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>
93 # include <afs/afsutil.h>
94 # include <WINNT\afsreg.h>
96 # include <sys/socket.h>
97 # include <sys/file.h>
99 # include <sys/stat.h>
100 # include <netinet/in.h>
101 # include <sys/time.h>
104 # include "rx_user.h"
105 # include "rx_clock.h"
106 # include "rx_queue.h"
107 # include "rx_globals.h"
108 # include "rx_trace.h"
109 # include <afs/rxgen_consts.h>
113 #ifdef AFS_PTHREAD_ENV
115 int (*registerProgram) (pid_t, char *) = 0;
116 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
119 int (*registerProgram) (PROCESS, char *) = 0;
120 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
124 /* Local static routines */
125 static void rxi_DestroyConnectionNoLock(register struct rx_connection *conn);
126 #ifdef RX_ENABLE_LOCKS
127 static void rxi_SetAcksInTransmitQueue(register struct rx_call *call);
130 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
132 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
133 afs_int32 rxi_start_in_error;
135 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
138 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
139 * currently allocated within rx. This number is used to allocate the
140 * memory required to return the statistics when queried.
143 static unsigned int rxi_rpc_peer_stat_cnt;
146 * rxi_rpc_process_stat_cnt counts the total number of local process stat
147 * structures currently allocated within rx. The number is used to allocate
148 * the memory required to return the statistics when queried.
151 static unsigned int rxi_rpc_process_stat_cnt;
153 #if !defined(offsetof)
154 #include <stddef.h> /* for definition of offsetof() */
157 #ifdef AFS_PTHREAD_ENV
161 * Use procedural initialization of mutexes/condition variables
165 extern pthread_mutex_t rx_stats_mutex;
166 extern pthread_mutex_t des_init_mutex;
167 extern pthread_mutex_t des_random_mutex;
168 extern pthread_mutex_t rx_clock_mutex;
169 extern pthread_mutex_t rxi_connCacheMutex;
170 extern pthread_mutex_t rx_event_mutex;
171 extern pthread_mutex_t osi_malloc_mutex;
172 extern pthread_mutex_t event_handler_mutex;
173 extern pthread_mutex_t listener_mutex;
174 extern pthread_mutex_t rx_if_init_mutex;
175 extern pthread_mutex_t rx_if_mutex;
176 extern pthread_mutex_t rxkad_client_uid_mutex;
177 extern pthread_mutex_t rxkad_random_mutex;
179 extern pthread_cond_t rx_event_handler_cond;
180 extern pthread_cond_t rx_listener_cond;
182 static pthread_mutex_t epoch_mutex;
183 static pthread_mutex_t rx_init_mutex;
184 static pthread_mutex_t rx_debug_mutex;
185 static pthread_mutex_t rx_rpc_stats;
188 rxi_InitPthread(void)
190 assert(pthread_mutex_init(&rx_clock_mutex, (const pthread_mutexattr_t *)0)
192 assert(pthread_mutex_init(&rx_stats_mutex, (const pthread_mutexattr_t *)0)
194 assert(pthread_mutex_init
195 (&rxi_connCacheMutex, (const pthread_mutexattr_t *)0) == 0);
196 assert(pthread_mutex_init(&rx_init_mutex, (const pthread_mutexattr_t *)0)
198 assert(pthread_mutex_init(&epoch_mutex, (const pthread_mutexattr_t *)0) ==
200 assert(pthread_mutex_init(&rx_event_mutex, (const pthread_mutexattr_t *)0)
202 assert(pthread_mutex_init(&des_init_mutex, (const pthread_mutexattr_t *)0)
204 assert(pthread_mutex_init
205 (&des_random_mutex, (const pthread_mutexattr_t *)0) == 0);
206 assert(pthread_mutex_init
207 (&osi_malloc_mutex, (const pthread_mutexattr_t *)0) == 0);
208 assert(pthread_mutex_init
209 (&event_handler_mutex, (const pthread_mutexattr_t *)0) == 0);
210 assert(pthread_mutex_init(&listener_mutex, (const pthread_mutexattr_t *)0)
212 assert(pthread_mutex_init
213 (&rx_if_init_mutex, (const pthread_mutexattr_t *)0) == 0);
214 assert(pthread_mutex_init(&rx_if_mutex, (const pthread_mutexattr_t *)0) ==
216 assert(pthread_mutex_init
217 (&rxkad_client_uid_mutex, (const pthread_mutexattr_t *)0) == 0);
218 assert(pthread_mutex_init
219 (&rxkad_random_mutex, (const pthread_mutexattr_t *)0) == 0);
220 assert(pthread_mutex_init(&rx_debug_mutex, (const pthread_mutexattr_t *)0)
223 assert(pthread_cond_init
224 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
225 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
227 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
228 assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
230 rxkad_global_stats_init();
232 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
233 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
234 #ifdef RX_ENABLE_LOCKS
237 #endif /* RX_LOCKS_DB */
238 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
239 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
241 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
243 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
245 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
247 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
248 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
249 #endif /* RX_ENABLE_LOCKS */
252 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
253 #define INIT_PTHREAD_LOCKS \
254 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
256 * The rx_stats_mutex mutex protects the following global variables:
261 * rxi_lowConnRefCount
262 * rxi_lowPeerRefCount
271 #define INIT_PTHREAD_LOCKS
275 /* Variables for handling the minProcs implementation. availProcs gives the
276 * number of threads available in the pool at this moment (not counting dudes
277 * executing right now). totalMin gives the total number of procs required
278 * for handling all minProcs requests. minDeficit is a dynamic variable
279 * tracking the # of procs required to satisfy all of the remaining minProcs
281 * For fine grain locking to work, the quota check and the reservation of
282 * a server thread has to come while rxi_availProcs and rxi_minDeficit
283 * are locked. To this end, the code has been modified under #ifdef
284 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
285 * same time. A new function, ReturnToServerPool() returns the allocation.
287 * A call can be on several queue's (but only one at a time). When
288 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
289 * that no one else is touching the queue. To this end, we store the address
290 * of the queue lock in the call structure (under the call lock) when we
291 * put the call on a queue, and we clear the call_queue_lock when the
292 * call is removed from a queue (once the call lock has been obtained).
293 * This allows rxi_ResetCall to safely synchronize with others wishing
294 * to manipulate the queue.
297 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
298 static afs_kmutex_t rx_rpc_stats;
299 void rxi_StartUnlocked(struct rxevent *event, void *call,
300 void *arg1, int istack);
303 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
304 ** pretty good that the next packet coming in is from the same connection
305 ** as the last packet, since we're send multiple packets in a transmit window.
307 struct rx_connection *rxLastConn = 0;
309 #ifdef RX_ENABLE_LOCKS
310 /* The locking hierarchy for rx fine grain locking is composed of these
313 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
314 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
315 * call->lock - locks call data fields.
316 * These are independent of each other:
317 * rx_freeCallQueue_lock
322 * serverQueueEntry->lock
324 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
325 * peer->lock - locks peer data fields.
326 * conn_data_lock - that more than one thread is not updating a conn data
327 * field at the same time.
335 * Do we need a lock to protect the peer field in the conn structure?
336 * conn->peer was previously a constant for all intents and so has no
337 * lock protecting this field. The multihomed client delta introduced
338 * a RX code change : change the peer field in the connection structure
339 * to that remote inetrface from which the last packet for this
340 * connection was sent out. This may become an issue if further changes
343 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
344 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
346 /* rxdb_fileID is used to identify the lock location, along with line#. */
347 static int rxdb_fileID = RXDB_FILE_RX;
348 #endif /* RX_LOCKS_DB */
349 #else /* RX_ENABLE_LOCKS */
350 #define SET_CALL_QUEUE_LOCK(C, L)
351 #define CLEAR_CALL_QUEUE_LOCK(C)
352 #endif /* RX_ENABLE_LOCKS */
353 struct rx_serverQueueEntry *rx_waitForPacket = 0;
354 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
356 /* ------------Exported Interfaces------------- */
358 /* This function allows rxkad to set the epoch to a suitably random number
359 * which rx_NewConnection will use in the future. The principle purpose is to
360 * get rxnull connections to use the same epoch as the rxkad connections do, at
361 * least once the first rxkad connection is established. This is important now
362 * that the host/port addresses aren't used in FindConnection: the uniqueness
363 * of epoch/cid matters and the start time won't do. */
365 #ifdef AFS_PTHREAD_ENV
367 * This mutex protects the following global variables:
371 #define LOCK_EPOCH assert(pthread_mutex_lock(&epoch_mutex)==0)
372 #define UNLOCK_EPOCH assert(pthread_mutex_unlock(&epoch_mutex)==0)
376 #endif /* AFS_PTHREAD_ENV */
379 rx_SetEpoch(afs_uint32 epoch)
386 /* Initialize rx. A port number may be mentioned, in which case this
387 * becomes the default port number for any service installed later.
388 * If 0 is provided for the port number, a random port will be chosen
389 * by the kernel. Whether this will ever overlap anything in
390 * /etc/services is anybody's guess... Returns 0 on success, -1 on
395 int rxinit_status = 1;
396 #ifdef AFS_PTHREAD_ENV
398 * This mutex protects the following global variables:
402 #define LOCK_RX_INIT assert(pthread_mutex_lock(&rx_init_mutex)==0)
403 #define UNLOCK_RX_INIT assert(pthread_mutex_unlock(&rx_init_mutex)==0)
406 #define UNLOCK_RX_INIT
410 rx_InitHost(u_int host, u_int port)
417 char *htable, *ptable;
420 #if defined(AFS_DJGPP_ENV) && !defined(DEBUG)
421 __djgpp_set_quiet_socket(1);
428 if (rxinit_status == 0) {
429 tmp_status = rxinit_status;
431 return tmp_status; /* Already started; return previous error code. */
437 if (afs_winsockInit() < 0)
443 * Initialize anything necessary to provide a non-premptive threading
446 rxi_InitializeThreadSupport();
449 /* Allocate and initialize a socket for client and perhaps server
452 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
453 if (rx_socket == OSI_NULLSOCKET) {
457 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
460 #endif /* RX_LOCKS_DB */
461 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
462 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
463 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
464 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
465 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
467 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
469 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
471 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
473 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
474 #if defined(AFS_HPUX110_ENV)
476 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
477 #endif /* AFS_HPUX110_ENV */
478 #endif /* RX_ENABLE_LOCKS && KERNEL */
481 rx_connDeadTime = 12;
482 rx_tranquil = 0; /* reset flag */
483 memset((char *)&rx_stats, 0, sizeof(struct rx_stats));
485 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
486 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
487 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
488 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
489 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
490 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
492 /* Malloc up a bunch of packets & buffers */
494 queue_Init(&rx_freePacketQueue);
495 rxi_NeedMorePackets = FALSE;
496 #ifdef RX_ENABLE_TSFPQ
497 rx_nPackets = 0; /* in TSFPQ version, rx_nPackets is managed by rxi_MorePackets* */
498 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
499 #else /* RX_ENABLE_TSFPQ */
500 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
501 rxi_MorePackets(rx_nPackets);
502 #endif /* RX_ENABLE_TSFPQ */
509 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
510 tv.tv_sec = clock_now.sec;
511 tv.tv_usec = clock_now.usec;
512 srand((unsigned int)tv.tv_usec);
519 #if defined(KERNEL) && !defined(UKERNEL)
520 /* Really, this should never happen in a real kernel */
523 struct sockaddr_in addr;
525 int addrlen = sizeof(addr);
527 socklen_t addrlen = sizeof(addr);
529 if (getsockname((int)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
533 rx_port = addr.sin_port;
536 rx_stats.minRtt.sec = 9999999;
538 rx_SetEpoch(tv.tv_sec | 0x80000000);
540 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
541 * will provide a randomer value. */
543 MUTEX_ENTER(&rx_stats_mutex);
544 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
545 MUTEX_EXIT(&rx_stats_mutex);
546 /* *Slightly* random start time for the cid. This is just to help
547 * out with the hashing function at the peer */
548 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
549 rx_connHashTable = (struct rx_connection **)htable;
550 rx_peerHashTable = (struct rx_peer **)ptable;
552 rx_lastAckDelay.sec = 0;
553 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
554 rx_hardAckDelay.sec = 0;
555 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
556 rx_softAckDelay.sec = 0;
557 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
559 rxevent_Init(20, rxi_ReScheduleEvents);
561 /* Initialize various global queues */
562 queue_Init(&rx_idleServerQueue);
563 queue_Init(&rx_incomingCallQueue);
564 queue_Init(&rx_freeCallQueue);
566 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
567 /* Initialize our list of usable IP addresses. */
571 /* Start listener process (exact function is dependent on the
572 * implementation environment--kernel or user space) */
576 tmp_status = rxinit_status = 0;
584 return rx_InitHost(htonl(INADDR_ANY), port);
587 /* called with unincremented nRequestsRunning to see if it is OK to start
588 * a new thread in this service. Could be "no" for two reasons: over the
589 * max quota, or would prevent others from reaching their min quota.
591 #ifdef RX_ENABLE_LOCKS
592 /* This verion of QuotaOK reserves quota if it's ok while the
593 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
596 QuotaOK(register struct rx_service *aservice)
598 /* check if over max quota */
599 if (aservice->nRequestsRunning >= aservice->maxProcs) {
603 /* under min quota, we're OK */
604 /* otherwise, can use only if there are enough to allow everyone
605 * to go to their min quota after this guy starts.
607 MUTEX_ENTER(&rx_stats_mutex);
608 if ((aservice->nRequestsRunning < aservice->minProcs)
609 || (rxi_availProcs > rxi_minDeficit)) {
610 aservice->nRequestsRunning++;
611 /* just started call in minProcs pool, need fewer to maintain
613 if (aservice->nRequestsRunning <= aservice->minProcs)
616 MUTEX_EXIT(&rx_stats_mutex);
619 MUTEX_EXIT(&rx_stats_mutex);
625 ReturnToServerPool(register struct rx_service *aservice)
627 aservice->nRequestsRunning--;
628 MUTEX_ENTER(&rx_stats_mutex);
629 if (aservice->nRequestsRunning < aservice->minProcs)
632 MUTEX_EXIT(&rx_stats_mutex);
635 #else /* RX_ENABLE_LOCKS */
637 QuotaOK(register struct rx_service *aservice)
640 /* under min quota, we're OK */
641 if (aservice->nRequestsRunning < aservice->minProcs)
644 /* check if over max quota */
645 if (aservice->nRequestsRunning >= aservice->maxProcs)
648 /* otherwise, can use only if there are enough to allow everyone
649 * to go to their min quota after this guy starts.
651 if (rxi_availProcs > rxi_minDeficit)
655 #endif /* RX_ENABLE_LOCKS */
658 /* Called by rx_StartServer to start up lwp's to service calls.
659 NExistingProcs gives the number of procs already existing, and which
660 therefore needn't be created. */
662 rxi_StartServerProcs(int nExistingProcs)
664 register struct rx_service *service;
669 /* For each service, reserve N processes, where N is the "minimum"
670 * number of processes that MUST be able to execute a request in parallel,
671 * at any time, for that process. Also compute the maximum difference
672 * between any service's maximum number of processes that can run
673 * (i.e. the maximum number that ever will be run, and a guarantee
674 * that this number will run if other services aren't running), and its
675 * minimum number. The result is the extra number of processes that
676 * we need in order to provide the latter guarantee */
677 for (i = 0; i < RX_MAX_SERVICES; i++) {
679 service = rx_services[i];
680 if (service == (struct rx_service *)0)
682 nProcs += service->minProcs;
683 diff = service->maxProcs - service->minProcs;
687 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
688 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
689 for (i = 0; i < nProcs; i++) {
690 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
696 /* This routine is only required on Windows */
698 rx_StartClientThread(void)
700 #ifdef AFS_PTHREAD_ENV
702 pid = pthread_self();
703 #endif /* AFS_PTHREAD_ENV */
705 #endif /* AFS_NT40_ENV */
707 /* This routine must be called if any services are exported. If the
708 * donateMe flag is set, the calling process is donated to the server
711 rx_StartServer(int donateMe)
713 register struct rx_service *service;
719 /* Start server processes, if necessary (exact function is dependent
720 * on the implementation environment--kernel or user space). DonateMe
721 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
722 * case, one less new proc will be created rx_StartServerProcs.
724 rxi_StartServerProcs(donateMe);
726 /* count up the # of threads in minProcs, and add set the min deficit to
727 * be that value, too.
729 for (i = 0; i < RX_MAX_SERVICES; i++) {
730 service = rx_services[i];
731 if (service == (struct rx_service *)0)
733 MUTEX_ENTER(&rx_stats_mutex);
734 rxi_totalMin += service->minProcs;
735 /* below works even if a thread is running, since minDeficit would
736 * still have been decremented and later re-incremented.
738 rxi_minDeficit += service->minProcs;
739 MUTEX_EXIT(&rx_stats_mutex);
742 /* Turn on reaping of idle server connections */
743 rxi_ReapConnections(NULL, NULL, NULL);
752 #ifdef AFS_PTHREAD_ENV
754 pid = (pid_t) pthread_self();
755 #else /* AFS_PTHREAD_ENV */
757 LWP_CurrentProcess(&pid);
758 #endif /* AFS_PTHREAD_ENV */
760 sprintf(name, "srv_%d", ++nProcs);
762 (*registerProgram) (pid, name);
764 #endif /* AFS_NT40_ENV */
765 rx_ServerProc(NULL); /* Never returns */
767 #ifdef RX_ENABLE_TSFPQ
768 /* no use leaving packets around in this thread's local queue if
769 * it isn't getting donated to the server thread pool.
771 rxi_FlushLocalPacketsTSFPQ();
772 #endif /* RX_ENABLE_TSFPQ */
776 /* Create a new client connection to the specified service, using the
777 * specified security object to implement the security model for this
779 struct rx_connection *
780 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
781 register struct rx_securityClass *securityObject,
782 int serviceSecurityIndex)
786 register struct rx_connection *conn;
791 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", ntohl(shost), ntohs(sport), sservice, securityObject, serviceSecurityIndex));
793 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
794 * the case of kmem_alloc? */
795 conn = rxi_AllocConnection();
796 #ifdef RX_ENABLE_LOCKS
797 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
798 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
799 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
802 MUTEX_ENTER(&rx_connHashTable_lock);
803 cid = (rx_nextCid += RX_MAXCALLS);
804 conn->type = RX_CLIENT_CONNECTION;
806 conn->epoch = rx_epoch;
807 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
808 conn->serviceId = sservice;
809 conn->securityObject = securityObject;
810 conn->securityData = (void *) 0;
811 conn->securityIndex = serviceSecurityIndex;
812 rx_SetConnDeadTime(conn, rx_connDeadTime);
813 conn->ackRate = RX_FAST_ACK_RATE;
815 conn->specific = NULL;
816 conn->challengeEvent = NULL;
817 conn->delayedAbortEvent = NULL;
818 conn->abortCount = 0;
820 for (i = 0; i < RX_MAXCALLS; i++) {
821 conn->twind[i] = rx_initSendWindow;
822 conn->rwind[i] = rx_initReceiveWindow;
825 RXS_NewConnection(securityObject, conn);
827 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
829 conn->refCount++; /* no lock required since only this thread knows... */
830 conn->next = rx_connHashTable[hashindex];
831 rx_connHashTable[hashindex] = conn;
832 rx_MutexIncrement(rx_stats.nClientConns, rx_stats_mutex);
833 MUTEX_EXIT(&rx_connHashTable_lock);
839 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
841 /* The idea is to set the dead time to a value that allows several
842 * keepalives to be dropped without timing out the connection. */
843 conn->secondsUntilDead = MAX(seconds, 6);
844 conn->secondsUntilPing = conn->secondsUntilDead / 6;
847 int rxi_lowPeerRefCount = 0;
848 int rxi_lowConnRefCount = 0;
851 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
852 * NOTE: must not be called with rx_connHashTable_lock held.
855 rxi_CleanupConnection(struct rx_connection *conn)
857 /* Notify the service exporter, if requested, that this connection
858 * is being destroyed */
859 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
860 (*conn->service->destroyConnProc) (conn);
862 /* Notify the security module that this connection is being destroyed */
863 RXS_DestroyConnection(conn->securityObject, conn);
865 /* If this is the last connection using the rx_peer struct, set its
866 * idle time to now. rxi_ReapConnections will reap it if it's still
867 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
869 MUTEX_ENTER(&rx_peerHashTable_lock);
870 if (conn->peer->refCount < 2) {
871 conn->peer->idleWhen = clock_Sec();
872 if (conn->peer->refCount < 1) {
873 conn->peer->refCount = 1;
874 MUTEX_ENTER(&rx_stats_mutex);
875 rxi_lowPeerRefCount++;
876 MUTEX_EXIT(&rx_stats_mutex);
879 conn->peer->refCount--;
880 MUTEX_EXIT(&rx_peerHashTable_lock);
882 if (conn->type == RX_SERVER_CONNECTION)
883 rx_MutexDecrement(rx_stats.nServerConns, rx_stats_mutex);
885 rx_MutexDecrement(rx_stats.nClientConns, rx_stats_mutex);
887 if (conn->specific) {
889 for (i = 0; i < conn->nSpecific; i++) {
890 if (conn->specific[i] && rxi_keyCreate_destructor[i])
891 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
892 conn->specific[i] = NULL;
894 free(conn->specific);
896 conn->specific = NULL;
900 MUTEX_DESTROY(&conn->conn_call_lock);
901 MUTEX_DESTROY(&conn->conn_data_lock);
902 CV_DESTROY(&conn->conn_call_cv);
904 rxi_FreeConnection(conn);
907 /* Destroy the specified connection */
909 rxi_DestroyConnection(register struct rx_connection *conn)
911 MUTEX_ENTER(&rx_connHashTable_lock);
912 rxi_DestroyConnectionNoLock(conn);
913 /* conn should be at the head of the cleanup list */
914 if (conn == rx_connCleanup_list) {
915 rx_connCleanup_list = rx_connCleanup_list->next;
916 MUTEX_EXIT(&rx_connHashTable_lock);
917 rxi_CleanupConnection(conn);
919 #ifdef RX_ENABLE_LOCKS
921 MUTEX_EXIT(&rx_connHashTable_lock);
923 #endif /* RX_ENABLE_LOCKS */
927 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
929 register struct rx_connection **conn_ptr;
930 register int havecalls = 0;
931 struct rx_packet *packet;
938 MUTEX_ENTER(&conn->conn_data_lock);
939 if (conn->refCount > 0)
942 MUTEX_ENTER(&rx_stats_mutex);
943 rxi_lowConnRefCount++;
944 MUTEX_EXIT(&rx_stats_mutex);
947 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
948 /* Busy; wait till the last guy before proceeding */
949 MUTEX_EXIT(&conn->conn_data_lock);
954 /* If the client previously called rx_NewCall, but it is still
955 * waiting, treat this as a running call, and wait to destroy the
956 * connection later when the call completes. */
957 if ((conn->type == RX_CLIENT_CONNECTION)
958 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
959 conn->flags |= RX_CONN_DESTROY_ME;
960 MUTEX_EXIT(&conn->conn_data_lock);
964 MUTEX_EXIT(&conn->conn_data_lock);
966 /* Check for extant references to this connection */
967 for (i = 0; i < RX_MAXCALLS; i++) {
968 register struct rx_call *call = conn->call[i];
971 if (conn->type == RX_CLIENT_CONNECTION) {
972 MUTEX_ENTER(&call->lock);
973 if (call->delayedAckEvent) {
974 /* Push the final acknowledgment out now--there
975 * won't be a subsequent call to acknowledge the
976 * last reply packets */
977 rxevent_Cancel(call->delayedAckEvent, call,
978 RX_CALL_REFCOUNT_DELAY);
979 if (call->state == RX_STATE_PRECALL
980 || call->state == RX_STATE_ACTIVE) {
981 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
983 rxi_AckAll(NULL, call, 0);
986 MUTEX_EXIT(&call->lock);
990 #ifdef RX_ENABLE_LOCKS
992 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
993 MUTEX_EXIT(&conn->conn_data_lock);
995 /* Someone is accessing a packet right now. */
999 #endif /* RX_ENABLE_LOCKS */
1002 /* Don't destroy the connection if there are any call
1003 * structures still in use */
1004 MUTEX_ENTER(&conn->conn_data_lock);
1005 conn->flags |= RX_CONN_DESTROY_ME;
1006 MUTEX_EXIT(&conn->conn_data_lock);
1011 if (conn->delayedAbortEvent) {
1012 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
1013 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1015 MUTEX_ENTER(&conn->conn_data_lock);
1016 rxi_SendConnectionAbort(conn, packet, 0, 1);
1017 MUTEX_EXIT(&conn->conn_data_lock);
1018 rxi_FreePacket(packet);
1022 /* Remove from connection hash table before proceeding */
1024 &rx_connHashTable[CONN_HASH
1025 (peer->host, peer->port, conn->cid, conn->epoch,
1027 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1028 if (*conn_ptr == conn) {
1029 *conn_ptr = conn->next;
1033 /* if the conn that we are destroying was the last connection, then we
1034 * clear rxLastConn as well */
1035 if (rxLastConn == conn)
1038 /* Make sure the connection is completely reset before deleting it. */
1039 /* get rid of pending events that could zap us later */
1040 if (conn->challengeEvent)
1041 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1042 if (conn->checkReachEvent)
1043 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1045 /* Add the connection to the list of destroyed connections that
1046 * need to be cleaned up. This is necessary to avoid deadlocks
1047 * in the routines we call to inform others that this connection is
1048 * being destroyed. */
1049 conn->next = rx_connCleanup_list;
1050 rx_connCleanup_list = conn;
1053 /* Externally available version */
1055 rx_DestroyConnection(register struct rx_connection *conn)
1060 rxi_DestroyConnection(conn);
1065 rx_GetConnection(register struct rx_connection *conn)
1070 MUTEX_ENTER(&conn->conn_data_lock);
1072 MUTEX_EXIT(&conn->conn_data_lock);
1076 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1077 /* Wait for the transmit queue to no longer be busy.
1078 * requires the call->lock to be held */
1079 static void rxi_WaitforTQBusy(struct rx_call *call) {
1080 while (call->flags & RX_CALL_TQ_BUSY) {
1081 call->flags |= RX_CALL_TQ_WAIT;
1083 #ifdef RX_ENABLE_LOCKS
1084 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1085 CV_WAIT(&call->cv_tq, &call->lock);
1086 #else /* RX_ENABLE_LOCKS */
1087 osi_rxSleep(&call->tq);
1088 #endif /* RX_ENABLE_LOCKS */
1090 if (call->tqWaiters == 0) {
1091 call->flags &= ~RX_CALL_TQ_WAIT;
1097 /* Start a new rx remote procedure call, on the specified connection.
1098 * If wait is set to 1, wait for a free call channel; otherwise return
1099 * 0. Maxtime gives the maximum number of seconds this call may take,
1100 * after rx_NewCall returns. After this time interval, a call to any
1101 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1102 * For fine grain locking, we hold the conn_call_lock in order to
1103 * to ensure that we don't get signalle after we found a call in an active
1104 * state and before we go to sleep.
1107 rx_NewCall(register struct rx_connection *conn)
1110 register struct rx_call *call;
1111 struct clock queueTime;
1115 dpf(("rx_NewCall(conn %x)\n", conn));
1118 clock_GetTime(&queueTime);
1119 MUTEX_ENTER(&conn->conn_call_lock);
1122 * Check if there are others waiting for a new call.
1123 * If so, let them go first to avoid starving them.
1124 * This is a fairly simple scheme, and might not be
1125 * a complete solution for large numbers of waiters.
1127 * makeCallWaiters keeps track of the number of
1128 * threads waiting to make calls and the
1129 * RX_CONN_MAKECALL_WAITING flag bit is used to
1130 * indicate that there are indeed calls waiting.
1131 * The flag is set when the waiter is incremented.
1132 * It is only cleared in rx_EndCall when
1133 * makeCallWaiters is 0. This prevents us from
1134 * accidently destroying the connection while it
1135 * is potentially about to be used.
1137 MUTEX_ENTER(&conn->conn_data_lock);
1138 if (conn->makeCallWaiters) {
1139 conn->flags |= RX_CONN_MAKECALL_WAITING;
1140 conn->makeCallWaiters++;
1141 MUTEX_EXIT(&conn->conn_data_lock);
1143 #ifdef RX_ENABLE_LOCKS
1144 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1148 MUTEX_ENTER(&conn->conn_data_lock);
1149 conn->makeCallWaiters--;
1151 MUTEX_EXIT(&conn->conn_data_lock);
1154 for (i = 0; i < RX_MAXCALLS; i++) {
1155 call = conn->call[i];
1157 MUTEX_ENTER(&call->lock);
1158 if (call->state == RX_STATE_DALLY) {
1159 rxi_ResetCall(call, 0);
1160 (*call->callNumber)++;
1163 MUTEX_EXIT(&call->lock);
1165 call = rxi_NewCall(conn, i);
1169 if (i < RX_MAXCALLS) {
1172 MUTEX_ENTER(&conn->conn_data_lock);
1173 conn->flags |= RX_CONN_MAKECALL_WAITING;
1174 conn->makeCallWaiters++;
1175 MUTEX_EXIT(&conn->conn_data_lock);
1177 #ifdef RX_ENABLE_LOCKS
1178 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1182 MUTEX_ENTER(&conn->conn_data_lock);
1183 conn->makeCallWaiters--;
1184 MUTEX_EXIT(&conn->conn_data_lock);
1187 * Wake up anyone else who might be giving us a chance to
1188 * run (see code above that avoids resource starvation).
1190 #ifdef RX_ENABLE_LOCKS
1191 CV_BROADCAST(&conn->conn_call_cv);
1196 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1198 /* Client is initially in send mode */
1199 call->state = RX_STATE_ACTIVE;
1200 call->error = conn->error;
1202 call->mode = RX_MODE_ERROR;
1204 call->mode = RX_MODE_SENDING;
1206 /* remember start time for call in case we have hard dead time limit */
1207 call->queueTime = queueTime;
1208 clock_GetTime(&call->startTime);
1209 hzero(call->bytesSent);
1210 hzero(call->bytesRcvd);
1212 /* Turn on busy protocol. */
1213 rxi_KeepAliveOn(call);
1215 MUTEX_EXIT(&call->lock);
1216 MUTEX_EXIT(&conn->conn_call_lock);
1219 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1220 /* Now, if TQ wasn't cleared earlier, do it now. */
1221 MUTEX_ENTER(&call->lock);
1222 rxi_WaitforTQBusy(call);
1223 if (call->flags & RX_CALL_TQ_CLEARME) {
1224 rxi_ClearTransmitQueue(call, 0);
1225 queue_Init(&call->tq);
1227 MUTEX_EXIT(&call->lock);
1228 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1230 dpf(("rx_NewCall(call %x)\n", call));
1235 rxi_HasActiveCalls(register struct rx_connection *aconn)
1238 register struct rx_call *tcall;
1242 for (i = 0; i < RX_MAXCALLS; i++) {
1243 if ((tcall = aconn->call[i])) {
1244 if ((tcall->state == RX_STATE_ACTIVE)
1245 || (tcall->state == RX_STATE_PRECALL)) {
1256 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1257 register afs_int32 * aint32s)
1260 register struct rx_call *tcall;
1264 for (i = 0; i < RX_MAXCALLS; i++) {
1265 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1266 aint32s[i] = aconn->callNumber[i] + 1;
1268 aint32s[i] = aconn->callNumber[i];
1275 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1276 register afs_int32 * aint32s)
1279 register struct rx_call *tcall;
1283 for (i = 0; i < RX_MAXCALLS; i++) {
1284 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1285 aconn->callNumber[i] = aint32s[i] - 1;
1287 aconn->callNumber[i] = aint32s[i];
1293 /* Advertise a new service. A service is named locally by a UDP port
1294 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1297 char *serviceName; Name for identification purposes (e.g. the
1298 service name might be used for probing for
1301 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1302 char *serviceName, struct rx_securityClass **securityObjects,
1303 int nSecurityObjects,
1304 afs_int32(*serviceProc) (struct rx_call * acall))
1306 osi_socket socket = OSI_NULLSOCKET;
1307 register struct rx_service *tservice;
1313 if (serviceId == 0) {
1315 "rx_NewService: service id for service %s is not non-zero.\n",
1322 "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",
1330 tservice = rxi_AllocService();
1332 for (i = 0; i < RX_MAX_SERVICES; i++) {
1333 register struct rx_service *service = rx_services[i];
1335 if (port == service->servicePort && host == service->serviceHost) {
1336 if (service->serviceId == serviceId) {
1337 /* The identical service has already been
1338 * installed; if the caller was intending to
1339 * change the security classes used by this
1340 * service, he/she loses. */
1342 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1343 serviceName, serviceId, service->serviceName);
1345 rxi_FreeService(tservice);
1348 /* Different service, same port: re-use the socket
1349 * which is bound to the same port */
1350 socket = service->socket;
1353 if (socket == OSI_NULLSOCKET) {
1354 /* If we don't already have a socket (from another
1355 * service on same port) get a new one */
1356 socket = rxi_GetHostUDPSocket(host, port);
1357 if (socket == OSI_NULLSOCKET) {
1359 rxi_FreeService(tservice);
1364 service->socket = socket;
1365 service->serviceHost = host;
1366 service->servicePort = port;
1367 service->serviceId = serviceId;
1368 service->serviceName = serviceName;
1369 service->nSecurityObjects = nSecurityObjects;
1370 service->securityObjects = securityObjects;
1371 service->minProcs = 0;
1372 service->maxProcs = 1;
1373 service->idleDeadTime = 60;
1374 service->idleDeadErr = 0;
1375 service->connDeadTime = rx_connDeadTime;
1376 service->executeRequestProc = serviceProc;
1377 service->checkReach = 0;
1378 rx_services[i] = service; /* not visible until now */
1384 rxi_FreeService(tservice);
1385 (osi_Msg "rx_NewService: cannot support > %d services\n",
1390 /* Set configuration options for all of a service's security objects */
1393 rx_SetSecurityConfiguration(struct rx_service *service,
1394 rx_securityConfigVariables type,
1398 for (i = 0; i<service->nSecurityObjects; i++) {
1399 if (service->securityObjects[i]) {
1400 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1408 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1409 struct rx_securityClass **securityObjects, int nSecurityObjects,
1410 afs_int32(*serviceProc) (struct rx_call * acall))
1412 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1415 /* Generic request processing loop. This routine should be called
1416 * by the implementation dependent rx_ServerProc. If socketp is
1417 * non-null, it will be set to the file descriptor that this thread
1418 * is now listening on. If socketp is null, this routine will never
1421 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1423 register struct rx_call *call;
1424 register afs_int32 code;
1425 register struct rx_service *tservice = NULL;
1432 call = rx_GetCall(threadID, tservice, socketp);
1433 if (socketp && *socketp != OSI_NULLSOCKET) {
1434 /* We are now a listener thread */
1439 /* if server is restarting( typically smooth shutdown) then do not
1440 * allow any new calls.
1443 if (rx_tranquil && (call != NULL)) {
1447 MUTEX_ENTER(&call->lock);
1449 rxi_CallError(call, RX_RESTARTING);
1450 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1452 MUTEX_EXIT(&call->lock);
1456 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1457 #ifdef RX_ENABLE_LOCKS
1459 #endif /* RX_ENABLE_LOCKS */
1460 afs_termState = AFSOP_STOP_AFS;
1461 afs_osi_Wakeup(&afs_termState);
1462 #ifdef RX_ENABLE_LOCKS
1464 #endif /* RX_ENABLE_LOCKS */
1469 tservice = call->conn->service;
1471 if (tservice->beforeProc)
1472 (*tservice->beforeProc) (call);
1474 code = call->conn->service->executeRequestProc(call);
1476 if (tservice->afterProc)
1477 (*tservice->afterProc) (call, code);
1479 rx_EndCall(call, code);
1480 MUTEX_ENTER(&rx_stats_mutex);
1482 MUTEX_EXIT(&rx_stats_mutex);
1488 rx_WakeupServerProcs(void)
1490 struct rx_serverQueueEntry *np, *tqp;
1494 MUTEX_ENTER(&rx_serverPool_lock);
1496 #ifdef RX_ENABLE_LOCKS
1497 if (rx_waitForPacket)
1498 CV_BROADCAST(&rx_waitForPacket->cv);
1499 #else /* RX_ENABLE_LOCKS */
1500 if (rx_waitForPacket)
1501 osi_rxWakeup(rx_waitForPacket);
1502 #endif /* RX_ENABLE_LOCKS */
1503 MUTEX_ENTER(&freeSQEList_lock);
1504 for (np = rx_FreeSQEList; np; np = tqp) {
1505 tqp = *(struct rx_serverQueueEntry **)np;
1506 #ifdef RX_ENABLE_LOCKS
1507 CV_BROADCAST(&np->cv);
1508 #else /* RX_ENABLE_LOCKS */
1510 #endif /* RX_ENABLE_LOCKS */
1512 MUTEX_EXIT(&freeSQEList_lock);
1513 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1514 #ifdef RX_ENABLE_LOCKS
1515 CV_BROADCAST(&np->cv);
1516 #else /* RX_ENABLE_LOCKS */
1518 #endif /* RX_ENABLE_LOCKS */
1520 MUTEX_EXIT(&rx_serverPool_lock);
1525 * One thing that seems to happen is that all the server threads get
1526 * tied up on some empty or slow call, and then a whole bunch of calls
1527 * arrive at once, using up the packet pool, so now there are more
1528 * empty calls. The most critical resources here are server threads
1529 * and the free packet pool. The "doreclaim" code seems to help in
1530 * general. I think that eventually we arrive in this state: there
1531 * are lots of pending calls which do have all their packets present,
1532 * so they won't be reclaimed, are multi-packet calls, so they won't
1533 * be scheduled until later, and thus are tying up most of the free
1534 * packet pool for a very long time.
1536 * 1. schedule multi-packet calls if all the packets are present.
1537 * Probably CPU-bound operation, useful to return packets to pool.
1538 * Do what if there is a full window, but the last packet isn't here?
1539 * 3. preserve one thread which *only* runs "best" calls, otherwise
1540 * it sleeps and waits for that type of call.
1541 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1542 * the current dataquota business is badly broken. The quota isn't adjusted
1543 * to reflect how many packets are presently queued for a running call.
1544 * So, when we schedule a queued call with a full window of packets queued
1545 * up for it, that *should* free up a window full of packets for other 2d-class
1546 * calls to be able to use from the packet pool. But it doesn't.
1548 * NB. Most of the time, this code doesn't run -- since idle server threads
1549 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1550 * as a new call arrives.
1552 /* Sleep until a call arrives. Returns a pointer to the call, ready
1553 * for an rx_Read. */
1554 #ifdef RX_ENABLE_LOCKS
1556 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1558 struct rx_serverQueueEntry *sq;
1559 register struct rx_call *call = (struct rx_call *)0;
1560 struct rx_service *service = NULL;
1563 MUTEX_ENTER(&freeSQEList_lock);
1565 if ((sq = rx_FreeSQEList)) {
1566 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1567 MUTEX_EXIT(&freeSQEList_lock);
1568 } else { /* otherwise allocate a new one and return that */
1569 MUTEX_EXIT(&freeSQEList_lock);
1570 sq = (struct rx_serverQueueEntry *)
1571 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1572 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1573 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1576 MUTEX_ENTER(&rx_serverPool_lock);
1577 if (cur_service != NULL) {
1578 ReturnToServerPool(cur_service);
1581 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1582 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1584 /* Scan for eligible incoming calls. A call is not eligible
1585 * if the maximum number of calls for its service type are
1586 * already executing */
1587 /* One thread will process calls FCFS (to prevent starvation),
1588 * while the other threads may run ahead looking for calls which
1589 * have all their input data available immediately. This helps
1590 * keep threads from blocking, waiting for data from the client. */
1591 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1592 service = tcall->conn->service;
1593 if (!QuotaOK(service)) {
1596 if (tno == rxi_fcfs_thread_num
1597 || !tcall->queue_item_header.next) {
1598 /* If we're the fcfs thread , then we'll just use
1599 * this call. If we haven't been able to find an optimal
1600 * choice, and we're at the end of the list, then use a
1601 * 2d choice if one has been identified. Otherwise... */
1602 call = (choice2 ? choice2 : tcall);
1603 service = call->conn->service;
1604 } else if (!queue_IsEmpty(&tcall->rq)) {
1605 struct rx_packet *rp;
1606 rp = queue_First(&tcall->rq, rx_packet);
1607 if (rp->header.seq == 1) {
1609 || (rp->header.flags & RX_LAST_PACKET)) {
1611 } else if (rxi_2dchoice && !choice2
1612 && !(tcall->flags & RX_CALL_CLEARED)
1613 && (tcall->rprev > rxi_HardAckRate)) {
1622 ReturnToServerPool(service);
1629 MUTEX_EXIT(&rx_serverPool_lock);
1630 MUTEX_ENTER(&call->lock);
1632 if (call->flags & RX_CALL_WAIT_PROC) {
1633 call->flags &= ~RX_CALL_WAIT_PROC;
1634 MUTEX_ENTER(&rx_stats_mutex);
1636 MUTEX_EXIT(&rx_stats_mutex);
1639 if (call->state != RX_STATE_PRECALL || call->error) {
1640 MUTEX_EXIT(&call->lock);
1641 MUTEX_ENTER(&rx_serverPool_lock);
1642 ReturnToServerPool(service);
1647 if (queue_IsEmpty(&call->rq)
1648 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1649 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1651 CLEAR_CALL_QUEUE_LOCK(call);
1654 /* If there are no eligible incoming calls, add this process
1655 * to the idle server queue, to wait for one */
1659 *socketp = OSI_NULLSOCKET;
1661 sq->socketp = socketp;
1662 queue_Append(&rx_idleServerQueue, sq);
1663 #ifndef AFS_AIX41_ENV
1664 rx_waitForPacket = sq;
1666 rx_waitingForPacket = sq;
1667 #endif /* AFS_AIX41_ENV */
1669 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1671 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1672 MUTEX_EXIT(&rx_serverPool_lock);
1673 return (struct rx_call *)0;
1676 } while (!(call = sq->newcall)
1677 && !(socketp && *socketp != OSI_NULLSOCKET));
1678 MUTEX_EXIT(&rx_serverPool_lock);
1680 MUTEX_ENTER(&call->lock);
1686 MUTEX_ENTER(&freeSQEList_lock);
1687 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1688 rx_FreeSQEList = sq;
1689 MUTEX_EXIT(&freeSQEList_lock);
1692 clock_GetTime(&call->startTime);
1693 call->state = RX_STATE_ACTIVE;
1694 call->mode = RX_MODE_RECEIVING;
1695 #ifdef RX_KERNEL_TRACE
1696 if (ICL_SETACTIVE(afs_iclSetp)) {
1697 int glockOwner = ISAFS_GLOCK();
1700 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1701 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1708 rxi_calltrace(RX_CALL_START, call);
1709 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1710 call->conn->service->servicePort, call->conn->service->serviceId,
1713 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1714 MUTEX_EXIT(&call->lock);
1716 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1721 #else /* RX_ENABLE_LOCKS */
1723 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1725 struct rx_serverQueueEntry *sq;
1726 register struct rx_call *call = (struct rx_call *)0, *choice2;
1727 struct rx_service *service = NULL;
1731 MUTEX_ENTER(&freeSQEList_lock);
1733 if ((sq = rx_FreeSQEList)) {
1734 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1735 MUTEX_EXIT(&freeSQEList_lock);
1736 } else { /* otherwise allocate a new one and return that */
1737 MUTEX_EXIT(&freeSQEList_lock);
1738 sq = (struct rx_serverQueueEntry *)
1739 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1740 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1741 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1743 MUTEX_ENTER(&sq->lock);
1745 if (cur_service != NULL) {
1746 cur_service->nRequestsRunning--;
1747 if (cur_service->nRequestsRunning < cur_service->minProcs)
1751 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1752 register struct rx_call *tcall, *ncall;
1753 /* Scan for eligible incoming calls. A call is not eligible
1754 * if the maximum number of calls for its service type are
1755 * already executing */
1756 /* One thread will process calls FCFS (to prevent starvation),
1757 * while the other threads may run ahead looking for calls which
1758 * have all their input data available immediately. This helps
1759 * keep threads from blocking, waiting for data from the client. */
1760 choice2 = (struct rx_call *)0;
1761 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1762 service = tcall->conn->service;
1763 if (QuotaOK(service)) {
1764 if (tno == rxi_fcfs_thread_num
1765 || !tcall->queue_item_header.next) {
1766 /* If we're the fcfs thread, then we'll just use
1767 * this call. If we haven't been able to find an optimal
1768 * choice, and we're at the end of the list, then use a
1769 * 2d choice if one has been identified. Otherwise... */
1770 call = (choice2 ? choice2 : tcall);
1771 service = call->conn->service;
1772 } else if (!queue_IsEmpty(&tcall->rq)) {
1773 struct rx_packet *rp;
1774 rp = queue_First(&tcall->rq, rx_packet);
1775 if (rp->header.seq == 1
1777 || (rp->header.flags & RX_LAST_PACKET))) {
1779 } else if (rxi_2dchoice && !choice2
1780 && !(tcall->flags & RX_CALL_CLEARED)
1781 && (tcall->rprev > rxi_HardAckRate)) {
1794 /* we can't schedule a call if there's no data!!! */
1795 /* send an ack if there's no data, if we're missing the
1796 * first packet, or we're missing something between first
1797 * and last -- there's a "hole" in the incoming data. */
1798 if (queue_IsEmpty(&call->rq)
1799 || queue_First(&call->rq, rx_packet)->header.seq != 1
1800 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1801 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1803 call->flags &= (~RX_CALL_WAIT_PROC);
1804 service->nRequestsRunning++;
1805 /* just started call in minProcs pool, need fewer to maintain
1807 if (service->nRequestsRunning <= service->minProcs)
1811 /* MUTEX_EXIT(&call->lock); */
1813 /* If there are no eligible incoming calls, add this process
1814 * to the idle server queue, to wait for one */
1817 *socketp = OSI_NULLSOCKET;
1819 sq->socketp = socketp;
1820 queue_Append(&rx_idleServerQueue, sq);
1824 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1826 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1827 return (struct rx_call *)0;
1830 } while (!(call = sq->newcall)
1831 && !(socketp && *socketp != OSI_NULLSOCKET));
1833 MUTEX_EXIT(&sq->lock);
1835 MUTEX_ENTER(&freeSQEList_lock);
1836 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1837 rx_FreeSQEList = sq;
1838 MUTEX_EXIT(&freeSQEList_lock);
1841 clock_GetTime(&call->startTime);
1842 call->state = RX_STATE_ACTIVE;
1843 call->mode = RX_MODE_RECEIVING;
1844 #ifdef RX_KERNEL_TRACE
1845 if (ICL_SETACTIVE(afs_iclSetp)) {
1846 int glockOwner = ISAFS_GLOCK();
1849 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1850 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1857 rxi_calltrace(RX_CALL_START, call);
1858 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1859 call->conn->service->servicePort, call->conn->service->serviceId,
1862 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1869 #endif /* RX_ENABLE_LOCKS */
1873 /* Establish a procedure to be called when a packet arrives for a
1874 * call. This routine will be called at most once after each call,
1875 * and will also be called if there is an error condition on the or
1876 * the call is complete. Used by multi rx to build a selection
1877 * function which determines which of several calls is likely to be a
1878 * good one to read from.
1879 * NOTE: the way this is currently implemented it is probably only a
1880 * good idea to (1) use it immediately after a newcall (clients only)
1881 * and (2) only use it once. Other uses currently void your warranty
1884 rx_SetArrivalProc(register struct rx_call *call,
1885 register void (*proc) (register struct rx_call * call,
1887 register int index),
1888 register void * handle, register int arg)
1890 call->arrivalProc = proc;
1891 call->arrivalProcHandle = handle;
1892 call->arrivalProcArg = arg;
1895 /* Call is finished (possibly prematurely). Return rc to the peer, if
1896 * appropriate, and return the final error code from the conversation
1900 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1902 register struct rx_connection *conn = call->conn;
1903 register struct rx_service *service;
1909 dpf(("rx_EndCall(call %x rc %d error %d abortCode %d)\n", call, rc, call->error, call->abortCode));
1912 MUTEX_ENTER(&call->lock);
1914 if (rc == 0 && call->error == 0) {
1915 call->abortCode = 0;
1916 call->abortCount = 0;
1919 call->arrivalProc = (void (*)())0;
1920 if (rc && call->error == 0) {
1921 rxi_CallError(call, rc);
1922 /* Send an abort message to the peer if this error code has
1923 * only just been set. If it was set previously, assume the
1924 * peer has already been sent the error code or will request it
1926 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1928 if (conn->type == RX_SERVER_CONNECTION) {
1929 /* Make sure reply or at least dummy reply is sent */
1930 if (call->mode == RX_MODE_RECEIVING) {
1931 rxi_WriteProc(call, 0, 0);
1933 if (call->mode == RX_MODE_SENDING) {
1934 rxi_FlushWrite(call);
1936 service = conn->service;
1937 rxi_calltrace(RX_CALL_END, call);
1938 /* Call goes to hold state until reply packets are acknowledged */
1939 if (call->tfirst + call->nSoftAcked < call->tnext) {
1940 call->state = RX_STATE_HOLD;
1942 call->state = RX_STATE_DALLY;
1943 rxi_ClearTransmitQueue(call, 0);
1944 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1945 rxevent_Cancel(call->keepAliveEvent, call,
1946 RX_CALL_REFCOUNT_ALIVE);
1948 } else { /* Client connection */
1950 /* Make sure server receives input packets, in the case where
1951 * no reply arguments are expected */
1952 if ((call->mode == RX_MODE_SENDING)
1953 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1954 (void)rxi_ReadProc(call, &dummy, 1);
1957 /* If we had an outstanding delayed ack, be nice to the server
1958 * and force-send it now.
1960 if (call->delayedAckEvent) {
1961 rxevent_Cancel(call->delayedAckEvent, call,
1962 RX_CALL_REFCOUNT_DELAY);
1963 call->delayedAckEvent = NULL;
1964 rxi_SendDelayedAck(NULL, call, NULL);
1967 /* We need to release the call lock since it's lower than the
1968 * conn_call_lock and we don't want to hold the conn_call_lock
1969 * over the rx_ReadProc call. The conn_call_lock needs to be held
1970 * here for the case where rx_NewCall is perusing the calls on
1971 * the connection structure. We don't want to signal until
1972 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1973 * have checked this call, found it active and by the time it
1974 * goes to sleep, will have missed the signal.
1976 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
1977 * there are threads waiting to use the conn object.
1979 MUTEX_EXIT(&call->lock);
1980 MUTEX_ENTER(&conn->conn_call_lock);
1981 MUTEX_ENTER(&call->lock);
1982 MUTEX_ENTER(&conn->conn_data_lock);
1983 conn->flags |= RX_CONN_BUSY;
1984 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1985 if (conn->makeCallWaiters == 0)
1986 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1987 MUTEX_EXIT(&conn->conn_data_lock);
1988 #ifdef RX_ENABLE_LOCKS
1989 CV_BROADCAST(&conn->conn_call_cv);
1994 #ifdef RX_ENABLE_LOCKS
1996 MUTEX_EXIT(&conn->conn_data_lock);
1998 #endif /* RX_ENABLE_LOCKS */
1999 call->state = RX_STATE_DALLY;
2001 error = call->error;
2003 /* currentPacket, nLeft, and NFree must be zeroed here, because
2004 * ResetCall cannot: ResetCall may be called at splnet(), in the
2005 * kernel version, and may interrupt the macros rx_Read or
2006 * rx_Write, which run at normal priority for efficiency. */
2007 if (call->currentPacket) {
2008 queue_Prepend(&call->iovq, call->currentPacket);
2009 call->currentPacket = (struct rx_packet *)0;
2012 call->nLeft = call->nFree = call->curlen = 0;
2014 /* Free any packets from the last call to ReadvProc/WritevProc */
2015 rxi_FreePackets(0, &call->iovq);
2017 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2018 MUTEX_EXIT(&call->lock);
2019 if (conn->type == RX_CLIENT_CONNECTION) {
2020 MUTEX_EXIT(&conn->conn_call_lock);
2021 conn->flags &= ~RX_CONN_BUSY;
2025 * Map errors to the local host's errno.h format.
2027 error = ntoh_syserr_conv(error);
2031 #if !defined(KERNEL)
2033 /* Call this routine when shutting down a server or client (especially
2034 * clients). This will allow Rx to gracefully garbage collect server
2035 * connections, and reduce the number of retries that a server might
2036 * make to a dead client.
2037 * This is not quite right, since some calls may still be ongoing and
2038 * we can't lock them to destroy them. */
2042 register struct rx_connection **conn_ptr, **conn_end;
2046 if (rxinit_status == 1) {
2048 return; /* Already shutdown. */
2050 rxi_DeleteCachedConnections();
2051 if (rx_connHashTable) {
2052 MUTEX_ENTER(&rx_connHashTable_lock);
2053 for (conn_ptr = &rx_connHashTable[0], conn_end =
2054 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2056 struct rx_connection *conn, *next;
2057 for (conn = *conn_ptr; conn; conn = next) {
2059 if (conn->type == RX_CLIENT_CONNECTION) {
2060 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2062 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2063 #ifdef RX_ENABLE_LOCKS
2064 rxi_DestroyConnectionNoLock(conn);
2065 #else /* RX_ENABLE_LOCKS */
2066 rxi_DestroyConnection(conn);
2067 #endif /* RX_ENABLE_LOCKS */
2071 #ifdef RX_ENABLE_LOCKS
2072 while (rx_connCleanup_list) {
2073 struct rx_connection *conn;
2074 conn = rx_connCleanup_list;
2075 rx_connCleanup_list = rx_connCleanup_list->next;
2076 MUTEX_EXIT(&rx_connHashTable_lock);
2077 rxi_CleanupConnection(conn);
2078 MUTEX_ENTER(&rx_connHashTable_lock);
2080 MUTEX_EXIT(&rx_connHashTable_lock);
2081 #endif /* RX_ENABLE_LOCKS */
2086 afs_winsockCleanup();
2094 /* if we wakeup packet waiter too often, can get in loop with two
2095 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2097 rxi_PacketsUnWait(void)
2099 if (!rx_waitingForPackets) {
2103 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2104 return; /* still over quota */
2107 rx_waitingForPackets = 0;
2108 #ifdef RX_ENABLE_LOCKS
2109 CV_BROADCAST(&rx_waitingForPackets_cv);
2111 osi_rxWakeup(&rx_waitingForPackets);
2117 /* ------------------Internal interfaces------------------------- */
2119 /* Return this process's service structure for the
2120 * specified socket and service */
2122 rxi_FindService(register osi_socket socket, register u_short serviceId)
2124 register struct rx_service **sp;
2125 for (sp = &rx_services[0]; *sp; sp++) {
2126 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2132 /* Allocate a call structure, for the indicated channel of the
2133 * supplied connection. The mode and state of the call must be set by
2134 * the caller. Returns the call with mutex locked. */
2136 rxi_NewCall(register struct rx_connection *conn, register int channel)
2138 register struct rx_call *call;
2139 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2140 register struct rx_call *cp; /* Call pointer temp */
2141 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2142 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2144 dpf(("rxi_NewCall(conn %x, channel %d)\n", conn, channel));
2146 /* Grab an existing call structure, or allocate a new one.
2147 * Existing call structures are assumed to have been left reset by
2149 MUTEX_ENTER(&rx_freeCallQueue_lock);
2151 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2153 * EXCEPT that the TQ might not yet be cleared out.
2154 * Skip over those with in-use TQs.
2157 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2158 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2164 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2165 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2166 call = queue_First(&rx_freeCallQueue, rx_call);
2167 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2169 rx_MutexDecrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2170 MUTEX_EXIT(&rx_freeCallQueue_lock);
2171 MUTEX_ENTER(&call->lock);
2172 CLEAR_CALL_QUEUE_LOCK(call);
2173 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2174 /* Now, if TQ wasn't cleared earlier, do it now. */
2175 if (call->flags & RX_CALL_TQ_CLEARME) {
2176 rxi_ClearTransmitQueue(call, 0);
2177 queue_Init(&call->tq);
2179 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2180 /* Bind the call to its connection structure */
2182 rxi_ResetCall(call, 1);
2185 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2187 MUTEX_EXIT(&rx_freeCallQueue_lock);
2188 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2189 MUTEX_ENTER(&call->lock);
2190 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2191 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2192 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2194 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2195 /* Initialize once-only items */
2196 queue_Init(&call->tq);
2197 queue_Init(&call->rq);
2198 queue_Init(&call->iovq);
2199 /* Bind the call to its connection structure (prereq for reset) */
2201 rxi_ResetCall(call, 1);
2203 call->channel = channel;
2204 call->callNumber = &conn->callNumber[channel];
2205 call->rwind = conn->rwind[channel];
2206 call->twind = conn->twind[channel];
2207 /* Note that the next expected call number is retained (in
2208 * conn->callNumber[i]), even if we reallocate the call structure
2210 conn->call[channel] = call;
2211 /* if the channel's never been used (== 0), we should start at 1, otherwise
2212 * the call number is valid from the last time this channel was used */
2213 if (*call->callNumber == 0)
2214 *call->callNumber = 1;
2219 /* A call has been inactive long enough that so we can throw away
2220 * state, including the call structure, which is placed on the call
2222 * Call is locked upon entry.
2223 * haveCTLock set if called from rxi_ReapConnections
2225 #ifdef RX_ENABLE_LOCKS
2227 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2228 #else /* RX_ENABLE_LOCKS */
2230 rxi_FreeCall(register struct rx_call *call)
2231 #endif /* RX_ENABLE_LOCKS */
2233 register int channel = call->channel;
2234 register struct rx_connection *conn = call->conn;
2237 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2238 (*call->callNumber)++;
2239 rxi_ResetCall(call, 0);
2240 call->conn->call[channel] = (struct rx_call *)0;
2242 MUTEX_ENTER(&rx_freeCallQueue_lock);
2243 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2244 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2245 /* A call may be free even though its transmit queue is still in use.
2246 * Since we search the call list from head to tail, put busy calls at
2247 * the head of the list, and idle calls at the tail.
2249 if (call->flags & RX_CALL_TQ_BUSY)
2250 queue_Prepend(&rx_freeCallQueue, call);
2252 queue_Append(&rx_freeCallQueue, call);
2253 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2254 queue_Append(&rx_freeCallQueue, call);
2255 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2256 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2257 MUTEX_EXIT(&rx_freeCallQueue_lock);
2259 /* Destroy the connection if it was previously slated for
2260 * destruction, i.e. the Rx client code previously called
2261 * rx_DestroyConnection (client connections), or
2262 * rxi_ReapConnections called the same routine (server
2263 * connections). Only do this, however, if there are no
2264 * outstanding calls. Note that for fine grain locking, there appears
2265 * to be a deadlock in that rxi_FreeCall has a call locked and
2266 * DestroyConnectionNoLock locks each call in the conn. But note a
2267 * few lines up where we have removed this call from the conn.
2268 * If someone else destroys a connection, they either have no
2269 * call lock held or are going through this section of code.
2271 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2272 MUTEX_ENTER(&conn->conn_data_lock);
2274 MUTEX_EXIT(&conn->conn_data_lock);
2275 #ifdef RX_ENABLE_LOCKS
2277 rxi_DestroyConnectionNoLock(conn);
2279 rxi_DestroyConnection(conn);
2280 #else /* RX_ENABLE_LOCKS */
2281 rxi_DestroyConnection(conn);
2282 #endif /* RX_ENABLE_LOCKS */
2286 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2288 rxi_Alloc(register size_t size)
2292 rx_MutexAdd1Increment2(rxi_Allocsize, (afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2295 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2296 afs_osi_Alloc_NoSleep(size);
2301 osi_Panic("rxi_Alloc error");
2307 rxi_Free(void *addr, register size_t size)
2309 rx_MutexAdd1Decrement2(rxi_Allocsize, -(afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2310 osi_Free(addr, size);
2314 rxi_SetPeerMtu(register afs_uint32 host, register afs_uint32 port, int mtu)
2316 struct rx_peer **peer_ptr, **peer_end;
2319 MUTEX_ENTER(&rx_peerHashTable_lock);
2321 for (peer_ptr = &rx_peerHashTable[0], peer_end =
2322 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
2324 struct rx_peer *peer, *next;
2325 for (peer = *peer_ptr; peer; peer = next) {
2327 if (host == peer->host) {
2328 MUTEX_ENTER(&peer->peer_lock);
2329 peer->ifMTU=MIN(mtu, peer->ifMTU);
2330 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2331 MUTEX_EXIT(&peer->peer_lock);
2336 struct rx_peer *peer;
2337 hashIndex = PEER_HASH(host, port);
2338 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2339 if ((peer->host == host) && (peer->port == port)) {
2340 MUTEX_ENTER(&peer->peer_lock);
2341 peer->ifMTU=MIN(mtu, peer->ifMTU);
2342 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2343 MUTEX_EXIT(&peer->peer_lock);
2347 MUTEX_EXIT(&rx_peerHashTable_lock);
2350 /* Find the peer process represented by the supplied (host,port)
2351 * combination. If there is no appropriate active peer structure, a
2352 * new one will be allocated and initialized
2353 * The origPeer, if set, is a pointer to a peer structure on which the
2354 * refcount will be be decremented. This is used to replace the peer
2355 * structure hanging off a connection structure */
2357 rxi_FindPeer(register afs_uint32 host, register u_short port,
2358 struct rx_peer *origPeer, int create)
2360 register struct rx_peer *pp;
2362 hashIndex = PEER_HASH(host, port);
2363 MUTEX_ENTER(&rx_peerHashTable_lock);
2364 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2365 if ((pp->host == host) && (pp->port == port))
2370 pp = rxi_AllocPeer(); /* This bzero's *pp */
2371 pp->host = host; /* set here or in InitPeerParams is zero */
2373 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2374 queue_Init(&pp->congestionQueue);
2375 queue_Init(&pp->rpcStats);
2376 pp->next = rx_peerHashTable[hashIndex];
2377 rx_peerHashTable[hashIndex] = pp;
2378 rxi_InitPeerParams(pp);
2379 rx_MutexIncrement(rx_stats.nPeerStructs, rx_stats_mutex);
2386 origPeer->refCount--;
2387 MUTEX_EXIT(&rx_peerHashTable_lock);
2392 /* Find the connection at (host, port) started at epoch, and with the
2393 * given connection id. Creates the server connection if necessary.
2394 * The type specifies whether a client connection or a server
2395 * connection is desired. In both cases, (host, port) specify the
2396 * peer's (host, pair) pair. Client connections are not made
2397 * automatically by this routine. The parameter socket gives the
2398 * socket descriptor on which the packet was received. This is used,
2399 * in the case of server connections, to check that *new* connections
2400 * come via a valid (port, serviceId). Finally, the securityIndex
2401 * parameter must match the existing index for the connection. If a
2402 * server connection is created, it will be created using the supplied
2403 * index, if the index is valid for this service */
2404 struct rx_connection *
2405 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2406 register u_short port, u_short serviceId, afs_uint32 cid,
2407 afs_uint32 epoch, int type, u_int securityIndex)
2409 int hashindex, flag, i;
2410 register struct rx_connection *conn;
2411 hashindex = CONN_HASH(host, port, cid, epoch, type);
2412 MUTEX_ENTER(&rx_connHashTable_lock);
2413 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2414 rx_connHashTable[hashindex],
2417 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2418 && (epoch == conn->epoch)) {
2419 register struct rx_peer *pp = conn->peer;
2420 if (securityIndex != conn->securityIndex) {
2421 /* this isn't supposed to happen, but someone could forge a packet
2422 * like this, and there seems to be some CM bug that makes this
2423 * happen from time to time -- in which case, the fileserver
2425 MUTEX_EXIT(&rx_connHashTable_lock);
2426 return (struct rx_connection *)0;
2428 if (pp->host == host && pp->port == port)
2430 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2432 /* So what happens when it's a callback connection? */
2433 if ( /*type == RX_CLIENT_CONNECTION && */
2434 (conn->epoch & 0x80000000))
2438 /* the connection rxLastConn that was used the last time is not the
2439 ** one we are looking for now. Hence, start searching in the hash */
2441 conn = rx_connHashTable[hashindex];
2446 struct rx_service *service;
2447 if (type == RX_CLIENT_CONNECTION) {
2448 MUTEX_EXIT(&rx_connHashTable_lock);
2449 return (struct rx_connection *)0;
2451 service = rxi_FindService(socket, serviceId);
2452 if (!service || (securityIndex >= service->nSecurityObjects)
2453 || (service->securityObjects[securityIndex] == 0)) {
2454 MUTEX_EXIT(&rx_connHashTable_lock);
2455 return (struct rx_connection *)0;
2457 conn = rxi_AllocConnection(); /* This bzero's the connection */
2458 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2459 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2460 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2461 conn->next = rx_connHashTable[hashindex];
2462 rx_connHashTable[hashindex] = conn;
2463 conn->peer = rxi_FindPeer(host, port, 0, 1);
2464 conn->type = RX_SERVER_CONNECTION;
2465 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2466 conn->epoch = epoch;
2467 conn->cid = cid & RX_CIDMASK;
2468 /* conn->serial = conn->lastSerial = 0; */
2469 /* conn->timeout = 0; */
2470 conn->ackRate = RX_FAST_ACK_RATE;
2471 conn->service = service;
2472 conn->serviceId = serviceId;
2473 conn->securityIndex = securityIndex;
2474 conn->securityObject = service->securityObjects[securityIndex];
2475 conn->nSpecific = 0;
2476 conn->specific = NULL;
2477 rx_SetConnDeadTime(conn, service->connDeadTime);
2478 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2479 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2480 for (i = 0; i < RX_MAXCALLS; i++) {
2481 conn->twind[i] = rx_initSendWindow;
2482 conn->rwind[i] = rx_initReceiveWindow;
2484 /* Notify security object of the new connection */
2485 RXS_NewConnection(conn->securityObject, conn);
2486 /* XXXX Connection timeout? */
2487 if (service->newConnProc)
2488 (*service->newConnProc) (conn);
2489 rx_MutexIncrement(rx_stats.nServerConns, rx_stats_mutex);
2492 MUTEX_ENTER(&conn->conn_data_lock);
2494 MUTEX_EXIT(&conn->conn_data_lock);
2496 rxLastConn = conn; /* store this connection as the last conn used */
2497 MUTEX_EXIT(&rx_connHashTable_lock);
2501 /* There are two packet tracing routines available for testing and monitoring
2502 * Rx. One is called just after every packet is received and the other is
2503 * called just before every packet is sent. Received packets, have had their
2504 * headers decoded, and packets to be sent have not yet had their headers
2505 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2506 * containing the network address. Both can be modified. The return value, if
2507 * non-zero, indicates that the packet should be dropped. */
2509 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
2510 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
2512 /* A packet has been received off the interface. Np is the packet, socket is
2513 * the socket number it was received from (useful in determining which service
2514 * this packet corresponds to), and (host, port) reflect the host,port of the
2515 * sender. This call returns the packet to the caller if it is finished with
2516 * it, rather than de-allocating it, just as a small performance hack */
2519 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2520 afs_uint32 host, u_short port, int *tnop,
2521 struct rx_call **newcallp)
2523 register struct rx_call *call;
2524 register struct rx_connection *conn;
2526 afs_uint32 currentCallNumber;
2532 struct rx_packet *tnp;
2535 /* We don't print out the packet until now because (1) the time may not be
2536 * accurate enough until now in the lwp implementation (rx_Listener only gets
2537 * the time after the packet is read) and (2) from a protocol point of view,
2538 * this is the first time the packet has been seen */
2539 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2540 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2541 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2542 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2543 np->header.epoch, np->header.cid, np->header.callNumber,
2544 np->header.seq, np->header.flags, np));
2547 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2548 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2551 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2552 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2555 /* If an input tracer function is defined, call it with the packet and
2556 * network address. Note this function may modify its arguments. */
2557 if (rx_justReceived) {
2558 struct sockaddr_in addr;
2560 addr.sin_family = AF_INET;
2561 addr.sin_port = port;
2562 addr.sin_addr.s_addr = host;
2563 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2564 addr.sin_len = sizeof(addr);
2565 #endif /* AFS_OSF_ENV */
2566 drop = (*rx_justReceived) (np, &addr);
2567 /* drop packet if return value is non-zero */
2570 port = addr.sin_port; /* in case fcn changed addr */
2571 host = addr.sin_addr.s_addr;
2575 /* If packet was not sent by the client, then *we* must be the client */
2576 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2577 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2579 /* Find the connection (or fabricate one, if we're the server & if
2580 * necessary) associated with this packet */
2582 rxi_FindConnection(socket, host, port, np->header.serviceId,
2583 np->header.cid, np->header.epoch, type,
2584 np->header.securityIndex);
2587 /* If no connection found or fabricated, just ignore the packet.
2588 * (An argument could be made for sending an abort packet for
2593 MUTEX_ENTER(&conn->conn_data_lock);
2594 if (conn->maxSerial < np->header.serial)
2595 conn->maxSerial = np->header.serial;
2596 MUTEX_EXIT(&conn->conn_data_lock);
2598 /* If the connection is in an error state, send an abort packet and ignore
2599 * the incoming packet */
2601 /* Don't respond to an abort packet--we don't want loops! */
2602 MUTEX_ENTER(&conn->conn_data_lock);
2603 if (np->header.type != RX_PACKET_TYPE_ABORT)
2604 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2606 MUTEX_EXIT(&conn->conn_data_lock);
2610 /* Check for connection-only requests (i.e. not call specific). */
2611 if (np->header.callNumber == 0) {
2612 switch (np->header.type) {
2613 case RX_PACKET_TYPE_ABORT: {
2614 /* What if the supplied error is zero? */
2615 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2616 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2617 rxi_ConnectionError(conn, errcode);
2618 MUTEX_ENTER(&conn->conn_data_lock);
2620 MUTEX_EXIT(&conn->conn_data_lock);
2623 case RX_PACKET_TYPE_CHALLENGE:
2624 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2625 MUTEX_ENTER(&conn->conn_data_lock);
2627 MUTEX_EXIT(&conn->conn_data_lock);
2629 case RX_PACKET_TYPE_RESPONSE:
2630 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2631 MUTEX_ENTER(&conn->conn_data_lock);
2633 MUTEX_EXIT(&conn->conn_data_lock);
2635 case RX_PACKET_TYPE_PARAMS:
2636 case RX_PACKET_TYPE_PARAMS + 1:
2637 case RX_PACKET_TYPE_PARAMS + 2:
2638 /* ignore these packet types for now */
2639 MUTEX_ENTER(&conn->conn_data_lock);
2641 MUTEX_EXIT(&conn->conn_data_lock);
2646 /* Should not reach here, unless the peer is broken: send an
2648 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2649 MUTEX_ENTER(&conn->conn_data_lock);
2650 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2652 MUTEX_EXIT(&conn->conn_data_lock);
2657 channel = np->header.cid & RX_CHANNELMASK;
2658 call = conn->call[channel];
2659 #ifdef RX_ENABLE_LOCKS
2661 MUTEX_ENTER(&call->lock);
2662 /* Test to see if call struct is still attached to conn. */
2663 if (call != conn->call[channel]) {
2665 MUTEX_EXIT(&call->lock);
2666 if (type == RX_SERVER_CONNECTION) {
2667 call = conn->call[channel];
2668 /* If we started with no call attached and there is one now,
2669 * another thread is also running this routine and has gotten
2670 * the connection channel. We should drop this packet in the tests
2671 * below. If there was a call on this connection and it's now
2672 * gone, then we'll be making a new call below.
2673 * If there was previously a call and it's now different then
2674 * the old call was freed and another thread running this routine
2675 * has created a call on this channel. One of these two threads
2676 * has a packet for the old call and the code below handles those
2680 MUTEX_ENTER(&call->lock);
2682 /* This packet can't be for this call. If the new call address is
2683 * 0 then no call is running on this channel. If there is a call
2684 * then, since this is a client connection we're getting data for
2685 * it must be for the previous call.
2687 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2688 MUTEX_ENTER(&conn->conn_data_lock);
2690 MUTEX_EXIT(&conn->conn_data_lock);
2695 currentCallNumber = conn->callNumber[channel];
2697 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2698 if (np->header.callNumber < currentCallNumber) {
2699 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2700 #ifdef RX_ENABLE_LOCKS
2702 MUTEX_EXIT(&call->lock);
2704 MUTEX_ENTER(&conn->conn_data_lock);
2706 MUTEX_EXIT(&conn->conn_data_lock);
2710 MUTEX_ENTER(&conn->conn_call_lock);
2711 call = rxi_NewCall(conn, channel);
2712 MUTEX_EXIT(&conn->conn_call_lock);
2713 *call->callNumber = np->header.callNumber;
2714 if (np->header.callNumber == 0)
2715 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));
2717 call->state = RX_STATE_PRECALL;
2718 clock_GetTime(&call->queueTime);
2719 hzero(call->bytesSent);
2720 hzero(call->bytesRcvd);
2722 * If the number of queued calls exceeds the overload
2723 * threshold then abort this call.
2725 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2726 struct rx_packet *tp;
2728 rxi_CallError(call, rx_BusyError);
2729 tp = rxi_SendCallAbort(call, np, 1, 0);
2730 MUTEX_EXIT(&call->lock);
2731 MUTEX_ENTER(&conn->conn_data_lock);
2733 MUTEX_EXIT(&conn->conn_data_lock);
2734 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2737 rxi_KeepAliveOn(call);
2738 } else if (np->header.callNumber != currentCallNumber) {
2739 /* Wait until the transmit queue is idle before deciding
2740 * whether to reset the current call. Chances are that the
2741 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2744 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2745 while ((call->state == RX_STATE_ACTIVE)
2746 && (call->flags & RX_CALL_TQ_BUSY)) {
2747 call->flags |= RX_CALL_TQ_WAIT;
2749 #ifdef RX_ENABLE_LOCKS
2750 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2751 CV_WAIT(&call->cv_tq, &call->lock);
2752 #else /* RX_ENABLE_LOCKS */
2753 osi_rxSleep(&call->tq);
2754 #endif /* RX_ENABLE_LOCKS */
2756 if (call->tqWaiters == 0)
2757 call->flags &= ~RX_CALL_TQ_WAIT;
2759 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2760 /* If the new call cannot be taken right now send a busy and set
2761 * the error condition in this call, so that it terminates as
2762 * quickly as possible */
2763 if (call->state == RX_STATE_ACTIVE) {
2764 struct rx_packet *tp;
2766 rxi_CallError(call, RX_CALL_DEAD);
2767 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2769 MUTEX_EXIT(&call->lock);
2770 MUTEX_ENTER(&conn->conn_data_lock);
2772 MUTEX_EXIT(&conn->conn_data_lock);
2775 rxi_ResetCall(call, 0);
2776 *call->callNumber = np->header.callNumber;
2777 if (np->header.callNumber == 0)
2778 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));
2780 call->state = RX_STATE_PRECALL;
2781 clock_GetTime(&call->queueTime);
2782 hzero(call->bytesSent);
2783 hzero(call->bytesRcvd);
2785 * If the number of queued calls exceeds the overload
2786 * threshold then abort this call.
2788 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2789 struct rx_packet *tp;
2791 rxi_CallError(call, rx_BusyError);
2792 tp = rxi_SendCallAbort(call, np, 1, 0);
2793 MUTEX_EXIT(&call->lock);
2794 MUTEX_ENTER(&conn->conn_data_lock);
2796 MUTEX_EXIT(&conn->conn_data_lock);
2797 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2800 rxi_KeepAliveOn(call);
2802 /* Continuing call; do nothing here. */
2804 } else { /* we're the client */
2805 /* Ignore all incoming acknowledgements for calls in DALLY state */
2806 if (call && (call->state == RX_STATE_DALLY)
2807 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2808 rx_MutexIncrement(rx_stats.ignorePacketDally, rx_stats_mutex);
2809 #ifdef RX_ENABLE_LOCKS
2811 MUTEX_EXIT(&call->lock);
2814 MUTEX_ENTER(&conn->conn_data_lock);
2816 MUTEX_EXIT(&conn->conn_data_lock);
2820 /* Ignore anything that's not relevant to the current call. If there
2821 * isn't a current call, then no packet is relevant. */
2822 if (!call || (np->header.callNumber != currentCallNumber)) {
2823 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2824 #ifdef RX_ENABLE_LOCKS
2826 MUTEX_EXIT(&call->lock);
2829 MUTEX_ENTER(&conn->conn_data_lock);
2831 MUTEX_EXIT(&conn->conn_data_lock);
2834 /* If the service security object index stamped in the packet does not
2835 * match the connection's security index, ignore the packet */
2836 if (np->header.securityIndex != conn->securityIndex) {
2837 #ifdef RX_ENABLE_LOCKS
2838 MUTEX_EXIT(&call->lock);
2840 MUTEX_ENTER(&conn->conn_data_lock);
2842 MUTEX_EXIT(&conn->conn_data_lock);
2846 /* If we're receiving the response, then all transmit packets are
2847 * implicitly acknowledged. Get rid of them. */
2848 if (np->header.type == RX_PACKET_TYPE_DATA) {
2849 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2850 /* XXX Hack. Because we must release the global rx lock when
2851 * sending packets (osi_NetSend) we drop all acks while we're
2852 * traversing the tq in rxi_Start sending packets out because
2853 * packets may move to the freePacketQueue as result of being here!
2854 * So we drop these packets until we're safely out of the
2855 * traversing. Really ugly!
2856 * For fine grain RX locking, we set the acked field in the
2857 * packets and let rxi_Start remove them from the transmit queue.
2859 if (call->flags & RX_CALL_TQ_BUSY) {
2860 #ifdef RX_ENABLE_LOCKS
2861 rxi_SetAcksInTransmitQueue(call);
2864 return np; /* xmitting; drop packet */
2867 rxi_ClearTransmitQueue(call, 0);
2869 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2870 rxi_ClearTransmitQueue(call, 0);
2871 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2873 if (np->header.type == RX_PACKET_TYPE_ACK) {
2874 /* now check to see if this is an ack packet acknowledging that the
2875 * server actually *lost* some hard-acked data. If this happens we
2876 * ignore this packet, as it may indicate that the server restarted in
2877 * the middle of a call. It is also possible that this is an old ack
2878 * packet. We don't abort the connection in this case, because this
2879 * *might* just be an old ack packet. The right way to detect a server
2880 * restart in the midst of a call is to notice that the server epoch
2882 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2883 * XXX unacknowledged. I think that this is off-by-one, but
2884 * XXX I don't dare change it just yet, since it will
2885 * XXX interact badly with the server-restart detection
2886 * XXX code in receiveackpacket. */
2887 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2888 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2889 MUTEX_EXIT(&call->lock);
2890 MUTEX_ENTER(&conn->conn_data_lock);
2892 MUTEX_EXIT(&conn->conn_data_lock);
2896 } /* else not a data packet */
2899 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2900 /* Set remote user defined status from packet */
2901 call->remoteStatus = np->header.userStatus;
2903 /* Note the gap between the expected next packet and the actual
2904 * packet that arrived, when the new packet has a smaller serial number
2905 * than expected. Rioses frequently reorder packets all by themselves,
2906 * so this will be quite important with very large window sizes.
2907 * Skew is checked against 0 here to avoid any dependence on the type of
2908 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2910 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2911 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2912 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2914 MUTEX_ENTER(&conn->conn_data_lock);
2915 skew = conn->lastSerial - np->header.serial;
2916 conn->lastSerial = np->header.serial;
2917 MUTEX_EXIT(&conn->conn_data_lock);
2919 register struct rx_peer *peer;
2921 if (skew > peer->inPacketSkew) {
2922 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2924 peer->inPacketSkew = skew;
2928 /* Now do packet type-specific processing */
2929 switch (np->header.type) {
2930 case RX_PACKET_TYPE_DATA:
2931 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2934 case RX_PACKET_TYPE_ACK:
2935 /* Respond immediately to ack packets requesting acknowledgement
2937 if (np->header.flags & RX_REQUEST_ACK) {
2939 (void)rxi_SendCallAbort(call, 0, 1, 0);
2941 (void)rxi_SendAck(call, 0, np->header.serial,
2942 RX_ACK_PING_RESPONSE, 1);
2944 np = rxi_ReceiveAckPacket(call, np, 1);
2946 case RX_PACKET_TYPE_ABORT: {
2947 /* An abort packet: reset the call, passing the error up to the user. */
2948 /* What if error is zero? */
2949 /* What if the error is -1? the application will treat it as a timeout. */
2950 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
2951 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
2952 rxi_CallError(call, errdata);
2953 MUTEX_EXIT(&call->lock);
2954 MUTEX_ENTER(&conn->conn_data_lock);
2956 MUTEX_EXIT(&conn->conn_data_lock);
2957 return np; /* xmitting; drop packet */
2959 case RX_PACKET_TYPE_BUSY:
2962 case RX_PACKET_TYPE_ACKALL:
2963 /* All packets acknowledged, so we can drop all packets previously
2964 * readied for sending */
2965 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2966 /* XXX Hack. We because we can't release the global rx lock when
2967 * sending packets (osi_NetSend) we drop all ack pkts while we're
2968 * traversing the tq in rxi_Start sending packets out because
2969 * packets may move to the freePacketQueue as result of being
2970 * here! So we drop these packets until we're safely out of the
2971 * traversing. Really ugly!
2972 * For fine grain RX locking, we set the acked field in the packets
2973 * and let rxi_Start remove the packets from the transmit queue.
2975 if (call->flags & RX_CALL_TQ_BUSY) {
2976 #ifdef RX_ENABLE_LOCKS
2977 rxi_SetAcksInTransmitQueue(call);
2979 #else /* RX_ENABLE_LOCKS */
2980 MUTEX_EXIT(&call->lock);
2981 MUTEX_ENTER(&conn->conn_data_lock);
2983 MUTEX_EXIT(&conn->conn_data_lock);
2984 return np; /* xmitting; drop packet */
2985 #endif /* RX_ENABLE_LOCKS */
2987 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2988 rxi_ClearTransmitQueue(call, 0);
2989 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
2992 /* Should not reach here, unless the peer is broken: send an abort
2994 rxi_CallError(call, RX_PROTOCOL_ERROR);
2995 np = rxi_SendCallAbort(call, np, 1, 0);
2998 /* Note when this last legitimate packet was received, for keep-alive
2999 * processing. Note, we delay getting the time until now in the hope that
3000 * the packet will be delivered to the user before any get time is required
3001 * (if not, then the time won't actually be re-evaluated here). */
3002 call->lastReceiveTime = clock_Sec();
3003 MUTEX_EXIT(&call->lock);
3004 MUTEX_ENTER(&conn->conn_data_lock);
3006 MUTEX_EXIT(&conn->conn_data_lock);
3010 /* return true if this is an "interesting" connection from the point of view
3011 of someone trying to debug the system */
3013 rxi_IsConnInteresting(struct rx_connection *aconn)
3016 register struct rx_call *tcall;
3018 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3020 for (i = 0; i < RX_MAXCALLS; i++) {
3021 tcall = aconn->call[i];
3023 if ((tcall->state == RX_STATE_PRECALL)
3024 || (tcall->state == RX_STATE_ACTIVE))
3026 if ((tcall->mode == RX_MODE_SENDING)
3027 || (tcall->mode == RX_MODE_RECEIVING))
3035 /* if this is one of the last few packets AND it wouldn't be used by the
3036 receiving call to immediately satisfy a read request, then drop it on
3037 the floor, since accepting it might prevent a lock-holding thread from
3038 making progress in its reading. If a call has been cleared while in
3039 the precall state then ignore all subsequent packets until the call
3040 is assigned to a thread. */
3043 TooLow(struct rx_packet *ap, struct rx_call *acall)
3046 MUTEX_ENTER(&rx_stats_mutex);
3047 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3048 && (acall->state == RX_STATE_PRECALL))
3049 || ((rx_nFreePackets < rxi_dataQuota + 2)
3050 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3051 && (acall->flags & RX_CALL_READER_WAIT)))) {
3054 MUTEX_EXIT(&rx_stats_mutex);
3060 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
3062 struct rx_connection *conn = arg1;
3063 struct rx_call *acall = arg2;
3064 struct rx_call *call = acall;
3065 struct clock when, now;
3068 MUTEX_ENTER(&conn->conn_data_lock);
3069 conn->checkReachEvent = NULL;
3070 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3073 MUTEX_EXIT(&conn->conn_data_lock);
3077 MUTEX_ENTER(&conn->conn_call_lock);
3078 MUTEX_ENTER(&conn->conn_data_lock);
3079 for (i = 0; i < RX_MAXCALLS; i++) {
3080 struct rx_call *tc = conn->call[i];
3081 if (tc && tc->state == RX_STATE_PRECALL) {
3087 /* Indicate that rxi_CheckReachEvent is no longer running by
3088 * clearing the flag. Must be atomic under conn_data_lock to
3089 * avoid a new call slipping by: rxi_CheckConnReach holds
3090 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3092 conn->flags &= ~RX_CONN_ATTACHWAIT;
3093 MUTEX_EXIT(&conn->conn_data_lock);
3094 MUTEX_EXIT(&conn->conn_call_lock);
3099 MUTEX_ENTER(&call->lock);
3100 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3102 MUTEX_EXIT(&call->lock);
3104 clock_GetTime(&now);
3106 when.sec += RX_CHECKREACH_TIMEOUT;
3107 MUTEX_ENTER(&conn->conn_data_lock);
3108 if (!conn->checkReachEvent) {
3110 conn->checkReachEvent =
3111 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3114 MUTEX_EXIT(&conn->conn_data_lock);
3120 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3122 struct rx_service *service = conn->service;
3123 struct rx_peer *peer = conn->peer;
3124 afs_uint32 now, lastReach;
3126 if (service->checkReach == 0)
3130 MUTEX_ENTER(&peer->peer_lock);
3131 lastReach = peer->lastReachTime;
3132 MUTEX_EXIT(&peer->peer_lock);
3133 if (now - lastReach < RX_CHECKREACH_TTL)
3136 MUTEX_ENTER(&conn->conn_data_lock);
3137 if (conn->flags & RX_CONN_ATTACHWAIT) {
3138 MUTEX_EXIT(&conn->conn_data_lock);
3141 conn->flags |= RX_CONN_ATTACHWAIT;
3142 MUTEX_EXIT(&conn->conn_data_lock);
3143 if (!conn->checkReachEvent)
3144 rxi_CheckReachEvent(NULL, conn, call);
3149 /* try to attach call, if authentication is complete */
3151 TryAttach(register struct rx_call *acall, register osi_socket socket,
3152 register int *tnop, register struct rx_call **newcallp,
3155 struct rx_connection *conn = acall->conn;
3157 if (conn->type == RX_SERVER_CONNECTION
3158 && acall->state == RX_STATE_PRECALL) {
3159 /* Don't attach until we have any req'd. authentication. */
3160 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3161 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3162 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3163 /* Note: this does not necessarily succeed; there
3164 * may not any proc available
3167 rxi_ChallengeOn(acall->conn);
3172 /* A data packet has been received off the interface. This packet is
3173 * appropriate to the call (the call is in the right state, etc.). This
3174 * routine can return a packet to the caller, for re-use */
3177 rxi_ReceiveDataPacket(register struct rx_call *call,
3178 register struct rx_packet *np, int istack,
3179 osi_socket socket, afs_uint32 host, u_short port,
3180 int *tnop, struct rx_call **newcallp)
3182 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3186 afs_uint32 seq, serial, flags;
3188 struct rx_packet *tnp;
3189 struct clock when, now;
3190 rx_MutexIncrement(rx_stats.dataPacketsRead, rx_stats_mutex);
3193 /* If there are no packet buffers, drop this new packet, unless we can find
3194 * packet buffers from inactive calls */
3196 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3197 MUTEX_ENTER(&rx_freePktQ_lock);
3198 rxi_NeedMorePackets = TRUE;
3199 MUTEX_EXIT(&rx_freePktQ_lock);
3200 rx_MutexIncrement(rx_stats.noPacketBuffersOnRead, rx_stats_mutex);
3201 call->rprev = np->header.serial;
3202 rxi_calltrace(RX_TRACE_DROP, call);
3203 dpf(("packet %x dropped on receipt - quota problems", np));
3205 rxi_ClearReceiveQueue(call);
3206 clock_GetTime(&now);
3208 clock_Add(&when, &rx_softAckDelay);
3209 if (!call->delayedAckEvent
3210 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3211 rxevent_Cancel(call->delayedAckEvent, call,
3212 RX_CALL_REFCOUNT_DELAY);
3213 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3214 call->delayedAckEvent =
3215 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3217 /* we've damaged this call already, might as well do it in. */
3223 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3224 * packet is one of several packets transmitted as a single
3225 * datagram. Do not send any soft or hard acks until all packets
3226 * in a jumbogram have been processed. Send negative acks right away.
3228 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3229 /* tnp is non-null when there are more packets in the
3230 * current jumbo gram */
3237 seq = np->header.seq;
3238 serial = np->header.serial;
3239 flags = np->header.flags;
3241 /* If the call is in an error state, send an abort message */
3243 return rxi_SendCallAbort(call, np, istack, 0);
3245 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3246 * AFS 3.5 jumbogram. */
3247 if (flags & RX_JUMBO_PACKET) {
3248 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3253 if (np->header.spare != 0) {
3254 MUTEX_ENTER(&call->conn->conn_data_lock);
3255 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3256 MUTEX_EXIT(&call->conn->conn_data_lock);
3259 /* The usual case is that this is the expected next packet */
3260 if (seq == call->rnext) {
3262 /* Check to make sure it is not a duplicate of one already queued */
3263 if (queue_IsNotEmpty(&call->rq)
3264 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3265 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3266 dpf(("packet %x dropped on receipt - duplicate", np));
3267 rxevent_Cancel(call->delayedAckEvent, call,
3268 RX_CALL_REFCOUNT_DELAY);
3269 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3275 /* It's the next packet. Stick it on the receive queue
3276 * for this call. Set newPackets to make sure we wake
3277 * the reader once all packets have been processed */
3278 queue_Prepend(&call->rq, np);
3280 np = NULL; /* We can't use this anymore */
3283 /* If an ack is requested then set a flag to make sure we
3284 * send an acknowledgement for this packet */
3285 if (flags & RX_REQUEST_ACK) {
3286 ackNeeded = RX_ACK_REQUESTED;
3289 /* Keep track of whether we have received the last packet */
3290 if (flags & RX_LAST_PACKET) {
3291 call->flags |= RX_CALL_HAVE_LAST;
3295 /* Check whether we have all of the packets for this call */
3296 if (call->flags & RX_CALL_HAVE_LAST) {
3297 afs_uint32 tseq; /* temporary sequence number */
3298 struct rx_packet *tp; /* Temporary packet pointer */
3299 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3301 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3302 if (tseq != tp->header.seq)
3304 if (tp->header.flags & RX_LAST_PACKET) {
3305 call->flags |= RX_CALL_RECEIVE_DONE;
3312 /* Provide asynchronous notification for those who want it
3313 * (e.g. multi rx) */
3314 if (call->arrivalProc) {
3315 (*call->arrivalProc) (call, call->arrivalProcHandle,
3316 call->arrivalProcArg);
3317 call->arrivalProc = (void (*)())0;
3320 /* Update last packet received */
3323 /* If there is no server process serving this call, grab
3324 * one, if available. We only need to do this once. If a
3325 * server thread is available, this thread becomes a server
3326 * thread and the server thread becomes a listener thread. */
3328 TryAttach(call, socket, tnop, newcallp, 0);
3331 /* This is not the expected next packet. */
3333 /* Determine whether this is a new or old packet, and if it's
3334 * a new one, whether it fits into the current receive window.
3335 * Also figure out whether the packet was delivered in sequence.
3336 * We use the prev variable to determine whether the new packet
3337 * is the successor of its immediate predecessor in the
3338 * receive queue, and the missing flag to determine whether
3339 * any of this packets predecessors are missing. */
3341 afs_uint32 prev; /* "Previous packet" sequence number */
3342 struct rx_packet *tp; /* Temporary packet pointer */
3343 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3344 int missing; /* Are any predecessors missing? */
3346 /* If the new packet's sequence number has been sent to the
3347 * application already, then this is a duplicate */
3348 if (seq < call->rnext) {
3349 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3350 rxevent_Cancel(call->delayedAckEvent, call,
3351 RX_CALL_REFCOUNT_DELAY);
3352 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3358 /* If the sequence number is greater than what can be
3359 * accomodated by the current window, then send a negative
3360 * acknowledge and drop the packet */
3361 if ((call->rnext + call->rwind) <= seq) {
3362 rxevent_Cancel(call->delayedAckEvent, call,
3363 RX_CALL_REFCOUNT_DELAY);
3364 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3371 /* Look for the packet in the queue of old received packets */
3372 for (prev = call->rnext - 1, missing =
3373 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3374 /*Check for duplicate packet */
3375 if (seq == tp->header.seq) {
3376 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3377 rxevent_Cancel(call->delayedAckEvent, call,
3378 RX_CALL_REFCOUNT_DELAY);
3379 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3385 /* If we find a higher sequence packet, break out and
3386 * insert the new packet here. */
3387 if (seq < tp->header.seq)
3389 /* Check for missing packet */
3390 if (tp->header.seq != prev + 1) {
3394 prev = tp->header.seq;
3397 /* Keep track of whether we have received the last packet. */
3398 if (flags & RX_LAST_PACKET) {
3399 call->flags |= RX_CALL_HAVE_LAST;
3402 /* It's within the window: add it to the the receive queue.
3403 * tp is left by the previous loop either pointing at the
3404 * packet before which to insert the new packet, or at the
3405 * queue head if the queue is empty or the packet should be
3407 queue_InsertBefore(tp, np);
3411 /* Check whether we have all of the packets for this call */
3412 if ((call->flags & RX_CALL_HAVE_LAST)
3413 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3414 afs_uint32 tseq; /* temporary sequence number */
3417 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3418 if (tseq != tp->header.seq)
3420 if (tp->header.flags & RX_LAST_PACKET) {
3421 call->flags |= RX_CALL_RECEIVE_DONE;
3428 /* We need to send an ack of the packet is out of sequence,
3429 * or if an ack was requested by the peer. */
3430 if (seq != prev + 1 || missing) {
3431 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3432 } else if (flags & RX_REQUEST_ACK) {
3433 ackNeeded = RX_ACK_REQUESTED;
3436 /* Acknowledge the last packet for each call */
3437 if (flags & RX_LAST_PACKET) {
3448 * If the receiver is waiting for an iovec, fill the iovec
3449 * using the data from the receive queue */
3450 if (call->flags & RX_CALL_IOVEC_WAIT) {
3451 didHardAck = rxi_FillReadVec(call, serial);
3452 /* the call may have been aborted */
3461 /* Wakeup the reader if any */
3462 if ((call->flags & RX_CALL_READER_WAIT)
3463 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3464 || (call->iovNext >= call->iovMax)
3465 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3466 call->flags &= ~RX_CALL_READER_WAIT;
3467 #ifdef RX_ENABLE_LOCKS
3468 CV_BROADCAST(&call->cv_rq);
3470 osi_rxWakeup(&call->rq);
3476 * Send an ack when requested by the peer, or once every
3477 * rxi_SoftAckRate packets until the last packet has been
3478 * received. Always send a soft ack for the last packet in
3479 * the server's reply. */
3481 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3482 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3483 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3484 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3485 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3486 } else if (call->nSoftAcks) {
3487 clock_GetTime(&now);
3489 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3490 clock_Add(&when, &rx_lastAckDelay);
3492 clock_Add(&when, &rx_softAckDelay);
3494 if (!call->delayedAckEvent
3495 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3496 rxevent_Cancel(call->delayedAckEvent, call,
3497 RX_CALL_REFCOUNT_DELAY);
3498 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3499 call->delayedAckEvent =
3500 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3502 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3503 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3510 static void rxi_ComputeRate();
3514 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3516 struct rx_peer *peer = conn->peer;
3518 MUTEX_ENTER(&peer->peer_lock);
3519 peer->lastReachTime = clock_Sec();
3520 MUTEX_EXIT(&peer->peer_lock);
3522 MUTEX_ENTER(&conn->conn_data_lock);
3523 if (conn->flags & RX_CONN_ATTACHWAIT) {
3526 conn->flags &= ~RX_CONN_ATTACHWAIT;
3527 MUTEX_EXIT(&conn->conn_data_lock);
3529 for (i = 0; i < RX_MAXCALLS; i++) {
3530 struct rx_call *call = conn->call[i];
3533 MUTEX_ENTER(&call->lock);
3534 /* tnop can be null if newcallp is null */
3535 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3537 MUTEX_EXIT(&call->lock);
3541 MUTEX_EXIT(&conn->conn_data_lock);
3545 rx_ack_reason(int reason)
3548 case RX_ACK_REQUESTED:
3550 case RX_ACK_DUPLICATE:
3552 case RX_ACK_OUT_OF_SEQUENCE:
3554 case RX_ACK_EXCEEDS_WINDOW:
3556 case RX_ACK_NOSPACE:
3560 case RX_ACK_PING_RESPONSE:
3572 /* rxi_ComputePeerNetStats
3574 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3575 * estimates (like RTT and throughput) based on ack packets. Caller
3576 * must ensure that the packet in question is the right one (i.e.
3577 * serial number matches).
3580 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3581 struct rx_ackPacket *ap, struct rx_packet *np)
3583 struct rx_peer *peer = call->conn->peer;
3585 /* Use RTT if not delayed by client. */
3586 if (ap->reason != RX_ACK_DELAY)
3587 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3589 rxi_ComputeRate(peer, call, p, np, ap->reason);
3593 /* The real smarts of the whole thing. */
3595 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3598 struct rx_ackPacket *ap;
3600 register struct rx_packet *tp;
3601 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3602 register struct rx_connection *conn = call->conn;
3603 struct rx_peer *peer = conn->peer;
3606 /* because there are CM's that are bogus, sending weird values for this. */
3607 afs_uint32 skew = 0;
3612 int newAckCount = 0;
3613 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3614 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3616 rx_MutexIncrement(rx_stats.ackPacketsRead, rx_stats_mutex);
3617 ap = (struct rx_ackPacket *)rx_DataOf(np);
3618 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3620 return np; /* truncated ack packet */
3622 /* depends on ack packet struct */
3623 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3624 first = ntohl(ap->firstPacket);
3625 serial = ntohl(ap->serial);
3626 /* temporarily disabled -- needs to degrade over time
3627 * skew = ntohs(ap->maxSkew); */
3629 /* Ignore ack packets received out of order */
3630 if (first < call->tfirst) {
3634 if (np->header.flags & RX_SLOW_START_OK) {
3635 call->flags |= RX_CALL_SLOW_START_OK;
3638 if (ap->reason == RX_ACK_PING_RESPONSE)
3639 rxi_UpdatePeerReach(conn, call);
3643 if (rxdebug_active) {
3647 len = _snprintf(msg, sizeof(msg),
3648 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3649 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3650 ntohl(ap->serial), ntohl(ap->previousPacket),
3651 (unsigned int)np->header.seq, (unsigned int)skew,
3652 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3656 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3657 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3661 OutputDebugString(msg);
3663 #else /* AFS_NT40_ENV */
3666 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3667 ap->reason, ntohl(ap->previousPacket),
3668 (unsigned int)np->header.seq, (unsigned int)serial,
3669 (unsigned int)skew, ntohl(ap->firstPacket));
3672 for (offset = 0; offset < nAcks; offset++)
3673 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3678 #endif /* AFS_NT40_ENV */
3681 /* Update the outgoing packet skew value to the latest value of
3682 * the peer's incoming packet skew value. The ack packet, of
3683 * course, could arrive out of order, but that won't affect things
3685 MUTEX_ENTER(&peer->peer_lock);
3686 peer->outPacketSkew = skew;
3688 /* Check for packets that no longer need to be transmitted, and
3689 * discard them. This only applies to packets positively
3690 * acknowledged as having been sent to the peer's upper level.
3691 * All other packets must be retained. So only packets with
3692 * sequence numbers < ap->firstPacket are candidates. */
3693 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3694 if (tp->header.seq >= first)
3696 call->tfirst = tp->header.seq + 1;
3698 && (tp->header.serial == serial || tp->firstSerial == serial))
3699 rxi_ComputePeerNetStats(call, tp, ap, np);
3700 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3703 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3704 /* XXX Hack. Because we have to release the global rx lock when sending
3705 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3706 * in rxi_Start sending packets out because packets may move to the
3707 * freePacketQueue as result of being here! So we drop these packets until
3708 * we're safely out of the traversing. Really ugly!
3709 * To make it even uglier, if we're using fine grain locking, we can
3710 * set the ack bits in the packets and have rxi_Start remove the packets
3711 * when it's done transmitting.
3713 if (call->flags & RX_CALL_TQ_BUSY) {
3714 #ifdef RX_ENABLE_LOCKS
3715 tp->flags |= RX_PKTFLAG_ACKED;
3716 call->flags |= RX_CALL_TQ_SOME_ACKED;
3717 #else /* RX_ENABLE_LOCKS */
3719 #endif /* RX_ENABLE_LOCKS */
3721 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3724 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3729 /* Give rate detector a chance to respond to ping requests */
3730 if (ap->reason == RX_ACK_PING_RESPONSE) {
3731 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3735 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3737 /* Now go through explicit acks/nacks and record the results in
3738 * the waiting packets. These are packets that can't be released
3739 * yet, even with a positive acknowledge. This positive
3740 * acknowledge only means the packet has been received by the
3741 * peer, not that it will be retained long enough to be sent to
3742 * the peer's upper level. In addition, reset the transmit timers
3743 * of any missing packets (those packets that must be missing
3744 * because this packet was out of sequence) */
3746 call->nSoftAcked = 0;
3747 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3748 /* Update round trip time if the ack was stimulated on receipt
3750 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3751 #ifdef RX_ENABLE_LOCKS
3752 if (tp->header.seq >= first)
3753 #endif /* RX_ENABLE_LOCKS */
3754 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3756 && (tp->header.serial == serial || tp->firstSerial == serial))
3757 rxi_ComputePeerNetStats(call, tp, ap, np);
3759 /* Set the acknowledge flag per packet based on the
3760 * information in the ack packet. An acknowlegded packet can
3761 * be downgraded when the server has discarded a packet it
3762 * soacked previously, or when an ack packet is received
3763 * out of sequence. */
3764 if (tp->header.seq < first) {
3765 /* Implicit ack information */
3766 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3769 tp->flags |= RX_PKTFLAG_ACKED;
3770 } else if (tp->header.seq < first + nAcks) {
3771 /* Explicit ack information: set it in the packet appropriately */
3772 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3773 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3775 tp->flags |= RX_PKTFLAG_ACKED;
3782 } else /* RX_ACK_TYPE_NACK */ {
3783 tp->flags &= ~RX_PKTFLAG_ACKED;
3787 tp->flags &= ~RX_PKTFLAG_ACKED;
3791 /* If packet isn't yet acked, and it has been transmitted at least
3792 * once, reset retransmit time using latest timeout
3793 * ie, this should readjust the retransmit timer for all outstanding
3794 * packets... So we don't just retransmit when we should know better*/
3796 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3797 tp->retryTime = tp->timeSent;
3798 clock_Add(&tp->retryTime, &peer->timeout);
3799 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3800 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3804 /* If the window has been extended by this acknowledge packet,
3805 * then wakeup a sender waiting in alloc for window space, or try
3806 * sending packets now, if he's been sitting on packets due to
3807 * lack of window space */
3808 if (call->tnext < (call->tfirst + call->twind)) {
3809 #ifdef RX_ENABLE_LOCKS
3810 CV_SIGNAL(&call->cv_twind);
3812 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3813 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3814 osi_rxWakeup(&call->twind);
3817 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3818 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3822 /* if the ack packet has a receivelen field hanging off it,
3823 * update our state */
3824 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3827 /* If the ack packet has a "recommended" size that is less than
3828 * what I am using now, reduce my size to match */
3829 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3830 (int)sizeof(afs_int32), &tSize);
3831 tSize = (afs_uint32) ntohl(tSize);
3832 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3834 /* Get the maximum packet size to send to this peer */
3835 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
3837 tSize = (afs_uint32) ntohl(tSize);
3838 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3839 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3841 /* sanity check - peer might have restarted with different params.
3842 * If peer says "send less", dammit, send less... Peer should never
3843 * be unable to accept packets of the size that prior AFS versions would
3844 * send without asking. */
3845 if (peer->maxMTU != tSize) {
3846 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
3848 peer->maxMTU = tSize;
3849 peer->MTU = MIN(tSize, peer->MTU);
3850 call->MTU = MIN(call->MTU, tSize);
3853 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3856 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3857 (int)sizeof(afs_int32), &tSize);
3858 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3859 if (tSize < call->twind) { /* smaller than our send */
3860 call->twind = tSize; /* window, we must send less... */
3861 call->ssthresh = MIN(call->twind, call->ssthresh);
3862 call->conn->twind[call->channel] = call->twind;
3865 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3866 * network MTU confused with the loopback MTU. Calculate the
3867 * maximum MTU here for use in the slow start code below.
3869 maxMTU = peer->maxMTU;
3870 /* Did peer restart with older RX version? */
3871 if (peer->maxDgramPackets > 1) {
3872 peer->maxDgramPackets = 1;
3874 } else if (np->length >=
3875 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3878 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3879 sizeof(afs_int32), &tSize);
3880 tSize = (afs_uint32) ntohl(tSize);
3882 * As of AFS 3.5 we set the send window to match the receive window.
3884 if (tSize < call->twind) {
3885 call->twind = tSize;
3886 call->conn->twind[call->channel] = call->twind;
3887 call->ssthresh = MIN(call->twind, call->ssthresh);
3888 } else if (tSize > call->twind) {
3889 call->twind = tSize;
3890 call->conn->twind[call->channel] = call->twind;
3894 * As of AFS 3.5, a jumbogram is more than one fixed size
3895 * packet transmitted in a single UDP datagram. If the remote
3896 * MTU is smaller than our local MTU then never send a datagram
3897 * larger than the natural MTU.
3900 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3901 sizeof(afs_int32), &tSize);
3902 maxDgramPackets = (afs_uint32) ntohl(tSize);
3903 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3905 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
3906 maxDgramPackets = MIN(maxDgramPackets, tSize);
3907 if (maxDgramPackets > 1) {
3908 peer->maxDgramPackets = maxDgramPackets;
3909 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3911 peer->maxDgramPackets = 1;
3912 call->MTU = peer->natMTU;
3914 } else if (peer->maxDgramPackets > 1) {
3915 /* Restarted with lower version of RX */
3916 peer->maxDgramPackets = 1;
3918 } else if (peer->maxDgramPackets > 1
3919 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3920 /* Restarted with lower version of RX */
3921 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3922 peer->natMTU = OLD_MAX_PACKET_SIZE;
3923 peer->MTU = OLD_MAX_PACKET_SIZE;
3924 peer->maxDgramPackets = 1;
3925 peer->nDgramPackets = 1;
3927 call->MTU = OLD_MAX_PACKET_SIZE;
3932 * Calculate how many datagrams were successfully received after
3933 * the first missing packet and adjust the negative ack counter
3938 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3939 if (call->nNacks < nNacked) {
3940 call->nNacks = nNacked;
3943 call->nAcks += newAckCount;
3947 if (call->flags & RX_CALL_FAST_RECOVER) {
3949 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3951 call->flags &= ~RX_CALL_FAST_RECOVER;
3952 call->cwind = call->nextCwind;
3953 call->nextCwind = 0;
3956 call->nCwindAcks = 0;
3957 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3958 /* Three negative acks in a row trigger congestion recovery */
3959 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3960 MUTEX_EXIT(&peer->peer_lock);
3961 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3962 /* someone else is waiting to start recovery */
3965 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3966 rxi_WaitforTQBusy(call);
3967 MUTEX_ENTER(&peer->peer_lock);
3968 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3969 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3970 call->flags |= RX_CALL_FAST_RECOVER;
3971 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3973 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3974 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3975 call->nextCwind = call->ssthresh;
3978 peer->MTU = call->MTU;
3979 peer->cwind = call->nextCwind;
3980 peer->nDgramPackets = call->nDgramPackets;
3982 call->congestSeq = peer->congestSeq;
3983 /* Reset the resend times on the packets that were nacked
3984 * so we will retransmit as soon as the window permits*/
3985 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3987 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3988 clock_Zero(&tp->retryTime);
3990 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3995 /* If cwind is smaller than ssthresh, then increase
3996 * the window one packet for each ack we receive (exponential
3998 * If cwind is greater than or equal to ssthresh then increase
3999 * the congestion window by one packet for each cwind acks we
4000 * receive (linear growth). */
4001 if (call->cwind < call->ssthresh) {
4003 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4004 call->nCwindAcks = 0;
4006 call->nCwindAcks += newAckCount;
4007 if (call->nCwindAcks >= call->cwind) {
4008 call->nCwindAcks = 0;
4009 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4013 * If we have received several acknowledgements in a row then
4014 * it is time to increase the size of our datagrams
4016 if ((int)call->nAcks > rx_nDgramThreshold) {
4017 if (peer->maxDgramPackets > 1) {
4018 if (call->nDgramPackets < peer->maxDgramPackets) {
4019 call->nDgramPackets++;
4021 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4022 } else if (call->MTU < peer->maxMTU) {
4023 call->MTU += peer->natMTU;
4024 call->MTU = MIN(call->MTU, peer->maxMTU);
4030 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4032 /* Servers need to hold the call until all response packets have
4033 * been acknowledged. Soft acks are good enough since clients
4034 * are not allowed to clear their receive queues. */
4035 if (call->state == RX_STATE_HOLD
4036 && call->tfirst + call->nSoftAcked >= call->tnext) {
4037 call->state = RX_STATE_DALLY;
4038 rxi_ClearTransmitQueue(call, 0);
4039 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4040 } else if (!queue_IsEmpty(&call->tq)) {
4041 rxi_Start(0, call, 0, istack);
4046 /* Received a response to a challenge packet */
4048 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
4049 register struct rx_packet *np, int istack)
4053 /* Ignore the packet if we're the client */
4054 if (conn->type == RX_CLIENT_CONNECTION)
4057 /* If already authenticated, ignore the packet (it's probably a retry) */
4058 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4061 /* Otherwise, have the security object evaluate the response packet */
4062 error = RXS_CheckResponse(conn->securityObject, conn, np);
4064 /* If the response is invalid, reset the connection, sending
4065 * an abort to the peer */
4069 rxi_ConnectionError(conn, error);
4070 MUTEX_ENTER(&conn->conn_data_lock);
4071 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4072 MUTEX_EXIT(&conn->conn_data_lock);
4075 /* If the response is valid, any calls waiting to attach
4076 * servers can now do so */
4079 for (i = 0; i < RX_MAXCALLS; i++) {
4080 struct rx_call *call = conn->call[i];
4082 MUTEX_ENTER(&call->lock);
4083 if (call->state == RX_STATE_PRECALL)
4084 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4085 /* tnop can be null if newcallp is null */
4086 MUTEX_EXIT(&call->lock);
4090 /* Update the peer reachability information, just in case
4091 * some calls went into attach-wait while we were waiting
4092 * for authentication..
4094 rxi_UpdatePeerReach(conn, NULL);
4099 /* A client has received an authentication challenge: the security
4100 * object is asked to cough up a respectable response packet to send
4101 * back to the server. The server is responsible for retrying the
4102 * challenge if it fails to get a response. */
4105 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
4106 register struct rx_packet *np, int istack)
4110 /* Ignore the challenge if we're the server */
4111 if (conn->type == RX_SERVER_CONNECTION)
4114 /* Ignore the challenge if the connection is otherwise idle; someone's
4115 * trying to use us as an oracle. */
4116 if (!rxi_HasActiveCalls(conn))
4119 /* Send the security object the challenge packet. It is expected to fill
4120 * in the response. */
4121 error = RXS_GetResponse(conn->securityObject, conn, np);
4123 /* If the security object is unable to return a valid response, reset the
4124 * connection and send an abort to the peer. Otherwise send the response
4125 * packet to the peer connection. */
4127 rxi_ConnectionError(conn, error);
4128 MUTEX_ENTER(&conn->conn_data_lock);
4129 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4130 MUTEX_EXIT(&conn->conn_data_lock);
4132 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4133 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4139 /* Find an available server process to service the current request in
4140 * the given call structure. If one isn't available, queue up this
4141 * call so it eventually gets one */
4143 rxi_AttachServerProc(register struct rx_call *call,
4144 register osi_socket socket, register int *tnop,
4145 register struct rx_call **newcallp)
4147 register struct rx_serverQueueEntry *sq;
4148 register struct rx_service *service = call->conn->service;
4149 register int haveQuota = 0;
4151 /* May already be attached */
4152 if (call->state == RX_STATE_ACTIVE)
4155 MUTEX_ENTER(&rx_serverPool_lock);
4157 haveQuota = QuotaOK(service);
4158 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4159 /* If there are no processes available to service this call,
4160 * put the call on the incoming call queue (unless it's
4161 * already on the queue).
4163 #ifdef RX_ENABLE_LOCKS
4165 ReturnToServerPool(service);
4166 #endif /* RX_ENABLE_LOCKS */
4168 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4169 call->flags |= RX_CALL_WAIT_PROC;
4170 MUTEX_ENTER(&rx_stats_mutex);
4173 MUTEX_EXIT(&rx_stats_mutex);
4174 rxi_calltrace(RX_CALL_ARRIVAL, call);
4175 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4176 queue_Append(&rx_incomingCallQueue, call);
4179 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4181 /* If hot threads are enabled, and both newcallp and sq->socketp
4182 * are non-null, then this thread will process the call, and the
4183 * idle server thread will start listening on this threads socket.
4186 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4189 *sq->socketp = socket;
4190 clock_GetTime(&call->startTime);
4191 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4195 if (call->flags & RX_CALL_WAIT_PROC) {
4196 /* Conservative: I don't think this should happen */
4197 call->flags &= ~RX_CALL_WAIT_PROC;
4198 if (queue_IsOnQueue(call)) {
4200 MUTEX_ENTER(&rx_stats_mutex);
4202 MUTEX_EXIT(&rx_stats_mutex);
4205 call->state = RX_STATE_ACTIVE;
4206 call->mode = RX_MODE_RECEIVING;
4207 #ifdef RX_KERNEL_TRACE
4209 int glockOwner = ISAFS_GLOCK();
4212 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4213 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4219 if (call->flags & RX_CALL_CLEARED) {
4220 /* send an ack now to start the packet flow up again */
4221 call->flags &= ~RX_CALL_CLEARED;
4222 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4224 #ifdef RX_ENABLE_LOCKS
4227 service->nRequestsRunning++;
4228 if (service->nRequestsRunning <= service->minProcs)
4234 MUTEX_EXIT(&rx_serverPool_lock);
4237 /* Delay the sending of an acknowledge event for a short while, while
4238 * a new call is being prepared (in the case of a client) or a reply
4239 * is being prepared (in the case of a server). Rather than sending
4240 * an ack packet, an ACKALL packet is sent. */
4242 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4244 #ifdef RX_ENABLE_LOCKS
4246 MUTEX_ENTER(&call->lock);
4247 call->delayedAckEvent = NULL;
4248 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4250 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4251 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4253 MUTEX_EXIT(&call->lock);
4254 #else /* RX_ENABLE_LOCKS */
4256 call->delayedAckEvent = NULL;
4257 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4258 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4259 #endif /* RX_ENABLE_LOCKS */
4263 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused)
4265 struct rx_call *call = arg1;
4266 #ifdef RX_ENABLE_LOCKS
4268 MUTEX_ENTER(&call->lock);
4269 if (event == call->delayedAckEvent)
4270 call->delayedAckEvent = NULL;
4271 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4273 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4275 MUTEX_EXIT(&call->lock);
4276 #else /* RX_ENABLE_LOCKS */
4278 call->delayedAckEvent = NULL;
4279 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4280 #endif /* RX_ENABLE_LOCKS */
4284 #ifdef RX_ENABLE_LOCKS
4285 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4286 * clearing them out.
4289 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4291 register struct rx_packet *p, *tp;
4294 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4295 p->flags |= RX_PKTFLAG_ACKED;
4299 call->flags |= RX_CALL_TQ_CLEARME;
4300 call->flags |= RX_CALL_TQ_SOME_ACKED;
4303 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4304 call->tfirst = call->tnext;
4305 call->nSoftAcked = 0;
4307 if (call->flags & RX_CALL_FAST_RECOVER) {
4308 call->flags &= ~RX_CALL_FAST_RECOVER;
4309 call->cwind = call->nextCwind;
4310 call->nextCwind = 0;
4313 CV_SIGNAL(&call->cv_twind);
4315 #endif /* RX_ENABLE_LOCKS */
4317 /* Clear out the transmit queue for the current call (all packets have
4318 * been received by peer) */
4320 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4322 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4323 register struct rx_packet *p, *tp;
4325 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4327 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4328 p->flags |= RX_PKTFLAG_ACKED;
4332 call->flags |= RX_CALL_TQ_CLEARME;
4333 call->flags |= RX_CALL_TQ_SOME_ACKED;
4336 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4337 rxi_FreePackets(0, &call->tq);
4338 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4339 call->flags &= ~RX_CALL_TQ_CLEARME;
4341 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4343 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4344 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4345 call->nSoftAcked = 0;
4347 if (call->flags & RX_CALL_FAST_RECOVER) {
4348 call->flags &= ~RX_CALL_FAST_RECOVER;
4349 call->cwind = call->nextCwind;
4351 #ifdef RX_ENABLE_LOCKS
4352 CV_SIGNAL(&call->cv_twind);
4354 osi_rxWakeup(&call->twind);
4359 rxi_ClearReceiveQueue(register struct rx_call *call)
4361 if (queue_IsNotEmpty(&call->rq)) {
4362 rx_packetReclaims += rxi_FreePackets(0, &call->rq);
4363 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4365 if (call->state == RX_STATE_PRECALL) {
4366 call->flags |= RX_CALL_CLEARED;
4370 /* Send an abort packet for the specified call */
4372 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4373 int istack, int force)
4376 struct clock when, now;
4381 /* Clients should never delay abort messages */
4382 if (rx_IsClientConn(call->conn))
4385 if (call->abortCode != call->error) {
4386 call->abortCode = call->error;
4387 call->abortCount = 0;
4390 if (force || rxi_callAbortThreshhold == 0
4391 || call->abortCount < rxi_callAbortThreshhold) {
4392 if (call->delayedAbortEvent) {
4393 rxevent_Cancel(call->delayedAbortEvent, call,
4394 RX_CALL_REFCOUNT_ABORT);
4396 error = htonl(call->error);
4399 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4400 (char *)&error, sizeof(error), istack);
4401 } else if (!call->delayedAbortEvent) {
4402 clock_GetTime(&now);
4404 clock_Addmsec(&when, rxi_callAbortDelay);
4405 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4406 call->delayedAbortEvent =
4407 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4412 /* Send an abort packet for the specified connection. Packet is an
4413 * optional pointer to a packet that can be used to send the abort.
4414 * Once the number of abort messages reaches the threshhold, an
4415 * event is scheduled to send the abort. Setting the force flag
4416 * overrides sending delayed abort messages.
4418 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4419 * to send the abort packet.
4422 rxi_SendConnectionAbort(register struct rx_connection *conn,
4423 struct rx_packet *packet, int istack, int force)
4426 struct clock when, now;
4431 /* Clients should never delay abort messages */
4432 if (rx_IsClientConn(conn))
4435 if (force || rxi_connAbortThreshhold == 0
4436 || conn->abortCount < rxi_connAbortThreshhold) {
4437 if (conn->delayedAbortEvent) {
4438 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4440 error = htonl(conn->error);
4442 MUTEX_EXIT(&conn->conn_data_lock);
4444 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4445 RX_PACKET_TYPE_ABORT, (char *)&error,
4446 sizeof(error), istack);
4447 MUTEX_ENTER(&conn->conn_data_lock);
4448 } else if (!conn->delayedAbortEvent) {
4449 clock_GetTime(&now);
4451 clock_Addmsec(&when, rxi_connAbortDelay);
4452 conn->delayedAbortEvent =
4453 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4458 /* Associate an error all of the calls owned by a connection. Called
4459 * with error non-zero. This is only for really fatal things, like
4460 * bad authentication responses. The connection itself is set in
4461 * error at this point, so that future packets received will be
4464 rxi_ConnectionError(register struct rx_connection *conn,
4465 register afs_int32 error)
4470 dpf(("rxi_ConnectionError conn %x error %d", conn, error));
4472 MUTEX_ENTER(&conn->conn_data_lock);
4473 if (conn->challengeEvent)
4474 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4475 if (conn->checkReachEvent) {
4476 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4477 conn->checkReachEvent = 0;
4478 conn->flags &= ~RX_CONN_ATTACHWAIT;
4481 MUTEX_EXIT(&conn->conn_data_lock);
4482 for (i = 0; i < RX_MAXCALLS; i++) {
4483 struct rx_call *call = conn->call[i];
4485 MUTEX_ENTER(&call->lock);
4486 rxi_CallError(call, error);
4487 MUTEX_EXIT(&call->lock);
4490 conn->error = error;
4491 rx_MutexIncrement(rx_stats.fatalErrors, rx_stats_mutex);
4496 rxi_CallError(register struct rx_call *call, afs_int32 error)
4498 dpf(("rxi_CallError call %x error %d call->error %d", call, error, call->error));
4500 error = call->error;
4502 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4503 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4504 rxi_ResetCall(call, 0);
4507 rxi_ResetCall(call, 0);
4509 call->error = error;
4510 call->mode = RX_MODE_ERROR;
4513 /* Reset various fields in a call structure, and wakeup waiting
4514 * processes. Some fields aren't changed: state & mode are not
4515 * touched (these must be set by the caller), and bufptr, nLeft, and
4516 * nFree are not reset, since these fields are manipulated by
4517 * unprotected macros, and may only be reset by non-interrupting code.
4520 /* this code requires that call->conn be set properly as a pre-condition. */
4521 #endif /* ADAPT_WINDOW */
4524 rxi_ResetCall(register struct rx_call *call, register int newcall)
4527 register struct rx_peer *peer;
4528 struct rx_packet *packet;
4530 dpf(("rxi_ResetCall(call %x, newcall %d)\n", call, newcall));
4532 /* Notify anyone who is waiting for asynchronous packet arrival */
4533 if (call->arrivalProc) {
4534 (*call->arrivalProc) (call, call->arrivalProcHandle,
4535 call->arrivalProcArg);
4536 call->arrivalProc = (void (*)())0;
4539 if (call->delayedAbortEvent) {
4540 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4541 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4543 rxi_SendCallAbort(call, packet, 0, 1);
4544 rxi_FreePacket(packet);
4549 * Update the peer with the congestion information in this call
4550 * so other calls on this connection can pick up where this call
4551 * left off. If the congestion sequence numbers don't match then
4552 * another call experienced a retransmission.
4554 peer = call->conn->peer;
4555 MUTEX_ENTER(&peer->peer_lock);
4557 if (call->congestSeq == peer->congestSeq) {
4558 peer->cwind = MAX(peer->cwind, call->cwind);
4559 peer->MTU = MAX(peer->MTU, call->MTU);
4560 peer->nDgramPackets =
4561 MAX(peer->nDgramPackets, call->nDgramPackets);
4564 call->abortCode = 0;
4565 call->abortCount = 0;
4567 if (peer->maxDgramPackets > 1) {
4568 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4570 call->MTU = peer->MTU;
4572 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4573 call->ssthresh = rx_maxSendWindow;
4574 call->nDgramPackets = peer->nDgramPackets;
4575 call->congestSeq = peer->congestSeq;
4576 MUTEX_EXIT(&peer->peer_lock);
4578 flags = call->flags;
4579 rxi_ClearReceiveQueue(call);
4580 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4581 if (flags & RX_CALL_TQ_BUSY) {
4582 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4583 call->flags |= (flags & RX_CALL_TQ_WAIT);
4585 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4587 rxi_ClearTransmitQueue(call, 0);
4588 queue_Init(&call->tq);
4589 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4590 dpf(("rcall %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4593 while (call->tqWaiters) {
4594 #ifdef RX_ENABLE_LOCKS
4595 CV_BROADCAST(&call->cv_tq);
4596 #else /* RX_ENABLE_LOCKS */
4597 osi_rxWakeup(&call->tq);
4598 #endif /* RX_ENABLE_LOCKS */
4602 queue_Init(&call->rq);
4604 call->twind = call->conn->twind[call->channel];
4605 call->rwind = call->conn->rwind[call->channel];
4606 call->nSoftAcked = 0;
4607 call->nextCwind = 0;
4610 call->nCwindAcks = 0;
4611 call->nSoftAcks = 0;
4612 call->nHardAcks = 0;
4614 call->tfirst = call->rnext = call->tnext = 1;
4616 call->lastAcked = 0;
4617 call->localStatus = call->remoteStatus = 0;
4619 if (flags & RX_CALL_READER_WAIT) {
4620 #ifdef RX_ENABLE_LOCKS
4621 CV_BROADCAST(&call->cv_rq);
4623 osi_rxWakeup(&call->rq);
4626 if (flags & RX_CALL_WAIT_PACKETS) {
4627 MUTEX_ENTER(&rx_freePktQ_lock);
4628 rxi_PacketsUnWait(); /* XXX */
4629 MUTEX_EXIT(&rx_freePktQ_lock);
4631 #ifdef RX_ENABLE_LOCKS
4632 CV_SIGNAL(&call->cv_twind);
4634 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4635 osi_rxWakeup(&call->twind);
4638 #ifdef RX_ENABLE_LOCKS
4639 /* The following ensures that we don't mess with any queue while some
4640 * other thread might also be doing so. The call_queue_lock field is
4641 * is only modified under the call lock. If the call is in the process
4642 * of being removed from a queue, the call is not locked until the
4643 * the queue lock is dropped and only then is the call_queue_lock field
4644 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4645 * Note that any other routine which removes a call from a queue has to
4646 * obtain the queue lock before examing the queue and removing the call.
4648 if (call->call_queue_lock) {
4649 MUTEX_ENTER(call->call_queue_lock);
4650 if (queue_IsOnQueue(call)) {
4652 if (flags & RX_CALL_WAIT_PROC) {
4653 MUTEX_ENTER(&rx_stats_mutex);
4655 MUTEX_EXIT(&rx_stats_mutex);
4658 MUTEX_EXIT(call->call_queue_lock);
4659 CLEAR_CALL_QUEUE_LOCK(call);
4661 #else /* RX_ENABLE_LOCKS */
4662 if (queue_IsOnQueue(call)) {
4664 if (flags & RX_CALL_WAIT_PROC)
4667 #endif /* RX_ENABLE_LOCKS */
4669 rxi_KeepAliveOff(call);
4670 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4673 /* Send an acknowledge for the indicated packet (seq,serial) of the
4674 * indicated call, for the indicated reason (reason). This
4675 * acknowledge will specifically acknowledge receiving the packet, and
4676 * will also specify which other packets for this call have been
4677 * received. This routine returns the packet that was used to the
4678 * caller. The caller is responsible for freeing it or re-using it.
4679 * This acknowledgement also returns the highest sequence number
4680 * actually read out by the higher level to the sender; the sender
4681 * promises to keep around packets that have not been read by the
4682 * higher level yet (unless, of course, the sender decides to abort
4683 * the call altogether). Any of p, seq, serial, pflags, or reason may
4684 * be set to zero without ill effect. That is, if they are zero, they
4685 * will not convey any information.
4686 * NOW there is a trailer field, after the ack where it will safely be
4687 * ignored by mundanes, which indicates the maximum size packet this
4688 * host can swallow. */
4690 register struct rx_packet *optionalPacket; use to send ack (or null)
4691 int seq; Sequence number of the packet we are acking
4692 int serial; Serial number of the packet
4693 int pflags; Flags field from packet header
4694 int reason; Reason an acknowledge was prompted
4698 rxi_SendAck(register struct rx_call *call,
4699 register struct rx_packet *optionalPacket, int serial, int reason,
4702 struct rx_ackPacket *ap;
4703 register struct rx_packet *rqp;
4704 register struct rx_packet *nxp; /* For queue_Scan */
4705 register struct rx_packet *p;
4708 #ifdef RX_ENABLE_TSFPQ
4709 struct rx_ts_info_t * rx_ts_info;
4713 * Open the receive window once a thread starts reading packets
4715 if (call->rnext > 1) {
4716 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
4719 call->nHardAcks = 0;
4720 call->nSoftAcks = 0;
4721 if (call->rnext > call->lastAcked)
4722 call->lastAcked = call->rnext;
4726 rx_computelen(p, p->length); /* reset length, you never know */
4727 } /* where that's been... */
4728 #ifdef RX_ENABLE_TSFPQ
4730 RX_TS_INFO_GET(rx_ts_info);
4731 if ((p = rx_ts_info->local_special_packet)) {
4732 rx_computelen(p, p->length);
4733 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4734 rx_ts_info->local_special_packet = p;
4735 } else { /* We won't send the ack, but don't panic. */
4736 return optionalPacket;
4740 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4741 /* We won't send the ack, but don't panic. */
4742 return optionalPacket;
4747 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4750 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4751 #ifndef RX_ENABLE_TSFPQ
4752 if (!optionalPacket)
4755 return optionalPacket;
4757 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4758 if (rx_Contiguous(p) < templ) {
4759 #ifndef RX_ENABLE_TSFPQ
4760 if (!optionalPacket)
4763 return optionalPacket;
4768 /* MTUXXX failing to send an ack is very serious. We should */
4769 /* try as hard as possible to send even a partial ack; it's */
4770 /* better than nothing. */
4771 ap = (struct rx_ackPacket *)rx_DataOf(p);
4772 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4773 ap->reason = reason;
4775 /* The skew computation used to be bogus, I think it's better now. */
4776 /* We should start paying attention to skew. XXX */
4777 ap->serial = htonl(serial);
4778 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4780 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4781 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4783 /* No fear of running out of ack packet here because there can only be at most
4784 * one window full of unacknowledged packets. The window size must be constrained
4785 * to be less than the maximum ack size, of course. Also, an ack should always
4786 * fit into a single packet -- it should not ever be fragmented. */
4787 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4788 if (!rqp || !call->rq.next
4789 || (rqp->header.seq > (call->rnext + call->rwind))) {
4790 #ifndef RX_ENABLE_TSFPQ
4791 if (!optionalPacket)
4794 rxi_CallError(call, RX_CALL_DEAD);
4795 return optionalPacket;
4798 while (rqp->header.seq > call->rnext + offset)
4799 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4800 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4802 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4803 #ifndef RX_ENABLE_TSFPQ
4804 if (!optionalPacket)
4807 rxi_CallError(call, RX_CALL_DEAD);
4808 return optionalPacket;
4813 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4815 /* these are new for AFS 3.3 */
4816 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4817 templ = htonl(templ);
4818 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4819 templ = htonl(call->conn->peer->ifMTU);
4820 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4821 sizeof(afs_int32), &templ);
4823 /* new for AFS 3.4 */
4824 templ = htonl(call->rwind);
4825 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4826 sizeof(afs_int32), &templ);
4828 /* new for AFS 3.5 */
4829 templ = htonl(call->conn->peer->ifDgramPackets);
4830 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4831 sizeof(afs_int32), &templ);
4833 p->header.serviceId = call->conn->serviceId;
4834 p->header.cid = (call->conn->cid | call->channel);
4835 p->header.callNumber = *call->callNumber;
4837 p->header.securityIndex = call->conn->securityIndex;
4838 p->header.epoch = call->conn->epoch;
4839 p->header.type = RX_PACKET_TYPE_ACK;
4840 p->header.flags = RX_SLOW_START_OK;
4841 if (reason == RX_ACK_PING) {
4842 p->header.flags |= RX_REQUEST_ACK;
4844 clock_GetTime(&call->pingRequestTime);
4847 if (call->conn->type == RX_CLIENT_CONNECTION)
4848 p->header.flags |= RX_CLIENT_INITIATED;
4852 if (rxdebug_active) {
4856 len = _snprintf(msg, sizeof(msg),
4857 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4858 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4859 ntohl(ap->serial), ntohl(ap->previousPacket),
4860 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
4861 ap->nAcks, ntohs(ap->bufferSpace) );
4865 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
4866 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4870 OutputDebugString(msg);
4872 #else /* AFS_NT40_ENV */
4874 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
4875 ap->reason, ntohl(ap->previousPacket),
4876 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4878 for (offset = 0; offset < ap->nAcks; offset++)
4879 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4884 #endif /* AFS_NT40_ENV */
4887 register int i, nbytes = p->length;
4889 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4890 if (nbytes <= p->wirevec[i].iov_len) {
4891 register int savelen, saven;
4893 savelen = p->wirevec[i].iov_len;
4895 p->wirevec[i].iov_len = nbytes;
4897 rxi_Send(call, p, istack);
4898 p->wirevec[i].iov_len = savelen;
4902 nbytes -= p->wirevec[i].iov_len;
4905 rx_MutexIncrement(rx_stats.ackPacketsSent, rx_stats_mutex);
4906 #ifndef RX_ENABLE_TSFPQ
4907 if (!optionalPacket)
4910 return optionalPacket; /* Return packet for re-use by caller */
4913 /* Send all of the packets in the list in single datagram */
4915 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4916 int istack, int moreFlag, struct clock *now,
4917 struct clock *retryTime, int resending)
4922 struct rx_connection *conn = call->conn;
4923 struct rx_peer *peer = conn->peer;
4925 MUTEX_ENTER(&peer->peer_lock);
4928 peer->reSends += len;
4929 rx_MutexIncrement(rx_stats.dataPacketsSent, rx_stats_mutex);
4930 MUTEX_EXIT(&peer->peer_lock);
4932 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4936 /* Set the packet flags and schedule the resend events */
4937 /* Only request an ack for the last packet in the list */
4938 for (i = 0; i < len; i++) {
4939 list[i]->retryTime = *retryTime;
4940 if (list[i]->header.serial) {
4941 /* Exponentially backoff retry times */
4942 if (list[i]->backoff < MAXBACKOFF) {
4943 /* so it can't stay == 0 */
4944 list[i]->backoff = (list[i]->backoff << 1) + 1;
4947 clock_Addmsec(&(list[i]->retryTime),
4948 ((afs_uint32) list[i]->backoff) << 8);
4951 /* Wait a little extra for the ack on the last packet */
4952 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4953 clock_Addmsec(&(list[i]->retryTime), 400);
4956 /* Record the time sent */
4957 list[i]->timeSent = *now;
4959 /* Ask for an ack on retransmitted packets, on every other packet
4960 * if the peer doesn't support slow start. Ask for an ack on every
4961 * packet until the congestion window reaches the ack rate. */
4962 if (list[i]->header.serial) {
4964 rx_MutexIncrement(rx_stats.dataPacketsReSent, rx_stats_mutex);
4966 /* improved RTO calculation- not Karn */
4967 list[i]->firstSent = *now;
4968 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4969 || (!(call->flags & RX_CALL_SLOW_START_OK)
4970 && (list[i]->header.seq & 1)))) {
4975 MUTEX_ENTER(&peer->peer_lock);
4979 rx_MutexIncrement(rx_stats.dataPacketsSent, rx_stats_mutex);
4980 MUTEX_EXIT(&peer->peer_lock);
4982 /* Tag this packet as not being the last in this group,
4983 * for the receiver's benefit */
4984 if (i < len - 1 || moreFlag) {
4985 list[i]->header.flags |= RX_MORE_PACKETS;
4988 /* Install the new retransmit time for the packet, and
4989 * record the time sent */
4990 list[i]->timeSent = *now;
4994 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4997 /* Since we're about to send a data packet to the peer, it's
4998 * safe to nuke any scheduled end-of-packets ack */
4999 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5001 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5002 MUTEX_EXIT(&call->lock);
5004 rxi_SendPacketList(call, conn, list, len, istack);
5006 rxi_SendPacket(call, conn, list[0], istack);
5008 MUTEX_ENTER(&call->lock);
5009 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5011 /* Update last send time for this call (for keep-alive
5012 * processing), and for the connection (so that we can discover
5013 * idle connections) */
5014 call->lastSendData = conn->lastSendTime = call->lastSendTime = clock_Sec();
5017 /* When sending packets we need to follow these rules:
5018 * 1. Never send more than maxDgramPackets in a jumbogram.
5019 * 2. Never send a packet with more than two iovecs in a jumbogram.
5020 * 3. Never send a retransmitted packet in a jumbogram.
5021 * 4. Never send more than cwind/4 packets in a jumbogram
5022 * We always keep the last list we should have sent so we
5023 * can set the RX_MORE_PACKETS flags correctly.
5026 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5027 int istack, struct clock *now, struct clock *retryTime,
5030 int i, cnt, lastCnt = 0;
5031 struct rx_packet **listP, **lastP = 0;
5032 struct rx_peer *peer = call->conn->peer;
5033 int morePackets = 0;
5035 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
5036 /* Does the current packet force us to flush the current list? */
5038 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5039 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5041 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
5043 /* If the call enters an error state stop sending, or if
5044 * we entered congestion recovery mode, stop sending */
5045 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5053 /* Add the current packet to the list if it hasn't been acked.
5054 * Otherwise adjust the list pointer to skip the current packet. */
5055 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5057 /* Do we need to flush the list? */
5058 if (cnt >= (int)peer->maxDgramPackets
5059 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5060 || list[i]->header.serial
5061 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5063 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5064 retryTime, resending);
5065 /* If the call enters an error state stop sending, or if
5066 * we entered congestion recovery mode, stop sending */
5068 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5073 listP = &list[i + 1];
5078 osi_Panic("rxi_SendList error");
5080 listP = &list[i + 1];
5084 /* Send the whole list when the call is in receive mode, when
5085 * the call is in eof mode, when we are in fast recovery mode,
5086 * and when we have the last packet */
5087 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5088 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5089 || (call->flags & RX_CALL_FAST_RECOVER)) {
5090 /* Check for the case where the current list contains
5091 * an acked packet. Since we always send retransmissions
5092 * in a separate packet, we only need to check the first
5093 * packet in the list */
5094 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5098 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5099 retryTime, resending);
5100 /* If the call enters an error state stop sending, or if
5101 * we entered congestion recovery mode, stop sending */
5102 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5106 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5109 } else if (lastCnt > 0) {
5110 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5115 #ifdef RX_ENABLE_LOCKS
5116 /* Call rxi_Start, below, but with the call lock held. */
5118 rxi_StartUnlocked(struct rxevent *event,
5119 void *arg0, void *arg1, int istack)
5121 struct rx_call *call = arg0;
5123 MUTEX_ENTER(&call->lock);
5124 rxi_Start(event, call, arg1, istack);
5125 MUTEX_EXIT(&call->lock);
5127 #endif /* RX_ENABLE_LOCKS */
5129 /* This routine is called when new packets are readied for
5130 * transmission and when retransmission may be necessary, or when the
5131 * transmission window or burst count are favourable. This should be
5132 * better optimized for new packets, the usual case, now that we've
5133 * got rid of queues of send packets. XXXXXXXXXXX */
5135 rxi_Start(struct rxevent *event,
5136 void *arg0, void *arg1, int istack)
5138 struct rx_call *call = arg0;
5140 struct rx_packet *p;
5141 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
5142 struct rx_peer *peer = call->conn->peer;
5143 struct clock now, usenow, retryTime;
5147 struct rx_packet **xmitList;
5150 /* If rxi_Start is being called as a result of a resend event,
5151 * then make sure that the event pointer is removed from the call
5152 * structure, since there is no longer a per-call retransmission
5154 if (event && event == call->resendEvent) {
5155 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5156 call->resendEvent = NULL;
5158 if (queue_IsEmpty(&call->tq)) {
5162 /* Timeouts trigger congestion recovery */
5163 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5164 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5165 /* someone else is waiting to start recovery */
5168 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5169 rxi_WaitforTQBusy(call);
5170 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5171 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5172 call->flags |= RX_CALL_FAST_RECOVER;
5173 if (peer->maxDgramPackets > 1) {
5174 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5176 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5178 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5179 call->nDgramPackets = 1;
5181 call->nextCwind = 1;
5184 MUTEX_ENTER(&peer->peer_lock);
5185 peer->MTU = call->MTU;
5186 peer->cwind = call->cwind;
5187 peer->nDgramPackets = 1;
5189 call->congestSeq = peer->congestSeq;
5190 MUTEX_EXIT(&peer->peer_lock);
5191 /* Clear retry times on packets. Otherwise, it's possible for
5192 * some packets in the queue to force resends at rates faster
5193 * than recovery rates.
5195 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5196 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5197 clock_Zero(&p->retryTime);
5202 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5203 rx_MutexIncrement(rx_tq_debug.rxi_start_in_error, rx_stats_mutex);
5208 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5209 /* Get clock to compute the re-transmit time for any packets
5210 * in this burst. Note, if we back off, it's reasonable to
5211 * back off all of the packets in the same manner, even if
5212 * some of them have been retransmitted more times than more
5214 * Do a dance to avoid blocking after setting now. */
5215 clock_Zero(&retryTime);
5216 MUTEX_ENTER(&peer->peer_lock);
5217 clock_Add(&retryTime, &peer->timeout);
5218 MUTEX_EXIT(&peer->peer_lock);
5219 clock_GetTime(&now);
5220 clock_Add(&retryTime, &now);
5222 /* Send (or resend) any packets that need it, subject to
5223 * window restrictions and congestion burst control
5224 * restrictions. Ask for an ack on the last packet sent in
5225 * this burst. For now, we're relying upon the window being
5226 * considerably bigger than the largest number of packets that
5227 * are typically sent at once by one initial call to
5228 * rxi_Start. This is probably bogus (perhaps we should ask
5229 * for an ack when we're half way through the current
5230 * window?). Also, for non file transfer applications, this
5231 * may end up asking for an ack for every packet. Bogus. XXXX
5234 * But check whether we're here recursively, and let the other guy
5237 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5238 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5239 call->flags |= RX_CALL_TQ_BUSY;
5241 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5243 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5244 call->flags &= ~RX_CALL_NEED_START;
5245 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5247 maxXmitPackets = MIN(call->twind, call->cwind);
5248 xmitList = (struct rx_packet **)
5249 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
5250 /* XXXX else we must drop any mtx we hold */
5251 afs_osi_Alloc_NoSleep(maxXmitPackets * sizeof(struct rx_packet *));
5253 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5255 if (xmitList == NULL)
5256 osi_Panic("rxi_Start, failed to allocate xmit list");
5257 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5258 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5259 /* We shouldn't be sending packets if a thread is waiting
5260 * to initiate congestion recovery */
5264 && (call->flags & RX_CALL_FAST_RECOVER)) {
5265 /* Only send one packet during fast recovery */
5268 if ((p->flags & RX_PKTFLAG_FREE)
5269 || (!queue_IsEnd(&call->tq, nxp)
5270 && (nxp->flags & RX_PKTFLAG_FREE))
5271 || (p == (struct rx_packet *)&rx_freePacketQueue)
5272 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5273 osi_Panic("rxi_Start: xmit queue clobbered");
5275 if (p->flags & RX_PKTFLAG_ACKED) {
5276 /* Since we may block, don't trust this */
5277 usenow.sec = usenow.usec = 0;
5278 rx_MutexIncrement(rx_stats.ignoreAckedPacket, rx_stats_mutex);
5279 continue; /* Ignore this packet if it has been acknowledged */
5282 /* Turn off all flags except these ones, which are the same
5283 * on each transmission */
5284 p->header.flags &= RX_PRESET_FLAGS;
5286 if (p->header.seq >=
5287 call->tfirst + MIN((int)call->twind,
5288 (int)(call->nSoftAcked +
5290 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5291 /* Note: if we're waiting for more window space, we can
5292 * still send retransmits; hence we don't return here, but
5293 * break out to schedule a retransmit event */
5294 dpf(("call %d waiting for window",
5295 *(call->callNumber)));
5299 /* Transmit the packet if it needs to be sent. */
5300 if (!clock_Lt(&now, &p->retryTime)) {
5301 if (nXmitPackets == maxXmitPackets) {
5302 rxi_SendXmitList(call, xmitList, nXmitPackets,
5303 istack, &now, &retryTime,
5305 osi_Free(xmitList, maxXmitPackets *
5306 sizeof(struct rx_packet *));
5309 xmitList[nXmitPackets++] = p;
5313 /* xmitList now hold pointers to all of the packets that are
5314 * ready to send. Now we loop to send the packets */
5315 if (nXmitPackets > 0) {
5316 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5317 &now, &retryTime, resending);
5320 maxXmitPackets * sizeof(struct rx_packet *));
5322 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5324 * TQ references no longer protected by this flag; they must remain
5325 * protected by the global lock.
5327 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5328 call->flags &= ~RX_CALL_TQ_BUSY;
5329 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5330 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5331 #ifdef RX_ENABLE_LOCKS
5332 osirx_AssertMine(&call->lock, "rxi_Start start");
5333 CV_BROADCAST(&call->cv_tq);
5334 #else /* RX_ENABLE_LOCKS */
5335 osi_rxWakeup(&call->tq);
5336 #endif /* RX_ENABLE_LOCKS */
5341 /* We went into the error state while sending packets. Now is
5342 * the time to reset the call. This will also inform the using
5343 * process that the call is in an error state.
5345 rx_MutexIncrement(rx_tq_debug.rxi_start_aborted, rx_stats_mutex);
5346 call->flags &= ~RX_CALL_TQ_BUSY;
5347 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5348 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5349 #ifdef RX_ENABLE_LOCKS
5350 osirx_AssertMine(&call->lock, "rxi_Start middle");
5351 CV_BROADCAST(&call->cv_tq);
5352 #else /* RX_ENABLE_LOCKS */
5353 osi_rxWakeup(&call->tq);
5354 #endif /* RX_ENABLE_LOCKS */
5356 rxi_CallError(call, call->error);
5359 #ifdef RX_ENABLE_LOCKS
5360 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5361 register int missing;
5362 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5363 /* Some packets have received acks. If they all have, we can clear
5364 * the transmit queue.
5367 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5368 if (p->header.seq < call->tfirst
5369 && (p->flags & RX_PKTFLAG_ACKED)) {
5376 call->flags |= RX_CALL_TQ_CLEARME;
5378 #endif /* RX_ENABLE_LOCKS */
5379 /* Don't bother doing retransmits if the TQ is cleared. */
5380 if (call->flags & RX_CALL_TQ_CLEARME) {
5381 rxi_ClearTransmitQueue(call, 1);
5383 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5386 /* Always post a resend event, if there is anything in the
5387 * queue, and resend is possible. There should be at least
5388 * one unacknowledged packet in the queue ... otherwise none
5389 * of these packets should be on the queue in the first place.
5391 if (call->resendEvent) {
5392 /* Cancel the existing event and post a new one */
5393 rxevent_Cancel(call->resendEvent, call,
5394 RX_CALL_REFCOUNT_RESEND);
5397 /* The retry time is the retry time on the first unacknowledged
5398 * packet inside the current window */
5400 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5401 /* Don't set timers for packets outside the window */
5402 if (p->header.seq >= call->tfirst + call->twind) {
5406 if (!(p->flags & RX_PKTFLAG_ACKED)
5407 && !clock_IsZero(&p->retryTime)) {
5409 retryTime = p->retryTime;
5414 /* Post a new event to re-run rxi_Start when retries may be needed */
5415 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5416 #ifdef RX_ENABLE_LOCKS
5417 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5419 rxevent_PostNow2(&retryTime, &usenow,
5421 (void *)call, 0, istack);
5422 #else /* RX_ENABLE_LOCKS */
5424 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5425 (void *)call, 0, istack);
5426 #endif /* RX_ENABLE_LOCKS */
5429 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5430 } while (call->flags & RX_CALL_NEED_START);
5432 * TQ references no longer protected by this flag; they must remain
5433 * protected by the global lock.
5435 call->flags &= ~RX_CALL_TQ_BUSY;
5436 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5437 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5438 #ifdef RX_ENABLE_LOCKS
5439 osirx_AssertMine(&call->lock, "rxi_Start end");
5440 CV_BROADCAST(&call->cv_tq);
5441 #else /* RX_ENABLE_LOCKS */
5442 osi_rxWakeup(&call->tq);
5443 #endif /* RX_ENABLE_LOCKS */
5446 call->flags |= RX_CALL_NEED_START;
5448 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5450 if (call->resendEvent) {
5451 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5456 /* Also adjusts the keep alive parameters for the call, to reflect
5457 * that we have just sent a packet (so keep alives aren't sent
5460 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5463 register struct rx_connection *conn = call->conn;
5465 /* Stamp each packet with the user supplied status */
5466 p->header.userStatus = call->localStatus;
5468 /* Allow the security object controlling this call's security to
5469 * make any last-minute changes to the packet */
5470 RXS_SendPacket(conn->securityObject, call, p);
5472 /* Since we're about to send SOME sort of packet to the peer, it's
5473 * safe to nuke any scheduled end-of-packets ack */
5474 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5476 /* Actually send the packet, filling in more connection-specific fields */
5477 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5478 MUTEX_EXIT(&call->lock);
5479 rxi_SendPacket(call, conn, p, istack);
5480 MUTEX_ENTER(&call->lock);
5481 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5483 /* Update last send time for this call (for keep-alive
5484 * processing), and for the connection (so that we can discover
5485 * idle connections) */
5486 conn->lastSendTime = call->lastSendTime = clock_Sec();
5487 /* Don't count keepalives here, so idleness can be tracked. */
5488 if (p->header.type != RX_PACKET_TYPE_ACK)
5489 call->lastSendData = call->lastSendTime;
5493 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5494 * that things are fine. Also called periodically to guarantee that nothing
5495 * falls through the cracks (e.g. (error + dally) connections have keepalive
5496 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5498 * haveCTLock Set if calling from rxi_ReapConnections
5500 #ifdef RX_ENABLE_LOCKS
5502 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5503 #else /* RX_ENABLE_LOCKS */
5505 rxi_CheckCall(register struct rx_call *call)
5506 #endif /* RX_ENABLE_LOCKS */
5508 register struct rx_connection *conn = call->conn;
5510 afs_uint32 deadTime;
5512 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5513 if (call->flags & RX_CALL_TQ_BUSY) {
5514 /* Call is active and will be reset by rxi_Start if it's
5515 * in an error state.
5520 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5522 (((afs_uint32) conn->secondsUntilDead << 10) +
5523 ((afs_uint32) conn->peer->rtt >> 3) +
5524 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5526 /* These are computed to the second (+- 1 second). But that's
5527 * good enough for these values, which should be a significant
5528 * number of seconds. */
5529 if (now > (call->lastReceiveTime + deadTime)) {
5530 if (call->state == RX_STATE_ACTIVE) {
5532 #if defined(KERNEL) && defined(AFS_SUN57_ENV)
5534 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5535 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
5536 ip_stack_t *ipst = ns->netstack_ip;
5538 ire = ire_cache_lookup(call->conn->peer->host
5539 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
5541 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
5543 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5550 if (ire && ire->ire_max_frag > 0)
5551 rxi_SetPeerMtu(call->conn->peer->host, 0, ire->ire_max_frag);
5552 #if defined(GLOBAL_NETSTACKID)
5556 #endif /* ADAPT_PMTU */
5557 rxi_CallError(call, RX_CALL_DEAD);
5560 #ifdef RX_ENABLE_LOCKS
5561 /* Cancel pending events */
5562 rxevent_Cancel(call->delayedAckEvent, call,
5563 RX_CALL_REFCOUNT_DELAY);
5564 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5565 rxevent_Cancel(call->keepAliveEvent, call,
5566 RX_CALL_REFCOUNT_ALIVE);
5567 if (call->refCount == 0) {
5568 rxi_FreeCall(call, haveCTLock);
5572 #else /* RX_ENABLE_LOCKS */
5575 #endif /* RX_ENABLE_LOCKS */
5577 /* Non-active calls are destroyed if they are not responding
5578 * to pings; active calls are simply flagged in error, so the
5579 * attached process can die reasonably gracefully. */
5581 /* see if we have a non-activity timeout */
5582 if (call->startWait && conn->idleDeadTime
5583 && ((call->startWait + conn->idleDeadTime) < now)) {
5584 if (call->state == RX_STATE_ACTIVE) {
5585 rxi_CallError(call, RX_CALL_TIMEOUT);
5589 if (call->lastSendData && conn->idleDeadTime && (conn->idleDeadErr != 0)
5590 && ((call->lastSendData + conn->idleDeadTime) < now)) {
5591 if (call->state == RX_STATE_ACTIVE) {
5592 rxi_CallError(call, conn->idleDeadErr);
5596 /* see if we have a hard timeout */
5597 if (conn->hardDeadTime
5598 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5599 if (call->state == RX_STATE_ACTIVE)
5600 rxi_CallError(call, RX_CALL_TIMEOUT);
5607 /* When a call is in progress, this routine is called occasionally to
5608 * make sure that some traffic has arrived (or been sent to) the peer.
5609 * If nothing has arrived in a reasonable amount of time, the call is
5610 * declared dead; if nothing has been sent for a while, we send a
5611 * keep-alive packet (if we're actually trying to keep the call alive)
5614 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
5616 struct rx_call *call = arg1;
5617 struct rx_connection *conn;
5620 MUTEX_ENTER(&call->lock);
5621 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5622 if (event == call->keepAliveEvent)
5623 call->keepAliveEvent = NULL;
5626 #ifdef RX_ENABLE_LOCKS
5627 if (rxi_CheckCall(call, 0)) {
5628 MUTEX_EXIT(&call->lock);
5631 #else /* RX_ENABLE_LOCKS */
5632 if (rxi_CheckCall(call))
5634 #endif /* RX_ENABLE_LOCKS */
5636 /* Don't try to keep alive dallying calls */
5637 if (call->state == RX_STATE_DALLY) {
5638 MUTEX_EXIT(&call->lock);
5643 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5644 /* Don't try to send keepalives if there is unacknowledged data */
5645 /* the rexmit code should be good enough, this little hack
5646 * doesn't quite work XXX */
5647 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5649 rxi_ScheduleKeepAliveEvent(call);
5650 MUTEX_EXIT(&call->lock);
5655 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5657 if (!call->keepAliveEvent) {
5658 struct clock when, now;
5659 clock_GetTime(&now);
5661 when.sec += call->conn->secondsUntilPing;
5662 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5663 call->keepAliveEvent =
5664 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
5668 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5670 rxi_KeepAliveOn(register struct rx_call *call)
5672 /* Pretend last packet received was received now--i.e. if another
5673 * packet isn't received within the keep alive time, then the call
5674 * will die; Initialize last send time to the current time--even
5675 * if a packet hasn't been sent yet. This will guarantee that a
5676 * keep-alive is sent within the ping time */
5677 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5678 rxi_ScheduleKeepAliveEvent(call);
5681 /* This routine is called to send connection abort messages
5682 * that have been delayed to throttle looping clients. */
5684 rxi_SendDelayedConnAbort(struct rxevent *event,
5685 void *arg1, void *unused)
5687 struct rx_connection *conn = arg1;
5690 struct rx_packet *packet;
5692 MUTEX_ENTER(&conn->conn_data_lock);
5693 conn->delayedAbortEvent = NULL;
5694 error = htonl(conn->error);
5696 MUTEX_EXIT(&conn->conn_data_lock);
5697 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5700 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5701 RX_PACKET_TYPE_ABORT, (char *)&error,
5703 rxi_FreePacket(packet);
5707 /* This routine is called to send call abort messages
5708 * that have been delayed to throttle looping clients. */
5710 rxi_SendDelayedCallAbort(struct rxevent *event,
5711 void *arg1, void *dummy)
5713 struct rx_call *call = arg1;
5716 struct rx_packet *packet;
5718 MUTEX_ENTER(&call->lock);
5719 call->delayedAbortEvent = NULL;
5720 error = htonl(call->error);
5722 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5725 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5726 (char *)&error, sizeof(error), 0);
5727 rxi_FreePacket(packet);
5729 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5730 MUTEX_EXIT(&call->lock);
5733 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5734 * seconds) to ask the client to authenticate itself. The routine
5735 * issues a challenge to the client, which is obtained from the
5736 * security object associated with the connection */
5738 rxi_ChallengeEvent(struct rxevent *event,
5739 void *arg0, void *arg1, int tries)
5741 struct rx_connection *conn = arg0;
5743 conn->challengeEvent = NULL;
5744 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5745 register struct rx_packet *packet;
5746 struct clock when, now;
5749 /* We've failed to authenticate for too long.
5750 * Reset any calls waiting for authentication;
5751 * they are all in RX_STATE_PRECALL.
5755 MUTEX_ENTER(&conn->conn_call_lock);
5756 for (i = 0; i < RX_MAXCALLS; i++) {
5757 struct rx_call *call = conn->call[i];
5759 MUTEX_ENTER(&call->lock);
5760 if (call->state == RX_STATE_PRECALL) {
5761 rxi_CallError(call, RX_CALL_DEAD);
5762 rxi_SendCallAbort(call, NULL, 0, 0);
5764 MUTEX_EXIT(&call->lock);
5767 MUTEX_EXIT(&conn->conn_call_lock);
5771 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5773 /* If there's no packet available, do this later. */
5774 RXS_GetChallenge(conn->securityObject, conn, packet);
5775 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5776 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5777 rxi_FreePacket(packet);
5779 clock_GetTime(&now);
5781 when.sec += RX_CHALLENGE_TIMEOUT;
5782 conn->challengeEvent =
5783 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
5788 /* Call this routine to start requesting the client to authenticate
5789 * itself. This will continue until authentication is established,
5790 * the call times out, or an invalid response is returned. The
5791 * security object associated with the connection is asked to create
5792 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5793 * defined earlier. */
5795 rxi_ChallengeOn(register struct rx_connection *conn)
5797 if (!conn->challengeEvent) {
5798 RXS_CreateChallenge(conn->securityObject, conn);
5799 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5804 /* Compute round trip time of the packet provided, in *rttp.
5807 /* rxi_ComputeRoundTripTime is called with peer locked. */
5808 /* sentp and/or peer may be null */
5810 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5811 register struct clock *sentp,
5812 register struct rx_peer *peer)
5814 struct clock thisRtt, *rttp = &thisRtt;
5816 register int rtt_timeout;
5818 clock_GetTime(rttp);
5820 if (clock_Lt(rttp, sentp)) {
5822 return; /* somebody set the clock back, don't count this time. */
5824 clock_Sub(rttp, sentp);
5825 MUTEX_ENTER(&rx_stats_mutex);
5826 if (clock_Lt(rttp, &rx_stats.minRtt))
5827 rx_stats.minRtt = *rttp;
5828 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5829 if (rttp->sec > 60) {
5830 MUTEX_EXIT(&rx_stats_mutex);
5831 return; /* somebody set the clock ahead */
5833 rx_stats.maxRtt = *rttp;
5835 clock_Add(&rx_stats.totalRtt, rttp);
5836 rx_stats.nRttSamples++;
5837 MUTEX_EXIT(&rx_stats_mutex);
5839 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5841 /* Apply VanJacobson round-trip estimations */
5846 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5847 * srtt is stored as fixed point with 3 bits after the binary
5848 * point (i.e., scaled by 8). The following magic is
5849 * equivalent to the smoothing algorithm in rfc793 with an
5850 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5851 * srtt*8 = srtt*8 + rtt - srtt
5852 * srtt = srtt + rtt/8 - srtt/8
5855 delta = MSEC(rttp) - (peer->rtt >> 3);
5859 * We accumulate a smoothed rtt variance (actually, a smoothed
5860 * mean difference), then set the retransmit timer to smoothed
5861 * rtt + 4 times the smoothed variance (was 2x in van's original
5862 * paper, but 4x works better for me, and apparently for him as
5864 * rttvar is stored as
5865 * fixed point with 2 bits after the binary point (scaled by
5866 * 4). The following is equivalent to rfc793 smoothing with
5867 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5868 * replaces rfc793's wired-in beta.
5869 * dev*4 = dev*4 + (|actual - expected| - dev)
5875 delta -= (peer->rtt_dev >> 2);
5876 peer->rtt_dev += delta;
5878 /* I don't have a stored RTT so I start with this value. Since I'm
5879 * probably just starting a call, and will be pushing more data down
5880 * this, I expect congestion to increase rapidly. So I fudge a
5881 * little, and I set deviance to half the rtt. In practice,
5882 * deviance tends to approach something a little less than
5883 * half the smoothed rtt. */
5884 peer->rtt = (MSEC(rttp) << 3) + 8;
5885 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5887 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5888 * the other of these connections is usually in a user process, and can
5889 * be switched and/or swapped out. So on fast, reliable networks, the
5890 * timeout would otherwise be too short.
5892 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5893 clock_Zero(&(peer->timeout));
5894 clock_Addmsec(&(peer->timeout), rtt_timeout);
5896 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)));
5900 /* Find all server connections that have not been active for a long time, and
5903 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2)
5905 struct clock now, when;
5906 clock_GetTime(&now);
5908 /* Find server connection structures that haven't been used for
5909 * greater than rx_idleConnectionTime */
5911 struct rx_connection **conn_ptr, **conn_end;
5912 int i, havecalls = 0;
5913 MUTEX_ENTER(&rx_connHashTable_lock);
5914 for (conn_ptr = &rx_connHashTable[0], conn_end =
5915 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5917 struct rx_connection *conn, *next;
5918 struct rx_call *call;
5922 for (conn = *conn_ptr; conn; conn = next) {
5923 /* XXX -- Shouldn't the connection be locked? */
5926 for (i = 0; i < RX_MAXCALLS; i++) {
5927 call = conn->call[i];
5930 MUTEX_ENTER(&call->lock);
5931 #ifdef RX_ENABLE_LOCKS
5932 result = rxi_CheckCall(call, 1);
5933 #else /* RX_ENABLE_LOCKS */
5934 result = rxi_CheckCall(call);
5935 #endif /* RX_ENABLE_LOCKS */
5936 MUTEX_EXIT(&call->lock);
5938 /* If CheckCall freed the call, it might
5939 * have destroyed the connection as well,
5940 * which screws up the linked lists.
5946 if (conn->type == RX_SERVER_CONNECTION) {
5947 /* This only actually destroys the connection if
5948 * there are no outstanding calls */
5949 MUTEX_ENTER(&conn->conn_data_lock);
5950 if (!havecalls && !conn->refCount
5951 && ((conn->lastSendTime + rx_idleConnectionTime) <
5953 conn->refCount++; /* it will be decr in rx_DestroyConn */
5954 MUTEX_EXIT(&conn->conn_data_lock);
5955 #ifdef RX_ENABLE_LOCKS
5956 rxi_DestroyConnectionNoLock(conn);
5957 #else /* RX_ENABLE_LOCKS */
5958 rxi_DestroyConnection(conn);
5959 #endif /* RX_ENABLE_LOCKS */
5961 #ifdef RX_ENABLE_LOCKS
5963 MUTEX_EXIT(&conn->conn_data_lock);
5965 #endif /* RX_ENABLE_LOCKS */
5969 #ifdef RX_ENABLE_LOCKS
5970 while (rx_connCleanup_list) {
5971 struct rx_connection *conn;
5972 conn = rx_connCleanup_list;
5973 rx_connCleanup_list = rx_connCleanup_list->next;
5974 MUTEX_EXIT(&rx_connHashTable_lock);
5975 rxi_CleanupConnection(conn);
5976 MUTEX_ENTER(&rx_connHashTable_lock);
5978 MUTEX_EXIT(&rx_connHashTable_lock);
5979 #endif /* RX_ENABLE_LOCKS */
5982 /* Find any peer structures that haven't been used (haven't had an
5983 * associated connection) for greater than rx_idlePeerTime */
5985 struct rx_peer **peer_ptr, **peer_end;
5987 MUTEX_ENTER(&rx_rpc_stats);
5988 MUTEX_ENTER(&rx_peerHashTable_lock);
5989 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5990 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5992 struct rx_peer *peer, *next, *prev;
5993 for (prev = peer = *peer_ptr; peer; peer = next) {
5995 code = MUTEX_TRYENTER(&peer->peer_lock);
5996 if ((code) && (peer->refCount == 0)
5997 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5998 rx_interface_stat_p rpc_stat, nrpc_stat;
6000 MUTEX_EXIT(&peer->peer_lock);
6001 MUTEX_DESTROY(&peer->peer_lock);
6003 (&peer->rpcStats, rpc_stat, nrpc_stat,
6004 rx_interface_stat)) {
6005 unsigned int num_funcs;
6008 queue_Remove(&rpc_stat->queue_header);
6009 queue_Remove(&rpc_stat->all_peers);
6010 num_funcs = rpc_stat->stats[0].func_total;
6012 sizeof(rx_interface_stat_t) +
6013 rpc_stat->stats[0].func_total *
6014 sizeof(rx_function_entry_v1_t);
6016 rxi_Free(rpc_stat, space);
6017 rxi_rpc_peer_stat_cnt -= num_funcs;
6020 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6021 if (peer == *peer_ptr) {
6028 MUTEX_EXIT(&peer->peer_lock);
6034 MUTEX_EXIT(&rx_peerHashTable_lock);
6035 MUTEX_EXIT(&rx_rpc_stats);
6038 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
6039 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
6040 * GC, just below. Really, we shouldn't have to keep moving packets from
6041 * one place to another, but instead ought to always know if we can
6042 * afford to hold onto a packet in its particular use. */
6043 MUTEX_ENTER(&rx_freePktQ_lock);
6044 if (rx_waitingForPackets) {
6045 rx_waitingForPackets = 0;
6046 #ifdef RX_ENABLE_LOCKS
6047 CV_BROADCAST(&rx_waitingForPackets_cv);
6049 osi_rxWakeup(&rx_waitingForPackets);
6052 MUTEX_EXIT(&rx_freePktQ_lock);
6055 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
6056 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
6060 /* rxs_Release - This isn't strictly necessary but, since the macro name from
6061 * rx.h is sort of strange this is better. This is called with a security
6062 * object before it is discarded. Each connection using a security object has
6063 * its own refcount to the object so it won't actually be freed until the last
6064 * connection is destroyed.
6066 * This is the only rxs module call. A hold could also be written but no one
6070 rxs_Release(struct rx_securityClass *aobj)
6072 return RXS_Close(aobj);
6076 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
6077 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
6078 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
6079 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
6081 /* Adjust our estimate of the transmission rate to this peer, given
6082 * that the packet p was just acked. We can adjust peer->timeout and
6083 * call->twind. Pragmatically, this is called
6084 * only with packets of maximal length.
6085 * Called with peer and call locked.
6089 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
6090 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
6092 afs_int32 xferSize, xferMs;
6093 register afs_int32 minTime;
6096 /* Count down packets */
6097 if (peer->rateFlag > 0)
6099 /* Do nothing until we're enabled */
6100 if (peer->rateFlag != 0)
6105 /* Count only when the ack seems legitimate */
6106 switch (ackReason) {
6107 case RX_ACK_REQUESTED:
6109 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6113 case RX_ACK_PING_RESPONSE:
6114 if (p) /* want the response to ping-request, not data send */
6116 clock_GetTime(&newTO);
6117 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6118 clock_Sub(&newTO, &call->pingRequestTime);
6119 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6123 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6130 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));
6132 /* Track only packets that are big enough. */
6133 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6137 /* absorb RTT data (in milliseconds) for these big packets */
6138 if (peer->smRtt == 0) {
6139 peer->smRtt = xferMs;
6141 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6146 if (peer->countDown) {
6150 peer->countDown = 10; /* recalculate only every so often */
6152 /* In practice, we can measure only the RTT for full packets,
6153 * because of the way Rx acks the data that it receives. (If it's
6154 * smaller than a full packet, it often gets implicitly acked
6155 * either by the call response (from a server) or by the next call
6156 * (from a client), and either case confuses transmission times
6157 * with processing times.) Therefore, replace the above
6158 * more-sophisticated processing with a simpler version, where the
6159 * smoothed RTT is kept for full-size packets, and the time to
6160 * transmit a windowful of full-size packets is simply RTT *
6161 * windowSize. Again, we take two steps:
6162 - ensure the timeout is large enough for a single packet's RTT;
6163 - ensure that the window is small enough to fit in the desired timeout.*/
6165 /* First, the timeout check. */
6166 minTime = peer->smRtt;
6167 /* Get a reasonable estimate for a timeout period */
6169 newTO.sec = minTime / 1000;
6170 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6172 /* Increase the timeout period so that we can always do at least
6173 * one packet exchange */
6174 if (clock_Gt(&newTO, &peer->timeout)) {
6176 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));
6178 peer->timeout = newTO;
6181 /* Now, get an estimate for the transmit window size. */
6182 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6183 /* Now, convert to the number of full packets that could fit in a
6184 * reasonable fraction of that interval */
6185 minTime /= (peer->smRtt << 1);
6186 xferSize = minTime; /* (make a copy) */
6188 /* Now clamp the size to reasonable bounds. */
6191 else if (minTime > rx_Window)
6192 minTime = rx_Window;
6193 /* if (minTime != peer->maxWindow) {
6194 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6195 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6196 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6198 peer->maxWindow = minTime;
6199 elide... call->twind = minTime;
6203 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6204 * Discern this by calculating the timeout necessary for rx_Window
6206 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6207 /* calculate estimate for transmission interval in milliseconds */
6208 minTime = rx_Window * peer->smRtt;
6209 if (minTime < 1000) {
6210 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
6211 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6212 peer->timeout.usec, peer->smRtt, peer->packetSize));
6214 newTO.sec = 0; /* cut back on timeout by half a second */
6215 newTO.usec = 500000;
6216 clock_Sub(&peer->timeout, &newTO);
6221 } /* end of rxi_ComputeRate */
6222 #endif /* ADAPT_WINDOW */
6230 #define TRACE_OPTION_DEBUGLOG 4
6238 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6239 0, KEY_QUERY_VALUE, &parmKey);
6240 if (code != ERROR_SUCCESS)
6243 dummyLen = sizeof(TraceOption);
6244 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6245 (BYTE *) &TraceOption, &dummyLen);
6246 if (code == ERROR_SUCCESS) {
6247 rxdebug_active = (TraceOption & TRACE_OPTION_DEBUGLOG) ? 1 : 0;
6249 RegCloseKey (parmKey);
6250 #endif /* AFS_NT40_ENV */
6255 rx_DebugOnOff(int on)
6257 rxdebug_active = on;
6259 #endif /* AFS_NT40_ENV */
6262 /* Don't call this debugging routine directly; use dpf */
6264 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
6265 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
6273 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6276 len = _snprintf(msg, sizeof(msg)-2,
6277 tformat, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10,
6278 a11, a12, a13, a14, a15);
6280 if (msg[len-1] != '\n') {
6284 OutputDebugString(msg);
6289 clock_GetTime(&now);
6290 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
6291 (unsigned int)now.usec / 1000);
6292 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
6299 * This function is used to process the rx_stats structure that is local
6300 * to a process as well as an rx_stats structure received from a remote
6301 * process (via rxdebug). Therefore, it needs to do minimal version
6305 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
6306 afs_int32 freePackets, char version)
6310 if (size != sizeof(struct rx_stats)) {
6312 "Unexpected size of stats structure: was %d, expected %lud\n",
6313 size, sizeof(struct rx_stats));
6316 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6319 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6320 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6321 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6322 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6323 s->specialPktAllocFailures);
6325 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6326 s->receivePktAllocFailures, s->sendPktAllocFailures,
6327 s->specialPktAllocFailures);
6331 " greedy %d, " "bogusReads %d (last from host %x), "
6332 "noPackets %d, " "noBuffers %d, " "selects %d, "
6333 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6334 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6335 s->selects, s->sendSelects);
6337 fprintf(file, " packets read: ");
6338 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6339 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6341 fprintf(file, "\n");
6344 " other read counters: data %d, " "ack %d, " "dup %d "
6345 "spurious %d " "dally %d\n", s->dataPacketsRead,
6346 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6347 s->ignorePacketDally);
6349 fprintf(file, " packets sent: ");
6350 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6351 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6353 fprintf(file, "\n");
6356 " other send counters: ack %d, " "data %d (not resends), "
6357 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6358 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6359 s->dataPacketsPushed, s->ignoreAckedPacket);
6362 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6363 s->netSendFailures, (int)s->fatalErrors);
6365 if (s->nRttSamples) {
6366 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6367 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6369 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6370 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6374 " %d server connections, " "%d client connections, "
6375 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6376 s->nServerConns, s->nClientConns, s->nPeerStructs,
6377 s->nCallStructs, s->nFreeCallStructs);
6379 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6380 fprintf(file, " %d clock updates\n", clock_nUpdates);
6385 /* for backward compatibility */
6387 rx_PrintStats(FILE * file)
6389 MUTEX_ENTER(&rx_stats_mutex);
6390 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6392 MUTEX_EXIT(&rx_stats_mutex);
6396 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6398 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6399 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6400 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6403 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6404 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6405 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6408 " Packet size %d, " "max in packet skew %d, "
6409 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6410 (int)peer->outPacketSkew);
6413 #ifdef AFS_PTHREAD_ENV
6415 * This mutex protects the following static variables:
6419 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6420 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6422 #define LOCK_RX_DEBUG
6423 #define UNLOCK_RX_DEBUG
6424 #endif /* AFS_PTHREAD_ENV */
6427 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6428 u_char type, void *inputData, size_t inputLength,
6429 void *outputData, size_t outputLength)
6431 static afs_int32 counter = 100;
6432 time_t waitTime, waitCount, startTime;
6433 struct rx_header theader;
6435 register afs_int32 code;
6436 struct timeval tv_now, tv_wake, tv_delta;
6437 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: