408 * packet inside the current window */
5410 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5411 /* Don't set timers for packets outside the window */
5412 if (p->header.seq >= call->tfirst + call->twind) {
5416 if (!(p->flags & RX_PKTFLAG_ACKED)
5417 && !clock_IsZero(&p->retryTime)) {
5419 retryTime = p->retryTime;
5424 /* Post a new event to re-run rxi_Start when retries may be needed */
5425 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5426 #ifdef RX_ENABLE_LOCKS
5427 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5429 rxevent_PostNow2(&retryTime, &usenow,
5431 (void *)call, 0, istack);
5432 #else /* RX_ENABLE_LOCKS */
5434 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5435 (void *)call, 0, istack);
5436 #endif /* RX_ENABLE_LOCKS */
5439 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5440 } while (call->flags & RX_CALL_NEED_START);
5442 * TQ references no longer protected by this flag; they must remain
5443 * protected by the global lock.
5445 call->flags &= ~RX_CALL_TQ_BUSY;
5446 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5447 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5448 #ifdef RX_ENABLE_LOCKS
5449 osirx_AssertMine(&call->lock, "rxi_Start end");
5450 CV_BROADCAST(&call->cv_tq);
5451 #else /* RX_ENABLE_LOCKS */
5452 osi_rxWakeup(&call->tq);
5453 #endif /* RX_ENABLE_LOCKS */
5456 call->flags |= RX_CALL_NEED_START;
5458 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5460 if (call->resendEvent) {
5461 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5466 /* Also adjusts the keep alive parameters for the call, to reflect
5467 * that we have just sent a packet (so keep alives aren't sent
5470 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5473 register struct rx_connection *conn = call->conn;
5475 /* Stamp each packet with the user supplied status */
5476 p->header.userStatus = call->localStatus;
5478 /* Allow the security object controlling this call's security to
5479 * make any last-minute changes to the packet */
5480 RXS_SendPacket(conn->securityObject, call, p);
5482 /* Since we're about to send SOME sort of packet to the peer, it's
5483 * safe to nuke any scheduled end-of-packets ack */
5484 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5486 /* Actually send the packet, filling in more connection-specific fields */
5487 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5488 MUTEX_EXIT(&call->lock);
5489 rxi_SendPacket(call, conn, p, istack);
5490 MUTEX_ENTER(&call->lock);
5491 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5493 /* Update last send time for this call (for keep-alive
5494 * processing), and for the connection (so that we can discover
5495 * idle connections) */
5496 conn->lastSendTime = call->lastSendTime = clock_Sec();
5497 /* Don't count keepalives here, so idleness can be tracked. */
5498 if (p->header.type != RX_PACKET_TYPE_ACK)
5499 call->lastSendData = call->lastSendTime;
5503 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5504 * that things are fine. Also called periodically to guarantee that nothing
5505 * falls through the cracks (e.g. (error + dally) connections have keepalive
5506 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5508 * haveCTLock Set if calling from rxi_ReapConnections
5510 #ifdef RX_ENABLE_LOCKS
5512 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5513 #else /* RX_ENABLE_LOCKS */
5515 rxi_CheckCall(register struct rx_call *call)
5516 #endif /* RX_ENABLE_LOCKS */
5518 register struct rx_connection *conn = call->conn;
5520 afs_uint32 deadTime;
5522 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5523 if (call->flags & RX_CALL_TQ_BUSY) {
5524 /* Call is active and will be reset by rxi_Start if it's
5525 * in an error state.
5530 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5532 (((afs_uint32) conn->secondsUntilDead << 10) +
5533 ((afs_uint32) conn->peer->rtt >> 3) +
5534 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5536 /* These are computed to the second (+- 1 second). But that's
5537 * good enough for these values, which should be a significant
5538 * number of seconds. */
5539 if (now > (call->lastReceiveTime + deadTime)) {
5540 if (call->state == RX_STATE_ACTIVE) {
5542 #if defined(KERNEL) && defined(AFS_SUN57_ENV)
5544 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5545 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
5546 ip_stack_t *ipst = ns->netstack_ip;
5548 ire = ire_cache_lookup(call->conn->peer->host
5549 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
5551 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
5553 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5560 if (ire && ire->ire_max_frag > 0)
5561 rxi_SetPeerMtu(call->conn->peer->host, 0, ire->ire_max_frag);
5562 #if defined(GLOBAL_NETSTACKID)
5566 #endif /* ADAPT_PMTU */
5567 rxi_CallError(call, RX_CALL_DEAD);
5570 #ifdef RX_ENABLE_LOCKS
5571 /* Cancel pending events */
5572 rxevent_Cancel(call->delayedAckEvent, call,
5573 RX_CALL_REFCOUNT_DELAY);
5574 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5575 rxevent_Cancel(call->keepAliveEvent, call,
5576 RX_CALL_REFCOUNT_ALIVE);
5577 if (call->refCount == 0) {
5578 rxi_FreeCall(call, haveCTLock);
5582 #else /* RX_ENABLE_LOCKS */
5585 #endif /* RX_ENABLE_LOCKS */
5587 /* Non-active calls are destroyed if they are not responding
5588 * to pings; active calls are simply flagged in error, so the
5589 * attached process can die reasonably gracefully. */
5591 /* see if we have a non-activity timeout */
5592 if (call->startWait && conn->idleDeadTime
5593 && ((call->startWait + conn->idleDeadTime) < now)) {
5594 if (call->state == RX_STATE_ACTIVE) {
5595 rxi_CallError(call, RX_CALL_TIMEOUT);
5599 if (call->lastSendData && conn->idleDeadTime && (conn->idleDeadErr != 0)
5600 && ((call->lastSendData + conn->idleDeadTime) < now)) {
5601 if (call->state == RX_STATE_ACTIVE) {
5602 rxi_CallError(call, conn->idleDeadErr);
5606 /* see if we have a hard timeout */
5607 if (conn->hardDeadTime
5608 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5609 if (call->state == RX_STATE_ACTIVE)
5610 rxi_CallError(call, RX_CALL_TIMEOUT);
5617 /* When a call is in progress, this routine is called occasionally to
5618 * make sure that some traffic has arrived (or been sent to) the peer.
5619 * If nothing has arrived in a reasonable amount of time, the call is
5620 * declared dead; if nothing has been sent for a while, we send a
5621 * keep-alive packet (if we're actually trying to keep the call alive)
5624 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5627 struct rx_connection *conn;
5630 MUTEX_ENTER(&call->lock);
5631 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5632 if (event == call->keepAliveEvent)
5633 call->keepAliveEvent = NULL;
5636 #ifdef RX_ENABLE_LOCKS
5637 if (rxi_CheckCall(call, 0)) {
5638 MUTEX_EXIT(&call->lock);
5641 #else /* RX_ENABLE_LOCKS */
5642 if (rxi_CheckCall(call))
5644 #endif /* RX_ENABLE_LOCKS */
5646 /* Don't try to keep alive dallying calls */
5647 if (call->state == RX_STATE_DALLY) {
5648 MUTEX_EXIT(&call->lock);
5653 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5654 /* Don't try to send keepalives if there is unacknowledged data */
5655 /* the rexmit code should be good enough, this little hack
5656 * doesn't quite work XXX */
5657 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5659 rxi_ScheduleKeepAliveEvent(call);
5660 MUTEX_EXIT(&call->lock);
5665 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5667 if (!call->keepAliveEvent) {
5668 struct clock when, now;
5669 clock_GetTime(&now);
5671 when.sec += call->conn->secondsUntilPing;
5672 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5673 call->keepAliveEvent =
5674 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
5678 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5680 rxi_KeepAliveOn(register struct rx_call *call)
5682 /* Pretend last packet received was received now--i.e. if another
5683 * packet isn't received within the keep alive time, then the call
5684 * will die; Initialize last send time to the current time--even
5685 * if a packet hasn't been sent yet. This will guarantee that a
5686 * keep-alive is sent within the ping time */
5687 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5688 rxi_ScheduleKeepAliveEvent(call);
5691 /* This routine is called to send connection abort messages
5692 * that have been delayed to throttle looping clients. */
5694 rxi_SendDelayedConnAbort(struct rxevent *event,
5695 register struct rx_connection *conn, char *dummy)
5698 struct rx_packet *packet;
5700 MUTEX_ENTER(&conn->conn_data_lock);
5701 conn->delayedAbortEvent = NULL;
5702 error = htonl(conn->error);
5704 MUTEX_EXIT(&conn->conn_data_lock);
5705 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5708 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5709 RX_PACKET_TYPE_ABORT, (char *)&error,
5711 rxi_FreePacket(packet);
5715 /* This routine is called to send call abort messages
5716 * that have been delayed to throttle looping clients. */
5718 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5722 struct rx_packet *packet;
5724 MUTEX_ENTER(&call->lock);
5725 call->delayedAbortEvent = NULL;
5726 error = htonl(call->error);
5728 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5731 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5732 (char *)&error, sizeof(error), 0);
5733 rxi_FreePacket(packet);
5735 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5736 MUTEX_EXIT(&call->lock);
5739 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5740 * seconds) to ask the client to authenticate itself. The routine
5741 * issues a challenge to the client, which is obtained from the
5742 * security object associated with the connection */
5744 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5745 void *arg1, int tries)
5747 conn->challengeEvent = NULL;
5748 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5749 register struct rx_packet *packet;
5750 struct clock when, now;
5753 /* We've failed to authenticate for too long.
5754 * Reset any calls waiting for authentication;
5755 * they are all in RX_STATE_PRECALL.
5759 MUTEX_ENTER(&conn->conn_call_lock);
5760 for (i = 0; i < RX_MAXCALLS; i++) {
5761 struct rx_call *call = conn->call[i];
5763 MUTEX_ENTER(&call->lock);
5764 if (call->state == RX_STATE_PRECALL) {
5765 rxi_CallError(call, RX_CALL_DEAD);
5766 rxi_SendCallAbort(call, NULL, 0, 0);
5768 MUTEX_EXIT(&call->lock);
5771 MUTEX_EXIT(&conn->conn_call_lock);
5775 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5777 /* If there's no packet available, do this later. */
5778 RXS_GetChallenge(conn->securityObject, conn, packet);
5779 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5780 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5781 rxi_FreePacket(packet);
5783 clock_GetTime(&now);
5785 when.sec += RX_CHALLENGE_TIMEOUT;
5786 conn->challengeEvent =
5787 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
5792 /* Call this routine to start requesting the client to authenticate
5793 * itself. This will continue until authentication is established,
5794 * the call times out, or an invalid response is returned. The
5795 * security object associated with the connection is asked to create
5796 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5797 * defined earlier. */
5799 rxi_ChallengeOn(register struct rx_connection *conn)
5801 if (!conn->challengeEvent) {
5802 RXS_CreateChallenge(conn->securityObject, conn);
5803 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5808 /* Compute round trip time of the packet provided, in *rttp.
5811 /* rxi_ComputeRoundTripTime is called with peer locked. */
5812 /* sentp and/or peer may be null */
5814 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5815 register struct clock *sentp,
5816 register struct rx_peer *peer)
5818 struct clock thisRtt, *rttp = &thisRtt;
5820 register int rtt_timeout;
5822 clock_GetTime(rttp);
5824 if (clock_Lt(rttp, sentp)) {
5826 return; /* somebody set the clock back, don't count this time. */
5828 clock_Sub(rttp, sentp);
5829 MUTEX_ENTER(&rx_stats_mutex);
5830 if (clock_Lt(rttp, &rx_stats.minRtt))
5831 rx_stats.minRtt = *rttp;
5832 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5833 if (rttp->sec > 60) {
5834 MUTEX_EXIT(&rx_stats_mutex);
5835 return; /* somebody set the clock ahead */
5837 rx_stats.maxRtt = *rttp;
5839 clock_Add(&rx_stats.totalRtt, rttp);
5840 rx_stats.nRttSamples++;
5841 MUTEX_EXIT(&rx_stats_mutex);
5843 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5845 /* Apply VanJacobson round-trip estimations */
5850 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5851 * srtt is stored as fixed point with 3 bits after the binary
5852 * point (i.e., scaled by 8). The following magic is
5853 * equivalent to the smoothing algorithm in rfc793 with an
5854 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5855 * srtt*8 = srtt*8 + rtt - srtt
5856 * srtt = srtt + rtt/8 - srtt/8
5859 delta = MSEC(rttp) - (peer->rtt >> 3);
5863 * We accumulate a smoothed rtt variance (actually, a smoothed
5864 * mean difference), then set the retransmit timer to smoothed
5865 * rtt + 4 times the smoothed variance (was 2x in van's original
5866 * paper, but 4x works better for me, and apparently for him as
5868 * rttvar is stored as
5869 * fixed point with 2 bits after the binary point (scaled by
5870 * 4). The following is equivalent to rfc793 smoothing with
5871 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5872 * replaces rfc793's wired-in beta.
5873 * dev*4 = dev*4 + (|actual - expected| - dev)
5879 delta -= (peer->rtt_dev >> 2);
5880 peer->rtt_dev += delta;
5882 /* I don't have a stored RTT so I start with this value. Since I'm
5883 * probably just starting a call, and will be pushing more data down
5884 * this, I expect congestion to increase rapidly. So I fudge a
5885 * little, and I set deviance to half the rtt. In practice,
5886 * deviance tends to approach something a little less than
5887 * half the smoothed rtt. */
5888 peer->rtt = (MSEC(rttp) << 3) + 8;
5889 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5891 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5892 * the other of these connections is usually in a user process, and can
5893 * be switched and/or swapped out. So on fast, reliable networks, the
5894 * timeout would otherwise be too short.
5896 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5897 clock_Zero(&(peer->timeout));
5898 clock_Addmsec(&(peer->timeout), rtt_timeout);
5900 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)));
5904 /* Find all server connections that have not been active for a long time, and
5907 rxi_ReapConnections(void)
5909 struct clock now, when;
5910 clock_GetTime(&now);
5912 /* Find server connection structures that haven't been used for
5913 * greater than rx_idleConnectionTime */
5915 struct rx_connection **conn_ptr, **conn_end;
5916 int i, havecalls = 0;
5917 MUTEX_ENTER(&rx_connHashTable_lock);
5918 for (conn_ptr = &rx_connHashTable[0], conn_end =
5919 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5921 struct rx_connection *conn, *next;
5922 struct rx_call *call;
5926 for (conn = *conn_ptr; conn; conn = next) {
5927 /* XXX -- Shouldn't the connection be locked? */
5930 for (i = 0; i < RX_MAXCALLS; i++) {
5931 call = conn->call[i];
5934 MUTEX_ENTER(&call->lock);
5935 #ifdef RX_ENABLE_LOCKS
5936 result = rxi_CheckCall(call, 1);
5937 #else /* RX_ENABLE_LOCKS */
5938 result = rxi_CheckCall(call);
5939 #endif /* RX_ENABLE_LOCKS */
5940 MUTEX_EXIT(&call->lock);
5942 /* If CheckCall freed the call, it might
5943 * have destroyed the connection as well,
5944 * which screws up the linked lists.
5950 if (conn->type == RX_SERVER_CONNECTION) {
5951 /* This only actually destroys the connection if
5952 * there are no outstanding calls */
5953 MUTEX_ENTER(&conn->conn_data_lock);
5954 if (!havecalls && !conn->refCount
5955 && ((conn->lastSendTime + rx_idleConnectionTime) <
5957 conn->refCount++; /* it will be decr in rx_DestroyConn */
5958 MUTEX_EXIT(&conn->conn_data_lock);
5959 #ifdef RX_ENABLE_LOCKS
5960 rxi_DestroyConnectionNoLock(conn);
5961 #else /* RX_ENABLE_LOCKS */
5962 rxi_DestroyConnection(conn);
5963 #endif /* RX_ENABLE_LOCKS */
5965 #ifdef RX_ENABLE_LOCKS
5967 MUTEX_EXIT(&conn->conn_data_lock);
5969 #endif /* RX_ENABLE_LOCKS */
5973 #ifdef RX_ENABLE_LOCKS
5974 while (rx_connCleanup_list) {
5975 struct rx_connection *conn;
5976 conn = rx_connCleanup_list;
5977 rx_connCleanup_list = rx_connCleanup_list->next;
5978 MUTEX_EXIT(&rx_connHashTable_lock);
5979 rxi_CleanupConnection(conn);
5980 MUTEX_ENTER(&rx_connHashTable_lock);
5982 MUTEX_EXIT(&rx_connHashTable_lock);
5983 #endif /* RX_ENABLE_LOCKS */
5986 /* Find any peer structures that haven't been used (haven't had an
5987 * associated connection) for greater than rx_idlePeerTime */
5989 struct rx_peer **peer_ptr, **peer_end;
5991 MUTEX_ENTER(&rx_rpc_stats);
5992 MUTEX_ENTER(&rx_peerHashTable_lock);
5993 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5994 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5996 struct rx_peer *peer, *next, *prev;
5997 for (prev = peer = *peer_ptr; peer; peer = next) {
5999 code = MUTEX_TRYENTER(&peer->peer_lock);
6000 if ((code) && (peer->refCount == 0)
6001 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6002 rx_interface_stat_p rpc_stat, nrpc_stat;
6004 MUTEX_EXIT(&peer->peer_lock);
6005 MUTEX_DESTROY(&peer->peer_lock);
6007 (&peer->rpcStats, rpc_stat, nrpc_stat,
6008 rx_interface_stat)) {
6009 unsigned int num_funcs;
6012 queue_Remove(&rpc_stat->queue_header);
6013 queue_Remove(&rpc_stat->all_peers);
6014 num_funcs = rpc_stat->stats[0].func_total;
6016 sizeof(rx_interface_stat_t) +
6017 rpc_stat->stats[0].func_total *
6018 sizeof(rx_function_entry_v1_t);
6020 rxi_Free(rpc_stat, space);
6021 rxi_rpc_peer_stat_cnt -= num_funcs;
6024 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6025 if (peer == *peer_ptr) {
6032 MUTEX_EXIT(&peer->peer_lock);
6038 MUTEX_EXIT(&rx_peerHashTable_lock);
6039 MUTEX_EXIT(&rx_rpc_stats);
6042 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
6043 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
6044 * GC, just below. Really, we shouldn't have to keep moving packets from
6045 * one place to another, but instead ought to always know if we can
6046 * afford to hold onto a packet in its particular use. */
6047 MUTEX_ENTER(&rx_freePktQ_lock);
6048 if (rx_waitingForPackets) {
6049 rx_waitingForPackets = 0;
6050 #ifdef RX_ENABLE_LOCKS
6051 CV_BROADCAST(&rx_waitingForPackets_cv);
6053 osi_rxWakeup(&rx_waitingForPackets);
6056 MUTEX_EXIT(&rx_freePktQ_lock);
6059 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
6060 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
6064 /* rxs_Release - This isn't strictly necessary but, since the macro name from
6065 * rx.h is sort of strange this is better. This is called with a security
6066 * object before it is discarded. Each connection using a security object has
6067 * its own refcount to the object so it won't actually be freed until the last
6068 * connection is destroyed.
6070 * This is the only rxs module call. A hold could also be written but no one
6074 rxs_Release(struct rx_securityClass *aobj)
6076 return RXS_Close(aobj);
6080 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
6081 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
6082 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
6083 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
6085 /* Adjust our estimate of the transmission rate to this peer, given
6086 * that the packet p was just acked. We can adjust peer->timeout and
6087 * call->twind. Pragmatically, this is called
6088 * only with packets of maximal length.
6089 * Called with peer and call locked.
6093 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
6094 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
6096 afs_int32 xferSize, xferMs;
6097 register afs_int32 minTime;
6100 /* Count down packets */
6101 if (peer->rateFlag > 0)
6103 /* Do nothing until we're enabled */
6104 if (peer->rateFlag != 0)
6109 /* Count only when the ack seems legitimate */
6110 switch (ackReason) {
6111 case RX_ACK_REQUESTED:
6113 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6117 case RX_ACK_PING_RESPONSE:
6118 if (p) /* want the response to ping-request, not data send */
6120 clock_GetTime(&newTO);
6121 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6122 clock_Sub(&newTO, &call->pingRequestTime);
6123 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6127 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6134 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));
6136 /* Track only packets that are big enough. */
6137 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6141 /* absorb RTT data (in milliseconds) for these big packets */
6142 if (peer->smRtt == 0) {
6143 peer->smRtt = xferMs;
6145 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6150 if (peer->countDown) {
6154 peer->countDown = 10; /* recalculate only every so often */
6156 /* In practice, we can measure only the RTT for full packets,
6157 * because of the way Rx acks the data that it receives. (If it's
6158 * smaller than a full packet, it often gets implicitly acked
6159 * either by the call response (from a server) or by the next call
6160 * (from a client), and either case confuses transmission times
6161 * with processing times.) Therefore, replace the above
6162 * more-sophisticated processing with a simpler version, where the
6163 * smoothed RTT is kept for full-size packets, and the time to
6164 * transmit a windowful of full-size packets is simply RTT *
6165 * windowSize. Again, we take two steps:
6166 - ensure the timeout is large enough for a single packet's RTT;
6167 - ensure that the window is small enough to fit in the desired timeout.*/
6169 /* First, the timeout check. */
6170 minTime = peer->smRtt;
6171 /* Get a reasonable estimate for a timeout period */
6173 newTO.sec = minTime / 1000;
6174 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6176 /* Increase the timeout period so that we can always do at least
6177 * one packet exchange */
6178 if (clock_Gt(&newTO, &peer->timeout)) {
6180 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));
6182 peer->timeout = newTO;
6185 /* Now, get an estimate for the transmit window size. */
6186 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6187 /* Now, convert to the number of full packets that could fit in a
6188 * reasonable fraction of that interval */
6189 minTime /= (peer->smRtt << 1);
6190 xferSize = minTime; /* (make a copy) */
6192 /* Now clamp the size to reasonable bounds. */
6195 else if (minTime > rx_Window)
6196 minTime = rx_Window;
6197 /* if (minTime != peer->maxWindow) {
6198 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6199 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6200 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6202 peer->maxWindow = minTime;
6203 elide... call->twind = minTime;
6207 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6208 * Discern this by calculating the timeout necessary for rx_Window
6210 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6211 /* calculate estimate for transmission interval in milliseconds */
6212 minTime = rx_Window * peer->smRtt;
6213 if (minTime < 1000) {
6214 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
6215 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6216 peer->timeout.usec, peer->smRtt, peer->packetSize));
6218 newTO.sec = 0; /* cut back on timeout by half a second */
6219 newTO.usec = 500000;
6220 clock_Sub(&peer->timeout, &newTO);
6225 } /* end of rxi_ComputeRate */
6226 #endif /* ADAPT_WINDOW */
6234 #define TRACE_OPTION_DEBUGLOG 4
6242 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6243 0, KEY_QUERY_VALUE, &parmKey);
6244 if (code != ERROR_SUCCESS)
6247 dummyLen = sizeof(TraceOption);
6248 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6249 (BYTE *) &TraceOption, &dummyLen);
6250 if (code == ERROR_SUCCESS) {
6251 rxdebug_active = (TraceOption & TRACE_OPTION_DEBUGLOG) ? 1 : 0;
6253 RegCloseKey (parmKey);
6254 #endif /* AFS_NT40_ENV */
6259 rx_DebugOnOff(int on)
6261 rxdebug_active = on;
6263 #endif /* AFS_NT40_ENV */
6266 /* Don't call this debugging routine directly; use dpf */
6268 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
6269 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
6277 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6280 len = _snprintf(msg, sizeof(msg)-2,
6281 tformat, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10,
6282 a11, a12, a13, a14, a15);
6284 if (msg[len-1] != '\n') {
6288 OutputDebugString(msg);
6293 clock_GetTime(&now);
6294 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
6295 (unsigned int)now.usec / 1000);
6296 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
6303 * This function is used to process the rx_stats structure that is local
6304 * to a process as well as an rx_stats structure received from a remote
6305 * process (via rxdebug). Therefore, it needs to do minimal version
6309 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
6310 afs_int32 freePackets, char version)
6314 if (size != sizeof(struct rx_stats)) {
6316 "Unexpected size of stats structure: was %d, expected %d\n",
6317 size, sizeof(struct rx_stats));
6320 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6323 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6324 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6325 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6326 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6327 s->specialPktAllocFailures);
6329 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6330 s->receivePktAllocFailures, s->sendPktAllocFailures,
6331 s->specialPktAllocFailures);
6335 " greedy %d, " "bogusReads %d (last from host %x), "
6336 "noPackets %d, " "noBuffers %d, " "selects %d, "
6337 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6338 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6339 s->selects, s->sendSelects);
6341 fprintf(file, " packets read: ");
6342 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6343 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6345 fprintf(file, "\n");
6348 " other read counters: data %d, " "ack %d, " "dup %d "
6349 "spurious %d " "dally %d\n", s->dataPacketsRead,
6350 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6351 s->ignorePacketDally);
6353 fprintf(file, " packets sent: ");
6354 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6355 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6357 fprintf(file, "\n");
6360 " other send counters: ack %d, " "data %d (not resends), "
6361 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6362 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6363 s->dataPacketsPushed, s->ignoreAckedPacket);
6366 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6367 s->netSendFailures, (int)s->fatalErrors);
6369 if (s->nRttSamples) {
6370 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6371 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6373 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6374 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6378 " %d server connections, " "%d client connections, "
6379 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6380 s->nServerConns, s->nClientConns, s->nPeerStructs,
6381 s->nCallStructs, s->nFreeCallStructs);
6383 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6384 fprintf(file, " %d clock updates\n", clock_nUpdates);
6389 /* for backward compatibility */
6391 rx_PrintStats(FILE * file)
6393 MUTEX_ENTER(&rx_stats_mutex);
6394 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6396 MUTEX_EXIT(&rx_stats_mutex);
6400 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6402 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6403 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6404 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6407 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6408 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6409 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6412 " Packet size %d, " "max in packet skew %d, "
6413 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6414 (int)peer->outPacketSkew);
6417 #ifdef AFS_PTHREAD_ENV
6419 * This mutex protects the following static variables:
6423 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6424 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6426 #define LOCK_RX_DEBUG
6427 #define UNLOCK_RX_DEBUG
6428 #endif /* AFS_PTHREAD_ENV */
6431 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6432 u_char type, void *inputData, size_t inputLength,
6433 void *outputData, size_t outputLength)
6435 static afs_int32 counter = 100;
6436 time_t waitTime, waitCount, startTime;
6437 struct rx_header theader;
6439 register afs_int32 code;
6440 struct timeval tv_now, tv_wake, tv_delta;
6441 struct sockaddr_in taddr, faddr;
6446 startTime = time(0);
6452 tp = &tbuffer[sizeof(struct rx_header)];
6453 taddr.sin_family = AF_INET;
6454 taddr.sin_port = remotePort;
6455 taddr.sin_addr.s_addr = remoteAddr;
6456 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6457 taddr.sin_len = sizeof(struct sockaddr_in);
6460 memset(&theader, 0, sizeof(theader));
6461 theader.epoch = htonl(999);
6463 theader.callNumber = htonl(counter);
6466 theader.type = type;
6467 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6468 theader.serviceId = 0;
6470 memcpy(tbuffer, &theader, sizeof(theader));
6471 memcpy(tp, inputData, inputLength);
6473 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6474 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6476 /* see if there's a packet available */
6477 gettimeofday(&tv_wake,0);
6478 tv_wake.tv_sec += waitTime;
6481 FD_SET(socket, &imask);
6482 tv_delta.tv_sec = tv_wake.tv_sec;
6483 tv_delta.tv_usec = tv_wake.tv_usec;
6484 gettimeofday(&tv_now, 0);
6486 if (tv_delta.tv_usec < tv_now.tv_usec) {
6488 tv_delta.tv_usec += 1000000;
6491 tv_delta.tv_usec -= tv_now.tv_usec;
6493 if (tv_delta.tv_sec < tv_now.tv_sec) {
6497 tv_delta.tv_sec -= tv_now.tv_sec;
6499 code = select(socket + 1, &imask, 0, 0, &tv_delta);
6500 if (code == 1 && FD_ISSET(socket, &imask)) {
6501 /* now receive a packet */
6502 faddrLen = sizeof(struct sockaddr_in);
6504 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6505 (struct sockaddr *)&faddr, &faddrLen);
6508 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6509 if (counter == ntohl(theader.callNumber))
6517 /* see if we've timed out */
6525 code -= sizeof(struct rx_header);
6526 if (code > outputLength)
6527 code = outputLength;
6528 memcpy(outputData, tp, code);
6533 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6534 afs_uint16 remotePort, struct rx_debugStats * stat,
6535 afs_uint32 * supportedValues)
6537 struct rx_debugIn in;
6540 *supportedValues = 0;
6541 in.type = htonl(RX_DEBUGI_GETSTATS);
6544 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6545 &in, sizeof(in), stat, sizeof(*stat));
6548 * If the call was successful, fixup the version and indicate
6549 * what contents of the stat structure are valid.
6550 * Also do net to host conversion of fields here.
6554 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6555 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6557 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6558 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6560 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6561 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6563 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6564 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6566 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6567 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6569 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6570 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6572 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6573 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6575 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6576 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6579 stat->nFreePackets = ntohl(stat->nFreePackets);
6580 stat->packetReclaims = ntohl(stat->packetReclaims);
6581 stat->callsExecuted = ntohl(stat->callsExecuted);
6582 stat->nWaiting = ntohl(stat->nWaiting);
6583 stat->idleThreads = ntohl(stat->idleThreads);
6590 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6591 afs_uint16 remotePort, struct rx_stats * stat,
6592 afs_uint32 * supportedValues)
6594 struct rx_debugIn in;
6595 afs_int32 *lp = (afs_int32 *) stat;
6600 * supportedValues is currently unused, but added to allow future
6601 * versioning of this function.
6604 *supportedValues = 0;
6605 in.type = htonl(RX_DEBUGI_RXSTATS);
6607 memset(stat, 0, sizeof(*stat));
6609 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6610 &in, sizeof(in), stat, sizeof(*stat));
6615 * Do net to host conversion here
6618 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6627 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6628 afs_uint16 remotePort, size_t version_length,
6632 return MakeDebugCall(socket, remoteAddr, remotePort,
6633 RX_PACKET_TYPE_VERSION, a, 1, version,
6638 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6639 afs_uint16 remotePort, afs_int32 * nextConnection,
6640 int allConnections, afs_uint32 debugSupportedValues,
6641 struct rx_debugConn * conn,
6642 afs_uint32 * supportedValues)
6644 struct rx_debugIn in;
6649 * supportedValues is currently unused, but added to allow future
6650 * versioning of this function.
6653 *supportedValues = 0;
6654 if (allConnections) {
6655 in.type = htonl(RX_DEBUGI_GETALLCONN);
6657 in.type = htonl(RX_DEBUGI_GETCONN);
6659 in.index = htonl(*nextConnection);
6660 memset(conn, 0, sizeof(*conn));
6662 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6663 &in, sizeof(in), conn, sizeof(*conn));
6666 *nextConnection += 1;
6669 * Convert old connection format to new structure.
6672 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6673 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6674 #define MOVEvL(a) (conn->a = vL->a)
6676 /* any old or unrecognized version... */
6677 for (i = 0; i < RX_MAXCALLS; i++) {
6678 MOVEvL(callState[i]);
6679 MOVEvL(callMode[i]);
6680 MOVEvL(callFlags[i]);
6681 MOVEvL(callOther[i]);
6683 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6684 MOVEvL(secStats.type);
6685 MOVEvL(secStats.level);
6686 MOVEvL(secStats.flags);
6687 MOVEvL(secStats.expires);
6688 MOVEvL(secStats.packetsReceived);
6689 MOVEvL(secStats.packetsSent);
6690 MOVEvL(secStats.bytesReceived);
6691 MOVEvL(secStats.bytesSent);
6696 * Do net to host conversion here
6698 * I don't convert host or port since we are most likely
6699 * going to want these in NBO.
6701 conn->cid = ntohl(conn->cid);
6702 conn->serial = ntohl(conn->serial);
6703 for (i = 0; i < RX_MAXCALLS; i++) {
6704 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6706 conn->error = ntohl(conn->error);
6707 conn->secStats.flags = ntohl(conn->secStats.flags);
6708 conn->secStats.expires = ntohl(conn->secStats.expires);
6709 conn->secStats.packetsReceived =
6710 ntohl(conn->secStats.packetsReceived);
6711 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6712 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6713 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6714 conn->epoch = ntohl(conn->epoch);
6715 conn->natMTU = ntohl(conn->natMTU);
6722 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6723 afs_uint16 remotePort, afs_int32 * nextPeer,
6724 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6725 afs_uint32 * supportedValues)
6727 struct rx_debugIn in;
6731 * supportedValues is currently unused, but added to allow future
6732 * versioning of this function.
6735 *supportedValues = 0;
6736 in.type = htonl(RX_DEBUGI_GETPEER);
6737 in.index = htonl(*nextPeer);
6738 memset(peer, 0, sizeof(*peer));
6740 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6741 &in, sizeof(in), peer, sizeof(*peer));
6747 * Do net to host conversion here
6749 * I don't convert host or port since we are most likely
6750 * going to want these in NBO.
6752 peer->ifMTU = ntohs(peer->ifMTU);
6753 peer->idleWhen = ntohl(peer->idleWhen);
6754 peer->refCount = ntohs(peer->refCount);
6755 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6756 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6757 peer->rtt = ntohl(peer->rtt);
6758 peer->rtt_dev = ntohl(peer->rtt_dev);
6759 peer->timeout.sec = ntohl(peer->timeout.sec);
6760 peer->timeout.usec = ntohl(peer->timeout.usec);
6761 peer->nSent = ntohl(peer->nSent);
6762 peer->reSends = ntohl(peer->reSends);
6763 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6764 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6765 peer->rateFlag = ntohl(peer->rateFlag);
6766 peer->natMTU = ntohs(peer->natMTU);
6767 peer->maxMTU = ntohs(peer->maxMTU);
6768 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6769 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6770 peer->MTU = ntohs(peer->MTU);
6771 peer->cwind = ntohs(peer->cwind);
6772 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6773 peer->congestSeq = ntohs(peer->congestSeq);
6774 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6775 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6776 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6777 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6782 #endif /* RXDEBUG */
6787 struct rx_serverQueueEntry *np;
6790 register struct rx_call *call;
6791 register struct rx_serverQueueEntry *sq;
6795 if (rxinit_status == 1) {
6797 return; /* Already shutdown. */
6801 #ifndef AFS_PTHREAD_ENV
6802 FD_ZERO(&rx_selectMask);
6803 #endif /* AFS_PTHREAD_ENV */
6804 rxi_dataQuota = RX_MAX_QUOTA;
6805 #ifndef AFS_PTHREAD_ENV
6807 #endif /* AFS_PTHREAD_ENV */
6810 #ifndef AFS_PTHREAD_ENV
6811 #ifndef AFS_USE_GETTIMEOFDAY
6813 #endif /* AFS_USE_GETTIMEOFDAY */
6814 #endif /* AFS_PTHREAD_ENV */
6816 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6817 call = queue_First(&rx_freeCallQueue, rx_call);
6819 rxi_Free(call, sizeof(struct rx_call));
6822 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6823 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6829 struct rx_peer **peer_ptr, **peer_end;
6830 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6831 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6833 struct rx_peer *peer, *next;
6834 for (peer = *peer_ptr; peer; peer = next) {
6835 rx_interface_stat_p rpc_stat, nrpc_stat;
6838 (&peer->rpcStats, rpc_stat, nrpc_stat,
6839 rx_interface_stat)) {
6840 unsigned int num_funcs;
6843 queue_Remove(&rpc_stat->queue_header);
6844 queue_Remove(&rpc_stat->all_peers);
6845 num_funcs = rpc_stat->stats[0].func_total;
6847 sizeof(rx_interface_stat_t) +
6848 rpc_stat->stats[0].func_total *
6849 sizeof(rx_function_entry_v1_t);
6851 rxi_Free(rpc_stat, space);
6852 MUTEX_ENTER(&rx_rpc_stats);
6853 rxi_rpc_peer_stat_cnt -= num_funcs;
6854 MUTEX_EXIT(&rx_rpc_stats);
6858 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6862 for (i = 0; i < RX_MAX_SERVICES; i++) {
6864 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6866 for (i = 0; i < rx_hashTableSize; i++) {
6867 register struct rx_connection *tc, *ntc;
6868 MUTEX_ENTER(&rx_connHashTable_lock);
6869 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6871 for (j = 0; j < RX_MAXCALLS; j++) {
6873 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6876 rxi_Free(tc, sizeof(*tc));
6878 MUTEX_EXIT(&rx_connHashTable_lock);
6881 MUTEX_ENTER(&freeSQEList_lock);
6883 while ((np = rx_FreeSQEList)) {
6884 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6885 MUTEX_DESTROY(&np->lock);
6886 rxi_Free(np, sizeof(*np));
6889 MUTEX_EXIT(&freeSQEList_lock);
6890 MUTEX_DESTROY(&freeSQEList_lock);
6891 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6892 MUTEX_DESTROY(&rx_connHashTable_lock);
6893 MUTEX_DESTROY(&rx_peerHashTable_lock);
6894 MUTEX_DESTROY(&rx_serverPool_lock);
6896 osi_Free(rx_connHashTable,
6897 rx_hashTableSize * sizeof(struct rx_connection *));
6898 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6900 UNPIN(rx_connHashTable,
6901 rx_hashTableSize * sizeof(struct rx_connection *));
6902 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6904 rxi_FreeAllPackets();
6906 MUTEX_ENTER(&rx_stats_mutex);
6907 rxi_dataQuota = RX_MAX_QUOTA;
6908 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6909 MUTEX_EXIT(&rx_stats_mutex);
6915 #ifdef RX_ENABLE_LOCKS
6917 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6919 if (!MUTEX_ISMINE(lockaddr))
6920 osi_Panic("Lock not held: %s", msg);
6922 #endif /* RX_ENABLE_LOCKS */
6927 * Routines to implement connection specific data.
6931 rx_KeyCreate(rx_destructor_t rtn)
6934 MUTEX_ENTER(&rxi_keyCreate_lock);
6935 key = rxi_keyCreate_counter++;
6936 rxi_keyCreate_destructor = (rx_destructor_t *)
6937 realloc((void *)rxi_keyCreate_destructor,
6938 (key + 1) * sizeof(rx_destructor_t));
6939 rxi_keyCreate_destructor[key] = rtn;
6940 MUTEX_EXIT(&rxi_keyCreate_lock);
6945 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6948 MUTEX_ENTER(&conn->conn_data_lock);
6949 if (!conn->specific) {
6950 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6951 for (i = 0; i < key; i++)
6952 conn->specific[i] = NULL;
6953 conn->nSpecific = key + 1;
6954 conn->specific[key] = ptr;
6955 } else if (key >= conn->nSpecific) {
6956 conn->specific = (void **)
6957 realloc(conn->specific, (key + 1) * sizeof(void *));
6958 for (i = conn->nSpecific; i < key; i++)
6959 conn->specific[i] = NULL;
6960 conn->nSpecific = key + 1;
6961 conn->specific[key] = ptr;
6963 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6964 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6965 conn->specific[key] = ptr;
6967 MUTEX_EXIT(&conn->conn_data_lock);
6971 rx_GetSpecific(struct rx_connection *conn, int key)
6974 MUTEX_ENTER(&conn->conn_data_lock);
6975 if (key >= conn->nSpecific)
6978 ptr = conn->specific[key];
6979 MUTEX_EXIT(&conn->conn_data_lock);
6983 #endif /* !KERNEL */
6986 * processStats is a queue used to store the statistics for the local
6987 * process. Its contents are similar to the contents of the rpcStats
6988 * queue on a rx_peer structure, but the actual data stored within
6989 * this queue contains totals across the lifetime of the process (assuming
6990 * the stats have not been reset) - unlike the per peer structures
6991 * which can come and go based upon the peer lifetime.
6994 static struct rx_queue processStats = { &processStats, &processStats };
6997 * peerStats is a queue used to store the statistics for all peer structs.
6998 * Its contents are the union of all the peer rpcStats queues.
7001 static struct rx_queue peerStats = { &peerStats, &peerStats };
7004 * rxi_monitor_processStats is used to turn process wide stat collection
7008 static int rxi_monitor_processStats = 0;
7011 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
7014 static int rxi_monitor_peerStats = 0;
7017 * rxi_AddRpcStat - given all of the information for a particular rpc
7018 * call, create (if needed) and update the stat totals for the rpc.
7022 * IN stats - the queue of stats that will be updated with the new value
7024 * IN rxInterface - a unique number that identifies the rpc interface
7026 * IN currentFunc - the index of the function being invoked
7028 * IN totalFunc - the total number of functions in this interface
7030 * IN queueTime - the amount of time this function waited for a thread
7032 * IN execTime - the amount of time this function invocation took to execute
7034 * IN bytesSent - the number bytes sent by this invocation
7036 * IN bytesRcvd - the number bytes received by this invocation
7038 * IN isServer - if true, this invocation was made to a server
7040 * IN remoteHost - the ip address of the remote host
7042 * IN remotePort - the port of the remote host
7044 * IN addToPeerList - if != 0, add newly created stat to the global peer list
7046 * INOUT counter - if a new stats structure is allocated, the counter will
7047 * be updated with the new number of allocated stat structures
7055 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
7056 afs_uint32 currentFunc, afs_uint32 totalFunc,
7057 struct clock *queueTime, struct clock *execTime,
7058 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
7059 afs_uint32 remoteHost, afs_uint32 remotePort,
7060 int addToPeerList, unsigned int *counter)
7063 rx_interface_stat_p rpc_stat, nrpc_stat;
7066 * See if there's already a structure for this interface
7069 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7070 if ((rpc_stat->stats[0].interfaceId == rxInterface)
7071 && (rpc_stat->stats[0].remote_is_server == isServer))
7076 * Didn't find a match so allocate a new structure and add it to the
7080 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
7081 || (rpc_stat->stats[0].interfaceId != rxInterface)
7082 || (rpc_stat->stats[0].remote_is_server != isServer)) {
7087 sizeof(rx_interface_stat_t) +
7088 totalFunc * sizeof(rx_function_entry_v1_t);
7090 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
7091 if (rpc_stat == NULL) {
7095 *counter += totalFunc;
7096 for (i = 0; i < totalFunc; i++) {
7097 rpc_stat->stats[i].remote_peer = remoteHost;
7098 rpc_stat->stats[i].remote_port = remotePort;
7099 rpc_stat->stats[i].remote_is_server = isServer;
7100 rpc_stat->stats[i].interfaceId = rxInterface;
7101 rpc_stat->stats[i].func_total = totalFunc;
7102 rpc_stat->stats[i].func_index = i;
7103 hzero(rpc_stat->stats[i].invocations);
7104 hzero(rpc_stat->stats[i].bytes_sent);
7105 hzero(rpc_stat->stats[i].bytes_rcvd);
7106 rpc_stat->stats[i].queue_time_sum.sec = 0;
7107 rpc_stat->stats[i].queue_time_sum.usec = 0;
7108 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7109 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7110 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7111 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7112 rpc_stat->stats[i].queue_time_max.sec = 0;
7113 rpc_stat->stats[i].queue_time_max.usec = 0;
7114 rpc_stat->stats[i].execution_time_sum.sec = 0;
7115 rpc_stat->stats[i].execution_time_sum.usec = 0;
7116 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7117 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7118 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7119 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7120 rpc_stat->stats[i].execution_time_max.sec = 0;
7121 rpc_stat->stats[i].execution_time_max.usec = 0;
7123 queue_Prepend(stats, rpc_stat);
7124 if (addToPeerList) {
7125 queue_Prepend(&peerStats, &rpc_stat->all_peers);
7130 * Increment the stats for this function
7133 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
7134 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
7135 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
7136 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
7137 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7138 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7139 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7141 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7142 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7144 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7145 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7147 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7148 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7150 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7151 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7159 * rx_IncrementTimeAndCount - increment the times and count for a particular
7164 * IN peer - the peer who invoked the rpc
7166 * IN rxInterface - a unique number that identifies the rpc interface
7168 * IN currentFunc - the index of the function being invoked
7170 * IN totalFunc - the total number of functions in this interface
7172 * IN queueTime - the amount of time this function waited for a thread
7174 * IN execTime - the amount of time this function invocation took to execute
7176 * IN bytesSent - the number bytes sent by this invocation
7178 * IN bytesRcvd - the number bytes received by this invocation
7180 * IN isServer - if true, this invocation was made to a server
7188 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7189 afs_uint32 currentFunc, afs_uint32 totalFunc,
7190 struct clock *queueTime, struct clock *execTime,
7191 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7195 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
7198 MUTEX_ENTER(&rx_rpc_stats);
7199 MUTEX_ENTER(&peer->peer_lock);
7201 if (rxi_monitor_peerStats) {
7202 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7203 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7204 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7207 if (rxi_monitor_processStats) {
7208 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7209 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7210 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7213 MUTEX_EXIT(&peer->peer_lock);
7214 MUTEX_EXIT(&rx_rpc_stats);
7219 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7223 * IN callerVersion - the rpc stat version of the caller.
7225 * IN count - the number of entries to marshall.
7227 * IN stats - pointer to stats to be marshalled.
7229 * OUT ptr - Where to store the marshalled data.
7236 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7237 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7243 * We only support the first version
7245 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7246 *(ptr++) = stats->remote_peer;
7247 *(ptr++) = stats->remote_port;
7248 *(ptr++) = stats->remote_is_server;
7249 *(ptr++) = stats->interfaceId;
7250 *(ptr++) = stats->func_total;
7251 *(ptr++) = stats->func_index;
7252 *(ptr++) = hgethi(stats->invocations);
7253 *(ptr++) = hgetlo(stats->invocations);
7254 *(ptr++) = hgethi(stats->bytes_sent);
7255 *(ptr++) = hgetlo(stats->bytes_sent);
7256 *(ptr++) = hgethi(stats->bytes_rcvd);
7257 *(ptr++) = hgetlo(stats->bytes_rcvd);
7258 *(ptr++) = stats->queue_time_sum.sec;
7259 *(ptr++) = stats->queue_time_sum.usec;
7260 *(ptr++) = stats->queue_time_sum_sqr.sec;
7261 *(ptr++) = stats->queue_time_sum_sqr.usec;
7262 *(ptr++) = stats->queue_time_min.sec;
7263 *(ptr++) = stats->queue_time_min.usec;
7264 *(ptr++) = stats->queue_time_max.sec;
7265 *(ptr++) = stats->queue_time_max.usec;
7266 *(ptr++) = stats->execution_time_sum.sec;
7267 *(ptr++) = stats->execution_time_sum.usec;
7268 *(ptr++) = stats->execution_time_sum_sqr.sec;
7269 *(ptr++) = stats->execution_time_sum_sqr.usec;
7270 *(ptr++) = stats->execution_time_min.sec;
7271 *(ptr++) = stats->execution_time_min.usec;
7272 *(ptr++) = stats->execution_time_max.sec;
7273 *(ptr++) = stats->execution_time_max.usec;
7279 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7284 * IN callerVersion - the rpc stat version of the caller
7286 * OUT myVersion - the rpc stat version of this function
7288 * OUT clock_sec - local time seconds
7290 * OUT clock_usec - local time microseconds
7292 * OUT allocSize - the number of bytes allocated to contain stats
7294 * OUT statCount - the number stats retrieved from this process.
7296 * OUT stats - the actual stats retrieved from this process.
7300 * Returns void. If successful, stats will != NULL.
7304 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7305 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7306 size_t * allocSize, afs_uint32 * statCount,
7307 afs_uint32 ** stats)
7317 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7320 * Check to see if stats are enabled
7323 MUTEX_ENTER(&rx_rpc_stats);
7324 if (!rxi_monitor_processStats) {
7325 MUTEX_EXIT(&rx_rpc_stats);
7329 clock_GetTime(&now);
7330 *clock_sec = now.sec;
7331 *clock_usec = now.usec;
7334 * Allocate the space based upon the caller version
7336 * If the client is at an older version than we are,
7337 * we return the statistic data in the older data format, but
7338 * we still return our version number so the client knows we
7339 * are maintaining more data than it can retrieve.
7342 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7343 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7344 *statCount = rxi_rpc_process_stat_cnt;
7347 * This can't happen yet, but in the future version changes
7348 * can be handled by adding additional code here
7352 if (space > (size_t) 0) {
7354 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7357 rx_interface_stat_p rpc_stat, nrpc_stat;
7361 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7363 * Copy the data based upon the caller version
7365 rx_MarshallProcessRPCStats(callerVersion,
7366 rpc_stat->stats[0].func_total,
7367 rpc_stat->stats, &ptr);
7373 MUTEX_EXIT(&rx_rpc_stats);
7378 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7382 * IN callerVersion - the rpc stat version of the caller
7384 * OUT myVersion - the rpc stat version of this function
7386 * OUT clock_sec - local time seconds
7388 * OUT clock_usec - local time microseconds
7390 * OUT allocSize - the number of bytes allocated to contain stats
7392 * OUT statCount - the number of stats retrieved from the individual
7395 * OUT stats - the actual stats retrieved from the individual peer structures.
7399 * Returns void. If successful, stats will != NULL.
7403 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7404 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7405 size_t * allocSize, afs_uint32 * statCount,
7406 afs_uint32 ** stats)
7416 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7419 * Check to see if stats are enabled
7422 MUTEX_ENTER(&rx_rpc_stats);
7423 if (!rxi_monitor_peerStats) {
7424 MUTEX_EXIT(&rx_rpc_stats);
7428 clock_GetTime(&now);
7429 *clock_sec = now.sec;
7430 *clock_usec = now.usec;
7433 * Allocate the space based upon the caller version
7435 * If the client is at an older version than we are,
7436 * we return the statistic data in the older data format, but
7437 * we still return our version number so the client knows we
7438 * are maintaining more data than it can retrieve.
7441 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7442 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7443 *statCount = rxi_rpc_peer_stat_cnt;
7446 * This can't happen yet, but in the future version changes
7447 * can be handled by adding additional code here
7451 if (space > (size_t) 0) {
7453 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7456 rx_interface_stat_p rpc_stat, nrpc_stat;
7460 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7462 * We have to fix the offset of rpc_stat since we are
7463 * keeping this structure on two rx_queues. The rx_queue
7464 * package assumes that the rx_queue member is the first
7465 * member of the structure. That is, rx_queue assumes that
7466 * any one item is only on one queue at a time. We are
7467 * breaking that assumption and so we have to do a little
7468 * math to fix our pointers.
7471 fix_offset = (char *)rpc_stat;
7472 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7473 rpc_stat = (rx_interface_stat_p) fix_offset;
7476 * Copy the data based upon the caller version
7478 rx_MarshallProcessRPCStats(callerVersion,
7479 rpc_stat->stats[0].func_total,
7480 rpc_stat->stats, &ptr);
7486 MUTEX_EXIT(&rx_rpc_stats);
7491 * rx_FreeRPCStats - free memory allocated by
7492 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7496 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7497 * rx_RetrievePeerRPCStats
7499 * IN allocSize - the number of bytes in stats.
7507 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7509 rxi_Free(stats, allocSize);
7513 * rx_queryProcessRPCStats - see if process rpc stat collection is
7514 * currently enabled.
7520 * Returns 0 if stats are not enabled != 0 otherwise
7524 rx_queryProcessRPCStats(void)
7527 MUTEX_ENTER(&rx_rpc_stats);
7528 rc = rxi_monitor_processStats;
7529 MUTEX_EXIT(&rx_rpc_stats);
7534 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7540 * Returns 0 if stats are not enabled != 0 otherwise
7544 rx_queryPeerRPCStats(void)
7547 MUTEX_ENTER(&rx_rpc_stats);
7548 rc = rxi_monitor_peerStats;
7549 MUTEX_EXIT(&rx_rpc_stats);
7554 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7564 rx_enableProcessRPCStats(void)
7566 MUTEX_ENTER(&rx_rpc_stats);
7567 rx_enable_stats = 1;
7568 rxi_monitor_processStats = 1;
7569 MUTEX_EXIT(&rx_rpc_stats);
7573 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7583 rx_enablePeerRPCStats(void)
7585 MUTEX_ENTER(&rx_rpc_stats);
7586 rx_enable_stats = 1;
7587 rxi_monitor_peerStats = 1;
7588 MUTEX_EXIT(&rx_rpc_stats);
7592 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7602 rx_disableProcessRPCStats(void)
7604 rx_interface_stat_p rpc_stat, nrpc_stat;
7607 MUTEX_ENTER(&rx_rpc_stats);
7610 * Turn off process statistics and if peer stats is also off, turn
7614 rxi_monitor_processStats = 0;
7615 if (rxi_monitor_peerStats == 0) {
7616 rx_enable_stats = 0;
7619 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7620 unsigned int num_funcs = 0;
7623 queue_Remove(rpc_stat);
7624 num_funcs = rpc_stat->stats[0].func_total;
7626 sizeof(rx_interface_stat_t) +
7627 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7629 rxi_Free(rpc_stat, space);
7630 rxi_rpc_process_stat_cnt -= num_funcs;
7632 MUTEX_EXIT(&rx_rpc_stats);
7636 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7646 rx_disablePeerRPCStats(void)
7648 struct rx_peer **peer_ptr, **peer_end;
7651 MUTEX_ENTER(&rx_rpc_stats);
7654 * Turn off peer statistics and if process stats is also off, turn
7658 rxi_monitor_peerStats = 0;
7659 if (rxi_monitor_processStats == 0) {
7660 rx_enable_stats = 0;
7663 MUTEX_ENTER(&rx_peerHashTable_lock);
7664 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7665 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7667 struct rx_peer *peer, *next, *prev;
7668 for (prev = peer = *peer_ptr; peer; peer = next) {
7670 code = MUTEX_TRYENTER(&peer->peer_lock);
7672 rx_interface_stat_p rpc_stat, nrpc_stat;
7675 (&peer->rpcStats, rpc_stat, nrpc_stat,
7676 rx_interface_stat)) {
7677 unsigned int num_funcs = 0;
7680 queue_Remove(&rpc_stat->queue_header);
7681 queue_Remove(&rpc_stat->all_peers);
7682 num_funcs = rpc_stat->stats[0].func_total;
7684 sizeof(rx_interface_stat_t) +
7685 rpc_stat->stats[0].func_total *
7686 sizeof(rx_function_entry_v1_t);
7688 rxi_Free(rpc_stat, space);
7689 rxi_rpc_peer_stat_cnt -= num_funcs;
7691 MUTEX_EXIT(&peer->peer_lock);
7692 if (prev == *peer_ptr) {
7702 MUTEX_EXIT(&rx_peerHashTable_lock);
7703 MUTEX_EXIT(&rx_rpc_stats);
7707 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7712 * IN clearFlag - flag indicating which stats to clear
7720 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7722 rx_interface_stat_p rpc_stat, nrpc_stat;
7724 MUTEX_ENTER(&rx_rpc_stats);
7726 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7727 unsigned int num_funcs = 0, i;
7728 num_funcs = rpc_stat->stats[0].func_total;
7729 for (i = 0; i < num_funcs; i++) {
7730 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7731 hzero(rpc_stat->stats[i].invocations);
7733 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7734 hzero(rpc_stat->stats[i].bytes_sent);
7736 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7737 hzero(rpc_stat->stats[i].bytes_rcvd);
7739 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7740 rpc_stat->stats[i].queue_time_sum.sec = 0;
7741 rpc_stat->stats[i].queue_time_sum.usec = 0;
7743 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7744 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7745 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7747 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7748 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7749 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7751 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7752 rpc_stat->stats[i].queue_time_max.sec = 0;
7753 rpc_stat->stats[i].queue_time_max.usec = 0;
7755 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7756 rpc_stat->stats[i].execution_time_sum.sec = 0;
7757 rpc_stat->stats[i].execution_time_sum.usec = 0;
7759 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7760 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7761 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7763 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7764 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7765 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7767 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7768 rpc_stat->stats[i].execution_time_max.sec = 0;
7769 rpc_stat->stats[i].execution_time_max.usec = 0;
7774 MUTEX_EXIT(&rx_rpc_stats);
7778 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7783 * IN clearFlag - flag indicating which stats to clear
7791 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7793 rx_interface_stat_p rpc_stat, nrpc_stat;
7795 MUTEX_ENTER(&rx_rpc_stats);
7797 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7798 unsigned int num_funcs = 0, i;
7801 * We have to fix the offset of rpc_stat since we are
7802 * keeping this structure on two rx_queues. The rx_queue
7803 * package assumes that the rx_queue member is the first
7804 * member of the structure. That is, rx_queue assumes that
7805 * any one item is only on one queue at a time. We are
7806 * breaking that assumption and so we have to do a little
7807 * math to fix our pointers.
7810 fix_offset = (char *)rpc_stat;
7811 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7812 rpc_stat = (rx_interface_stat_p) fix_offset;
7814 num_funcs = rpc_stat->stats[0].func_total;
7815 for (i = 0; i < num_funcs; i++) {
7816 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7817 hzero(rpc_stat->stats[i].invocations);
7819 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7820 hzero(rpc_stat->stats[i].bytes_sent);
7822 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7823 hzero(rpc_stat->stats[i].bytes_rcvd);
7825 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7826 rpc_stat->stats[i].queue_time_sum.sec = 0;
7827 rpc_stat->stats[i].queue_time_sum.usec = 0;
7829 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7830 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7831 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7833 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7834 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7835 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7837 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7838 rpc_stat->stats[i].queue_time_max.sec = 0;
7839 rpc_stat->stats[i].queue_time_max.usec = 0;
7841 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7842 rpc_stat->stats[i].execution_time_sum.sec = 0;
7843 rpc_stat->stats[i].execution_time_sum.usec = 0;
7845 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7846 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7847 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7849 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7850 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7851 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7853 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7854 rpc_stat->stats[i].execution_time_max.sec = 0;
7855 rpc_stat->stats[i].execution_time_max.usec = 0;
7860 MUTEX_EXIT(&rx_rpc_stats);
7864 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7865 * is authorized to enable/disable/clear RX statistics.
7867 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7870 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7872 rxi_rxstat_userok = proc;
7876 rx_RxStatUserOk(struct rx_call *call)
7878 if (!rxi_rxstat_userok)
7880 return rxi_rxstat_userok(call);
7885 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
7886 * function in the MSVC runtime DLL (msvcrt.dll).
7888 * Note: the system serializes calls to this function.
7891 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
7892 DWORD reason, /* reason function is being called */
7893 LPVOID reserved) /* reserved for future use */
7896 case DLL_PROCESS_ATTACH:
7897 /* library is being attached to a process */
7901 case DLL_PROCESS_DETACH: