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"
36 #include "afs/afs_args.h"
37 #include "afs/afs_osi.h"
38 #ifdef RX_KERNEL_TRACE
39 #include "rx_kcommon.h"
41 #if (defined(AFS_AUX_ENV) || defined(AFS_AIX_ENV))
45 #undef RXDEBUG /* turn off debugging */
47 #if defined(AFS_SGI_ENV)
48 #include "sys/debug.h"
57 #endif /* AFS_OSF_ENV */
59 #include "afs/sysincludes.h"
60 #include "afsincludes.h"
63 #include "rx_kmutex.h"
64 #include "rx_kernel.h"
68 #include "rx_globals.h"
70 #define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
71 #define AFSOP_STOP_AFS 211 /* Stop AFS process */
72 #define AFSOP_STOP_BKG 212 /* Stop BKG process */
74 extern afs_int32 afs_termState;
76 #include "sys/lockl.h"
77 #include "sys/lock_def.h"
78 #endif /* AFS_AIX41_ENV */
79 # include "rxgen_consts.h"
81 # include <sys/types.h>
86 # include <afs/afsutil.h>
87 # include <WINNT\afsreg.h>
89 # include <sys/socket.h>
90 # include <sys/file.h>
92 # include <sys/stat.h>
93 # include <netinet/in.h>
94 # 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>
112 int (*registerProgram) () = 0;
113 int (*swapNameProgram) () = 0;
115 /* Local static routines */
116 static void rxi_DestroyConnectionNoLock(register struct rx_connection *conn);
117 #ifdef RX_ENABLE_LOCKS
118 static void rxi_SetAcksInTransmitQueue(register struct rx_call *call);
121 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
123 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
124 afs_int32 rxi_start_in_error;
126 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
129 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
130 * currently allocated within rx. This number is used to allocate the
131 * memory required to return the statistics when queried.
134 static unsigned int rxi_rpc_peer_stat_cnt;
137 * rxi_rpc_process_stat_cnt counts the total number of local process stat
138 * structures currently allocated within rx. The number is used to allocate
139 * the memory required to return the statistics when queried.
142 static unsigned int rxi_rpc_process_stat_cnt;
144 #if !defined(offsetof)
145 #include <stddef.h> /* for definition of offsetof() */
148 #ifdef AFS_PTHREAD_ENV
152 * Use procedural initialization of mutexes/condition variables
156 extern pthread_mutex_t rx_stats_mutex;
157 extern pthread_mutex_t des_init_mutex;
158 extern pthread_mutex_t des_random_mutex;
159 extern pthread_mutex_t rx_clock_mutex;
160 extern pthread_mutex_t rxi_connCacheMutex;
161 extern pthread_mutex_t rx_event_mutex;
162 extern pthread_mutex_t osi_malloc_mutex;
163 extern pthread_mutex_t event_handler_mutex;
164 extern pthread_mutex_t listener_mutex;
165 extern pthread_mutex_t rx_if_init_mutex;
166 extern pthread_mutex_t rx_if_mutex;
167 extern pthread_mutex_t rxkad_client_uid_mutex;
168 extern pthread_mutex_t rxkad_random_mutex;
170 extern pthread_cond_t rx_event_handler_cond;
171 extern pthread_cond_t rx_listener_cond;
173 static pthread_mutex_t epoch_mutex;
174 static pthread_mutex_t rx_init_mutex;
175 static pthread_mutex_t rx_debug_mutex;
178 rxi_InitPthread(void)
180 assert(pthread_mutex_init(&rx_clock_mutex, (const pthread_mutexattr_t *)0)
182 assert(pthread_mutex_init(&rx_stats_mutex, (const pthread_mutexattr_t *)0)
184 assert(pthread_mutex_init
185 (&rxi_connCacheMutex, (const pthread_mutexattr_t *)0) == 0);
186 assert(pthread_mutex_init(&rx_init_mutex, (const pthread_mutexattr_t *)0)
188 assert(pthread_mutex_init(&epoch_mutex, (const pthread_mutexattr_t *)0) ==
190 assert(pthread_mutex_init(&rx_event_mutex, (const pthread_mutexattr_t *)0)
192 assert(pthread_mutex_init(&des_init_mutex, (const pthread_mutexattr_t *)0)
194 assert(pthread_mutex_init
195 (&des_random_mutex, (const pthread_mutexattr_t *)0) == 0);
196 assert(pthread_mutex_init
197 (&osi_malloc_mutex, (const pthread_mutexattr_t *)0) == 0);
198 assert(pthread_mutex_init
199 (&event_handler_mutex, (const pthread_mutexattr_t *)0) == 0);
200 assert(pthread_mutex_init(&listener_mutex, (const pthread_mutexattr_t *)0)
202 assert(pthread_mutex_init
203 (&rx_if_init_mutex, (const pthread_mutexattr_t *)0) == 0);
204 assert(pthread_mutex_init(&rx_if_mutex, (const pthread_mutexattr_t *)0) ==
206 assert(pthread_mutex_init
207 (&rxkad_client_uid_mutex, (const pthread_mutexattr_t *)0) == 0);
208 assert(pthread_mutex_init
209 (&rxkad_random_mutex, (const pthread_mutexattr_t *)0) == 0);
210 assert(pthread_mutex_init(&rx_debug_mutex, (const pthread_mutexattr_t *)0)
213 assert(pthread_cond_init
214 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
215 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
217 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
218 assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
220 rxkad_global_stats_init();
223 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
224 #define INIT_PTHREAD_LOCKS \
225 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
227 * The rx_stats_mutex mutex protects the following global variables:
232 * rxi_lowConnRefCount
233 * rxi_lowPeerRefCount
242 #define INIT_PTHREAD_LOCKS
246 /* Variables for handling the minProcs implementation. availProcs gives the
247 * number of threads available in the pool at this moment (not counting dudes
248 * executing right now). totalMin gives the total number of procs required
249 * for handling all minProcs requests. minDeficit is a dynamic variable
250 * tracking the # of procs required to satisfy all of the remaining minProcs
252 * For fine grain locking to work, the quota check and the reservation of
253 * a server thread has to come while rxi_availProcs and rxi_minDeficit
254 * are locked. To this end, the code has been modified under #ifdef
255 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
256 * same time. A new function, ReturnToServerPool() returns the allocation.
258 * A call can be on several queue's (but only one at a time). When
259 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
260 * that no one else is touching the queue. To this end, we store the address
261 * of the queue lock in the call structure (under the call lock) when we
262 * put the call on a queue, and we clear the call_queue_lock when the
263 * call is removed from a queue (once the call lock has been obtained).
264 * This allows rxi_ResetCall to safely synchronize with others wishing
265 * to manipulate the queue.
268 #ifdef RX_ENABLE_LOCKS
269 static afs_kmutex_t rx_rpc_stats;
270 void rxi_StartUnlocked();
273 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
274 ** pretty good that the next packet coming in is from the same connection
275 ** as the last packet, since we're send multiple packets in a transmit window.
277 struct rx_connection *rxLastConn = 0;
279 #ifdef RX_ENABLE_LOCKS
280 /* The locking hierarchy for rx fine grain locking is composed of these
283 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
284 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
285 * call->lock - locks call data fields.
286 * These are independent of each other:
287 * rx_freeCallQueue_lock
292 * serverQueueEntry->lock
294 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
295 * peer->lock - locks peer data fields.
296 * conn_data_lock - that more than one thread is not updating a conn data
297 * field at the same time.
305 * Do we need a lock to protect the peer field in the conn structure?
306 * conn->peer was previously a constant for all intents and so has no
307 * lock protecting this field. The multihomed client delta introduced
308 * a RX code change : change the peer field in the connection structure
309 * to that remote inetrface from which the last packet for this
310 * connection was sent out. This may become an issue if further changes
313 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
314 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
316 /* rxdb_fileID is used to identify the lock location, along with line#. */
317 static int rxdb_fileID = RXDB_FILE_RX;
318 #endif /* RX_LOCKS_DB */
319 #else /* RX_ENABLE_LOCKS */
320 #define SET_CALL_QUEUE_LOCK(C, L)
321 #define CLEAR_CALL_QUEUE_LOCK(C)
322 #endif /* RX_ENABLE_LOCKS */
323 struct rx_serverQueueEntry *rx_waitForPacket = 0;
324 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
326 /* ------------Exported Interfaces------------- */
328 /* This function allows rxkad to set the epoch to a suitably random number
329 * which rx_NewConnection will use in the future. The principle purpose is to
330 * get rxnull connections to use the same epoch as the rxkad connections do, at
331 * least once the first rxkad connection is established. This is important now
332 * that the host/port addresses aren't used in FindConnection: the uniqueness
333 * of epoch/cid matters and the start time won't do. */
335 #ifdef AFS_PTHREAD_ENV
337 * This mutex protects the following global variables:
341 #define LOCK_EPOCH assert(pthread_mutex_lock(&epoch_mutex)==0)
342 #define UNLOCK_EPOCH assert(pthread_mutex_unlock(&epoch_mutex)==0)
346 #endif /* AFS_PTHREAD_ENV */
349 rx_SetEpoch(afs_uint32 epoch)
356 /* Initialize rx. A port number may be mentioned, in which case this
357 * becomes the default port number for any service installed later.
358 * If 0 is provided for the port number, a random port will be chosen
359 * by the kernel. Whether this will ever overlap anything in
360 * /etc/services is anybody's guess... Returns 0 on success, -1 on
362 static int rxinit_status = 1;
363 #ifdef AFS_PTHREAD_ENV
365 * This mutex protects the following global variables:
369 #define LOCK_RX_INIT assert(pthread_mutex_lock(&rx_init_mutex)==0)
370 #define UNLOCK_RX_INIT assert(pthread_mutex_unlock(&rx_init_mutex)==0)
373 #define UNLOCK_RX_INIT
377 rx_InitHost(u_int host, u_int port)
384 char *htable, *ptable;
391 if (rxinit_status == 0) {
392 tmp_status = rxinit_status;
394 return tmp_status; /* Already started; return previous error code. */
400 if (afs_winsockInit() < 0)
406 * Initialize anything necessary to provide a non-premptive threading
409 rxi_InitializeThreadSupport();
412 /* Allocate and initialize a socket for client and perhaps server
415 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
416 if (rx_socket == OSI_NULLSOCKET) {
420 #ifdef RX_ENABLE_LOCKS
423 #endif /* RX_LOCKS_DB */
424 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
425 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
426 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
427 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
428 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
430 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
432 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
434 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
436 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
438 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
440 #if defined(KERNEL) && defined(AFS_HPUX110_ENV)
442 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
443 #endif /* KERNEL && AFS_HPUX110_ENV */
444 #endif /* RX_ENABLE_LOCKS */
447 rx_connDeadTime = 12;
448 rx_tranquil = 0; /* reset flag */
449 memset((char *)&rx_stats, 0, sizeof(struct rx_stats));
451 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
452 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
453 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
454 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
455 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
456 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
458 /* Malloc up a bunch of packets & buffers */
460 queue_Init(&rx_freePacketQueue);
461 rxi_NeedMorePackets = FALSE;
462 #ifdef RX_ENABLE_TSFPQ
463 rx_nPackets = 0; /* in TSFPQ version, rx_nPackets is managed by rxi_MorePackets* */
464 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
465 #else /* RX_ENABLE_TSFPQ */
466 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
467 rxi_MorePackets(rx_nPackets);
468 #endif /* RX_ENABLE_TSFPQ */
475 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
476 tv.tv_sec = clock_now.sec;
477 tv.tv_usec = clock_now.usec;
478 srand((unsigned int)tv.tv_usec);
485 #if defined(KERNEL) && !defined(UKERNEL)
486 /* Really, this should never happen in a real kernel */
489 struct sockaddr_in addr;
490 int addrlen = sizeof(addr);
491 if (getsockname((int)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
495 rx_port = addr.sin_port;
498 rx_stats.minRtt.sec = 9999999;
500 rx_SetEpoch(tv.tv_sec | 0x80000000);
502 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
503 * will provide a randomer value. */
505 MUTEX_ENTER(&rx_stats_mutex);
506 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
507 MUTEX_EXIT(&rx_stats_mutex);
508 /* *Slightly* random start time for the cid. This is just to help
509 * out with the hashing function at the peer */
510 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
511 rx_connHashTable = (struct rx_connection **)htable;
512 rx_peerHashTable = (struct rx_peer **)ptable;
514 rx_lastAckDelay.sec = 0;
515 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
516 rx_hardAckDelay.sec = 0;
517 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
518 rx_softAckDelay.sec = 0;
519 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
521 rxevent_Init(20, rxi_ReScheduleEvents);
523 /* Initialize various global queues */
524 queue_Init(&rx_idleServerQueue);
525 queue_Init(&rx_incomingCallQueue);
526 queue_Init(&rx_freeCallQueue);
528 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
529 /* Initialize our list of usable IP addresses. */
533 /* Start listener process (exact function is dependent on the
534 * implementation environment--kernel or user space) */
538 tmp_status = rxinit_status = 0;
546 return rx_InitHost(htonl(INADDR_ANY), port);
549 /* called with unincremented nRequestsRunning to see if it is OK to start
550 * a new thread in this service. Could be "no" for two reasons: over the
551 * max quota, or would prevent others from reaching their min quota.
553 #ifdef RX_ENABLE_LOCKS
554 /* This verion of QuotaOK reserves quota if it's ok while the
555 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
558 QuotaOK(register struct rx_service *aservice)
560 /* check if over max quota */
561 if (aservice->nRequestsRunning >= aservice->maxProcs) {
565 /* under min quota, we're OK */
566 /* otherwise, can use only if there are enough to allow everyone
567 * to go to their min quota after this guy starts.
569 MUTEX_ENTER(&rx_stats_mutex);
570 if ((aservice->nRequestsRunning < aservice->minProcs)
571 || (rxi_availProcs > rxi_minDeficit)) {
572 aservice->nRequestsRunning++;
573 /* just started call in minProcs pool, need fewer to maintain
575 if (aservice->nRequestsRunning <= aservice->minProcs)
578 MUTEX_EXIT(&rx_stats_mutex);
581 MUTEX_EXIT(&rx_stats_mutex);
587 ReturnToServerPool(register struct rx_service *aservice)
589 aservice->nRequestsRunning--;
590 MUTEX_ENTER(&rx_stats_mutex);
591 if (aservice->nRequestsRunning < aservice->minProcs)
594 MUTEX_EXIT(&rx_stats_mutex);
597 #else /* RX_ENABLE_LOCKS */
599 QuotaOK(register struct rx_service *aservice)
602 /* under min quota, we're OK */
603 if (aservice->nRequestsRunning < aservice->minProcs)
606 /* check if over max quota */
607 if (aservice->nRequestsRunning >= aservice->maxProcs)
610 /* otherwise, can use only if there are enough to allow everyone
611 * to go to their min quota after this guy starts.
613 if (rxi_availProcs > rxi_minDeficit)
617 #endif /* RX_ENABLE_LOCKS */
620 /* Called by rx_StartServer to start up lwp's to service calls.
621 NExistingProcs gives the number of procs already existing, and which
622 therefore needn't be created. */
624 rxi_StartServerProcs(int nExistingProcs)
626 register struct rx_service *service;
631 /* For each service, reserve N processes, where N is the "minimum"
632 * number of processes that MUST be able to execute a request in parallel,
633 * at any time, for that process. Also compute the maximum difference
634 * between any service's maximum number of processes that can run
635 * (i.e. the maximum number that ever will be run, and a guarantee
636 * that this number will run if other services aren't running), and its
637 * minimum number. The result is the extra number of processes that
638 * we need in order to provide the latter guarantee */
639 for (i = 0; i < RX_MAX_SERVICES; i++) {
641 service = rx_services[i];
642 if (service == (struct rx_service *)0)
644 nProcs += service->minProcs;
645 diff = service->maxProcs - service->minProcs;
649 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
650 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
651 for (i = 0; i < nProcs; i++) {
652 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
658 /* This routine is only required on Windows */
660 rx_StartClientThread(void)
662 #ifdef AFS_PTHREAD_ENV
664 pid = (int) pthread_self();
665 #endif /* AFS_PTHREAD_ENV */
667 #endif /* AFS_NT40_ENV */
669 /* This routine must be called if any services are exported. If the
670 * donateMe flag is set, the calling process is donated to the server
673 rx_StartServer(int donateMe)
675 register struct rx_service *service;
681 /* Start server processes, if necessary (exact function is dependent
682 * on the implementation environment--kernel or user space). DonateMe
683 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
684 * case, one less new proc will be created rx_StartServerProcs.
686 rxi_StartServerProcs(donateMe);
688 /* count up the # of threads in minProcs, and add set the min deficit to
689 * be that value, too.
691 for (i = 0; i < RX_MAX_SERVICES; i++) {
692 service = rx_services[i];
693 if (service == (struct rx_service *)0)
695 MUTEX_ENTER(&rx_stats_mutex);
696 rxi_totalMin += service->minProcs;
697 /* below works even if a thread is running, since minDeficit would
698 * still have been decremented and later re-incremented.
700 rxi_minDeficit += service->minProcs;
701 MUTEX_EXIT(&rx_stats_mutex);
704 /* Turn on reaping of idle server connections */
705 rxi_ReapConnections();
714 #ifdef AFS_PTHREAD_ENV
716 pid = (pid_t) pthread_self();
717 #else /* AFS_PTHREAD_ENV */
719 LWP_CurrentProcess(&pid);
720 #endif /* AFS_PTHREAD_ENV */
722 sprintf(name, "srv_%d", ++nProcs);
724 (*registerProgram) (pid, name);
726 #endif /* AFS_NT40_ENV */
727 rx_ServerProc(); /* Never returns */
729 #ifdef RX_ENABLE_TSFPQ
730 /* no use leaving packets around in this thread's local queue if
731 * it isn't getting donated to the server thread pool.
733 rxi_FlushLocalPacketsTSFPQ();
734 #endif /* RX_ENABLE_TSFPQ */
738 /* Create a new client connection to the specified service, using the
739 * specified security object to implement the security model for this
741 struct rx_connection *
742 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
743 register struct rx_securityClass *securityObject,
744 int serviceSecurityIndex)
748 register struct rx_connection *conn;
753 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", ntohl(shost), ntohs(sport), sservice, securityObject, serviceSecurityIndex));
755 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
756 * the case of kmem_alloc? */
757 conn = rxi_AllocConnection();
758 #ifdef RX_ENABLE_LOCKS
759 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
760 MUTEX_INIT(&conn->conn_data_lock, "conn call lock", MUTEX_DEFAULT, 0);
761 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
764 MUTEX_ENTER(&rx_connHashTable_lock);
765 cid = (rx_nextCid += RX_MAXCALLS);
766 conn->type = RX_CLIENT_CONNECTION;
768 conn->epoch = rx_epoch;
769 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
770 conn->serviceId = sservice;
771 conn->securityObject = securityObject;
772 /* This doesn't work in all compilers with void (they're buggy), so fake it
774 conn->securityData = (VOID *) 0;
775 conn->securityIndex = serviceSecurityIndex;
776 rx_SetConnDeadTime(conn, rx_connDeadTime);
777 conn->ackRate = RX_FAST_ACK_RATE;
779 conn->specific = NULL;
780 conn->challengeEvent = NULL;
781 conn->delayedAbortEvent = NULL;
782 conn->abortCount = 0;
785 RXS_NewConnection(securityObject, conn);
787 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
789 conn->refCount++; /* no lock required since only this thread knows... */
790 conn->next = rx_connHashTable[hashindex];
791 rx_connHashTable[hashindex] = conn;
792 MUTEX_ENTER(&rx_stats_mutex);
793 rx_stats.nClientConns++;
794 MUTEX_EXIT(&rx_stats_mutex);
796 MUTEX_EXIT(&rx_connHashTable_lock);
802 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
804 /* The idea is to set the dead time to a value that allows several
805 * keepalives to be dropped without timing out the connection. */
806 conn->secondsUntilDead = MAX(seconds, 6);
807 conn->secondsUntilPing = conn->secondsUntilDead / 6;
810 int rxi_lowPeerRefCount = 0;
811 int rxi_lowConnRefCount = 0;
814 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
815 * NOTE: must not be called with rx_connHashTable_lock held.
818 rxi_CleanupConnection(struct rx_connection *conn)
820 /* Notify the service exporter, if requested, that this connection
821 * is being destroyed */
822 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
823 (*conn->service->destroyConnProc) (conn);
825 /* Notify the security module that this connection is being destroyed */
826 RXS_DestroyConnection(conn->securityObject, conn);
828 /* If this is the last connection using the rx_peer struct, set its
829 * idle time to now. rxi_ReapConnections will reap it if it's still
830 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
832 MUTEX_ENTER(&rx_peerHashTable_lock);
833 if (conn->peer->refCount < 2) {
834 conn->peer->idleWhen = clock_Sec();
835 if (conn->peer->refCount < 1) {
836 conn->peer->refCount = 1;
837 MUTEX_ENTER(&rx_stats_mutex);
838 rxi_lowPeerRefCount++;
839 MUTEX_EXIT(&rx_stats_mutex);
842 conn->peer->refCount--;
843 MUTEX_EXIT(&rx_peerHashTable_lock);
845 MUTEX_ENTER(&rx_stats_mutex);
846 if (conn->type == RX_SERVER_CONNECTION)
847 rx_stats.nServerConns--;
849 rx_stats.nClientConns--;
850 MUTEX_EXIT(&rx_stats_mutex);
853 if (conn->specific) {
855 for (i = 0; i < conn->nSpecific; i++) {
856 if (conn->specific[i] && rxi_keyCreate_destructor[i])
857 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
858 conn->specific[i] = NULL;
860 free(conn->specific);
862 conn->specific = NULL;
866 MUTEX_DESTROY(&conn->conn_call_lock);
867 MUTEX_DESTROY(&conn->conn_data_lock);
868 CV_DESTROY(&conn->conn_call_cv);
870 rxi_FreeConnection(conn);
873 /* Destroy the specified connection */
875 rxi_DestroyConnection(register struct rx_connection *conn)
877 MUTEX_ENTER(&rx_connHashTable_lock);
878 rxi_DestroyConnectionNoLock(conn);
879 /* conn should be at the head of the cleanup list */
880 if (conn == rx_connCleanup_list) {
881 rx_connCleanup_list = rx_connCleanup_list->next;
882 MUTEX_EXIT(&rx_connHashTable_lock);
883 rxi_CleanupConnection(conn);
885 #ifdef RX_ENABLE_LOCKS
887 MUTEX_EXIT(&rx_connHashTable_lock);
889 #endif /* RX_ENABLE_LOCKS */
893 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
895 register struct rx_connection **conn_ptr;
896 register int havecalls = 0;
897 struct rx_packet *packet;
904 MUTEX_ENTER(&conn->conn_data_lock);
905 if (conn->refCount > 0)
908 MUTEX_ENTER(&rx_stats_mutex);
909 rxi_lowConnRefCount++;
910 MUTEX_EXIT(&rx_stats_mutex);
913 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
914 /* Busy; wait till the last guy before proceeding */
915 MUTEX_EXIT(&conn->conn_data_lock);
920 /* If the client previously called rx_NewCall, but it is still
921 * waiting, treat this as a running call, and wait to destroy the
922 * connection later when the call completes. */
923 if ((conn->type == RX_CLIENT_CONNECTION)
924 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
925 conn->flags |= RX_CONN_DESTROY_ME;
926 MUTEX_EXIT(&conn->conn_data_lock);
930 MUTEX_EXIT(&conn->conn_data_lock);
932 /* Check for extant references to this connection */
933 for (i = 0; i < RX_MAXCALLS; i++) {
934 register struct rx_call *call = conn->call[i];
937 if (conn->type == RX_CLIENT_CONNECTION) {
938 MUTEX_ENTER(&call->lock);
939 if (call->delayedAckEvent) {
940 /* Push the final acknowledgment out now--there
941 * won't be a subsequent call to acknowledge the
942 * last reply packets */
943 rxevent_Cancel(call->delayedAckEvent, call,
944 RX_CALL_REFCOUNT_DELAY);
945 if (call->state == RX_STATE_PRECALL
946 || call->state == RX_STATE_ACTIVE) {
947 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
949 rxi_AckAll(NULL, call, 0);
952 MUTEX_EXIT(&call->lock);
956 #ifdef RX_ENABLE_LOCKS
958 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
959 MUTEX_EXIT(&conn->conn_data_lock);
961 /* Someone is accessing a packet right now. */
965 #endif /* RX_ENABLE_LOCKS */
968 /* Don't destroy the connection if there are any call
969 * structures still in use */
970 MUTEX_ENTER(&conn->conn_data_lock);
971 conn->flags |= RX_CONN_DESTROY_ME;
972 MUTEX_EXIT(&conn->conn_data_lock);
977 if (conn->delayedAbortEvent) {
978 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
979 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
981 MUTEX_ENTER(&conn->conn_data_lock);
982 rxi_SendConnectionAbort(conn, packet, 0, 1);
983 MUTEX_EXIT(&conn->conn_data_lock);
984 rxi_FreePacket(packet);
988 /* Remove from connection hash table before proceeding */
990 &rx_connHashTable[CONN_HASH
991 (peer->host, peer->port, conn->cid, conn->epoch,
993 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
994 if (*conn_ptr == conn) {
995 *conn_ptr = conn->next;
999 /* if the conn that we are destroying was the last connection, then we
1000 * clear rxLastConn as well */
1001 if (rxLastConn == conn)
1004 /* Make sure the connection is completely reset before deleting it. */
1005 /* get rid of pending events that could zap us later */
1006 if (conn->challengeEvent)
1007 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1008 if (conn->checkReachEvent)
1009 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1011 /* Add the connection to the list of destroyed connections that
1012 * need to be cleaned up. This is necessary to avoid deadlocks
1013 * in the routines we call to inform others that this connection is
1014 * being destroyed. */
1015 conn->next = rx_connCleanup_list;
1016 rx_connCleanup_list = conn;
1019 /* Externally available version */
1021 rx_DestroyConnection(register struct rx_connection *conn)
1026 rxi_DestroyConnection(conn);
1031 rx_GetConnection(register struct rx_connection *conn)
1036 MUTEX_ENTER(&conn->conn_data_lock);
1038 MUTEX_EXIT(&conn->conn_data_lock);
1042 /* Start a new rx remote procedure call, on the specified connection.
1043 * If wait is set to 1, wait for a free call channel; otherwise return
1044 * 0. Maxtime gives the maximum number of seconds this call may take,
1045 * after rx_MakeCall returns. After this time interval, a call to any
1046 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1047 * For fine grain locking, we hold the conn_call_lock in order to
1048 * to ensure that we don't get signalle after we found a call in an active
1049 * state and before we go to sleep.
1052 rx_NewCall(register struct rx_connection *conn)
1055 register struct rx_call *call;
1056 struct clock queueTime;
1060 dpf(("rx_MakeCall(conn %x)\n", conn));
1063 clock_GetTime(&queueTime);
1064 MUTEX_ENTER(&conn->conn_call_lock);
1067 * Check if there are others waiting for a new call.
1068 * If so, let them go first to avoid starving them.
1069 * This is a fairly simple scheme, and might not be
1070 * a complete solution for large numbers of waiters.
1072 * makeCallWaiters keeps track of the number of
1073 * threads waiting to make calls and the
1074 * RX_CONN_MAKECALL_WAITING flag bit is used to
1075 * indicate that there are indeed calls waiting.
1076 * The flag is set when the waiter is incremented.
1077 * It is only cleared in rx_EndCall when
1078 * makeCallWaiters is 0. This prevents us from
1079 * accidently destroying the connection while it
1080 * is potentially about to be used.
1082 MUTEX_ENTER(&conn->conn_data_lock);
1083 if (conn->makeCallWaiters) {
1084 conn->flags |= RX_CONN_MAKECALL_WAITING;
1085 conn->makeCallWaiters++;
1086 MUTEX_EXIT(&conn->conn_data_lock);
1088 #ifdef RX_ENABLE_LOCKS
1089 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1093 MUTEX_ENTER(&conn->conn_data_lock);
1094 conn->makeCallWaiters--;
1096 MUTEX_EXIT(&conn->conn_data_lock);
1099 for (i = 0; i < RX_MAXCALLS; i++) {
1100 call = conn->call[i];
1102 MUTEX_ENTER(&call->lock);
1103 if (call->state == RX_STATE_DALLY) {
1104 rxi_ResetCall(call, 0);
1105 (*call->callNumber)++;
1108 MUTEX_EXIT(&call->lock);
1110 call = rxi_NewCall(conn, i);
1114 if (i < RX_MAXCALLS) {
1117 MUTEX_ENTER(&conn->conn_data_lock);
1118 conn->flags |= RX_CONN_MAKECALL_WAITING;
1119 conn->makeCallWaiters++;
1120 MUTEX_EXIT(&conn->conn_data_lock);
1122 #ifdef RX_ENABLE_LOCKS
1123 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1127 MUTEX_ENTER(&conn->conn_data_lock);
1128 conn->makeCallWaiters--;
1129 MUTEX_EXIT(&conn->conn_data_lock);
1132 * Wake up anyone else who might be giving us a chance to
1133 * run (see code above that avoids resource starvation).
1135 #ifdef RX_ENABLE_LOCKS
1136 CV_BROADCAST(&conn->conn_call_cv);
1141 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1143 /* Client is initially in send mode */
1144 call->state = RX_STATE_ACTIVE;
1145 call->error = conn->error;
1147 call->mode = RX_MODE_ERROR;
1149 call->mode = RX_MODE_SENDING;
1151 /* remember start time for call in case we have hard dead time limit */
1152 call->queueTime = queueTime;
1153 clock_GetTime(&call->startTime);
1154 hzero(call->bytesSent);
1155 hzero(call->bytesRcvd);
1157 /* Turn on busy protocol. */
1158 rxi_KeepAliveOn(call);
1160 MUTEX_EXIT(&call->lock);
1161 MUTEX_EXIT(&conn->conn_call_lock);
1164 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1165 /* Now, if TQ wasn't cleared earlier, do it now. */
1166 MUTEX_ENTER(&call->lock);
1167 while (call->flags & RX_CALL_TQ_BUSY) {
1168 call->flags |= RX_CALL_TQ_WAIT;
1170 #ifdef RX_ENABLE_LOCKS
1171 osirx_AssertMine(&call->lock, "rxi_Start lock4");
1172 CV_WAIT(&call->cv_tq, &call->lock);
1173 #else /* RX_ENABLE_LOCKS */
1174 osi_rxSleep(&call->tq);
1175 #endif /* RX_ENABLE_LOCKS */
1177 if (call->tqWaiters == 0) {
1178 call->flags &= ~RX_CALL_TQ_WAIT;
1181 if (call->flags & RX_CALL_TQ_CLEARME) {
1182 rxi_ClearTransmitQueue(call, 0);
1183 queue_Init(&call->tq);
1185 MUTEX_EXIT(&call->lock);
1186 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1192 rxi_HasActiveCalls(register struct rx_connection *aconn)
1195 register struct rx_call *tcall;
1199 for (i = 0; i < RX_MAXCALLS; i++) {
1200 if ((tcall = aconn->call[i])) {
1201 if ((tcall->state == RX_STATE_ACTIVE)
1202 || (tcall->state == RX_STATE_PRECALL)) {
1213 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1214 register afs_int32 * aint32s)
1217 register struct rx_call *tcall;
1221 for (i = 0; i < RX_MAXCALLS; i++) {
1222 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1223 aint32s[i] = aconn->callNumber[i] + 1;
1225 aint32s[i] = aconn->callNumber[i];
1232 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1233 register afs_int32 * aint32s)
1236 register struct rx_call *tcall;
1240 for (i = 0; i < RX_MAXCALLS; i++) {
1241 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1242 aconn->callNumber[i] = aint32s[i] - 1;
1244 aconn->callNumber[i] = aint32s[i];
1250 /* Advertise a new service. A service is named locally by a UDP port
1251 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1254 char *serviceName; Name for identification purposes (e.g. the
1255 service name might be used for probing for
1258 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1259 char *serviceName, struct rx_securityClass **securityObjects,
1260 int nSecurityObjects,
1261 afs_int32(*serviceProc) (struct rx_call * acall))
1263 osi_socket socket = OSI_NULLSOCKET;
1264 register struct rx_service *tservice;
1270 if (serviceId == 0) {
1272 "rx_NewService: service id for service %s is not non-zero.\n",
1279 "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",
1287 tservice = rxi_AllocService();
1289 for (i = 0; i < RX_MAX_SERVICES; i++) {
1290 register struct rx_service *service = rx_services[i];
1292 if (port == service->servicePort && host == service->serviceHost) {
1293 if (service->serviceId == serviceId) {
1294 /* The identical service has already been
1295 * installed; if the caller was intending to
1296 * change the security classes used by this
1297 * service, he/she loses. */
1299 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1300 serviceName, serviceId, service->serviceName);
1302 rxi_FreeService(tservice);
1305 /* Different service, same port: re-use the socket
1306 * which is bound to the same port */
1307 socket = service->socket;
1310 if (socket == OSI_NULLSOCKET) {
1311 /* If we don't already have a socket (from another
1312 * service on same port) get a new one */
1313 socket = rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
1314 if (socket == OSI_NULLSOCKET) {
1316 rxi_FreeService(tservice);
1321 service->socket = socket;
1322 service->serviceHost = host;
1323 service->servicePort = port;
1324 service->serviceId = serviceId;
1325 service->serviceName = serviceName;
1326 service->nSecurityObjects = nSecurityObjects;
1327 service->securityObjects = securityObjects;
1328 service->minProcs = 0;
1329 service->maxProcs = 1;
1330 service->idleDeadTime = 60;
1331 service->connDeadTime = rx_connDeadTime;
1332 service->executeRequestProc = serviceProc;
1333 service->checkReach = 0;
1334 rx_services[i] = service; /* not visible until now */
1340 rxi_FreeService(tservice);
1341 (osi_Msg "rx_NewService: cannot support > %d services\n",
1347 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1348 struct rx_securityClass **securityObjects, int nSecurityObjects,
1349 afs_int32(*serviceProc) (struct rx_call * acall))
1351 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1354 /* Generic request processing loop. This routine should be called
1355 * by the implementation dependent rx_ServerProc. If socketp is
1356 * non-null, it will be set to the file descriptor that this thread
1357 * is now listening on. If socketp is null, this routine will never
1360 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1362 register struct rx_call *call;
1363 register afs_int32 code;
1364 register struct rx_service *tservice = NULL;
1371 call = rx_GetCall(threadID, tservice, socketp);
1372 if (socketp && *socketp != OSI_NULLSOCKET) {
1373 /* We are now a listener thread */
1378 /* if server is restarting( typically smooth shutdown) then do not
1379 * allow any new calls.
1382 if (rx_tranquil && (call != NULL)) {
1386 MUTEX_ENTER(&call->lock);
1388 rxi_CallError(call, RX_RESTARTING);
1389 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1391 MUTEX_EXIT(&call->lock);
1395 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1396 #ifdef RX_ENABLE_LOCKS
1398 #endif /* RX_ENABLE_LOCKS */
1399 afs_termState = AFSOP_STOP_AFS;
1400 afs_osi_Wakeup(&afs_termState);
1401 #ifdef RX_ENABLE_LOCKS
1403 #endif /* RX_ENABLE_LOCKS */
1408 tservice = call->conn->service;
1410 if (tservice->beforeProc)
1411 (*tservice->beforeProc) (call);
1413 code = call->conn->service->executeRequestProc(call);
1415 if (tservice->afterProc)
1416 (*tservice->afterProc) (call, code);
1418 rx_EndCall(call, code);
1419 MUTEX_ENTER(&rx_stats_mutex);
1421 MUTEX_EXIT(&rx_stats_mutex);
1427 rx_WakeupServerProcs(void)
1429 struct rx_serverQueueEntry *np, *tqp;
1433 MUTEX_ENTER(&rx_serverPool_lock);
1435 #ifdef RX_ENABLE_LOCKS
1436 if (rx_waitForPacket)
1437 CV_BROADCAST(&rx_waitForPacket->cv);
1438 #else /* RX_ENABLE_LOCKS */
1439 if (rx_waitForPacket)
1440 osi_rxWakeup(rx_waitForPacket);
1441 #endif /* RX_ENABLE_LOCKS */
1442 MUTEX_ENTER(&freeSQEList_lock);
1443 for (np = rx_FreeSQEList; np; np = tqp) {
1444 tqp = *(struct rx_serverQueueEntry **)np;
1445 #ifdef RX_ENABLE_LOCKS
1446 CV_BROADCAST(&np->cv);
1447 #else /* RX_ENABLE_LOCKS */
1449 #endif /* RX_ENABLE_LOCKS */
1451 MUTEX_EXIT(&freeSQEList_lock);
1452 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1453 #ifdef RX_ENABLE_LOCKS
1454 CV_BROADCAST(&np->cv);
1455 #else /* RX_ENABLE_LOCKS */
1457 #endif /* RX_ENABLE_LOCKS */
1459 MUTEX_EXIT(&rx_serverPool_lock);
1464 * One thing that seems to happen is that all the server threads get
1465 * tied up on some empty or slow call, and then a whole bunch of calls
1466 * arrive at once, using up the packet pool, so now there are more
1467 * empty calls. The most critical resources here are server threads
1468 * and the free packet pool. The "doreclaim" code seems to help in
1469 * general. I think that eventually we arrive in this state: there
1470 * are lots of pending calls which do have all their packets present,
1471 * so they won't be reclaimed, are multi-packet calls, so they won't
1472 * be scheduled until later, and thus are tying up most of the free
1473 * packet pool for a very long time.
1475 * 1. schedule multi-packet calls if all the packets are present.
1476 * Probably CPU-bound operation, useful to return packets to pool.
1477 * Do what if there is a full window, but the last packet isn't here?
1478 * 3. preserve one thread which *only* runs "best" calls, otherwise
1479 * it sleeps and waits for that type of call.
1480 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1481 * the current dataquota business is badly broken. The quota isn't adjusted
1482 * to reflect how many packets are presently queued for a running call.
1483 * So, when we schedule a queued call with a full window of packets queued
1484 * up for it, that *should* free up a window full of packets for other 2d-class
1485 * calls to be able to use from the packet pool. But it doesn't.
1487 * NB. Most of the time, this code doesn't run -- since idle server threads
1488 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1489 * as a new call arrives.
1491 /* Sleep until a call arrives. Returns a pointer to the call, ready
1492 * for an rx_Read. */
1493 #ifdef RX_ENABLE_LOCKS
1495 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1497 struct rx_serverQueueEntry *sq;
1498 register struct rx_call *call = (struct rx_call *)0;
1499 struct rx_service *service = NULL;
1502 MUTEX_ENTER(&freeSQEList_lock);
1504 if ((sq = rx_FreeSQEList)) {
1505 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1506 MUTEX_EXIT(&freeSQEList_lock);
1507 } else { /* otherwise allocate a new one and return that */
1508 MUTEX_EXIT(&freeSQEList_lock);
1509 sq = (struct rx_serverQueueEntry *)
1510 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1511 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1512 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1515 MUTEX_ENTER(&rx_serverPool_lock);
1516 if (cur_service != NULL) {
1517 ReturnToServerPool(cur_service);
1520 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1521 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1523 /* Scan for eligible incoming calls. A call is not eligible
1524 * if the maximum number of calls for its service type are
1525 * already executing */
1526 /* One thread will process calls FCFS (to prevent starvation),
1527 * while the other threads may run ahead looking for calls which
1528 * have all their input data available immediately. This helps
1529 * keep threads from blocking, waiting for data from the client. */
1530 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1531 service = tcall->conn->service;
1532 if (!QuotaOK(service)) {
1535 if (tno == rxi_fcfs_thread_num
1536 || !tcall->queue_item_header.next) {
1537 /* If we're the fcfs thread , then we'll just use
1538 * this call. If we haven't been able to find an optimal
1539 * choice, and we're at the end of the list, then use a
1540 * 2d choice if one has been identified. Otherwise... */
1541 call = (choice2 ? choice2 : tcall);
1542 service = call->conn->service;
1543 } else if (!queue_IsEmpty(&tcall->rq)) {
1544 struct rx_packet *rp;
1545 rp = queue_First(&tcall->rq, rx_packet);
1546 if (rp->header.seq == 1) {
1548 || (rp->header.flags & RX_LAST_PACKET)) {
1550 } else if (rxi_2dchoice && !choice2
1551 && !(tcall->flags & RX_CALL_CLEARED)
1552 && (tcall->rprev > rxi_HardAckRate)) {
1561 ReturnToServerPool(service);
1568 MUTEX_EXIT(&rx_serverPool_lock);
1569 MUTEX_ENTER(&call->lock);
1571 if (call->flags & RX_CALL_WAIT_PROC) {
1572 call->flags &= ~RX_CALL_WAIT_PROC;
1573 MUTEX_ENTER(&rx_stats_mutex);
1575 MUTEX_EXIT(&rx_stats_mutex);
1578 if (call->state != RX_STATE_PRECALL || call->error) {
1579 MUTEX_EXIT(&call->lock);
1580 MUTEX_ENTER(&rx_serverPool_lock);
1581 ReturnToServerPool(service);
1586 if (queue_IsEmpty(&call->rq)
1587 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1588 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1590 CLEAR_CALL_QUEUE_LOCK(call);
1593 /* If there are no eligible incoming calls, add this process
1594 * to the idle server queue, to wait for one */
1598 *socketp = OSI_NULLSOCKET;
1600 sq->socketp = socketp;
1601 queue_Append(&rx_idleServerQueue, sq);
1602 #ifndef AFS_AIX41_ENV
1603 rx_waitForPacket = sq;
1605 rx_waitingForPacket = sq;
1606 #endif /* AFS_AIX41_ENV */
1608 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1610 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1611 MUTEX_EXIT(&rx_serverPool_lock);
1612 return (struct rx_call *)0;
1615 } while (!(call = sq->newcall)
1616 && !(socketp && *socketp != OSI_NULLSOCKET));
1617 MUTEX_EXIT(&rx_serverPool_lock);
1619 MUTEX_ENTER(&call->lock);
1625 MUTEX_ENTER(&freeSQEList_lock);
1626 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1627 rx_FreeSQEList = sq;
1628 MUTEX_EXIT(&freeSQEList_lock);
1631 clock_GetTime(&call->startTime);
1632 call->state = RX_STATE_ACTIVE;
1633 call->mode = RX_MODE_RECEIVING;
1634 #ifdef RX_KERNEL_TRACE
1635 if (ICL_SETACTIVE(afs_iclSetp)) {
1636 int glockOwner = ISAFS_GLOCK();
1639 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1640 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1647 rxi_calltrace(RX_CALL_START, call);
1648 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1649 call->conn->service->servicePort, call->conn->service->serviceId,
1652 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1653 MUTEX_EXIT(&call->lock);
1655 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1660 #else /* RX_ENABLE_LOCKS */
1662 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1664 struct rx_serverQueueEntry *sq;
1665 register struct rx_call *call = (struct rx_call *)0, *choice2;
1666 struct rx_service *service = NULL;
1670 MUTEX_ENTER(&freeSQEList_lock);
1672 if ((sq = rx_FreeSQEList)) {
1673 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1674 MUTEX_EXIT(&freeSQEList_lock);
1675 } else { /* otherwise allocate a new one and return that */
1676 MUTEX_EXIT(&freeSQEList_lock);
1677 sq = (struct rx_serverQueueEntry *)
1678 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1679 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1680 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1682 MUTEX_ENTER(&sq->lock);
1684 if (cur_service != NULL) {
1685 cur_service->nRequestsRunning--;
1686 if (cur_service->nRequestsRunning < cur_service->minProcs)
1690 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1691 register struct rx_call *tcall, *ncall;
1692 /* Scan for eligible incoming calls. A call is not eligible
1693 * if the maximum number of calls for its service type are
1694 * already executing */
1695 /* One thread will process calls FCFS (to prevent starvation),
1696 * while the other threads may run ahead looking for calls which
1697 * have all their input data available immediately. This helps
1698 * keep threads from blocking, waiting for data from the client. */
1699 choice2 = (struct rx_call *)0;
1700 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1701 service = tcall->conn->service;
1702 if (QuotaOK(service)) {
1703 if (tno == rxi_fcfs_thread_num
1704 || !tcall->queue_item_header.next) {
1705 /* If we're the fcfs thread, then we'll just use
1706 * this call. If we haven't been able to find an optimal
1707 * choice, and we're at the end of the list, then use a
1708 * 2d choice if one has been identified. Otherwise... */
1709 call = (choice2 ? choice2 : tcall);
1710 service = call->conn->service;
1711 } else if (!queue_IsEmpty(&tcall->rq)) {
1712 struct rx_packet *rp;
1713 rp = queue_First(&tcall->rq, rx_packet);
1714 if (rp->header.seq == 1
1716 || (rp->header.flags & RX_LAST_PACKET))) {
1718 } else if (rxi_2dchoice && !choice2
1719 && !(tcall->flags & RX_CALL_CLEARED)
1720 && (tcall->rprev > rxi_HardAckRate)) {
1733 /* we can't schedule a call if there's no data!!! */
1734 /* send an ack if there's no data, if we're missing the
1735 * first packet, or we're missing something between first
1736 * and last -- there's a "hole" in the incoming data. */
1737 if (queue_IsEmpty(&call->rq)
1738 || queue_First(&call->rq, rx_packet)->header.seq != 1
1739 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1740 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1742 call->flags &= (~RX_CALL_WAIT_PROC);
1743 service->nRequestsRunning++;
1744 /* just started call in minProcs pool, need fewer to maintain
1746 if (service->nRequestsRunning <= service->minProcs)
1750 /* MUTEX_EXIT(&call->lock); */
1752 /* If there are no eligible incoming calls, add this process
1753 * to the idle server queue, to wait for one */
1756 *socketp = OSI_NULLSOCKET;
1758 sq->socketp = socketp;
1759 queue_Append(&rx_idleServerQueue, sq);
1763 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1765 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1766 return (struct rx_call *)0;
1769 } while (!(call = sq->newcall)
1770 && !(socketp && *socketp != OSI_NULLSOCKET));
1772 MUTEX_EXIT(&sq->lock);
1774 MUTEX_ENTER(&freeSQEList_lock);
1775 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1776 rx_FreeSQEList = sq;
1777 MUTEX_EXIT(&freeSQEList_lock);
1780 clock_GetTime(&call->startTime);
1781 call->state = RX_STATE_ACTIVE;
1782 call->mode = RX_MODE_RECEIVING;
1783 #ifdef RX_KERNEL_TRACE
1784 if (ICL_SETACTIVE(afs_iclSetp)) {
1785 int glockOwner = ISAFS_GLOCK();
1788 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1789 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1796 rxi_calltrace(RX_CALL_START, call);
1797 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1798 call->conn->service->servicePort, call->conn->service->serviceId,
1801 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1808 #endif /* RX_ENABLE_LOCKS */
1812 /* Establish a procedure to be called when a packet arrives for a
1813 * call. This routine will be called at most once after each call,
1814 * and will also be called if there is an error condition on the or
1815 * the call is complete. Used by multi rx to build a selection
1816 * function which determines which of several calls is likely to be a
1817 * good one to read from.
1818 * NOTE: the way this is currently implemented it is probably only a
1819 * good idea to (1) use it immediately after a newcall (clients only)
1820 * and (2) only use it once. Other uses currently void your warranty
1823 rx_SetArrivalProc(register struct rx_call *call,
1824 register void (*proc) (register struct rx_call * call,
1826 register int index),
1827 register VOID * handle, register int arg)
1829 call->arrivalProc = proc;
1830 call->arrivalProcHandle = handle;
1831 call->arrivalProcArg = arg;
1834 /* Call is finished (possibly prematurely). Return rc to the peer, if
1835 * appropriate, and return the final error code from the conversation
1839 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1841 register struct rx_connection *conn = call->conn;
1842 register struct rx_service *service;
1848 dpf(("rx_EndCall(call %x rc %d error %d abortCode %d)\n", call, rc, call->error, call->abortCode));
1851 MUTEX_ENTER(&call->lock);
1853 if (rc == 0 && call->error == 0) {
1854 call->abortCode = 0;
1855 call->abortCount = 0;
1858 call->arrivalProc = (void (*)())0;
1859 if (rc && call->error == 0) {
1860 rxi_CallError(call, rc);
1861 /* Send an abort message to the peer if this error code has
1862 * only just been set. If it was set previously, assume the
1863 * peer has already been sent the error code or will request it
1865 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1867 if (conn->type == RX_SERVER_CONNECTION) {
1868 /* Make sure reply or at least dummy reply is sent */
1869 if (call->mode == RX_MODE_RECEIVING) {
1870 rxi_WriteProc(call, 0, 0);
1872 if (call->mode == RX_MODE_SENDING) {
1873 rxi_FlushWrite(call);
1875 service = conn->service;
1876 rxi_calltrace(RX_CALL_END, call);
1877 /* Call goes to hold state until reply packets are acknowledged */
1878 if (call->tfirst + call->nSoftAcked < call->tnext) {
1879 call->state = RX_STATE_HOLD;
1881 call->state = RX_STATE_DALLY;
1882 rxi_ClearTransmitQueue(call, 0);
1883 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1884 rxevent_Cancel(call->keepAliveEvent, call,
1885 RX_CALL_REFCOUNT_ALIVE);
1887 } else { /* Client connection */
1889 /* Make sure server receives input packets, in the case where
1890 * no reply arguments are expected */
1891 if ((call->mode == RX_MODE_SENDING)
1892 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1893 (void)rxi_ReadProc(call, &dummy, 1);
1896 /* If we had an outstanding delayed ack, be nice to the server
1897 * and force-send it now.
1899 if (call->delayedAckEvent) {
1900 rxevent_Cancel(call->delayedAckEvent, call,
1901 RX_CALL_REFCOUNT_DELAY);
1902 call->delayedAckEvent = NULL;
1903 rxi_SendDelayedAck(NULL, call, NULL);
1906 /* We need to release the call lock since it's lower than the
1907 * conn_call_lock and we don't want to hold the conn_call_lock
1908 * over the rx_ReadProc call. The conn_call_lock needs to be held
1909 * here for the case where rx_NewCall is perusing the calls on
1910 * the connection structure. We don't want to signal until
1911 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1912 * have checked this call, found it active and by the time it
1913 * goes to sleep, will have missed the signal.
1915 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
1916 * there are threads waiting to use the conn object.
1918 MUTEX_EXIT(&call->lock);
1919 MUTEX_ENTER(&conn->conn_call_lock);
1920 MUTEX_ENTER(&call->lock);
1921 MUTEX_ENTER(&conn->conn_data_lock);
1922 conn->flags |= RX_CONN_BUSY;
1923 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1924 if (conn->makeCallWaiters == 0)
1925 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1926 MUTEX_EXIT(&conn->conn_data_lock);
1927 #ifdef RX_ENABLE_LOCKS
1928 CV_BROADCAST(&conn->conn_call_cv);
1933 #ifdef RX_ENABLE_LOCKS
1935 MUTEX_EXIT(&conn->conn_data_lock);
1937 #endif /* RX_ENABLE_LOCKS */
1938 call->state = RX_STATE_DALLY;
1940 error = call->error;
1942 /* currentPacket, nLeft, and NFree must be zeroed here, because
1943 * ResetCall cannot: ResetCall may be called at splnet(), in the
1944 * kernel version, and may interrupt the macros rx_Read or
1945 * rx_Write, which run at normal priority for efficiency. */
1946 if (call->currentPacket) {
1947 queue_Prepend(&call->iovq, call->currentPacket);
1948 call->currentPacket = (struct rx_packet *)0;
1951 call->nLeft = call->nFree = call->curlen = 0;
1953 /* Free any packets from the last call to ReadvProc/WritevProc */
1954 rxi_FreePackets(0, &call->iovq);
1956 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1957 MUTEX_EXIT(&call->lock);
1958 if (conn->type == RX_CLIENT_CONNECTION) {
1959 MUTEX_EXIT(&conn->conn_call_lock);
1960 conn->flags &= ~RX_CONN_BUSY;
1964 * Map errors to the local host's errno.h format.
1966 error = ntoh_syserr_conv(error);
1970 #if !defined(KERNEL)
1972 /* Call this routine when shutting down a server or client (especially
1973 * clients). This will allow Rx to gracefully garbage collect server
1974 * connections, and reduce the number of retries that a server might
1975 * make to a dead client.
1976 * This is not quite right, since some calls may still be ongoing and
1977 * we can't lock them to destroy them. */
1981 register struct rx_connection **conn_ptr, **conn_end;
1985 if (rxinit_status == 1) {
1987 return; /* Already shutdown. */
1989 rxi_DeleteCachedConnections();
1990 if (rx_connHashTable) {
1991 MUTEX_ENTER(&rx_connHashTable_lock);
1992 for (conn_ptr = &rx_connHashTable[0], conn_end =
1993 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
1995 struct rx_connection *conn, *next;
1996 for (conn = *conn_ptr; conn; conn = next) {
1998 if (conn->type == RX_CLIENT_CONNECTION) {
1999 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2001 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2002 #ifdef RX_ENABLE_LOCKS
2003 rxi_DestroyConnectionNoLock(conn);
2004 #else /* RX_ENABLE_LOCKS */
2005 rxi_DestroyConnection(conn);
2006 #endif /* RX_ENABLE_LOCKS */
2010 #ifdef RX_ENABLE_LOCKS
2011 while (rx_connCleanup_list) {
2012 struct rx_connection *conn;
2013 conn = rx_connCleanup_list;
2014 rx_connCleanup_list = rx_connCleanup_list->next;
2015 MUTEX_EXIT(&rx_connHashTable_lock);
2016 rxi_CleanupConnection(conn);
2017 MUTEX_ENTER(&rx_connHashTable_lock);
2019 MUTEX_EXIT(&rx_connHashTable_lock);
2020 #endif /* RX_ENABLE_LOCKS */
2025 afs_winsockCleanup();
2033 /* if we wakeup packet waiter too often, can get in loop with two
2034 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2036 rxi_PacketsUnWait(void)
2038 if (!rx_waitingForPackets) {
2042 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2043 return; /* still over quota */
2046 rx_waitingForPackets = 0;
2047 #ifdef RX_ENABLE_LOCKS
2048 CV_BROADCAST(&rx_waitingForPackets_cv);
2050 osi_rxWakeup(&rx_waitingForPackets);
2056 /* ------------------Internal interfaces------------------------- */
2058 /* Return this process's service structure for the
2059 * specified socket and service */
2061 rxi_FindService(register osi_socket socket, register u_short serviceId)
2063 register struct rx_service **sp;
2064 for (sp = &rx_services[0]; *sp; sp++) {
2065 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2071 /* Allocate a call structure, for the indicated channel of the
2072 * supplied connection. The mode and state of the call must be set by
2073 * the caller. Returns the call with mutex locked. */
2075 rxi_NewCall(register struct rx_connection *conn, register int channel)
2077 register struct rx_call *call;
2078 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2079 register struct rx_call *cp; /* Call pointer temp */
2080 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2081 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2083 /* Grab an existing call structure, or allocate a new one.
2084 * Existing call structures are assumed to have been left reset by
2086 MUTEX_ENTER(&rx_freeCallQueue_lock);
2088 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2090 * EXCEPT that the TQ might not yet be cleared out.
2091 * Skip over those with in-use TQs.
2094 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2095 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2101 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2102 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2103 call = queue_First(&rx_freeCallQueue, rx_call);
2104 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2106 MUTEX_ENTER(&rx_stats_mutex);
2107 rx_stats.nFreeCallStructs--;
2108 MUTEX_EXIT(&rx_stats_mutex);
2109 MUTEX_EXIT(&rx_freeCallQueue_lock);
2110 MUTEX_ENTER(&call->lock);
2111 CLEAR_CALL_QUEUE_LOCK(call);
2112 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2113 /* Now, if TQ wasn't cleared earlier, do it now. */
2114 if (call->flags & RX_CALL_TQ_CLEARME) {
2115 rxi_ClearTransmitQueue(call, 0);
2116 queue_Init(&call->tq);
2118 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2119 /* Bind the call to its connection structure */
2121 rxi_ResetCall(call, 1);
2123 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2125 MUTEX_EXIT(&rx_freeCallQueue_lock);
2126 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2127 MUTEX_ENTER(&call->lock);
2128 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2129 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2130 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2132 MUTEX_ENTER(&rx_stats_mutex);
2133 rx_stats.nCallStructs++;
2134 MUTEX_EXIT(&rx_stats_mutex);
2135 /* Initialize once-only items */
2136 queue_Init(&call->tq);
2137 queue_Init(&call->rq);
2138 queue_Init(&call->iovq);
2139 /* Bind the call to its connection structure (prereq for reset) */
2141 rxi_ResetCall(call, 1);
2143 call->channel = channel;
2144 call->callNumber = &conn->callNumber[channel];
2145 /* Note that the next expected call number is retained (in
2146 * conn->callNumber[i]), even if we reallocate the call structure
2148 conn->call[channel] = call;
2149 /* if the channel's never been used (== 0), we should start at 1, otherwise
2150 * the call number is valid from the last time this channel was used */
2151 if (*call->callNumber == 0)
2152 *call->callNumber = 1;
2157 /* A call has been inactive long enough that so we can throw away
2158 * state, including the call structure, which is placed on the call
2160 * Call is locked upon entry.
2161 * haveCTLock set if called from rxi_ReapConnections
2163 #ifdef RX_ENABLE_LOCKS
2165 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2166 #else /* RX_ENABLE_LOCKS */
2168 rxi_FreeCall(register struct rx_call *call)
2169 #endif /* RX_ENABLE_LOCKS */
2171 register int channel = call->channel;
2172 register struct rx_connection *conn = call->conn;
2175 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2176 (*call->callNumber)++;
2177 rxi_ResetCall(call, 0);
2178 call->conn->call[channel] = (struct rx_call *)0;
2180 MUTEX_ENTER(&rx_freeCallQueue_lock);
2181 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2182 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2183 /* A call may be free even though its transmit queue is still in use.
2184 * Since we search the call list from head to tail, put busy calls at
2185 * the head of the list, and idle calls at the tail.
2187 if (call->flags & RX_CALL_TQ_BUSY)
2188 queue_Prepend(&rx_freeCallQueue, call);
2190 queue_Append(&rx_freeCallQueue, call);
2191 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2192 queue_Append(&rx_freeCallQueue, call);
2193 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2194 MUTEX_ENTER(&rx_stats_mutex);
2195 rx_stats.nFreeCallStructs++;
2196 MUTEX_EXIT(&rx_stats_mutex);
2198 MUTEX_EXIT(&rx_freeCallQueue_lock);
2200 /* Destroy the connection if it was previously slated for
2201 * destruction, i.e. the Rx client code previously called
2202 * rx_DestroyConnection (client connections), or
2203 * rxi_ReapConnections called the same routine (server
2204 * connections). Only do this, however, if there are no
2205 * outstanding calls. Note that for fine grain locking, there appears
2206 * to be a deadlock in that rxi_FreeCall has a call locked and
2207 * DestroyConnectionNoLock locks each call in the conn. But note a
2208 * few lines up where we have removed this call from the conn.
2209 * If someone else destroys a connection, they either have no
2210 * call lock held or are going through this section of code.
2212 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2213 MUTEX_ENTER(&conn->conn_data_lock);
2215 MUTEX_EXIT(&conn->conn_data_lock);
2216 #ifdef RX_ENABLE_LOCKS
2218 rxi_DestroyConnectionNoLock(conn);
2220 rxi_DestroyConnection(conn);
2221 #else /* RX_ENABLE_LOCKS */
2222 rxi_DestroyConnection(conn);
2223 #endif /* RX_ENABLE_LOCKS */
2227 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2229 rxi_Alloc(register size_t size)
2233 MUTEX_ENTER(&rx_stats_mutex);
2235 rxi_Allocsize += (afs_int32)size;
2236 MUTEX_EXIT(&rx_stats_mutex);
2238 p = (char *)osi_Alloc(size);
2241 osi_Panic("rxi_Alloc error");
2247 rxi_Free(void *addr, register size_t size)
2249 MUTEX_ENTER(&rx_stats_mutex);
2251 rxi_Allocsize -= (afs_int32)size;
2252 MUTEX_EXIT(&rx_stats_mutex);
2254 osi_Free(addr, size);
2257 /* Find the peer process represented by the supplied (host,port)
2258 * combination. If there is no appropriate active peer structure, a
2259 * new one will be allocated and initialized
2260 * The origPeer, if set, is a pointer to a peer structure on which the
2261 * refcount will be be decremented. This is used to replace the peer
2262 * structure hanging off a connection structure */
2264 rxi_FindPeer(register afs_uint32 host, register u_short port,
2265 struct rx_peer *origPeer, int create)
2267 register struct rx_peer *pp;
2269 hashIndex = PEER_HASH(host, port);
2270 MUTEX_ENTER(&rx_peerHashTable_lock);
2271 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2272 if ((pp->host == host) && (pp->port == port))
2277 pp = rxi_AllocPeer(); /* This bzero's *pp */
2278 pp->host = host; /* set here or in InitPeerParams is zero */
2280 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2281 queue_Init(&pp->congestionQueue);
2282 queue_Init(&pp->rpcStats);
2283 pp->next = rx_peerHashTable[hashIndex];
2284 rx_peerHashTable[hashIndex] = pp;
2285 rxi_InitPeerParams(pp);
2286 MUTEX_ENTER(&rx_stats_mutex);
2287 rx_stats.nPeerStructs++;
2288 MUTEX_EXIT(&rx_stats_mutex);
2295 origPeer->refCount--;
2296 MUTEX_EXIT(&rx_peerHashTable_lock);
2301 /* Find the connection at (host, port) started at epoch, and with the
2302 * given connection id. Creates the server connection if necessary.
2303 * The type specifies whether a client connection or a server
2304 * connection is desired. In both cases, (host, port) specify the
2305 * peer's (host, pair) pair. Client connections are not made
2306 * automatically by this routine. The parameter socket gives the
2307 * socket descriptor on which the packet was received. This is used,
2308 * in the case of server connections, to check that *new* connections
2309 * come via a valid (port, serviceId). Finally, the securityIndex
2310 * parameter must match the existing index for the connection. If a
2311 * server connection is created, it will be created using the supplied
2312 * index, if the index is valid for this service */
2313 struct rx_connection *
2314 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2315 register u_short port, u_short serviceId, afs_uint32 cid,
2316 afs_uint32 epoch, int type, u_int securityIndex)
2318 int hashindex, flag;
2319 register struct rx_connection *conn;
2320 hashindex = CONN_HASH(host, port, cid, epoch, type);
2321 MUTEX_ENTER(&rx_connHashTable_lock);
2322 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2323 rx_connHashTable[hashindex],
2326 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2327 && (epoch == conn->epoch)) {
2328 register struct rx_peer *pp = conn->peer;
2329 if (securityIndex != conn->securityIndex) {
2330 /* this isn't supposed to happen, but someone could forge a packet
2331 * like this, and there seems to be some CM bug that makes this
2332 * happen from time to time -- in which case, the fileserver
2334 MUTEX_EXIT(&rx_connHashTable_lock);
2335 return (struct rx_connection *)0;
2337 if (pp->host == host && pp->port == port)
2339 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2341 /* So what happens when it's a callback connection? */
2342 if ( /*type == RX_CLIENT_CONNECTION && */
2343 (conn->epoch & 0x80000000))
2347 /* the connection rxLastConn that was used the last time is not the
2348 ** one we are looking for now. Hence, start searching in the hash */
2350 conn = rx_connHashTable[hashindex];
2355 struct rx_service *service;
2356 if (type == RX_CLIENT_CONNECTION) {
2357 MUTEX_EXIT(&rx_connHashTable_lock);
2358 return (struct rx_connection *)0;
2360 service = rxi_FindService(socket, serviceId);
2361 if (!service || (securityIndex >= service->nSecurityObjects)
2362 || (service->securityObjects[securityIndex] == 0)) {
2363 MUTEX_EXIT(&rx_connHashTable_lock);
2364 return (struct rx_connection *)0;
2366 conn = rxi_AllocConnection(); /* This bzero's the connection */
2367 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2368 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2369 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2370 conn->next = rx_connHashTable[hashindex];
2371 rx_connHashTable[hashindex] = conn;
2372 conn->peer = rxi_FindPeer(host, port, 0, 1);
2373 conn->type = RX_SERVER_CONNECTION;
2374 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2375 conn->epoch = epoch;
2376 conn->cid = cid & RX_CIDMASK;
2377 /* conn->serial = conn->lastSerial = 0; */
2378 /* conn->timeout = 0; */
2379 conn->ackRate = RX_FAST_ACK_RATE;
2380 conn->service = service;
2381 conn->serviceId = serviceId;
2382 conn->securityIndex = securityIndex;
2383 conn->securityObject = service->securityObjects[securityIndex];
2384 conn->nSpecific = 0;
2385 conn->specific = NULL;
2386 rx_SetConnDeadTime(conn, service->connDeadTime);
2387 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2388 /* Notify security object of the new connection */
2389 RXS_NewConnection(conn->securityObject, conn);
2390 /* XXXX Connection timeout? */
2391 if (service->newConnProc)
2392 (*service->newConnProc) (conn);
2393 MUTEX_ENTER(&rx_stats_mutex);
2394 rx_stats.nServerConns++;
2395 MUTEX_EXIT(&rx_stats_mutex);
2398 MUTEX_ENTER(&conn->conn_data_lock);
2400 MUTEX_EXIT(&conn->conn_data_lock);
2402 rxLastConn = conn; /* store this connection as the last conn used */
2403 MUTEX_EXIT(&rx_connHashTable_lock);
2407 /* There are two packet tracing routines available for testing and monitoring
2408 * Rx. One is called just after every packet is received and the other is
2409 * called just before every packet is sent. Received packets, have had their
2410 * headers decoded, and packets to be sent have not yet had their headers
2411 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2412 * containing the network address. Both can be modified. The return value, if
2413 * non-zero, indicates that the packet should be dropped. */
2415 int (*rx_justReceived) () = 0;
2416 int (*rx_almostSent) () = 0;
2418 /* A packet has been received off the interface. Np is the packet, socket is
2419 * the socket number it was received from (useful in determining which service
2420 * this packet corresponds to), and (host, port) reflect the host,port of the
2421 * sender. This call returns the packet to the caller if it is finished with
2422 * it, rather than de-allocating it, just as a small performance hack */
2425 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2426 afs_uint32 host, u_short port, int *tnop,
2427 struct rx_call **newcallp)
2429 register struct rx_call *call;
2430 register struct rx_connection *conn;
2432 afs_uint32 currentCallNumber;
2438 struct rx_packet *tnp;
2441 /* We don't print out the packet until now because (1) the time may not be
2442 * accurate enough until now in the lwp implementation (rx_Listener only gets
2443 * the time after the packet is read) and (2) from a protocol point of view,
2444 * this is the first time the packet has been seen */
2445 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2446 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2447 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2448 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2449 np->header.epoch, np->header.cid, np->header.callNumber,
2450 np->header.seq, np->header.flags, np));
2453 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2454 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2457 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2458 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2461 /* If an input tracer function is defined, call it with the packet and
2462 * network address. Note this function may modify its arguments. */
2463 if (rx_justReceived) {
2464 struct sockaddr_in addr;
2466 addr.sin_family = AF_INET;
2467 addr.sin_port = port;
2468 addr.sin_addr.s_addr = host;
2469 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2470 addr.sin_len = sizeof(addr);
2471 #endif /* AFS_OSF_ENV */
2472 drop = (*rx_justReceived) (np, &addr);
2473 /* drop packet if return value is non-zero */
2476 port = addr.sin_port; /* in case fcn changed addr */
2477 host = addr.sin_addr.s_addr;
2481 /* If packet was not sent by the client, then *we* must be the client */
2482 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2483 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2485 /* Find the connection (or fabricate one, if we're the server & if
2486 * necessary) associated with this packet */
2488 rxi_FindConnection(socket, host, port, np->header.serviceId,
2489 np->header.cid, np->header.epoch, type,
2490 np->header.securityIndex);
2493 /* If no connection found or fabricated, just ignore the packet.
2494 * (An argument could be made for sending an abort packet for
2499 MUTEX_ENTER(&conn->conn_data_lock);
2500 if (conn->maxSerial < np->header.serial)
2501 conn->maxSerial = np->header.serial;
2502 MUTEX_EXIT(&conn->conn_data_lock);
2504 /* If the connection is in an error state, send an abort packet and ignore
2505 * the incoming packet */
2507 /* Don't respond to an abort packet--we don't want loops! */
2508 MUTEX_ENTER(&conn->conn_data_lock);
2509 if (np->header.type != RX_PACKET_TYPE_ABORT)
2510 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2512 MUTEX_EXIT(&conn->conn_data_lock);
2516 /* Check for connection-only requests (i.e. not call specific). */
2517 if (np->header.callNumber == 0) {
2518 switch (np->header.type) {
2519 case RX_PACKET_TYPE_ABORT: {
2520 /* What if the supplied error is zero? */
2521 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2522 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2523 rxi_ConnectionError(conn, errcode);
2524 MUTEX_ENTER(&conn->conn_data_lock);
2526 MUTEX_EXIT(&conn->conn_data_lock);
2529 case RX_PACKET_TYPE_CHALLENGE:
2530 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2531 MUTEX_ENTER(&conn->conn_data_lock);
2533 MUTEX_EXIT(&conn->conn_data_lock);
2535 case RX_PACKET_TYPE_RESPONSE:
2536 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2537 MUTEX_ENTER(&conn->conn_data_lock);
2539 MUTEX_EXIT(&conn->conn_data_lock);
2541 case RX_PACKET_TYPE_PARAMS:
2542 case RX_PACKET_TYPE_PARAMS + 1:
2543 case RX_PACKET_TYPE_PARAMS + 2:
2544 /* ignore these packet types for now */
2545 MUTEX_ENTER(&conn->conn_data_lock);
2547 MUTEX_EXIT(&conn->conn_data_lock);
2552 /* Should not reach here, unless the peer is broken: send an
2554 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2555 MUTEX_ENTER(&conn->conn_data_lock);
2556 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2558 MUTEX_EXIT(&conn->conn_data_lock);
2563 channel = np->header.cid & RX_CHANNELMASK;
2564 call = conn->call[channel];
2565 #ifdef RX_ENABLE_LOCKS
2567 MUTEX_ENTER(&call->lock);
2568 /* Test to see if call struct is still attached to conn. */
2569 if (call != conn->call[channel]) {
2571 MUTEX_EXIT(&call->lock);
2572 if (type == RX_SERVER_CONNECTION) {
2573 call = conn->call[channel];
2574 /* If we started with no call attached and there is one now,
2575 * another thread is also running this routine and has gotten
2576 * the connection channel. We should drop this packet in the tests
2577 * below. If there was a call on this connection and it's now
2578 * gone, then we'll be making a new call below.
2579 * If there was previously a call and it's now different then
2580 * the old call was freed and another thread running this routine
2581 * has created a call on this channel. One of these two threads
2582 * has a packet for the old call and the code below handles those
2586 MUTEX_ENTER(&call->lock);
2588 /* This packet can't be for this call. If the new call address is
2589 * 0 then no call is running on this channel. If there is a call
2590 * then, since this is a client connection we're getting data for
2591 * it must be for the previous call.
2593 MUTEX_ENTER(&rx_stats_mutex);
2594 rx_stats.spuriousPacketsRead++;
2595 MUTEX_EXIT(&rx_stats_mutex);
2596 MUTEX_ENTER(&conn->conn_data_lock);
2598 MUTEX_EXIT(&conn->conn_data_lock);
2603 currentCallNumber = conn->callNumber[channel];
2605 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2606 if (np->header.callNumber < currentCallNumber) {
2607 MUTEX_ENTER(&rx_stats_mutex);
2608 rx_stats.spuriousPacketsRead++;
2609 MUTEX_EXIT(&rx_stats_mutex);
2610 #ifdef RX_ENABLE_LOCKS
2612 MUTEX_EXIT(&call->lock);
2614 MUTEX_ENTER(&conn->conn_data_lock);
2616 MUTEX_EXIT(&conn->conn_data_lock);
2620 MUTEX_ENTER(&conn->conn_call_lock);
2621 call = rxi_NewCall(conn, channel);
2622 MUTEX_EXIT(&conn->conn_call_lock);
2623 *call->callNumber = np->header.callNumber;
2624 if (np->header.callNumber == 0)
2625 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));
2627 call->state = RX_STATE_PRECALL;
2628 clock_GetTime(&call->queueTime);
2629 hzero(call->bytesSent);
2630 hzero(call->bytesRcvd);
2632 * If the number of queued calls exceeds the overload
2633 * threshold then abort this call.
2635 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2636 struct rx_packet *tp;
2638 rxi_CallError(call, rx_BusyError);
2639 tp = rxi_SendCallAbort(call, np, 1, 0);
2640 MUTEX_EXIT(&call->lock);
2641 MUTEX_ENTER(&conn->conn_data_lock);
2643 MUTEX_EXIT(&conn->conn_data_lock);
2644 MUTEX_ENTER(&rx_stats_mutex);
2646 MUTEX_EXIT(&rx_stats_mutex);
2649 rxi_KeepAliveOn(call);
2650 } else if (np->header.callNumber != currentCallNumber) {
2651 /* Wait until the transmit queue is idle before deciding
2652 * whether to reset the current call. Chances are that the
2653 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2656 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2657 while ((call->state == RX_STATE_ACTIVE)
2658 && (call->flags & RX_CALL_TQ_BUSY)) {
2659 call->flags |= RX_CALL_TQ_WAIT;
2661 #ifdef RX_ENABLE_LOCKS
2662 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2663 CV_WAIT(&call->cv_tq, &call->lock);
2664 #else /* RX_ENABLE_LOCKS */
2665 osi_rxSleep(&call->tq);
2666 #endif /* RX_ENABLE_LOCKS */
2668 if (call->tqWaiters == 0)
2669 call->flags &= ~RX_CALL_TQ_WAIT;
2671 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2672 /* If the new call cannot be taken right now send a busy and set
2673 * the error condition in this call, so that it terminates as
2674 * quickly as possible */
2675 if (call->state == RX_STATE_ACTIVE) {
2676 struct rx_packet *tp;
2678 rxi_CallError(call, RX_CALL_DEAD);
2679 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2681 MUTEX_EXIT(&call->lock);
2682 MUTEX_ENTER(&conn->conn_data_lock);
2684 MUTEX_EXIT(&conn->conn_data_lock);
2687 rxi_ResetCall(call, 0);
2688 *call->callNumber = np->header.callNumber;
2689 if (np->header.callNumber == 0)
2690 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));
2692 call->state = RX_STATE_PRECALL;
2693 clock_GetTime(&call->queueTime);
2694 hzero(call->bytesSent);
2695 hzero(call->bytesRcvd);
2697 * If the number of queued calls exceeds the overload
2698 * threshold then abort this call.
2700 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2701 struct rx_packet *tp;
2703 rxi_CallError(call, rx_BusyError);
2704 tp = rxi_SendCallAbort(call, np, 1, 0);
2705 MUTEX_EXIT(&call->lock);
2706 MUTEX_ENTER(&conn->conn_data_lock);
2708 MUTEX_EXIT(&conn->conn_data_lock);
2709 MUTEX_ENTER(&rx_stats_mutex);
2711 MUTEX_EXIT(&rx_stats_mutex);
2714 rxi_KeepAliveOn(call);
2716 /* Continuing call; do nothing here. */
2718 } else { /* we're the client */
2719 /* Ignore all incoming acknowledgements for calls in DALLY state */
2720 if (call && (call->state == RX_STATE_DALLY)
2721 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2722 MUTEX_ENTER(&rx_stats_mutex);
2723 rx_stats.ignorePacketDally++;
2724 MUTEX_EXIT(&rx_stats_mutex);
2725 #ifdef RX_ENABLE_LOCKS
2727 MUTEX_EXIT(&call->lock);
2730 MUTEX_ENTER(&conn->conn_data_lock);
2732 MUTEX_EXIT(&conn->conn_data_lock);
2736 /* Ignore anything that's not relevant to the current call. If there
2737 * isn't a current call, then no packet is relevant. */
2738 if (!call || (np->header.callNumber != currentCallNumber)) {
2739 MUTEX_ENTER(&rx_stats_mutex);
2740 rx_stats.spuriousPacketsRead++;
2741 MUTEX_EXIT(&rx_stats_mutex);
2742 #ifdef RX_ENABLE_LOCKS
2744 MUTEX_EXIT(&call->lock);
2747 MUTEX_ENTER(&conn->conn_data_lock);
2749 MUTEX_EXIT(&conn->conn_data_lock);
2752 /* If the service security object index stamped in the packet does not
2753 * match the connection's security index, ignore the packet */
2754 if (np->header.securityIndex != conn->securityIndex) {
2755 #ifdef RX_ENABLE_LOCKS
2756 MUTEX_EXIT(&call->lock);
2758 MUTEX_ENTER(&conn->conn_data_lock);
2760 MUTEX_EXIT(&conn->conn_data_lock);
2764 /* If we're receiving the response, then all transmit packets are
2765 * implicitly acknowledged. Get rid of them. */
2766 if (np->header.type == RX_PACKET_TYPE_DATA) {
2767 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2768 /* XXX Hack. Because we must release the global rx lock when
2769 * sending packets (osi_NetSend) we drop all acks while we're
2770 * traversing the tq in rxi_Start sending packets out because
2771 * packets may move to the freePacketQueue as result of being here!
2772 * So we drop these packets until we're safely out of the
2773 * traversing. Really ugly!
2774 * For fine grain RX locking, we set the acked field in the
2775 * packets and let rxi_Start remove them from the transmit queue.
2777 if (call->flags & RX_CALL_TQ_BUSY) {
2778 #ifdef RX_ENABLE_LOCKS
2779 rxi_SetAcksInTransmitQueue(call);
2782 return np; /* xmitting; drop packet */
2785 rxi_ClearTransmitQueue(call, 0);
2787 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2788 rxi_ClearTransmitQueue(call, 0);
2789 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2791 if (np->header.type == RX_PACKET_TYPE_ACK) {
2792 /* now check to see if this is an ack packet acknowledging that the
2793 * server actually *lost* some hard-acked data. If this happens we
2794 * ignore this packet, as it may indicate that the server restarted in
2795 * the middle of a call. It is also possible that this is an old ack
2796 * packet. We don't abort the connection in this case, because this
2797 * *might* just be an old ack packet. The right way to detect a server
2798 * restart in the midst of a call is to notice that the server epoch
2800 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2801 * XXX unacknowledged. I think that this is off-by-one, but
2802 * XXX I don't dare change it just yet, since it will
2803 * XXX interact badly with the server-restart detection
2804 * XXX code in receiveackpacket. */
2805 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2806 MUTEX_ENTER(&rx_stats_mutex);
2807 rx_stats.spuriousPacketsRead++;
2808 MUTEX_EXIT(&rx_stats_mutex);
2809 MUTEX_EXIT(&call->lock);
2810 MUTEX_ENTER(&conn->conn_data_lock);
2812 MUTEX_EXIT(&conn->conn_data_lock);
2816 } /* else not a data packet */
2819 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2820 /* Set remote user defined status from packet */
2821 call->remoteStatus = np->header.userStatus;
2823 /* Note the gap between the expected next packet and the actual
2824 * packet that arrived, when the new packet has a smaller serial number
2825 * than expected. Rioses frequently reorder packets all by themselves,
2826 * so this will be quite important with very large window sizes.
2827 * Skew is checked against 0 here to avoid any dependence on the type of
2828 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2830 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2831 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2832 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2834 MUTEX_ENTER(&conn->conn_data_lock);
2835 skew = conn->lastSerial - np->header.serial;
2836 conn->lastSerial = np->header.serial;
2837 MUTEX_EXIT(&conn->conn_data_lock);
2839 register struct rx_peer *peer;
2841 if (skew > peer->inPacketSkew) {
2842 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2844 peer->inPacketSkew = skew;
2848 /* Now do packet type-specific processing */
2849 switch (np->header.type) {
2850 case RX_PACKET_TYPE_DATA:
2851 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2854 case RX_PACKET_TYPE_ACK:
2855 /* Respond immediately to ack packets requesting acknowledgement
2857 if (np->header.flags & RX_REQUEST_ACK) {
2859 (void)rxi_SendCallAbort(call, 0, 1, 0);
2861 (void)rxi_SendAck(call, 0, np->header.serial,
2862 RX_ACK_PING_RESPONSE, 1);
2864 np = rxi_ReceiveAckPacket(call, np, 1);
2866 case RX_PACKET_TYPE_ABORT: {
2867 /* An abort packet: reset the call, passing the error up to the user. */
2868 /* What if error is zero? */
2869 /* What if the error is -1? the application will treat it as a timeout. */
2870 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
2871 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
2872 rxi_CallError(call, errdata);
2873 MUTEX_EXIT(&call->lock);
2874 MUTEX_ENTER(&conn->conn_data_lock);
2876 MUTEX_EXIT(&conn->conn_data_lock);
2877 return np; /* xmitting; drop packet */
2879 case RX_PACKET_TYPE_BUSY:
2882 case RX_PACKET_TYPE_ACKALL:
2883 /* All packets acknowledged, so we can drop all packets previously
2884 * readied for sending */
2885 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2886 /* XXX Hack. We because we can't release the global rx lock when
2887 * sending packets (osi_NetSend) we drop all ack pkts while we're
2888 * traversing the tq in rxi_Start sending packets out because
2889 * packets may move to the freePacketQueue as result of being
2890 * here! So we drop these packets until we're safely out of the
2891 * traversing. Really ugly!
2892 * For fine grain RX locking, we set the acked field in the packets
2893 * and let rxi_Start remove the packets from the transmit queue.
2895 if (call->flags & RX_CALL_TQ_BUSY) {
2896 #ifdef RX_ENABLE_LOCKS
2897 rxi_SetAcksInTransmitQueue(call);
2899 #else /* RX_ENABLE_LOCKS */
2900 MUTEX_EXIT(&call->lock);
2901 MUTEX_ENTER(&conn->conn_data_lock);
2903 MUTEX_EXIT(&conn->conn_data_lock);
2904 return np; /* xmitting; drop packet */
2905 #endif /* RX_ENABLE_LOCKS */
2907 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2908 rxi_ClearTransmitQueue(call, 0);
2911 /* Should not reach here, unless the peer is broken: send an abort
2913 rxi_CallError(call, RX_PROTOCOL_ERROR);
2914 np = rxi_SendCallAbort(call, np, 1, 0);
2917 /* Note when this last legitimate packet was received, for keep-alive
2918 * processing. Note, we delay getting the time until now in the hope that
2919 * the packet will be delivered to the user before any get time is required
2920 * (if not, then the time won't actually be re-evaluated here). */
2921 call->lastReceiveTime = clock_Sec();
2922 MUTEX_EXIT(&call->lock);
2923 MUTEX_ENTER(&conn->conn_data_lock);
2925 MUTEX_EXIT(&conn->conn_data_lock);
2929 /* return true if this is an "interesting" connection from the point of view
2930 of someone trying to debug the system */
2932 rxi_IsConnInteresting(struct rx_connection *aconn)
2935 register struct rx_call *tcall;
2937 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2939 for (i = 0; i < RX_MAXCALLS; i++) {
2940 tcall = aconn->call[i];
2942 if ((tcall->state == RX_STATE_PRECALL)
2943 || (tcall->state == RX_STATE_ACTIVE))
2945 if ((tcall->mode == RX_MODE_SENDING)
2946 || (tcall->mode == RX_MODE_RECEIVING))
2954 /* if this is one of the last few packets AND it wouldn't be used by the
2955 receiving call to immediately satisfy a read request, then drop it on
2956 the floor, since accepting it might prevent a lock-holding thread from
2957 making progress in its reading. If a call has been cleared while in
2958 the precall state then ignore all subsequent packets until the call
2959 is assigned to a thread. */
2962 TooLow(struct rx_packet *ap, struct rx_call *acall)
2965 MUTEX_ENTER(&rx_stats_mutex);
2966 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
2967 && (acall->state == RX_STATE_PRECALL))
2968 || ((rx_nFreePackets < rxi_dataQuota + 2)
2969 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
2970 && (acall->flags & RX_CALL_READER_WAIT)))) {
2973 MUTEX_EXIT(&rx_stats_mutex);
2979 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
2980 struct rx_call *acall)
2982 struct rx_call *call = acall;
2986 MUTEX_ENTER(&conn->conn_data_lock);
2987 conn->checkReachEvent = NULL;
2988 waiting = conn->flags & RX_CONN_ATTACHWAIT;
2991 MUTEX_EXIT(&conn->conn_data_lock);
2995 MUTEX_ENTER(&conn->conn_call_lock);
2996 MUTEX_ENTER(&conn->conn_data_lock);
2997 for (i = 0; i < RX_MAXCALLS; i++) {
2998 struct rx_call *tc = conn->call[i];
2999 if (tc && tc->state == RX_STATE_PRECALL) {
3005 /* Indicate that rxi_CheckReachEvent is no longer running by
3006 * clearing the flag. Must be atomic under conn_data_lock to
3007 * avoid a new call slipping by: rxi_CheckConnReach holds
3008 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3010 conn->flags &= ~RX_CONN_ATTACHWAIT;
3011 MUTEX_EXIT(&conn->conn_data_lock);
3012 MUTEX_EXIT(&conn->conn_call_lock);
3017 MUTEX_ENTER(&call->lock);
3018 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3020 MUTEX_EXIT(&call->lock);
3022 clock_GetTime(&when);
3023 when.sec += RX_CHECKREACH_TIMEOUT;
3024 MUTEX_ENTER(&conn->conn_data_lock);
3025 if (!conn->checkReachEvent) {
3027 conn->checkReachEvent =
3028 rxevent_Post(&when, rxi_CheckReachEvent, conn, NULL);
3030 MUTEX_EXIT(&conn->conn_data_lock);
3036 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3038 struct rx_service *service = conn->service;
3039 struct rx_peer *peer = conn->peer;
3040 afs_uint32 now, lastReach;
3042 if (service->checkReach == 0)
3046 MUTEX_ENTER(&peer->peer_lock);
3047 lastReach = peer->lastReachTime;
3048 MUTEX_EXIT(&peer->peer_lock);
3049 if (now - lastReach < RX_CHECKREACH_TTL)
3052 MUTEX_ENTER(&conn->conn_data_lock);
3053 if (conn->flags & RX_CONN_ATTACHWAIT) {
3054 MUTEX_EXIT(&conn->conn_data_lock);
3057 conn->flags |= RX_CONN_ATTACHWAIT;
3058 MUTEX_EXIT(&conn->conn_data_lock);
3059 if (!conn->checkReachEvent)
3060 rxi_CheckReachEvent(NULL, conn, call);
3065 /* try to attach call, if authentication is complete */
3067 TryAttach(register struct rx_call *acall, register osi_socket socket,
3068 register int *tnop, register struct rx_call **newcallp,
3071 struct rx_connection *conn = acall->conn;
3073 if (conn->type == RX_SERVER_CONNECTION
3074 && acall->state == RX_STATE_PRECALL) {
3075 /* Don't attach until we have any req'd. authentication. */
3076 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3077 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3078 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3079 /* Note: this does not necessarily succeed; there
3080 * may not any proc available
3083 rxi_ChallengeOn(acall->conn);
3088 /* A data packet has been received off the interface. This packet is
3089 * appropriate to the call (the call is in the right state, etc.). This
3090 * routine can return a packet to the caller, for re-use */
3093 rxi_ReceiveDataPacket(register struct rx_call *call,
3094 register struct rx_packet *np, int istack,
3095 osi_socket socket, afs_uint32 host, u_short port,
3096 int *tnop, struct rx_call **newcallp)
3098 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3102 afs_uint32 seq, serial, flags;
3104 struct rx_packet *tnp;
3106 MUTEX_ENTER(&rx_stats_mutex);
3107 rx_stats.dataPacketsRead++;
3108 MUTEX_EXIT(&rx_stats_mutex);
3111 /* If there are no packet buffers, drop this new packet, unless we can find
3112 * packet buffers from inactive calls */
3114 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3115 MUTEX_ENTER(&rx_freePktQ_lock);
3116 rxi_NeedMorePackets = TRUE;
3117 MUTEX_EXIT(&rx_freePktQ_lock);
3118 MUTEX_ENTER(&rx_stats_mutex);
3119 rx_stats.noPacketBuffersOnRead++;
3120 MUTEX_EXIT(&rx_stats_mutex);
3121 call->rprev = np->header.serial;
3122 rxi_calltrace(RX_TRACE_DROP, call);
3123 dpf(("packet %x dropped on receipt - quota problems", np));
3125 rxi_ClearReceiveQueue(call);
3126 clock_GetTime(&when);
3127 clock_Add(&when, &rx_softAckDelay);
3128 if (!call->delayedAckEvent
3129 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3130 rxevent_Cancel(call->delayedAckEvent, call,
3131 RX_CALL_REFCOUNT_DELAY);
3132 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3133 call->delayedAckEvent =
3134 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3136 /* we've damaged this call already, might as well do it in. */
3142 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3143 * packet is one of several packets transmitted as a single
3144 * datagram. Do not send any soft or hard acks until all packets
3145 * in a jumbogram have been processed. Send negative acks right away.
3147 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3148 /* tnp is non-null when there are more packets in the
3149 * current jumbo gram */
3156 seq = np->header.seq;
3157 serial = np->header.serial;
3158 flags = np->header.flags;
3160 /* If the call is in an error state, send an abort message */
3162 return rxi_SendCallAbort(call, np, istack, 0);
3164 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3165 * AFS 3.5 jumbogram. */
3166 if (flags & RX_JUMBO_PACKET) {
3167 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3172 if (np->header.spare != 0) {
3173 MUTEX_ENTER(&call->conn->conn_data_lock);
3174 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3175 MUTEX_EXIT(&call->conn->conn_data_lock);
3178 /* The usual case is that this is the expected next packet */
3179 if (seq == call->rnext) {
3181 /* Check to make sure it is not a duplicate of one already queued */
3182 if (queue_IsNotEmpty(&call->rq)
3183 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3184 MUTEX_ENTER(&rx_stats_mutex);
3185 rx_stats.dupPacketsRead++;
3186 MUTEX_EXIT(&rx_stats_mutex);
3187 dpf(("packet %x dropped on receipt - duplicate", np));
3188 rxevent_Cancel(call->delayedAckEvent, call,
3189 RX_CALL_REFCOUNT_DELAY);
3190 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3196 /* It's the next packet. Stick it on the receive queue
3197 * for this call. Set newPackets to make sure we wake
3198 * the reader once all packets have been processed */
3199 queue_Prepend(&call->rq, np);
3201 np = NULL; /* We can't use this anymore */
3204 /* If an ack is requested then set a flag to make sure we
3205 * send an acknowledgement for this packet */
3206 if (flags & RX_REQUEST_ACK) {
3207 ackNeeded = RX_ACK_REQUESTED;
3210 /* Keep track of whether we have received the last packet */
3211 if (flags & RX_LAST_PACKET) {
3212 call->flags |= RX_CALL_HAVE_LAST;
3216 /* Check whether we have all of the packets for this call */
3217 if (call->flags & RX_CALL_HAVE_LAST) {
3218 afs_uint32 tseq; /* temporary sequence number */
3219 struct rx_packet *tp; /* Temporary packet pointer */
3220 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3222 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3223 if (tseq != tp->header.seq)
3225 if (tp->header.flags & RX_LAST_PACKET) {
3226 call->flags |= RX_CALL_RECEIVE_DONE;
3233 /* Provide asynchronous notification for those who want it
3234 * (e.g. multi rx) */
3235 if (call->arrivalProc) {
3236 (*call->arrivalProc) (call, call->arrivalProcHandle,
3237 call->arrivalProcArg);
3238 call->arrivalProc = (void (*)())0;
3241 /* Update last packet received */
3244 /* If there is no server process serving this call, grab
3245 * one, if available. We only need to do this once. If a
3246 * server thread is available, this thread becomes a server
3247 * thread and the server thread becomes a listener thread. */
3249 TryAttach(call, socket, tnop, newcallp, 0);
3252 /* This is not the expected next packet. */
3254 /* Determine whether this is a new or old packet, and if it's
3255 * a new one, whether it fits into the current receive window.
3256 * Also figure out whether the packet was delivered in sequence.
3257 * We use the prev variable to determine whether the new packet
3258 * is the successor of its immediate predecessor in the
3259 * receive queue, and the missing flag to determine whether
3260 * any of this packets predecessors are missing. */
3262 afs_uint32 prev; /* "Previous packet" sequence number */
3263 struct rx_packet *tp; /* Temporary packet pointer */
3264 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3265 int missing; /* Are any predecessors missing? */
3267 /* If the new packet's sequence number has been sent to the
3268 * application already, then this is a duplicate */
3269 if (seq < call->rnext) {
3270 MUTEX_ENTER(&rx_stats_mutex);
3271 rx_stats.dupPacketsRead++;
3272 MUTEX_EXIT(&rx_stats_mutex);
3273 rxevent_Cancel(call->delayedAckEvent, call,
3274 RX_CALL_REFCOUNT_DELAY);
3275 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3281 /* If the sequence number is greater than what can be
3282 * accomodated by the current window, then send a negative
3283 * acknowledge and drop the packet */
3284 if ((call->rnext + call->rwind) <= seq) {
3285 rxevent_Cancel(call->delayedAckEvent, call,
3286 RX_CALL_REFCOUNT_DELAY);
3287 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3294 /* Look for the packet in the queue of old received packets */
3295 for (prev = call->rnext - 1, missing =
3296 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3297 /*Check for duplicate packet */
3298 if (seq == tp->header.seq) {
3299 MUTEX_ENTER(&rx_stats_mutex);
3300 rx_stats.dupPacketsRead++;
3301 MUTEX_EXIT(&rx_stats_mutex);
3302 rxevent_Cancel(call->delayedAckEvent, call,
3303 RX_CALL_REFCOUNT_DELAY);
3304 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3310 /* If we find a higher sequence packet, break out and
3311 * insert the new packet here. */
3312 if (seq < tp->header.seq)
3314 /* Check for missing packet */
3315 if (tp->header.seq != prev + 1) {
3319 prev = tp->header.seq;
3322 /* Keep track of whether we have received the last packet. */
3323 if (flags & RX_LAST_PACKET) {
3324 call->flags |= RX_CALL_HAVE_LAST;
3327 /* It's within the window: add it to the the receive queue.
3328 * tp is left by the previous loop either pointing at the
3329 * packet before which to insert the new packet, or at the
3330 * queue head if the queue is empty or the packet should be
3332 queue_InsertBefore(tp, np);
3336 /* Check whether we have all of the packets for this call */
3337 if ((call->flags & RX_CALL_HAVE_LAST)
3338 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3339 afs_uint32 tseq; /* temporary sequence number */
3342 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3343 if (tseq != tp->header.seq)
3345 if (tp->header.flags & RX_LAST_PACKET) {
3346 call->flags |= RX_CALL_RECEIVE_DONE;
3353 /* We need to send an ack of the packet is out of sequence,
3354 * or if an ack was requested by the peer. */
3355 if (seq != prev + 1 || missing || (flags & RX_REQUEST_ACK)) {
3356 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3359 /* Acknowledge the last packet for each call */
3360 if (flags & RX_LAST_PACKET) {
3371 * If the receiver is waiting for an iovec, fill the iovec
3372 * using the data from the receive queue */
3373 if (call->flags & RX_CALL_IOVEC_WAIT) {
3374 didHardAck = rxi_FillReadVec(call, serial);
3375 /* the call may have been aborted */
3384 /* Wakeup the reader if any */
3385 if ((call->flags & RX_CALL_READER_WAIT)
3386 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3387 || (call->iovNext >= call->iovMax)
3388 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3389 call->flags &= ~RX_CALL_READER_WAIT;
3390 #ifdef RX_ENABLE_LOCKS
3391 CV_BROADCAST(&call->cv_rq);
3393 osi_rxWakeup(&call->rq);
3399 * Send an ack when requested by the peer, or once every
3400 * rxi_SoftAckRate packets until the last packet has been
3401 * received. Always send a soft ack for the last packet in
3402 * the server's reply. */
3404 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3405 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3406 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3407 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3408 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3409 } else if (call->nSoftAcks) {
3410 clock_GetTime(&when);
3411 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3412 clock_Add(&when, &rx_lastAckDelay);
3414 clock_Add(&when, &rx_softAckDelay);
3416 if (!call->delayedAckEvent
3417 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3418 rxevent_Cancel(call->delayedAckEvent, call,
3419 RX_CALL_REFCOUNT_DELAY);
3420 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3421 call->delayedAckEvent =
3422 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3424 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3425 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3432 static void rxi_ComputeRate();
3436 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3438 struct rx_peer *peer = conn->peer;
3440 MUTEX_ENTER(&peer->peer_lock);
3441 peer->lastReachTime = clock_Sec();
3442 MUTEX_EXIT(&peer->peer_lock);
3444 MUTEX_ENTER(&conn->conn_data_lock);
3445 if (conn->flags & RX_CONN_ATTACHWAIT) {
3448 conn->flags &= ~RX_CONN_ATTACHWAIT;
3449 MUTEX_EXIT(&conn->conn_data_lock);
3451 for (i = 0; i < RX_MAXCALLS; i++) {
3452 struct rx_call *call = conn->call[i];
3455 MUTEX_ENTER(&call->lock);
3456 /* tnop can be null if newcallp is null */
3457 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3459 MUTEX_EXIT(&call->lock);
3463 MUTEX_EXIT(&conn->conn_data_lock);
3467 rx_ack_reason(int reason)
3470 case RX_ACK_REQUESTED:
3472 case RX_ACK_DUPLICATE:
3474 case RX_ACK_OUT_OF_SEQUENCE:
3476 case RX_ACK_EXCEEDS_WINDOW:
3478 case RX_ACK_NOSPACE:
3482 case RX_ACK_PING_RESPONSE:
3494 /* rxi_ComputePeerNetStats
3496 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3497 * estimates (like RTT and throughput) based on ack packets. Caller
3498 * must ensure that the packet in question is the right one (i.e.
3499 * serial number matches).
3502 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3503 struct rx_ackPacket *ap, struct rx_packet *np)
3505 struct rx_peer *peer = call->conn->peer;
3507 /* Use RTT if not delayed by client. */
3508 if (ap->reason != RX_ACK_DELAY)
3509 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3511 rxi_ComputeRate(peer, call, p, np, ap->reason);
3515 /* The real smarts of the whole thing. */
3517 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3520 struct rx_ackPacket *ap;
3522 register struct rx_packet *tp;
3523 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3524 register struct rx_connection *conn = call->conn;
3525 struct rx_peer *peer = conn->peer;
3528 /* because there are CM's that are bogus, sending weird values for this. */
3529 afs_uint32 skew = 0;
3534 int newAckCount = 0;
3535 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3536 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3538 MUTEX_ENTER(&rx_stats_mutex);
3539 rx_stats.ackPacketsRead++;
3540 MUTEX_EXIT(&rx_stats_mutex);
3541 ap = (struct rx_ackPacket *)rx_DataOf(np);
3542 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3544 return np; /* truncated ack packet */
3546 /* depends on ack packet struct */
3547 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3548 first = ntohl(ap->firstPacket);
3549 serial = ntohl(ap->serial);
3550 /* temporarily disabled -- needs to degrade over time
3551 * skew = ntohs(ap->maxSkew); */
3553 /* Ignore ack packets received out of order */
3554 if (first < call->tfirst) {
3558 if (np->header.flags & RX_SLOW_START_OK) {
3559 call->flags |= RX_CALL_SLOW_START_OK;
3562 if (ap->reason == RX_ACK_PING_RESPONSE)
3563 rxi_UpdatePeerReach(conn, call);
3567 if (rxdebug_active) {
3571 len = _snprintf(msg, sizeof(msg),
3572 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3573 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3574 ntohl(ap->serial), ntohl(ap->previousPacket),
3575 (unsigned int)np->header.seq, (unsigned int)skew,
3576 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3580 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3581 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3585 OutputDebugString(msg);
3587 #else /* AFS_NT40_ENV */
3590 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3591 ap->reason, ntohl(ap->previousPacket),
3592 (unsigned int)np->header.seq, (unsigned int)serial,
3593 (unsigned int)skew, ntohl(ap->firstPacket));
3596 for (offset = 0; offset < nAcks; offset++)
3597 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3602 #endif /* AFS_NT40_ENV */
3605 /* Update the outgoing packet skew value to the latest value of
3606 * the peer's incoming packet skew value. The ack packet, of
3607 * course, could arrive out of order, but that won't affect things
3609 MUTEX_ENTER(&peer->peer_lock);
3610 peer->outPacketSkew = skew;
3612 /* Check for packets that no longer need to be transmitted, and
3613 * discard them. This only applies to packets positively
3614 * acknowledged as having been sent to the peer's upper level.
3615 * All other packets must be retained. So only packets with
3616 * sequence numbers < ap->firstPacket are candidates. */
3617 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3618 if (tp->header.seq >= first)
3620 call->tfirst = tp->header.seq + 1;
3622 && (tp->header.serial == serial || tp->firstSerial == serial))
3623 rxi_ComputePeerNetStats(call, tp, ap, np);
3624 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3627 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3628 /* XXX Hack. Because we have to release the global rx lock when sending
3629 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3630 * in rxi_Start sending packets out because packets may move to the
3631 * freePacketQueue as result of being here! So we drop these packets until
3632 * we're safely out of the traversing. Really ugly!
3633 * To make it even uglier, if we're using fine grain locking, we can
3634 * set the ack bits in the packets and have rxi_Start remove the packets
3635 * when it's done transmitting.
3637 if (call->flags & RX_CALL_TQ_BUSY) {
3638 #ifdef RX_ENABLE_LOCKS
3639 tp->flags |= RX_PKTFLAG_ACKED;
3640 call->flags |= RX_CALL_TQ_SOME_ACKED;
3641 #else /* RX_ENABLE_LOCKS */
3643 #endif /* RX_ENABLE_LOCKS */
3645 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3648 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3653 /* Give rate detector a chance to respond to ping requests */
3654 if (ap->reason == RX_ACK_PING_RESPONSE) {
3655 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3659 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3661 /* Now go through explicit acks/nacks and record the results in
3662 * the waiting packets. These are packets that can't be released
3663 * yet, even with a positive acknowledge. This positive
3664 * acknowledge only means the packet has been received by the
3665 * peer, not that it will be retained long enough to be sent to
3666 * the peer's upper level. In addition, reset the transmit timers
3667 * of any missing packets (those packets that must be missing
3668 * because this packet was out of sequence) */
3670 call->nSoftAcked = 0;
3671 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3672 /* Update round trip time if the ack was stimulated on receipt
3674 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3675 #ifdef RX_ENABLE_LOCKS
3676 if (tp->header.seq >= first)
3677 #endif /* RX_ENABLE_LOCKS */
3678 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3680 && (tp->header.serial == serial || tp->firstSerial == serial))
3681 rxi_ComputePeerNetStats(call, tp, ap, np);
3683 /* Set the acknowledge flag per packet based on the
3684 * information in the ack packet. An acknowlegded packet can
3685 * be downgraded when the server has discarded a packet it
3686 * soacked previously, or when an ack packet is received
3687 * out of sequence. */
3688 if (tp->header.seq < first) {
3689 /* Implicit ack information */
3690 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3693 tp->flags |= RX_PKTFLAG_ACKED;
3694 } else if (tp->header.seq < first + nAcks) {
3695 /* Explicit ack information: set it in the packet appropriately */
3696 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3697 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3699 tp->flags |= RX_PKTFLAG_ACKED;
3706 } else /* RX_ACK_TYPE_NACK */ {
3707 tp->flags &= ~RX_PKTFLAG_ACKED;
3711 tp->flags &= ~RX_PKTFLAG_ACKED;
3715 /* If packet isn't yet acked, and it has been transmitted at least
3716 * once, reset retransmit time using latest timeout
3717 * ie, this should readjust the retransmit timer for all outstanding
3718 * packets... So we don't just retransmit when we should know better*/
3720 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3721 tp->retryTime = tp->timeSent;
3722 clock_Add(&tp->retryTime, &peer->timeout);
3723 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3724 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3728 /* If the window has been extended by this acknowledge packet,
3729 * then wakeup a sender waiting in alloc for window space, or try
3730 * sending packets now, if he's been sitting on packets due to
3731 * lack of window space */
3732 if (call->tnext < (call->tfirst + call->twind)) {
3733 #ifdef RX_ENABLE_LOCKS
3734 CV_SIGNAL(&call->cv_twind);
3736 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3737 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3738 osi_rxWakeup(&call->twind);
3741 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3742 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3746 /* if the ack packet has a receivelen field hanging off it,
3747 * update our state */
3748 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3751 /* If the ack packet has a "recommended" size that is less than
3752 * what I am using now, reduce my size to match */
3753 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3754 (int)sizeof(afs_int32), &tSize);
3755 tSize = (afs_uint32) ntohl(tSize);
3756 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3758 /* Get the maximum packet size to send to this peer */
3759 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
3761 tSize = (afs_uint32) ntohl(tSize);
3762 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3763 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3765 /* sanity check - peer might have restarted with different params.
3766 * If peer says "send less", dammit, send less... Peer should never
3767 * be unable to accept packets of the size that prior AFS versions would
3768 * send without asking. */
3769 if (peer->maxMTU != tSize) {
3770 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
3772 peer->maxMTU = tSize;
3773 peer->MTU = MIN(tSize, peer->MTU);
3774 call->MTU = MIN(call->MTU, tSize);
3777 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3780 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3781 (int)sizeof(afs_int32), &tSize);
3782 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3783 if (tSize < call->twind) { /* smaller than our send */
3784 call->twind = tSize; /* window, we must send less... */
3785 call->ssthresh = MIN(call->twind, call->ssthresh);
3788 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3789 * network MTU confused with the loopback MTU. Calculate the
3790 * maximum MTU here for use in the slow start code below.
3792 maxMTU = peer->maxMTU;
3793 /* Did peer restart with older RX version? */
3794 if (peer->maxDgramPackets > 1) {
3795 peer->maxDgramPackets = 1;
3797 } else if (np->length >=
3798 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3801 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3802 sizeof(afs_int32), &tSize);
3803 tSize = (afs_uint32) ntohl(tSize);
3805 * As of AFS 3.5 we set the send window to match the receive window.
3807 if (tSize < call->twind) {
3808 call->twind = tSize;
3809 call->ssthresh = MIN(call->twind, call->ssthresh);
3810 } else if (tSize > call->twind) {
3811 call->twind = tSize;
3815 * As of AFS 3.5, a jumbogram is more than one fixed size
3816 * packet transmitted in a single UDP datagram. If the remote
3817 * MTU is smaller than our local MTU then never send a datagram
3818 * larger than the natural MTU.
3821 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3822 sizeof(afs_int32), &tSize);
3823 maxDgramPackets = (afs_uint32) ntohl(tSize);
3824 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3825 maxDgramPackets = MIN(maxDgramPackets, peer->ifDgramPackets);
3826 if (peer->natMTU < peer->ifMTU)
3827 maxDgramPackets = MIN(maxDgramPackets, rxi_AdjustDgramPackets(1, peer->natMTU));
3828 if (maxDgramPackets > 1) {
3829 peer->maxDgramPackets = maxDgramPackets;
3830 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3832 peer->maxDgramPackets = 1;
3833 call->MTU = peer->natMTU;
3835 } else if (peer->maxDgramPackets > 1) {
3836 /* Restarted with lower version of RX */
3837 peer->maxDgramPackets = 1;
3839 } else if (peer->maxDgramPackets > 1
3840 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3841 /* Restarted with lower version of RX */
3842 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3843 peer->natMTU = OLD_MAX_PACKET_SIZE;
3844 peer->MTU = OLD_MAX_PACKET_SIZE;
3845 peer->maxDgramPackets = 1;
3846 peer->nDgramPackets = 1;
3848 call->MTU = OLD_MAX_PACKET_SIZE;
3853 * Calculate how many datagrams were successfully received after
3854 * the first missing packet and adjust the negative ack counter
3859 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3860 if (call->nNacks < nNacked) {
3861 call->nNacks = nNacked;
3870 if (call->flags & RX_CALL_FAST_RECOVER) {
3872 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3874 call->flags &= ~RX_CALL_FAST_RECOVER;
3875 call->cwind = call->nextCwind;
3876 call->nextCwind = 0;
3879 call->nCwindAcks = 0;
3880 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3881 /* Three negative acks in a row trigger congestion recovery */
3882 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3883 MUTEX_EXIT(&peer->peer_lock);
3884 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3885 /* someone else is waiting to start recovery */
3888 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3889 while (call->flags & RX_CALL_TQ_BUSY) {
3890 call->flags |= RX_CALL_TQ_WAIT;
3892 #ifdef RX_ENABLE_LOCKS
3893 osirx_AssertMine(&call->lock, "rxi_Start lock2");
3894 CV_WAIT(&call->cv_tq, &call->lock);
3895 #else /* RX_ENABLE_LOCKS */
3896 osi_rxSleep(&call->tq);
3897 #endif /* RX_ENABLE_LOCKS */
3899 if (call->tqWaiters == 0)
3900 call->flags &= ~RX_CALL_TQ_WAIT;
3902 MUTEX_ENTER(&peer->peer_lock);
3903 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3904 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3905 call->flags |= RX_CALL_FAST_RECOVER;
3906 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3908 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3909 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3910 call->nextCwind = call->ssthresh;
3913 peer->MTU = call->MTU;
3914 peer->cwind = call->nextCwind;
3915 peer->nDgramPackets = call->nDgramPackets;
3917 call->congestSeq = peer->congestSeq;
3918 /* Reset the resend times on the packets that were nacked
3919 * so we will retransmit as soon as the window permits*/
3920 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3922 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3923 clock_Zero(&tp->retryTime);
3925 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3930 /* If cwind is smaller than ssthresh, then increase
3931 * the window one packet for each ack we receive (exponential
3933 * If cwind is greater than or equal to ssthresh then increase
3934 * the congestion window by one packet for each cwind acks we
3935 * receive (linear growth). */
3936 if (call->cwind < call->ssthresh) {
3938 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3939 call->nCwindAcks = 0;
3941 call->nCwindAcks += newAckCount;
3942 if (call->nCwindAcks >= call->cwind) {
3943 call->nCwindAcks = 0;
3944 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3948 * If we have received several acknowledgements in a row then
3949 * it is time to increase the size of our datagrams
3951 if ((int)call->nAcks > rx_nDgramThreshold) {
3952 if (peer->maxDgramPackets > 1) {
3953 if (call->nDgramPackets < peer->maxDgramPackets) {
3954 call->nDgramPackets++;
3956 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3957 } else if (call->MTU < peer->maxMTU) {
3958 call->MTU += peer->natMTU;
3959 call->MTU = MIN(call->MTU, peer->maxMTU);
3965 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3967 /* Servers need to hold the call until all response packets have
3968 * been acknowledged. Soft acks are good enough since clients
3969 * are not allowed to clear their receive queues. */
3970 if (call->state == RX_STATE_HOLD
3971 && call->tfirst + call->nSoftAcked >= call->tnext) {
3972 call->state = RX_STATE_DALLY;
3973 rxi_ClearTransmitQueue(call, 0);
3974 } else if (!queue_IsEmpty(&call->tq)) {
3975 rxi_Start(0, call, 0, istack);
3980 /* Received a response to a challenge packet */
3982 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
3983 register struct rx_packet *np, int istack)
3987 /* Ignore the packet if we're the client */
3988 if (conn->type == RX_CLIENT_CONNECTION)
3991 /* If already authenticated, ignore the packet (it's probably a retry) */
3992 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
3995 /* Otherwise, have the security object evaluate the response packet */
3996 error = RXS_CheckResponse(conn->securityObject, conn, np);
3998 /* If the response is invalid, reset the connection, sending
3999 * an abort to the peer */
4003 rxi_ConnectionError(conn, error);
4004 MUTEX_ENTER(&conn->conn_data_lock);
4005 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4006 MUTEX_EXIT(&conn->conn_data_lock);
4009 /* If the response is valid, any calls waiting to attach
4010 * servers can now do so */
4013 for (i = 0; i < RX_MAXCALLS; i++) {
4014 struct rx_call *call = conn->call[i];
4016 MUTEX_ENTER(&call->lock);
4017 if (call->state == RX_STATE_PRECALL)
4018 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4019 /* tnop can be null if newcallp is null */
4020 MUTEX_EXIT(&call->lock);
4024 /* Update the peer reachability information, just in case
4025 * some calls went into attach-wait while we were waiting
4026 * for authentication..
4028 rxi_UpdatePeerReach(conn, NULL);
4033 /* A client has received an authentication challenge: the security
4034 * object is asked to cough up a respectable response packet to send
4035 * back to the server. The server is responsible for retrying the
4036 * challenge if it fails to get a response. */
4039 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
4040 register struct rx_packet *np, int istack)
4044 /* Ignore the challenge if we're the server */
4045 if (conn->type == RX_SERVER_CONNECTION)
4048 /* Ignore the challenge if the connection is otherwise idle; someone's
4049 * trying to use us as an oracle. */
4050 if (!rxi_HasActiveCalls(conn))
4053 /* Send the security object the challenge packet. It is expected to fill
4054 * in the response. */
4055 error = RXS_GetResponse(conn->securityObject, conn, np);
4057 /* If the security object is unable to return a valid response, reset the
4058 * connection and send an abort to the peer. Otherwise send the response
4059 * packet to the peer connection. */
4061 rxi_ConnectionError(conn, error);
4062 MUTEX_ENTER(&conn->conn_data_lock);
4063 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4064 MUTEX_EXIT(&conn->conn_data_lock);
4066 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4067 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4073 /* Find an available server process to service the current request in
4074 * the given call structure. If one isn't available, queue up this
4075 * call so it eventually gets one */
4077 rxi_AttachServerProc(register struct rx_call *call,
4078 register osi_socket socket, register int *tnop,
4079 register struct rx_call **newcallp)
4081 register struct rx_serverQueueEntry *sq;
4082 register struct rx_service *service = call->conn->service;
4083 register int haveQuota = 0;
4085 /* May already be attached */
4086 if (call->state == RX_STATE_ACTIVE)
4089 MUTEX_ENTER(&rx_serverPool_lock);
4091 haveQuota = QuotaOK(service);
4092 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4093 /* If there are no processes available to service this call,
4094 * put the call on the incoming call queue (unless it's
4095 * already on the queue).
4097 #ifdef RX_ENABLE_LOCKS
4099 ReturnToServerPool(service);
4100 #endif /* RX_ENABLE_LOCKS */
4102 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4103 call->flags |= RX_CALL_WAIT_PROC;
4104 MUTEX_ENTER(&rx_stats_mutex);
4107 MUTEX_EXIT(&rx_stats_mutex);
4108 rxi_calltrace(RX_CALL_ARRIVAL, call);
4109 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4110 queue_Append(&rx_incomingCallQueue, call);
4113 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4115 /* If hot threads are enabled, and both newcallp and sq->socketp
4116 * are non-null, then this thread will process the call, and the
4117 * idle server thread will start listening on this threads socket.
4120 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4123 *sq->socketp = socket;
4124 clock_GetTime(&call->startTime);
4125 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4129 if (call->flags & RX_CALL_WAIT_PROC) {
4130 /* Conservative: I don't think this should happen */
4131 call->flags &= ~RX_CALL_WAIT_PROC;
4132 if (queue_IsOnQueue(call)) {
4134 MUTEX_ENTER(&rx_stats_mutex);
4136 MUTEX_EXIT(&rx_stats_mutex);
4139 call->state = RX_STATE_ACTIVE;
4140 call->mode = RX_MODE_RECEIVING;
4141 #ifdef RX_KERNEL_TRACE
4143 int glockOwner = ISAFS_GLOCK();
4146 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4147 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4153 if (call->flags & RX_CALL_CLEARED) {
4154 /* send an ack now to start the packet flow up again */
4155 call->flags &= ~RX_CALL_CLEARED;
4156 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4158 #ifdef RX_ENABLE_LOCKS
4161 service->nRequestsRunning++;
4162 if (service->nRequestsRunning <= service->minProcs)
4168 MUTEX_EXIT(&rx_serverPool_lock);
4171 /* Delay the sending of an acknowledge event for a short while, while
4172 * a new call is being prepared (in the case of a client) or a reply
4173 * is being prepared (in the case of a server). Rather than sending
4174 * an ack packet, an ACKALL packet is sent. */
4176 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4178 #ifdef RX_ENABLE_LOCKS
4180 MUTEX_ENTER(&call->lock);
4181 call->delayedAckEvent = NULL;
4182 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4184 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4185 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4187 MUTEX_EXIT(&call->lock);
4188 #else /* RX_ENABLE_LOCKS */
4190 call->delayedAckEvent = NULL;
4191 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4192 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4193 #endif /* RX_ENABLE_LOCKS */
4197 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4200 #ifdef RX_ENABLE_LOCKS
4202 MUTEX_ENTER(&call->lock);
4203 if (event == call->delayedAckEvent)
4204 call->delayedAckEvent = NULL;
4205 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4207 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4209 MUTEX_EXIT(&call->lock);
4210 #else /* RX_ENABLE_LOCKS */
4212 call->delayedAckEvent = NULL;
4213 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4214 #endif /* RX_ENABLE_LOCKS */
4218 #ifdef RX_ENABLE_LOCKS
4219 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4220 * clearing them out.
4223 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4225 register struct rx_packet *p, *tp;
4228 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4229 p->flags |= RX_PKTFLAG_ACKED;
4233 call->flags |= RX_CALL_TQ_CLEARME;
4234 call->flags |= RX_CALL_TQ_SOME_ACKED;
4237 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4238 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4239 call->tfirst = call->tnext;
4240 call->nSoftAcked = 0;
4242 if (call->flags & RX_CALL_FAST_RECOVER) {
4243 call->flags &= ~RX_CALL_FAST_RECOVER;
4244 call->cwind = call->nextCwind;
4245 call->nextCwind = 0;
4248 CV_SIGNAL(&call->cv_twind);
4250 #endif /* RX_ENABLE_LOCKS */
4252 /* Clear out the transmit queue for the current call (all packets have
4253 * been received by peer) */
4255 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4257 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4258 register struct rx_packet *p, *tp;
4260 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4262 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4263 p->flags |= RX_PKTFLAG_ACKED;
4267 call->flags |= RX_CALL_TQ_CLEARME;
4268 call->flags |= RX_CALL_TQ_SOME_ACKED;
4271 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4272 rxi_FreePackets(0, &call->tq);
4273 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4274 call->flags &= ~RX_CALL_TQ_CLEARME;
4276 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4278 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4279 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4280 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4281 call->nSoftAcked = 0;
4283 if (call->flags & RX_CALL_FAST_RECOVER) {
4284 call->flags &= ~RX_CALL_FAST_RECOVER;
4285 call->cwind = call->nextCwind;
4287 #ifdef RX_ENABLE_LOCKS
4288 CV_SIGNAL(&call->cv_twind);
4290 osi_rxWakeup(&call->twind);
4295 rxi_ClearReceiveQueue(register struct rx_call *call)
4297 if (queue_IsNotEmpty(&call->rq)) {
4298 rx_packetReclaims += rxi_FreePackets(0, &call->rq);
4299 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4301 if (call->state == RX_STATE_PRECALL) {
4302 call->flags |= RX_CALL_CLEARED;
4306 /* Send an abort packet for the specified call */
4308 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4309 int istack, int force)
4317 /* Clients should never delay abort messages */
4318 if (rx_IsClientConn(call->conn))
4321 if (call->abortCode != call->error) {
4322 call->abortCode = call->error;
4323 call->abortCount = 0;
4326 if (force || rxi_callAbortThreshhold == 0
4327 || call->abortCount < rxi_callAbortThreshhold) {
4328 if (call->delayedAbortEvent) {
4329 rxevent_Cancel(call->delayedAbortEvent, call,
4330 RX_CALL_REFCOUNT_ABORT);
4332 error = htonl(call->error);
4335 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4336 (char *)&error, sizeof(error), istack);
4337 } else if (!call->delayedAbortEvent) {
4338 clock_GetTime(&when);
4339 clock_Addmsec(&when, rxi_callAbortDelay);
4340 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4341 call->delayedAbortEvent =
4342 rxevent_Post(&when, rxi_SendDelayedCallAbort, call, 0);
4347 /* Send an abort packet for the specified connection. Packet is an
4348 * optional pointer to a packet that can be used to send the abort.
4349 * Once the number of abort messages reaches the threshhold, an
4350 * event is scheduled to send the abort. Setting the force flag
4351 * overrides sending delayed abort messages.
4353 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4354 * to send the abort packet.
4357 rxi_SendConnectionAbort(register struct rx_connection *conn,
4358 struct rx_packet *packet, int istack, int force)
4366 /* Clients should never delay abort messages */
4367 if (rx_IsClientConn(conn))
4370 if (force || rxi_connAbortThreshhold == 0
4371 || conn->abortCount < rxi_connAbortThreshhold) {
4372 if (conn->delayedAbortEvent) {
4373 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4375 error = htonl(conn->error);
4377 MUTEX_EXIT(&conn->conn_data_lock);
4379 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4380 RX_PACKET_TYPE_ABORT, (char *)&error,
4381 sizeof(error), istack);
4382 MUTEX_ENTER(&conn->conn_data_lock);
4383 } else if (!conn->delayedAbortEvent) {
4384 clock_GetTime(&when);
4385 clock_Addmsec(&when, rxi_connAbortDelay);
4386 conn->delayedAbortEvent =
4387 rxevent_Post(&when, rxi_SendDelayedConnAbort, conn, 0);
4392 /* Associate an error all of the calls owned by a connection. Called
4393 * with error non-zero. This is only for really fatal things, like
4394 * bad authentication responses. The connection itself is set in
4395 * error at this point, so that future packets received will be
4398 rxi_ConnectionError(register struct rx_connection *conn,
4399 register afs_int32 error)
4404 dpf(("rxi_ConnectionError conn %x error %d", conn, error));
4406 MUTEX_ENTER(&conn->conn_data_lock);
4407 if (conn->challengeEvent)
4408 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4409 if (conn->checkReachEvent) {
4410 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4411 conn->checkReachEvent = 0;
4412 conn->flags &= ~RX_CONN_ATTACHWAIT;
4415 MUTEX_EXIT(&conn->conn_data_lock);
4416 for (i = 0; i < RX_MAXCALLS; i++) {
4417 struct rx_call *call = conn->call[i];
4419 MUTEX_ENTER(&call->lock);
4420 rxi_CallError(call, error);
4421 MUTEX_EXIT(&call->lock);
4424 conn->error = error;
4425 MUTEX_ENTER(&rx_stats_mutex);
4426 rx_stats.fatalErrors++;
4427 MUTEX_EXIT(&rx_stats_mutex);
4432 rxi_CallError(register struct rx_call *call, afs_int32 error)
4434 dpf(("rxi_CallError call %x error %d call->error %d", call, error, call->error));
4436 error = call->error;
4438 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4439 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4440 rxi_ResetCall(call, 0);
4443 rxi_ResetCall(call, 0);
4445 call->error = error;
4446 call->mode = RX_MODE_ERROR;
4449 /* Reset various fields in a call structure, and wakeup waiting
4450 * processes. Some fields aren't changed: state & mode are not
4451 * touched (these must be set by the caller), and bufptr, nLeft, and
4452 * nFree are not reset, since these fields are manipulated by
4453 * unprotected macros, and may only be reset by non-interrupting code.
4456 /* this code requires that call->conn be set properly as a pre-condition. */
4457 #endif /* ADAPT_WINDOW */
4460 rxi_ResetCall(register struct rx_call *call, register int newcall)
4463 register struct rx_peer *peer;
4464 struct rx_packet *packet;
4466 /* Notify anyone who is waiting for asynchronous packet arrival */
4467 if (call->arrivalProc) {
4468 (*call->arrivalProc) (call, call->arrivalProcHandle,
4469 call->arrivalProcArg);
4470 call->arrivalProc = (void (*)())0;
4473 if (call->delayedAbortEvent) {
4474 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4475 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4477 rxi_SendCallAbort(call, packet, 0, 1);
4478 rxi_FreePacket(packet);
4483 * Update the peer with the congestion information in this call
4484 * so other calls on this connection can pick up where this call
4485 * left off. If the congestion sequence numbers don't match then
4486 * another call experienced a retransmission.
4488 peer = call->conn->peer;
4489 MUTEX_ENTER(&peer->peer_lock);
4491 if (call->congestSeq == peer->congestSeq) {
4492 peer->cwind = MAX(peer->cwind, call->cwind);
4493 peer->MTU = MAX(peer->MTU, call->MTU);
4494 peer->nDgramPackets =
4495 MAX(peer->nDgramPackets, call->nDgramPackets);
4498 call->abortCode = 0;
4499 call->abortCount = 0;
4501 if (peer->maxDgramPackets > 1) {
4502 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4504 call->MTU = peer->MTU;
4506 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4507 call->ssthresh = rx_maxSendWindow;
4508 call->nDgramPackets = peer->nDgramPackets;
4509 call->congestSeq = peer->congestSeq;
4510 MUTEX_EXIT(&peer->peer_lock);
4512 flags = call->flags;
4513 rxi_ClearReceiveQueue(call);
4514 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4515 if (flags & RX_CALL_TQ_BUSY) {
4516 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4517 call->flags |= (flags & RX_CALL_TQ_WAIT);
4519 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4521 rxi_ClearTransmitQueue(call, 0);
4522 queue_Init(&call->tq);
4523 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4524 dpf(("rcall %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4527 while (call->tqWaiters) {
4528 #ifdef RX_ENABLE_LOCKS
4529 CV_BROADCAST(&call->cv_tq);
4530 #else /* RX_ENABLE_LOCKS */
4531 osi_rxWakeup(&call->tq);
4532 #endif /* RX_ENABLE_LOCKS */
4536 queue_Init(&call->rq);
4538 call->rwind = rx_initReceiveWindow;
4539 call->twind = rx_initSendWindow;
4540 call->nSoftAcked = 0;
4541 call->nextCwind = 0;
4544 call->nCwindAcks = 0;
4545 call->nSoftAcks = 0;
4546 call->nHardAcks = 0;
4548 call->tfirst = call->rnext = call->tnext = 1;
4550 call->lastAcked = 0;
4551 call->localStatus = call->remoteStatus = 0;
4553 if (flags & RX_CALL_READER_WAIT) {
4554 #ifdef RX_ENABLE_LOCKS
4555 CV_BROADCAST(&call->cv_rq);
4557 osi_rxWakeup(&call->rq);
4560 if (flags & RX_CALL_WAIT_PACKETS) {
4561 MUTEX_ENTER(&rx_freePktQ_lock);
4562 rxi_PacketsUnWait(); /* XXX */
4563 MUTEX_EXIT(&rx_freePktQ_lock);
4565 #ifdef RX_ENABLE_LOCKS
4566 CV_SIGNAL(&call->cv_twind);
4568 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4569 osi_rxWakeup(&call->twind);
4572 #ifdef RX_ENABLE_LOCKS
4573 /* The following ensures that we don't mess with any queue while some
4574 * other thread might also be doing so. The call_queue_lock field is
4575 * is only modified under the call lock. If the call is in the process
4576 * of being removed from a queue, the call is not locked until the
4577 * the queue lock is dropped and only then is the call_queue_lock field
4578 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4579 * Note that any other routine which removes a call from a queue has to
4580 * obtain the queue lock before examing the queue and removing the call.
4582 if (call->call_queue_lock) {
4583 MUTEX_ENTER(call->call_queue_lock);
4584 if (queue_IsOnQueue(call)) {
4586 if (flags & RX_CALL_WAIT_PROC) {
4587 MUTEX_ENTER(&rx_stats_mutex);
4589 MUTEX_EXIT(&rx_stats_mutex);
4592 MUTEX_EXIT(call->call_queue_lock);
4593 CLEAR_CALL_QUEUE_LOCK(call);
4595 #else /* RX_ENABLE_LOCKS */
4596 if (queue_IsOnQueue(call)) {
4598 if (flags & RX_CALL_WAIT_PROC)
4601 #endif /* RX_ENABLE_LOCKS */
4603 rxi_KeepAliveOff(call);
4604 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4607 /* Send an acknowledge for the indicated packet (seq,serial) of the
4608 * indicated call, for the indicated reason (reason). This
4609 * acknowledge will specifically acknowledge receiving the packet, and
4610 * will also specify which other packets for this call have been
4611 * received. This routine returns the packet that was used to the
4612 * caller. The caller is responsible for freeing it or re-using it.
4613 * This acknowledgement also returns the highest sequence number
4614 * actually read out by the higher level to the sender; the sender
4615 * promises to keep around packets that have not been read by the
4616 * higher level yet (unless, of course, the sender decides to abort
4617 * the call altogether). Any of p, seq, serial, pflags, or reason may
4618 * be set to zero without ill effect. That is, if they are zero, they
4619 * will not convey any information.
4620 * NOW there is a trailer field, after the ack where it will safely be
4621 * ignored by mundanes, which indicates the maximum size packet this
4622 * host can swallow. */
4624 register struct rx_packet *optionalPacket; use to send ack (or null)
4625 int seq; Sequence number of the packet we are acking
4626 int serial; Serial number of the packet
4627 int pflags; Flags field from packet header
4628 int reason; Reason an acknowledge was prompted
4632 rxi_SendAck(register struct rx_call *call,
4633 register struct rx_packet *optionalPacket, int serial, int reason,
4636 struct rx_ackPacket *ap;
4637 register struct rx_packet *rqp;
4638 register struct rx_packet *nxp; /* For queue_Scan */
4639 register struct rx_packet *p;
4642 #ifdef RX_ENABLE_TSFPQ
4643 struct rx_ts_info_t * rx_ts_info;
4647 * Open the receive window once a thread starts reading packets
4649 if (call->rnext > 1) {
4650 call->rwind = rx_maxReceiveWindow;
4653 call->nHardAcks = 0;
4654 call->nSoftAcks = 0;
4655 if (call->rnext > call->lastAcked)
4656 call->lastAcked = call->rnext;
4660 rx_computelen(p, p->length); /* reset length, you never know */
4661 } /* where that's been... */
4662 #ifdef RX_ENABLE_TSFPQ
4664 RX_TS_INFO_GET(rx_ts_info);
4665 if ((p = rx_ts_info->local_special_packet)) {
4666 rx_computelen(p, p->length);
4667 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4668 rx_ts_info->local_special_packet = p;
4669 } else { /* We won't send the ack, but don't panic. */
4670 return optionalPacket;
4674 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4675 /* We won't send the ack, but don't panic. */
4676 return optionalPacket;
4681 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4684 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4685 #ifndef RX_ENABLE_TSFPQ
4686 if (!optionalPacket)
4689 return optionalPacket;
4691 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4692 if (rx_Contiguous(p) < templ) {
4693 #ifndef RX_ENABLE_TSFPQ
4694 if (!optionalPacket)
4697 return optionalPacket;
4702 /* MTUXXX failing to send an ack is very serious. We should */
4703 /* try as hard as possible to send even a partial ack; it's */
4704 /* better than nothing. */
4705 ap = (struct rx_ackPacket *)rx_DataOf(p);
4706 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4707 ap->reason = reason;
4709 /* The skew computation used to be bogus, I think it's better now. */
4710 /* We should start paying attention to skew. XXX */
4711 ap->serial = htonl(serial);
4712 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4714 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4715 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4717 /* No fear of running out of ack packet here because there can only be at most
4718 * one window full of unacknowledged packets. The window size must be constrained
4719 * to be less than the maximum ack size, of course. Also, an ack should always
4720 * fit into a single packet -- it should not ever be fragmented. */
4721 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4722 if (!rqp || !call->rq.next
4723 || (rqp->header.seq > (call->rnext + call->rwind))) {
4724 #ifndef RX_ENABLE_TSFPQ
4725 if (!optionalPacket)
4728 rxi_CallError(call, RX_CALL_DEAD);
4729 return optionalPacket;
4732 while (rqp->header.seq > call->rnext + offset)
4733 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4734 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4736 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4737 #ifndef RX_ENABLE_TSFPQ
4738 if (!optionalPacket)
4741 rxi_CallError(call, RX_CALL_DEAD);
4742 return optionalPacket;
4747 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4749 /* these are new for AFS 3.3 */
4750 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4751 templ = htonl(templ);
4752 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4753 templ = htonl(call->conn->peer->ifMTU);
4754 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4755 sizeof(afs_int32), &templ);
4757 /* new for AFS 3.4 */
4758 templ = htonl(call->rwind);
4759 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4760 sizeof(afs_int32), &templ);
4762 /* new for AFS 3.5 */
4763 templ = htonl(call->conn->peer->ifDgramPackets);
4764 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4765 sizeof(afs_int32), &templ);
4767 p->header.serviceId = call->conn->serviceId;
4768 p->header.cid = (call->conn->cid | call->channel);
4769 p->header.callNumber = *call->callNumber;
4771 p->header.securityIndex = call->conn->securityIndex;
4772 p->header.epoch = call->conn->epoch;
4773 p->header.type = RX_PACKET_TYPE_ACK;
4774 p->header.flags = RX_SLOW_START_OK;
4775 if (reason == RX_ACK_PING) {
4776 p->header.flags |= RX_REQUEST_ACK;
4778 clock_GetTime(&call->pingRequestTime);
4781 if (call->conn->type == RX_CLIENT_CONNECTION)
4782 p->header.flags |= RX_CLIENT_INITIATED;
4786 if (rxdebug_active) {
4790 len = _snprintf(msg, sizeof(msg),
4791 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4792 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4793 ntohl(ap->serial), ntohl(ap->previousPacket),
4794 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
4795 ap->nAcks, ntohs(ap->bufferSpace) );
4799 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
4800 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4804 OutputDebugString(msg);
4806 #else /* AFS_NT40_ENV */
4808 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
4809 ap->reason, ntohl(ap->previousPacket),
4810 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4812 for (offset = 0; offset < ap->nAcks; offset++)
4813 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4818 #endif /* AFS_NT40_ENV */
4821 register int i, nbytes = p->length;
4823 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4824 if (nbytes <= p->wirevec[i].iov_len) {
4825 register int savelen, saven;
4827 savelen = p->wirevec[i].iov_len;
4829 p->wirevec[i].iov_len = nbytes;
4831 rxi_Send(call, p, istack);
4832 p->wirevec[i].iov_len = savelen;
4836 nbytes -= p->wirevec[i].iov_len;
4839 MUTEX_ENTER(&rx_stats_mutex);
4840 rx_stats.ackPacketsSent++;
4841 MUTEX_EXIT(&rx_stats_mutex);
4842 #ifndef RX_ENABLE_TSFPQ
4843 if (!optionalPacket)
4846 return optionalPacket; /* Return packet for re-use by caller */
4849 /* Send all of the packets in the list in single datagram */
4851 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4852 int istack, int moreFlag, struct clock *now,
4853 struct clock *retryTime, int resending)
4858 struct rx_connection *conn = call->conn;
4859 struct rx_peer *peer = conn->peer;
4861 MUTEX_ENTER(&peer->peer_lock);
4864 peer->reSends += len;
4865 MUTEX_ENTER(&rx_stats_mutex);
4866 rx_stats.dataPacketsSent += len;
4867 MUTEX_EXIT(&rx_stats_mutex);
4868 MUTEX_EXIT(&peer->peer_lock);
4870 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4874 /* Set the packet flags and schedule the resend events */
4875 /* Only request an ack for the last packet in the list */
4876 for (i = 0; i < len; i++) {
4877 list[i]->retryTime = *retryTime;
4878 if (list[i]->header.serial) {
4879 /* Exponentially backoff retry times */
4880 if (list[i]->backoff < MAXBACKOFF) {
4881 /* so it can't stay == 0 */
4882 list[i]->backoff = (list[i]->backoff << 1) + 1;
4885 clock_Addmsec(&(list[i]->retryTime),
4886 ((afs_uint32) list[i]->backoff) << 8);
4889 /* Wait a little extra for the ack on the last packet */
4890 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4891 clock_Addmsec(&(list[i]->retryTime), 400);
4894 /* Record the time sent */
4895 list[i]->timeSent = *now;
4897 /* Ask for an ack on retransmitted packets, on every other packet
4898 * if the peer doesn't support slow start. Ask for an ack on every
4899 * packet until the congestion window reaches the ack rate. */
4900 if (list[i]->header.serial) {
4902 MUTEX_ENTER(&rx_stats_mutex);
4903 rx_stats.dataPacketsReSent++;
4904 MUTEX_EXIT(&rx_stats_mutex);
4906 /* improved RTO calculation- not Karn */
4907 list[i]->firstSent = *now;
4908 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4909 || (!(call->flags & RX_CALL_SLOW_START_OK)
4910 && (list[i]->header.seq & 1)))) {
4915 MUTEX_ENTER(&peer->peer_lock);
4919 MUTEX_ENTER(&rx_stats_mutex);
4920 rx_stats.dataPacketsSent++;
4921 MUTEX_EXIT(&rx_stats_mutex);
4922 MUTEX_EXIT(&peer->peer_lock);
4924 /* Tag this packet as not being the last in this group,
4925 * for the receiver's benefit */
4926 if (i < len - 1 || moreFlag) {
4927 list[i]->header.flags |= RX_MORE_PACKETS;
4930 /* Install the new retransmit time for the packet, and
4931 * record the time sent */
4932 list[i]->timeSent = *now;
4936 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4939 /* Since we're about to send a data packet to the peer, it's
4940 * safe to nuke any scheduled end-of-packets ack */
4941 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4943 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4944 MUTEX_EXIT(&call->lock);
4946 rxi_SendPacketList(call, conn, list, len, istack);
4948 rxi_SendPacket(call, conn, list[0], istack);
4950 MUTEX_ENTER(&call->lock);
4951 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4953 /* Update last send time for this call (for keep-alive
4954 * processing), and for the connection (so that we can discover
4955 * idle connections) */
4956 conn->lastSendTime = call->lastSendTime = clock_Sec();
4959 /* When sending packets we need to follow these rules:
4960 * 1. Never send more than maxDgramPackets in a jumbogram.
4961 * 2. Never send a packet with more than two iovecs in a jumbogram.
4962 * 3. Never send a retransmitted packet in a jumbogram.
4963 * 4. Never send more than cwind/4 packets in a jumbogram
4964 * We always keep the last list we should have sent so we
4965 * can set the RX_MORE_PACKETS flags correctly.
4968 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4969 int istack, struct clock *now, struct clock *retryTime,
4972 int i, cnt, lastCnt = 0;
4973 struct rx_packet **listP, **lastP = 0;
4974 struct rx_peer *peer = call->conn->peer;
4975 int morePackets = 0;
4977 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4978 /* Does the current packet force us to flush the current list? */
4980 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4981 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4983 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
4985 /* If the call enters an error state stop sending, or if
4986 * we entered congestion recovery mode, stop sending */
4987 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4995 /* Add the current packet to the list if it hasn't been acked.
4996 * Otherwise adjust the list pointer to skip the current packet. */
4997 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
4999 /* Do we need to flush the list? */
5000 if (cnt >= (int)peer->maxDgramPackets
5001 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5002 || list[i]->header.serial
5003 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5005 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5006 retryTime, resending);
5007 /* If the call enters an error state stop sending, or if
5008 * we entered congestion recovery mode, stop sending */
5010 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5015 listP = &list[i + 1];
5020 osi_Panic("rxi_SendList error");
5022 listP = &list[i + 1];
5026 /* Send the whole list when the call is in receive mode, when
5027 * the call is in eof mode, when we are in fast recovery mode,
5028 * and when we have the last packet */
5029 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5030 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5031 || (call->flags & RX_CALL_FAST_RECOVER)) {
5032 /* Check for the case where the current list contains
5033 * an acked packet. Since we always send retransmissions
5034 * in a separate packet, we only need to check the first
5035 * packet in the list */
5036 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5040 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5041 retryTime, resending);
5042 /* If the call enters an error state stop sending, or if
5043 * we entered congestion recovery mode, stop sending */
5044 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5048 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5051 } else if (lastCnt > 0) {
5052 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5057 #ifdef RX_ENABLE_LOCKS
5058 /* Call rxi_Start, below, but with the call lock held. */
5060 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
5061 void *arg1, int istack)
5063 MUTEX_ENTER(&call->lock);
5064 rxi_Start(event, call, arg1, istack);
5065 MUTEX_EXIT(&call->lock);
5067 #endif /* RX_ENABLE_LOCKS */
5069 /* This routine is called when new packets are readied for
5070 * transmission and when retransmission may be necessary, or when the
5071 * transmission window or burst count are favourable. This should be
5072 * better optimized for new packets, the usual case, now that we've
5073 * got rid of queues of send packets. XXXXXXXXXXX */
5075 rxi_Start(struct rxevent *event, register struct rx_call *call,
5076 void *arg1, int istack)
5078 struct rx_packet *p;
5079 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
5080 struct rx_peer *peer = call->conn->peer;
5081 struct clock now, retryTime;
5085 struct rx_packet **xmitList;
5088 /* If rxi_Start is being called as a result of a resend event,
5089 * then make sure that the event pointer is removed from the call
5090 * structure, since there is no longer a per-call retransmission
5092 if (event && event == call->resendEvent) {
5093 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5094 call->resendEvent = NULL;
5096 if (queue_IsEmpty(&call->tq)) {
5100 /* Timeouts trigger congestion recovery */
5101 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5102 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5103 /* someone else is waiting to start recovery */
5106 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5107 while (call->flags & RX_CALL_TQ_BUSY) {
5108 call->flags |= RX_CALL_TQ_WAIT;
5110 #ifdef RX_ENABLE_LOCKS
5111 osirx_AssertMine(&call->lock, "rxi_Start lock1");
5112 CV_WAIT(&call->cv_tq, &call->lock);
5113 #else /* RX_ENABLE_LOCKS */
5114 osi_rxSleep(&call->tq);
5115 #endif /* RX_ENABLE_LOCKS */
5117 if (call->tqWaiters == 0)
5118 call->flags &= ~RX_CALL_TQ_WAIT;
5120 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5121 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5122 call->flags |= RX_CALL_FAST_RECOVER;
5123 if (peer->maxDgramPackets > 1) {
5124 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5126 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5128 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5129 call->nDgramPackets = 1;
5131 call->nextCwind = 1;
5134 MUTEX_ENTER(&peer->peer_lock);
5135 peer->MTU = call->MTU;
5136 peer->cwind = call->cwind;
5137 peer->nDgramPackets = 1;
5139 call->congestSeq = peer->congestSeq;
5140 MUTEX_EXIT(&peer->peer_lock);
5141 /* Clear retry times on packets. Otherwise, it's possible for
5142 * some packets in the queue to force resends at rates faster
5143 * than recovery rates.
5145 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5146 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5147 clock_Zero(&p->retryTime);
5152 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5153 MUTEX_ENTER(&rx_stats_mutex);
5154 rx_tq_debug.rxi_start_in_error++;
5155 MUTEX_EXIT(&rx_stats_mutex);
5160 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5161 /* Get clock to compute the re-transmit time for any packets
5162 * in this burst. Note, if we back off, it's reasonable to
5163 * back off all of the packets in the same manner, even if
5164 * some of them have been retransmitted more times than more
5165 * recent additions */
5166 clock_GetTime(&now);
5167 retryTime = now; /* initialize before use */
5168 MUTEX_ENTER(&peer->peer_lock);
5169 clock_Add(&retryTime, &peer->timeout);
5170 MUTEX_EXIT(&peer->peer_lock);
5172 /* Send (or resend) any packets that need it, subject to
5173 * window restrictions and congestion burst control
5174 * restrictions. Ask for an ack on the last packet sent in
5175 * this burst. For now, we're relying upon the window being
5176 * considerably bigger than the largest number of packets that
5177 * are typically sent at once by one initial call to
5178 * rxi_Start. This is probably bogus (perhaps we should ask
5179 * for an ack when we're half way through the current
5180 * window?). Also, for non file transfer applications, this
5181 * may end up asking for an ack for every packet. Bogus. XXXX
5184 * But check whether we're here recursively, and let the other guy
5187 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5188 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5189 call->flags |= RX_CALL_TQ_BUSY;
5191 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5193 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5194 call->flags &= ~RX_CALL_NEED_START;
5195 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5197 maxXmitPackets = MIN(call->twind, call->cwind);
5198 xmitList = (struct rx_packet **)
5199 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5200 if (xmitList == NULL)
5201 osi_Panic("rxi_Start, failed to allocate xmit list");
5202 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5203 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5204 /* We shouldn't be sending packets if a thread is waiting
5205 * to initiate congestion recovery */
5209 && (call->flags & RX_CALL_FAST_RECOVER)) {
5210 /* Only send one packet during fast recovery */
5213 if ((p->flags & RX_PKTFLAG_FREE)
5214 || (!queue_IsEnd(&call->tq, nxp)
5215 && (nxp->flags & RX_PKTFLAG_FREE))
5216 || (p == (struct rx_packet *)&rx_freePacketQueue)
5217 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5218 osi_Panic("rxi_Start: xmit queue clobbered");
5220 if (p->flags & RX_PKTFLAG_ACKED) {
5221 MUTEX_ENTER(&rx_stats_mutex);
5222 rx_stats.ignoreAckedPacket++;
5223 MUTEX_EXIT(&rx_stats_mutex);
5224 continue; /* Ignore this packet if it has been acknowledged */
5227 /* Turn off all flags except these ones, which are the same
5228 * on each transmission */
5229 p->header.flags &= RX_PRESET_FLAGS;
5231 if (p->header.seq >=
5232 call->tfirst + MIN((int)call->twind,
5233 (int)(call->nSoftAcked +
5235 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5236 /* Note: if we're waiting for more window space, we can
5237 * still send retransmits; hence we don't return here, but
5238 * break out to schedule a retransmit event */
5239 dpf(("call %d waiting for window",
5240 *(call->callNumber)));
5244 /* Transmit the packet if it needs to be sent. */
5245 if (!clock_Lt(&now, &p->retryTime)) {
5246 if (nXmitPackets == maxXmitPackets) {
5247 rxi_SendXmitList(call, xmitList, nXmitPackets,
5248 istack, &now, &retryTime,
5250 osi_Free(xmitList, maxXmitPackets *
5251 sizeof(struct rx_packet *));
5254 xmitList[nXmitPackets++] = p;
5258 /* xmitList now hold pointers to all of the packets that are
5259 * ready to send. Now we loop to send the packets */
5260 if (nXmitPackets > 0) {
5261 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5262 &now, &retryTime, resending);
5265 maxXmitPackets * sizeof(struct rx_packet *));
5267 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5269 * TQ references no longer protected by this flag; they must remain
5270 * protected by the global lock.
5272 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5273 call->flags &= ~RX_CALL_TQ_BUSY;
5274 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5275 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5276 #ifdef RX_ENABLE_LOCKS
5277 osirx_AssertMine(&call->lock, "rxi_Start start");
5278 CV_BROADCAST(&call->cv_tq);
5279 #else /* RX_ENABLE_LOCKS */
5280 osi_rxWakeup(&call->tq);
5281 #endif /* RX_ENABLE_LOCKS */
5286 /* We went into the error state while sending packets. Now is
5287 * the time to reset the call. This will also inform the using
5288 * process that the call is in an error state.
5290 MUTEX_ENTER(&rx_stats_mutex);
5291 rx_tq_debug.rxi_start_aborted++;
5292 MUTEX_EXIT(&rx_stats_mutex);
5293 call->flags &= ~RX_CALL_TQ_BUSY;
5294 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5295 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5296 #ifdef RX_ENABLE_LOCKS
5297 osirx_AssertMine(&call->lock, "rxi_Start middle");
5298 CV_BROADCAST(&call->cv_tq);
5299 #else /* RX_ENABLE_LOCKS */
5300 osi_rxWakeup(&call->tq);
5301 #endif /* RX_ENABLE_LOCKS */
5303 rxi_CallError(call, call->error);
5306 #ifdef RX_ENABLE_LOCKS
5307 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5308 register int missing;
5309 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5310 /* Some packets have received acks. If they all have, we can clear
5311 * the transmit queue.
5314 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5315 if (p->header.seq < call->tfirst
5316 && (p->flags & RX_PKTFLAG_ACKED)) {
5323 call->flags |= RX_CALL_TQ_CLEARME;
5325 #endif /* RX_ENABLE_LOCKS */
5326 /* Don't bother doing retransmits if the TQ is cleared. */
5327 if (call->flags & RX_CALL_TQ_CLEARME) {
5328 rxi_ClearTransmitQueue(call, 1);
5330 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5333 /* Always post a resend event, if there is anything in the
5334 * queue, and resend is possible. There should be at least
5335 * one unacknowledged packet in the queue ... otherwise none
5336 * of these packets should be on the queue in the first place.
5338 if (call->resendEvent) {
5339 /* Cancel the existing event and post a new one */
5340 rxevent_Cancel(call->resendEvent, call,
5341 RX_CALL_REFCOUNT_RESEND);
5344 /* The retry time is the retry time on the first unacknowledged
5345 * packet inside the current window */
5347 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5348 /* Don't set timers for packets outside the window */
5349 if (p->header.seq >= call->tfirst + call->twind) {
5353 if (!(p->flags & RX_PKTFLAG_ACKED)
5354 && !clock_IsZero(&p->retryTime)) {
5356 retryTime = p->retryTime;
5361 /* Post a new event to re-run rxi_Start when retries may be needed */
5362 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5363 #ifdef RX_ENABLE_LOCKS
5364 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5366 rxevent_Post2(&retryTime, rxi_StartUnlocked,
5367 (void *)call, 0, istack);
5368 #else /* RX_ENABLE_LOCKS */
5370 rxevent_Post2(&retryTime, rxi_Start, (void *)call,
5372 #endif /* RX_ENABLE_LOCKS */
5375 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5376 } while (call->flags & RX_CALL_NEED_START);
5378 * TQ references no longer protected by this flag; they must remain
5379 * protected by the global lock.
5381 call->flags &= ~RX_CALL_TQ_BUSY;
5382 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5383 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5384 #ifdef RX_ENABLE_LOCKS
5385 osirx_AssertMine(&call->lock, "rxi_Start end");
5386 CV_BROADCAST(&call->cv_tq);
5387 #else /* RX_ENABLE_LOCKS */
5388 osi_rxWakeup(&call->tq);
5389 #endif /* RX_ENABLE_LOCKS */
5392 call->flags |= RX_CALL_NEED_START;
5394 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5396 if (call->resendEvent) {
5397 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5402 /* Also adjusts the keep alive parameters for the call, to reflect
5403 * that we have just sent a packet (so keep alives aren't sent
5406 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5409 register struct rx_connection *conn = call->conn;
5411 /* Stamp each packet with the user supplied status */
5412 p->header.userStatus = call->localStatus;
5414 /* Allow the security object controlling this call's security to
5415 * make any last-minute changes to the packet */
5416 RXS_SendPacket(conn->securityObject, call, p);
5418 /* Since we're about to send SOME sort of packet to the peer, it's
5419 * safe to nuke any scheduled end-of-packets ack */
5420 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5422 /* Actually send the packet, filling in more connection-specific fields */
5423 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5424 MUTEX_EXIT(&call->lock);
5425 rxi_SendPacket(call, conn, p, istack);
5426 MUTEX_ENTER(&call->lock);
5427 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5429 /* Update last send time for this call (for keep-alive
5430 * processing), and for the connection (so that we can discover
5431 * idle connections) */
5432 conn->lastSendTime = call->lastSendTime = clock_Sec();
5436 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5437 * that things are fine. Also called periodically to guarantee that nothing
5438 * falls through the cracks (e.g. (error + dally) connections have keepalive
5439 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5441 * haveCTLock Set if calling from rxi_ReapConnections
5443 #ifdef RX_ENABLE_LOCKS
5445 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5446 #else /* RX_ENABLE_LOCKS */
5448 rxi_CheckCall(register struct rx_call *call)
5449 #endif /* RX_ENABLE_LOCKS */
5451 register struct rx_connection *conn = call->conn;
5453 afs_uint32 deadTime;
5455 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5456 if (call->flags & RX_CALL_TQ_BUSY) {
5457 /* Call is active and will be reset by rxi_Start if it's
5458 * in an error state.
5463 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5465 (((afs_uint32) conn->secondsUntilDead << 10) +
5466 ((afs_uint32) conn->peer->rtt >> 3) +
5467 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5469 /* These are computed to the second (+- 1 second). But that's
5470 * good enough for these values, which should be a significant
5471 * number of seconds. */
5472 if (now > (call->lastReceiveTime + deadTime)) {
5473 if (call->state == RX_STATE_ACTIVE) {
5474 rxi_CallError(call, RX_CALL_DEAD);
5477 #ifdef RX_ENABLE_LOCKS
5478 /* Cancel pending events */
5479 rxevent_Cancel(call->delayedAckEvent, call,
5480 RX_CALL_REFCOUNT_DELAY);
5481 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5482 rxevent_Cancel(call->keepAliveEvent, call,
5483 RX_CALL_REFCOUNT_ALIVE);
5484 if (call->refCount == 0) {
5485 rxi_FreeCall(call, haveCTLock);
5489 #else /* RX_ENABLE_LOCKS */
5492 #endif /* RX_ENABLE_LOCKS */
5494 /* Non-active calls are destroyed if they are not responding
5495 * to pings; active calls are simply flagged in error, so the
5496 * attached process can die reasonably gracefully. */
5498 /* see if we have a non-activity timeout */
5499 if (call->startWait && conn->idleDeadTime
5500 && ((call->startWait + conn->idleDeadTime) < now)) {
5501 if (call->state == RX_STATE_ACTIVE) {
5502 rxi_CallError(call, RX_CALL_TIMEOUT);
5506 /* see if we have a hard timeout */
5507 if (conn->hardDeadTime
5508 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5509 if (call->state == RX_STATE_ACTIVE)
5510 rxi_CallError(call, RX_CALL_TIMEOUT);
5517 /* When a call is in progress, this routine is called occasionally to
5518 * make sure that some traffic has arrived (or been sent to) the peer.
5519 * If nothing has arrived in a reasonable amount of time, the call is
5520 * declared dead; if nothing has been sent for a while, we send a
5521 * keep-alive packet (if we're actually trying to keep the call alive)
5524 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5527 struct rx_connection *conn;
5530 MUTEX_ENTER(&call->lock);
5531 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5532 if (event == call->keepAliveEvent)
5533 call->keepAliveEvent = NULL;
5536 #ifdef RX_ENABLE_LOCKS
5537 if (rxi_CheckCall(call, 0)) {
5538 MUTEX_EXIT(&call->lock);
5541 #else /* RX_ENABLE_LOCKS */
5542 if (rxi_CheckCall(call))
5544 #endif /* RX_ENABLE_LOCKS */
5546 /* Don't try to keep alive dallying calls */
5547 if (call->state == RX_STATE_DALLY) {
5548 MUTEX_EXIT(&call->lock);
5553 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5554 /* Don't try to send keepalives if there is unacknowledged data */
5555 /* the rexmit code should be good enough, this little hack
5556 * doesn't quite work XXX */
5557 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5559 rxi_ScheduleKeepAliveEvent(call);
5560 MUTEX_EXIT(&call->lock);
5565 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5567 if (!call->keepAliveEvent) {
5569 clock_GetTime(&when);
5570 when.sec += call->conn->secondsUntilPing;
5571 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5572 call->keepAliveEvent =
5573 rxevent_Post(&when, rxi_KeepAliveEvent, call, 0);
5577 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5579 rxi_KeepAliveOn(register struct rx_call *call)
5581 /* Pretend last packet received was received now--i.e. if another
5582 * packet isn't received within the keep alive time, then the call
5583 * will die; Initialize last send time to the current time--even
5584 * if a packet hasn't been sent yet. This will guarantee that a
5585 * keep-alive is sent within the ping time */
5586 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5587 rxi_ScheduleKeepAliveEvent(call);
5590 /* This routine is called to send connection abort messages
5591 * that have been delayed to throttle looping clients. */
5593 rxi_SendDelayedConnAbort(struct rxevent *event,
5594 register struct rx_connection *conn, char *dummy)
5597 struct rx_packet *packet;
5599 MUTEX_ENTER(&conn->conn_data_lock);
5600 conn->delayedAbortEvent = NULL;
5601 error = htonl(conn->error);
5603 MUTEX_EXIT(&conn->conn_data_lock);
5604 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5607 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5608 RX_PACKET_TYPE_ABORT, (char *)&error,
5610 rxi_FreePacket(packet);
5614 /* This routine is called to send call abort messages
5615 * that have been delayed to throttle looping clients. */
5617 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5621 struct rx_packet *packet;
5623 MUTEX_ENTER(&call->lock);
5624 call->delayedAbortEvent = NULL;
5625 error = htonl(call->error);
5627 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5630 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5631 (char *)&error, sizeof(error), 0);
5632 rxi_FreePacket(packet);
5634 MUTEX_EXIT(&call->lock);
5637 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5638 * seconds) to ask the client to authenticate itself. The routine
5639 * issues a challenge to the client, which is obtained from the
5640 * security object associated with the connection */
5642 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5643 void *arg1, int tries)
5645 conn->challengeEvent = NULL;
5646 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5647 register struct rx_packet *packet;
5651 /* We've failed to authenticate for too long.
5652 * Reset any calls waiting for authentication;
5653 * they are all in RX_STATE_PRECALL.
5657 MUTEX_ENTER(&conn->conn_call_lock);
5658 for (i = 0; i < RX_MAXCALLS; i++) {
5659 struct rx_call *call = conn->call[i];
5661 MUTEX_ENTER(&call->lock);
5662 if (call->state == RX_STATE_PRECALL) {
5663 rxi_CallError(call, RX_CALL_DEAD);
5664 rxi_SendCallAbort(call, NULL, 0, 0);
5666 MUTEX_EXIT(&call->lock);
5669 MUTEX_EXIT(&conn->conn_call_lock);
5673 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5675 /* If there's no packet available, do this later. */
5676 RXS_GetChallenge(conn->securityObject, conn, packet);
5677 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5678 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5679 rxi_FreePacket(packet);
5681 clock_GetTime(&when);
5682 when.sec += RX_CHALLENGE_TIMEOUT;
5683 conn->challengeEvent =
5684 rxevent_Post2(&when, rxi_ChallengeEvent, conn, 0,
5689 /* Call this routine to start requesting the client to authenticate
5690 * itself. This will continue until authentication is established,
5691 * the call times out, or an invalid response is returned. The
5692 * security object associated with the connection is asked to create
5693 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5694 * defined earlier. */
5696 rxi_ChallengeOn(register struct rx_connection *conn)
5698 if (!conn->challengeEvent) {
5699 RXS_CreateChallenge(conn->securityObject, conn);
5700 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5705 /* Compute round trip time of the packet provided, in *rttp.
5708 /* rxi_ComputeRoundTripTime is called with peer locked. */
5709 /* sentp and/or peer may be null */
5711 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5712 register struct clock *sentp,
5713 register struct rx_peer *peer)
5715 struct clock thisRtt, *rttp = &thisRtt;
5717 register int rtt_timeout;
5719 clock_GetTime(rttp);
5721 if (clock_Lt(rttp, sentp)) {
5723 return; /* somebody set the clock back, don't count this time. */
5725 clock_Sub(rttp, sentp);
5726 MUTEX_ENTER(&rx_stats_mutex);
5727 if (clock_Lt(rttp, &rx_stats.minRtt))
5728 rx_stats.minRtt = *rttp;
5729 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5730 if (rttp->sec > 60) {
5731 MUTEX_EXIT(&rx_stats_mutex);
5732 return; /* somebody set the clock ahead */
5734 rx_stats.maxRtt = *rttp;
5736 clock_Add(&rx_stats.totalRtt, rttp);
5737 rx_stats.nRttSamples++;
5738 MUTEX_EXIT(&rx_stats_mutex);
5740 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5742 /* Apply VanJacobson round-trip estimations */
5747 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5748 * srtt is stored as fixed point with 3 bits after the binary
5749 * point (i.e., scaled by 8). The following magic is
5750 * equivalent to the smoothing algorithm in rfc793 with an
5751 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5752 * srtt*8 = srtt*8 + rtt - srtt
5753 * srtt = srtt + rtt/8 - srtt/8
5756 delta = MSEC(rttp) - (peer->rtt >> 3);
5760 * We accumulate a smoothed rtt variance (actually, a smoothed
5761 * mean difference), then set the retransmit timer to smoothed
5762 * rtt + 4 times the smoothed variance (was 2x in van's original
5763 * paper, but 4x works better for me, and apparently for him as
5765 * rttvar is stored as
5766 * fixed point with 2 bits after the binary point (scaled by
5767 * 4). The following is equivalent to rfc793 smoothing with
5768 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5769 * replaces rfc793's wired-in beta.
5770 * dev*4 = dev*4 + (|actual - expected| - dev)
5776 delta -= (peer->rtt_dev >> 2);
5777 peer->rtt_dev += delta;
5779 /* I don't have a stored RTT so I start with this value. Since I'm
5780 * probably just starting a call, and will be pushing more data down
5781 * this, I expect congestion to increase rapidly. So I fudge a
5782 * little, and I set deviance to half the rtt. In practice,
5783 * deviance tends to approach something a little less than
5784 * half the smoothed rtt. */
5785 peer->rtt = (MSEC(rttp) << 3) + 8;
5786 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5788 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5789 * the other of these connections is usually in a user process, and can
5790 * be switched and/or swapped out. So on fast, reliable networks, the
5791 * timeout would otherwise be too short.
5793 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5794 clock_Zero(&(peer->timeout));
5795 clock_Addmsec(&(peer->timeout), rtt_timeout);
5797 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)));
5801 /* Find all server connections that have not been active for a long time, and
5804 rxi_ReapConnections(void)
5807 clock_GetTime(&now);
5809 /* Find server connection structures that haven't been used for
5810 * greater than rx_idleConnectionTime */
5812 struct rx_connection **conn_ptr, **conn_end;
5813 int i, havecalls = 0;
5814 MUTEX_ENTER(&rx_connHashTable_lock);
5815 for (conn_ptr = &rx_connHashTable[0], conn_end =
5816 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5818 struct rx_connection *conn, *next;
5819 struct rx_call *call;
5823 for (conn = *conn_ptr; conn; conn = next) {
5824 /* XXX -- Shouldn't the connection be locked? */
5827 for (i = 0; i < RX_MAXCALLS; i++) {
5828 call = conn->call[i];
5831 MUTEX_ENTER(&call->lock);
5832 #ifdef RX_ENABLE_LOCKS
5833 result = rxi_CheckCall(call, 1);
5834 #else /* RX_ENABLE_LOCKS */
5835 result = rxi_CheckCall(call);
5836 #endif /* RX_ENABLE_LOCKS */
5837 MUTEX_EXIT(&call->lock);
5839 /* If CheckCall freed the call, it might
5840 * have destroyed the connection as well,
5841 * which screws up the linked lists.
5847 if (conn->type == RX_SERVER_CONNECTION) {
5848 /* This only actually destroys the connection if
5849 * there are no outstanding calls */
5850 MUTEX_ENTER(&conn->conn_data_lock);
5851 if (!havecalls && !conn->refCount
5852 && ((conn->lastSendTime + rx_idleConnectionTime) <
5854 conn->refCount++; /* it will be decr in rx_DestroyConn */
5855 MUTEX_EXIT(&conn->conn_data_lock);
5856 #ifdef RX_ENABLE_LOCKS
5857 rxi_DestroyConnectionNoLock(conn);
5858 #else /* RX_ENABLE_LOCKS */
5859 rxi_DestroyConnection(conn);
5860 #endif /* RX_ENABLE_LOCKS */
5862 #ifdef RX_ENABLE_LOCKS
5864 MUTEX_EXIT(&conn->conn_data_lock);
5866 #endif /* RX_ENABLE_LOCKS */
5870 #ifdef RX_ENABLE_LOCKS
5871 while (rx_connCleanup_list) {
5872 struct rx_connection *conn;
5873 conn = rx_connCleanup_list;
5874 rx_connCleanup_list = rx_connCleanup_list->next;
5875 MUTEX_EXIT(&rx_connHashTable_lock);
5876 rxi_CleanupConnection(conn);
5877 MUTEX_ENTER(&rx_connHashTable_lock);
5879 MUTEX_EXIT(&rx_connHashTable_lock);
5880 #endif /* RX_ENABLE_LOCKS */
5883 /* Find any peer structures that haven't been used (haven't had an
5884 * associated connection) for greater than rx_idlePeerTime */
5886 struct rx_peer **peer_ptr, **peer_end;
5888 MUTEX_ENTER(&rx_rpc_stats);
5889 MUTEX_ENTER(&rx_peerHashTable_lock);
5890 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5891 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5893 struct rx_peer *peer, *next, *prev;
5894 for (prev = peer = *peer_ptr; peer; peer = next) {
5896 code = MUTEX_TRYENTER(&peer->peer_lock);
5897 if ((code) && (peer->refCount == 0)
5898 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5899 rx_interface_stat_p rpc_stat, nrpc_stat;
5901 MUTEX_EXIT(&peer->peer_lock);
5902 MUTEX_DESTROY(&peer->peer_lock);
5904 (&peer->rpcStats, rpc_stat, nrpc_stat,
5905 rx_interface_stat)) {
5906 unsigned int num_funcs;
5909 queue_Remove(&rpc_stat->queue_header);
5910 queue_Remove(&rpc_stat->all_peers);
5911 num_funcs = rpc_stat->stats[0].func_total;
5913 sizeof(rx_interface_stat_t) +
5914 rpc_stat->stats[0].func_total *
5915 sizeof(rx_function_entry_v1_t);
5917 rxi_Free(rpc_stat, space);
5918 rxi_rpc_peer_stat_cnt -= num_funcs;
5921 MUTEX_ENTER(&rx_stats_mutex);
5922 rx_stats.nPeerStructs--;
5923 MUTEX_EXIT(&rx_stats_mutex);
5924 if (peer == *peer_ptr) {
5931 MUTEX_EXIT(&peer->peer_lock);
5937 MUTEX_EXIT(&rx_peerHashTable_lock);
5938 MUTEX_EXIT(&rx_rpc_stats);
5941 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5942 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5943 * GC, just below. Really, we shouldn't have to keep moving packets from
5944 * one place to another, but instead ought to always know if we can
5945 * afford to hold onto a packet in its particular use. */
5946 MUTEX_ENTER(&rx_freePktQ_lock);
5947 if (rx_waitingForPackets) {
5948 rx_waitingForPackets = 0;
5949 #ifdef RX_ENABLE_LOCKS
5950 CV_BROADCAST(&rx_waitingForPackets_cv);
5952 osi_rxWakeup(&rx_waitingForPackets);
5955 MUTEX_EXIT(&rx_freePktQ_lock);
5957 now.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5958 rxevent_Post(&now, rxi_ReapConnections, 0, 0);
5962 /* rxs_Release - This isn't strictly necessary but, since the macro name from
5963 * rx.h is sort of strange this is better. This is called with a security
5964 * object before it is discarded. Each connection using a security object has
5965 * its own refcount to the object so it won't actually be freed until the last
5966 * connection is destroyed.
5968 * This is the only rxs module call. A hold could also be written but no one
5972 rxs_Release(struct rx_securityClass *aobj)
5974 return RXS_Close(aobj);
5978 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
5979 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
5980 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
5981 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
5983 /* Adjust our estimate of the transmission rate to this peer, given
5984 * that the packet p was just acked. We can adjust peer->timeout and
5985 * call->twind. Pragmatically, this is called
5986 * only with packets of maximal length.
5987 * Called with peer and call locked.
5991 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
5992 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
5994 afs_int32 xferSize, xferMs;
5995 register afs_int32 minTime;
5998 /* Count down packets */
5999 if (peer->rateFlag > 0)
6001 /* Do nothing until we're enabled */
6002 if (peer->rateFlag != 0)
6007 /* Count only when the ack seems legitimate */
6008 switch (ackReason) {
6009 case RX_ACK_REQUESTED:
6011 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6015 case RX_ACK_PING_RESPONSE:
6016 if (p) /* want the response to ping-request, not data send */
6018 clock_GetTime(&newTO);
6019 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6020 clock_Sub(&newTO, &call->pingRequestTime);
6021 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6025 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6032 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));
6034 /* Track only packets that are big enough. */
6035 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6039 /* absorb RTT data (in milliseconds) for these big packets */
6040 if (peer->smRtt == 0) {
6041 peer->smRtt = xferMs;
6043 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6048 if (peer->countDown) {
6052 peer->countDown = 10; /* recalculate only every so often */
6054 /* In practice, we can measure only the RTT for full packets,
6055 * because of the way Rx acks the data that it receives. (If it's
6056 * smaller than a full packet, it often gets implicitly acked
6057 * either by the call response (from a server) or by the next call
6058 * (from a client), and either case confuses transmission times
6059 * with processing times.) Therefore, replace the above
6060 * more-sophisticated processing with a simpler version, where the
6061 * smoothed RTT is kept for full-size packets, and the time to
6062 * transmit a windowful of full-size packets is simply RTT *
6063 * windowSize. Again, we take two steps:
6064 - ensure the timeout is large enough for a single packet's RTT;
6065 - ensure that the window is small enough to fit in the desired timeout.*/
6067 /* First, the timeout check. */
6068 minTime = peer->smRtt;
6069 /* Get a reasonable estimate for a timeout period */
6071 newTO.sec = minTime / 1000;
6072 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6074 /* Increase the timeout period so that we can always do at least
6075 * one packet exchange */
6076 if (clock_Gt(&newTO, &peer->timeout)) {
6078 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));
6080 peer->timeout = newTO;
6083 /* Now, get an estimate for the transmit window size. */
6084 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6085 /* Now, convert to the number of full packets that could fit in a
6086 * reasonable fraction of that interval */
6087 minTime /= (peer->smRtt << 1);
6088 xferSize = minTime; /* (make a copy) */
6090 /* Now clamp the size to reasonable bounds. */
6093 else if (minTime > rx_Window)
6094 minTime = rx_Window;
6095 /* if (minTime != peer->maxWindow) {
6096 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6097 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6098 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6100 peer->maxWindow = minTime;
6101 elide... call->twind = minTime;
6105 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6106 * Discern this by calculating the timeout necessary for rx_Window
6108 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6109 /* calculate estimate for transmission interval in milliseconds */
6110 minTime = rx_Window * peer->smRtt;
6111 if (minTime < 1000) {
6112 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
6113 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6114 peer->timeout.usec, peer->smRtt, peer->packetSize));
6116 newTO.sec = 0; /* cut back on timeout by half a second */
6117 newTO.usec = 500000;
6118 clock_Sub(&peer->timeout, &newTO);
6123 } /* end of rxi_ComputeRate */
6124 #endif /* ADAPT_WINDOW */
6132 #define TRACE_OPTION_DEBUGLOG 4
6140 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6141 0, KEY_QUERY_VALUE, &parmKey);
6142 if (code != ERROR_SUCCESS)
6145 dummyLen = sizeof(TraceOption);
6146 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6147 (BYTE *) &TraceOption, &dummyLen);
6148 if (code == ERROR_SUCCESS) {
6149 rxdebug_active = (TraceOption & TRACE_OPTION_DEBUGLOG) ? 1 : 0;
6151 RegCloseKey (parmKey);
6152 #endif /* AFS_NT40_ENV */
6157 rx_DebugOnOff(int on)
6159 rxdebug_active = on;
6161 #endif /* AFS_NT40_ENV */
6164 /* Don't call this debugging routine directly; use dpf */
6166 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
6167 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
6175 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6178 len = _snprintf(msg, sizeof(msg)-2,
6179 tformat, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10,
6180 a11, a12, a13, a14, a15);
6182 if (msg[len-1] != '\n') {
6186 OutputDebugString(msg);
6191 clock_GetTime(&now);
6192 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
6193 (unsigned int)now.usec / 1000);
6194 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
6201 * This function is used to process the rx_stats structure that is local
6202 * to a process as well as an rx_stats structure received from a remote
6203 * process (via rxdebug). Therefore, it needs to do minimal version
6207 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
6208 afs_int32 freePackets, char version)
6212 if (size != sizeof(struct rx_stats)) {
6214 "Unexpected size of stats structure: was %d, expected %d\n",
6215 size, sizeof(struct rx_stats));
6218 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6221 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6222 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6223 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6224 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6225 s->specialPktAllocFailures);
6227 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6228 s->receivePktAllocFailures, s->sendPktAllocFailures,
6229 s->specialPktAllocFailures);
6233 " greedy %d, " "bogusReads %d (last from host %x), "
6234 "noPackets %d, " "noBuffers %d, " "selects %d, "
6235 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6236 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6237 s->selects, s->sendSelects);
6239 fprintf(file, " packets read: ");
6240 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6241 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6243 fprintf(file, "\n");
6246 " other read counters: data %d, " "ack %d, " "dup %d "
6247 "spurious %d " "dally %d\n", s->dataPacketsRead,
6248 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6249 s->ignorePacketDally);
6251 fprintf(file, " packets sent: ");
6252 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6253 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6255 fprintf(file, "\n");
6258 " other send counters: ack %d, " "data %d (not resends), "
6259 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6260 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6261 s->dataPacketsPushed, s->ignoreAckedPacket);
6264 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6265 s->netSendFailures, (int)s->fatalErrors);
6267 if (s->nRttSamples) {
6268 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6269 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6271 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6272 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6276 " %d server connections, " "%d client connections, "
6277 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6278 s->nServerConns, s->nClientConns, s->nPeerStructs,
6279 s->nCallStructs, s->nFreeCallStructs);
6281 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6282 fprintf(file, " %d clock updates\n", clock_nUpdates);
6287 /* for backward compatibility */
6289 rx_PrintStats(FILE * file)
6291 MUTEX_ENTER(&rx_stats_mutex);
6292 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6294 MUTEX_EXIT(&rx_stats_mutex);
6298 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6300 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6301 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6302 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6305 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6306 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6307 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6310 " Packet size %d, " "max in packet skew %d, "
6311 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6312 (int)peer->outPacketSkew);
6315 #ifdef AFS_PTHREAD_ENV
6317 * This mutex protects the following static variables:
6321 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6322 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6324 #define LOCK_RX_DEBUG
6325 #define UNLOCK_RX_DEBUG
6326 #endif /* AFS_PTHREAD_ENV */
6329 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6330 u_char type, void *inputData, size_t inputLength,
6331 void *outputData, size_t outputLength)
6333 static afs_int32 counter = 100;
6335 struct rx_header theader;
6337 register afs_int32 code;
6339 struct sockaddr_in taddr, faddr;
6344 endTime = time(0) + 20; /* try for 20 seconds */
6348 tp = &tbuffer[sizeof(struct rx_header)];
6349 taddr.sin_family = AF_INET;
6350 taddr.sin_port = remotePort;
6351 taddr.sin_addr.s_addr = remoteAddr;
6352 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6353 taddr.sin_len = sizeof(struct sockaddr_in);
6356 memset(&theader, 0, sizeof(theader));
6357 theader.epoch = htonl(999);
6359 theader.callNumber = htonl(counter);
6362 theader.type = type;
6363 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6364 theader.serviceId = 0;
6366 memcpy(tbuffer, &theader, sizeof(theader));
6367 memcpy(tp, inputData, inputLength);
6369 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6370 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6372 /* see if there's a packet available */
6374 FD_SET(socket, &imask);
6377 code = select((int)(socket + 1), &imask, 0, 0, &tv);
6378 if (code == 1 && FD_ISSET(socket, &imask)) {
6379 /* now receive a packet */
6380 faddrLen = sizeof(struct sockaddr_in);
6382 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6383 (struct sockaddr *)&faddr, &faddrLen);
6386 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6387 if (counter == ntohl(theader.callNumber))
6392 /* see if we've timed out */
6393 if (endTime < time(0))
6396 code -= sizeof(struct rx_header);
6397 if (code > outputLength)
6398 code = outputLength;
6399 memcpy(outputData, tp, code);
6404 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6405 afs_uint16 remotePort, struct rx_debugStats * stat,
6406 afs_uint32 * supportedValues)
6408 struct rx_debugIn in;
6411 *supportedValues = 0;
6412 in.type = htonl(RX_DEBUGI_GETSTATS);
6415 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6416 &in, sizeof(in), stat, sizeof(*stat));
6419 * If the call was successful, fixup the version and indicate
6420 * what contents of the stat structure are valid.
6421 * Also do net to host conversion of fields here.
6425 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6426 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6428 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6429 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6431 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6432 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6434 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6435 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6437 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6438 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6440 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6441 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6443 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6444 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6446 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6447 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6450 stat->nFreePackets = ntohl(stat->nFreePackets);
6451 stat->packetReclaims = ntohl(stat->packetReclaims);
6452 stat->callsExecuted = ntohl(stat->callsExecuted);
6453 stat->nWaiting = ntohl(stat->nWaiting);
6454 stat->idleThreads = ntohl(stat->idleThreads);
6461 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6462 afs_uint16 remotePort, struct rx_stats * stat,
6463 afs_uint32 * supportedValues)
6465 struct rx_debugIn in;
6466 afs_int32 *lp = (afs_int32 *) stat;
6471 * supportedValues is currently unused, but added to allow future
6472 * versioning of this function.
6475 *supportedValues = 0;
6476 in.type = htonl(RX_DEBUGI_RXSTATS);
6478 memset(stat, 0, sizeof(*stat));
6480 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6481 &in, sizeof(in), stat, sizeof(*stat));
6486 * Do net to host conversion here
6489 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6498 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6499 afs_uint16 remotePort, size_t version_length,
6503 return MakeDebugCall(socket, remoteAddr, remotePort,
6504 RX_PACKET_TYPE_VERSION, a, 1, version,
6509 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6510 afs_uint16 remotePort, afs_int32 * nextConnection,
6511 int allConnections, afs_uint32 debugSupportedValues,
6512 struct rx_debugConn * conn,
6513 afs_uint32 * supportedValues)
6515 struct rx_debugIn in;
6520 * supportedValues is currently unused, but added to allow future
6521 * versioning of this function.
6524 *supportedValues = 0;
6525 if (allConnections) {
6526 in.type = htonl(RX_DEBUGI_GETALLCONN);
6528 in.type = htonl(RX_DEBUGI_GETCONN);
6530 in.index = htonl(*nextConnection);
6531 memset(conn, 0, sizeof(*conn));
6533 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6534 &in, sizeof(in), conn, sizeof(*conn));
6537 *nextConnection += 1;
6540 * Convert old connection format to new structure.
6543 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6544 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6545 #define MOVEvL(a) (conn->a = vL->a)
6547 /* any old or unrecognized version... */
6548 for (i = 0; i < RX_MAXCALLS; i++) {
6549 MOVEvL(callState[i]);
6550 MOVEvL(callMode[i]);
6551 MOVEvL(callFlags[i]);
6552 MOVEvL(callOther[i]);
6554 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6555 MOVEvL(secStats.type);
6556 MOVEvL(secStats.level);
6557 MOVEvL(secStats.flags);
6558 MOVEvL(secStats.expires);
6559 MOVEvL(secStats.packetsReceived);
6560 MOVEvL(secStats.packetsSent);
6561 MOVEvL(secStats.bytesReceived);
6562 MOVEvL(secStats.bytesSent);
6567 * Do net to host conversion here
6569 * I don't convert host or port since we are most likely
6570 * going to want these in NBO.
6572 conn->cid = ntohl(conn->cid);
6573 conn->serial = ntohl(conn->serial);
6574 for (i = 0; i < RX_MAXCALLS; i++) {
6575 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6577 conn->error = ntohl(conn->error);
6578 conn->secStats.flags = ntohl(conn->secStats.flags);
6579 conn->secStats.expires = ntohl(conn->secStats.expires);
6580 conn->secStats.packetsReceived =
6581 ntohl(conn->secStats.packetsReceived);
6582 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6583 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6584 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6585 conn->epoch = ntohl(conn->epoch);
6586 conn->natMTU = ntohl(conn->natMTU);
6593 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6594 afs_uint16 remotePort, afs_int32 * nextPeer,
6595 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6596 afs_uint32 * supportedValues)
6598 struct rx_debugIn in;
6602 * supportedValues is currently unused, but added to allow future
6603 * versioning of this function.
6606 *supportedValues = 0;
6607 in.type = htonl(RX_DEBUGI_GETPEER);
6608 in.index = htonl(*nextPeer);
6609 memset(peer, 0, sizeof(*peer));
6611 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6612 &in, sizeof(in), peer, sizeof(*peer));
6618 * Do net to host conversion here
6620 * I don't convert host or port since we are most likely
6621 * going to want these in NBO.
6623 peer->ifMTU = ntohs(peer->ifMTU);
6624 peer->idleWhen = ntohl(peer->idleWhen);
6625 peer->refCount = ntohs(peer->refCount);
6626 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6627 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6628 peer->rtt = ntohl(peer->rtt);
6629 peer->rtt_dev = ntohl(peer->rtt_dev);
6630 peer->timeout.sec = ntohl(peer->timeout.sec);
6631 peer->timeout.usec = ntohl(peer->timeout.usec);
6632 peer->nSent = ntohl(peer->nSent);
6633 peer->reSends = ntohl(peer->reSends);
6634 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6635 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6636 peer->rateFlag = ntohl(peer->rateFlag);
6637 peer->natMTU = ntohs(peer->natMTU);
6638 peer->maxMTU = ntohs(peer->maxMTU);
6639 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6640 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6641 peer->MTU = ntohs(peer->MTU);
6642 peer->cwind = ntohs(peer->cwind);
6643 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6644 peer->congestSeq = ntohs(peer->congestSeq);
6645 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6646 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6647 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6648 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6653 #endif /* RXDEBUG */
6658 struct rx_serverQueueEntry *np;
6661 register struct rx_call *call;
6662 register struct rx_serverQueueEntry *sq;
6666 if (rxinit_status == 1) {
6668 return; /* Already shutdown. */
6672 #ifndef AFS_PTHREAD_ENV
6673 FD_ZERO(&rx_selectMask);
6674 #endif /* AFS_PTHREAD_ENV */
6675 rxi_dataQuota = RX_MAX_QUOTA;
6676 #ifndef AFS_PTHREAD_ENV
6678 #endif /* AFS_PTHREAD_ENV */
6681 #ifndef AFS_PTHREAD_ENV
6682 #ifndef AFS_USE_GETTIMEOFDAY
6684 #endif /* AFS_USE_GETTIMEOFDAY */
6685 #endif /* AFS_PTHREAD_ENV */
6687 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6688 call = queue_First(&rx_freeCallQueue, rx_call);
6690 rxi_Free(call, sizeof(struct rx_call));
6693 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6694 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6700 struct rx_peer **peer_ptr, **peer_end;
6701 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6702 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6704 struct rx_peer *peer, *next;
6705 for (peer = *peer_ptr; peer; peer = next) {
6706 rx_interface_stat_p rpc_stat, nrpc_stat;
6709 (&peer->rpcStats, rpc_stat, nrpc_stat,
6710 rx_interface_stat)) {
6711 unsigned int num_funcs;
6714 queue_Remove(&rpc_stat->queue_header);
6715 queue_Remove(&rpc_stat->all_peers);
6716 num_funcs = rpc_stat->stats[0].func_total;
6718 sizeof(rx_interface_stat_t) +
6719 rpc_stat->stats[0].func_total *
6720 sizeof(rx_function_entry_v1_t);
6722 rxi_Free(rpc_stat, space);
6723 MUTEX_ENTER(&rx_rpc_stats);
6724 rxi_rpc_peer_stat_cnt -= num_funcs;
6725 MUTEX_EXIT(&rx_rpc_stats);
6729 MUTEX_ENTER(&rx_stats_mutex);
6730 rx_stats.nPeerStructs--;
6731 MUTEX_EXIT(&rx_stats_mutex);
6735 for (i = 0; i < RX_MAX_SERVICES; i++) {
6737 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6739 for (i = 0; i < rx_hashTableSize; i++) {
6740 register struct rx_connection *tc, *ntc;
6741 MUTEX_ENTER(&rx_connHashTable_lock);
6742 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6744 for (j = 0; j < RX_MAXCALLS; j++) {
6746 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6749 rxi_Free(tc, sizeof(*tc));
6751 MUTEX_EXIT(&rx_connHashTable_lock);
6754 MUTEX_ENTER(&freeSQEList_lock);
6756 while ((np = rx_FreeSQEList)) {
6757 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6758 MUTEX_DESTROY(&np->lock);
6759 rxi_Free(np, sizeof(*np));
6762 MUTEX_EXIT(&freeSQEList_lock);
6763 MUTEX_DESTROY(&freeSQEList_lock);
6764 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6765 MUTEX_DESTROY(&rx_connHashTable_lock);
6766 MUTEX_DESTROY(&rx_peerHashTable_lock);
6767 MUTEX_DESTROY(&rx_serverPool_lock);
6769 osi_Free(rx_connHashTable,
6770 rx_hashTableSize * sizeof(struct rx_connection *));
6771 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6773 UNPIN(rx_connHashTable,
6774 rx_hashTableSize * sizeof(struct rx_connection *));
6775 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6777 rxi_FreeAllPackets();
6779 MUTEX_ENTER(&rx_stats_mutex);
6780 rxi_dataQuota = RX_MAX_QUOTA;
6781 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6782 MUTEX_EXIT(&rx_stats_mutex);
6788 #ifdef RX_ENABLE_LOCKS
6790 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6792 if (!MUTEX_ISMINE(lockaddr))
6793 osi_Panic("Lock not held: %s", msg);
6795 #endif /* RX_ENABLE_LOCKS */
6800 * Routines to implement connection specific data.
6804 rx_KeyCreate(rx_destructor_t rtn)
6807 MUTEX_ENTER(&rxi_keyCreate_lock);
6808 key = rxi_keyCreate_counter++;
6809 rxi_keyCreate_destructor = (rx_destructor_t *)
6810 realloc((void *)rxi_keyCreate_destructor,
6811 (key + 1) * sizeof(rx_destructor_t));
6812 rxi_keyCreate_destructor[key] = rtn;
6813 MUTEX_EXIT(&rxi_keyCreate_lock);
6818 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6821 MUTEX_ENTER(&conn->conn_data_lock);
6822 if (!conn->specific) {
6823 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6824 for (i = 0; i < key; i++)
6825 conn->specific[i] = NULL;
6826 conn->nSpecific = key + 1;
6827 conn->specific[key] = ptr;
6828 } else if (key >= conn->nSpecific) {
6829 conn->specific = (void **)
6830 realloc(conn->specific, (key + 1) * sizeof(void *));
6831 for (i = conn->nSpecific; i < key; i++)
6832 conn->specific[i] = NULL;
6833 conn->nSpecific = key + 1;
6834 conn->specific[key] = ptr;
6836 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6837 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6838 conn->specific[key] = ptr;
6840 MUTEX_EXIT(&conn->conn_data_lock);
6844 rx_GetSpecific(struct rx_connection *conn, int key)
6847 MUTEX_ENTER(&conn->conn_data_lock);
6848 if (key >= conn->nSpecific)
6851 ptr = conn->specific[key];
6852 MUTEX_EXIT(&conn->conn_data_lock);
6856 #endif /* !KERNEL */
6859 * processStats is a queue used to store the statistics for the local
6860 * process. Its contents are similar to the contents of the rpcStats
6861 * queue on a rx_peer structure, but the actual data stored within
6862 * this queue contains totals across the lifetime of the process (assuming
6863 * the stats have not been reset) - unlike the per peer structures
6864 * which can come and go based upon the peer lifetime.
6867 static struct rx_queue processStats = { &processStats, &processStats };
6870 * peerStats is a queue used to store the statistics for all peer structs.
6871 * Its contents are the union of all the peer rpcStats queues.
6874 static struct rx_queue peerStats = { &peerStats, &peerStats };
6877 * rxi_monitor_processStats is used to turn process wide stat collection
6881 static int rxi_monitor_processStats = 0;
6884 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6887 static int rxi_monitor_peerStats = 0;
6890 * rxi_AddRpcStat - given all of the information for a particular rpc
6891 * call, create (if needed) and update the stat totals for the rpc.
6895 * IN stats - the queue of stats that will be updated with the new value
6897 * IN rxInterface - a unique number that identifies the rpc interface
6899 * IN currentFunc - the index of the function being invoked
6901 * IN totalFunc - the total number of functions in this interface
6903 * IN queueTime - the amount of time this function waited for a thread
6905 * IN execTime - the amount of time this function invocation took to execute
6907 * IN bytesSent - the number bytes sent by this invocation
6909 * IN bytesRcvd - the number bytes received by this invocation
6911 * IN isServer - if true, this invocation was made to a server
6913 * IN remoteHost - the ip address of the remote host
6915 * IN remotePort - the port of the remote host
6917 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6919 * INOUT counter - if a new stats structure is allocated, the counter will
6920 * be updated with the new number of allocated stat structures
6928 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6929 afs_uint32 currentFunc, afs_uint32 totalFunc,
6930 struct clock *queueTime, struct clock *execTime,
6931 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6932 afs_uint32 remoteHost, afs_uint32 remotePort,
6933 int addToPeerList, unsigned int *counter)
6936 rx_interface_stat_p rpc_stat, nrpc_stat;
6939 * See if there's already a structure for this interface
6942 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6943 if ((rpc_stat->stats[0].interfaceId == rxInterface)
6944 && (rpc_stat->stats[0].remote_is_server == isServer))
6949 * Didn't find a match so allocate a new structure and add it to the
6953 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
6954 || (rpc_stat->stats[0].interfaceId != rxInterface)
6955 || (rpc_stat->stats[0].remote_is_server != isServer)) {
6960 sizeof(rx_interface_stat_t) +
6961 totalFunc * sizeof(rx_function_entry_v1_t);
6963 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
6964 if (rpc_stat == NULL) {
6968 *counter += totalFunc;
6969 for (i = 0; i < totalFunc; i++) {
6970 rpc_stat->stats[i].remote_peer = remoteHost;
6971 rpc_stat->stats[i].remote_port = remotePort;
6972 rpc_stat->stats[i].remote_is_server = isServer;
6973 rpc_stat->stats[i].interfaceId = rxInterface;
6974 rpc_stat->stats[i].func_total = totalFunc;
6975 rpc_stat->stats[i].func_index = i;
6976 hzero(rpc_stat->stats[i].invocations);
6977 hzero(rpc_stat->stats[i].bytes_sent);
6978 hzero(rpc_stat->stats[i].bytes_rcvd);
6979 rpc_stat->stats[i].queue_time_sum.sec = 0;
6980 rpc_stat->stats[i].queue_time_sum.usec = 0;
6981 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
6982 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
6983 rpc_stat->stats[i].queue_time_min.sec = 9999999;
6984 rpc_stat->stats[i].queue_time_min.usec = 9999999;
6985 rpc_stat->stats[i].queue_time_max.sec = 0;
6986 rpc_stat->stats[i].queue_time_max.usec = 0;
6987 rpc_stat->stats[i].execution_time_sum.sec = 0;
6988 rpc_stat->stats[i].execution_time_sum.usec = 0;
6989 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
6990 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
6991 rpc_stat->stats[i].execution_time_min.sec = 9999999;
6992 rpc_stat->stats[i].execution_time_min.usec = 9999999;
6993 rpc_stat->stats[i].execution_time_max.sec = 0;
6994 rpc_stat->stats[i].execution_time_max.usec = 0;
6996 queue_Prepend(stats, rpc_stat);
6997 if (addToPeerList) {
6998 queue_Prepend(&peerStats, &rpc_stat->all_peers);
7003 * Increment the stats for this function
7006 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
7007 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
7008 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
7009 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
7010 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7011 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7012 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7014 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7015 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7017 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7018 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7020 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7021 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7023 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7024 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7032 * rx_IncrementTimeAndCount - increment the times and count for a particular
7037 * IN peer - the peer who invoked the rpc
7039 * IN rxInterface - a unique number that identifies the rpc interface
7041 * IN currentFunc - the index of the function being invoked
7043 * IN totalFunc - the total number of functions in this interface
7045 * IN queueTime - the amount of time this function waited for a thread
7047 * IN execTime - the amount of time this function invocation took to execute
7049 * IN bytesSent - the number bytes sent by this invocation
7051 * IN bytesRcvd - the number bytes received by this invocation
7053 * IN isServer - if true, this invocation was made to a server
7061 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7062 afs_uint32 currentFunc, afs_uint32 totalFunc,
7063 struct clock *queueTime, struct clock *execTime,
7064 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7068 MUTEX_ENTER(&rx_rpc_stats);
7069 MUTEX_ENTER(&peer->peer_lock);
7071 if (rxi_monitor_peerStats) {
7072 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7073 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7074 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7077 if (rxi_monitor_processStats) {
7078 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7079 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7080 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7083 MUTEX_EXIT(&peer->peer_lock);
7084 MUTEX_EXIT(&rx_rpc_stats);
7089 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7093 * IN callerVersion - the rpc stat version of the caller.
7095 * IN count - the number of entries to marshall.
7097 * IN stats - pointer to stats to be marshalled.
7099 * OUT ptr - Where to store the marshalled data.
7106 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7107 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7113 * We only support the first version
7115 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7116 *(ptr++) = stats->remote_peer;
7117 *(ptr++) = stats->remote_port;
7118 *(ptr++) = stats->remote_is_server;
7119 *(ptr++) = stats->interfaceId;
7120 *(ptr++) = stats->func_total;
7121 *(ptr++) = stats->func_index;
7122 *(ptr++) = hgethi(stats->invocations);
7123 *(ptr++) = hgetlo(stats->invocations);
7124 *(ptr++) = hgethi(stats->bytes_sent);
7125 *(ptr++) = hgetlo(stats->bytes_sent);
7126 *(ptr++) = hgethi(stats->bytes_rcvd);
7127 *(ptr++) = hgetlo(stats->bytes_rcvd);
7128 *(ptr++) = stats->queue_time_sum.sec;
7129 *(ptr++) = stats->queue_time_sum.usec;
7130 *(ptr++) = stats->queue_time_sum_sqr.sec;
7131 *(ptr++) = stats->queue_time_sum_sqr.usec;
7132 *(ptr++) = stats->queue_time_min.sec;
7133 *(ptr++) = stats->queue_time_min.usec;
7134 *(ptr++) = stats->queue_time_max.sec;
7135 *(ptr++) = stats->queue_time_max.usec;
7136 *(ptr++) = stats->execution_time_sum.sec;
7137 *(ptr++) = stats->execution_time_sum.usec;
7138 *(ptr++) = stats->execution_time_sum_sqr.sec;
7139 *(ptr++) = stats->execution_time_sum_sqr.usec;
7140 *(ptr++) = stats->execution_time_min.sec;
7141 *(ptr++) = stats->execution_time_min.usec;
7142 *(ptr++) = stats->execution_time_max.sec;
7143 *(ptr++) = stats->execution_time_max.usec;
7149 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7154 * IN callerVersion - the rpc stat version of the caller
7156 * OUT myVersion - the rpc stat version of this function
7158 * OUT clock_sec - local time seconds
7160 * OUT clock_usec - local time microseconds
7162 * OUT allocSize - the number of bytes allocated to contain stats
7164 * OUT statCount - the number stats retrieved from this process.
7166 * OUT stats - the actual stats retrieved from this process.
7170 * Returns void. If successful, stats will != NULL.
7174 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7175 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7176 size_t * allocSize, afs_uint32 * statCount,
7177 afs_uint32 ** stats)
7187 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7190 * Check to see if stats are enabled
7193 MUTEX_ENTER(&rx_rpc_stats);
7194 if (!rxi_monitor_processStats) {
7195 MUTEX_EXIT(&rx_rpc_stats);
7199 clock_GetTime(&now);
7200 *clock_sec = now.sec;
7201 *clock_usec = now.usec;
7204 * Allocate the space based upon the caller version
7206 * If the client is at an older version than we are,
7207 * we return the statistic data in the older data format, but
7208 * we still return our version number so the client knows we
7209 * are maintaining more data than it can retrieve.
7212 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7213 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7214 *statCount = rxi_rpc_process_stat_cnt;
7217 * This can't happen yet, but in the future version changes
7218 * can be handled by adding additional code here
7222 if (space > (size_t) 0) {
7224 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7227 rx_interface_stat_p rpc_stat, nrpc_stat;
7231 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7233 * Copy the data based upon the caller version
7235 rx_MarshallProcessRPCStats(callerVersion,
7236 rpc_stat->stats[0].func_total,
7237 rpc_stat->stats, &ptr);
7243 MUTEX_EXIT(&rx_rpc_stats);
7248 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7252 * IN callerVersion - the rpc stat version of the caller
7254 * OUT myVersion - the rpc stat version of this function
7256 * OUT clock_sec - local time seconds
7258 * OUT clock_usec - local time microseconds
7260 * OUT allocSize - the number of bytes allocated to contain stats
7262 * OUT statCount - the number of stats retrieved from the individual
7265 * OUT stats - the actual stats retrieved from the individual peer structures.
7269 * Returns void. If successful, stats will != NULL.
7273 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7274 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7275 size_t * allocSize, afs_uint32 * statCount,
7276 afs_uint32 ** stats)
7286 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7289 * Check to see if stats are enabled
7292 MUTEX_ENTER(&rx_rpc_stats);
7293 if (!rxi_monitor_peerStats) {
7294 MUTEX_EXIT(&rx_rpc_stats);
7298 clock_GetTime(&now);
7299 *clock_sec = now.sec;
7300 *clock_usec = now.usec;
7303 * Allocate the space based upon the caller version
7305 * If the client is at an older version than we are,
7306 * we return the statistic data in the older data format, but
7307 * we still return our version number so the client knows we
7308 * are maintaining more data than it can retrieve.
7311 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7312 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7313 *statCount = rxi_rpc_peer_stat_cnt;
7316 * This can't happen yet, but in the future version changes
7317 * can be handled by adding additional code here
7321 if (space > (size_t) 0) {
7323 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7326 rx_interface_stat_p rpc_stat, nrpc_stat;
7330 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7332 * We have to fix the offset of rpc_stat since we are
7333 * keeping this structure on two rx_queues. The rx_queue
7334 * package assumes that the rx_queue member is the first
7335 * member of the structure. That is, rx_queue assumes that
7336 * any one item is only on one queue at a time. We are
7337 * breaking that assumption and so we have to do a little
7338 * math to fix our pointers.
7341 fix_offset = (char *)rpc_stat;
7342 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7343 rpc_stat = (rx_interface_stat_p) fix_offset;
7346 * Copy the data based upon the caller version
7348 rx_MarshallProcessRPCStats(callerVersion,
7349 rpc_stat->stats[0].func_total,
7350 rpc_stat->stats, &ptr);
7356 MUTEX_EXIT(&rx_rpc_stats);
7361 * rx_FreeRPCStats - free memory allocated by
7362 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7366 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7367 * rx_RetrievePeerRPCStats
7369 * IN allocSize - the number of bytes in stats.
7377 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7379 rxi_Free(stats, allocSize);
7383 * rx_queryProcessRPCStats - see if process rpc stat collection is
7384 * currently enabled.
7390 * Returns 0 if stats are not enabled != 0 otherwise
7394 rx_queryProcessRPCStats(void)
7397 MUTEX_ENTER(&rx_rpc_stats);
7398 rc = rxi_monitor_processStats;
7399 MUTEX_EXIT(&rx_rpc_stats);
7404 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7410 * Returns 0 if stats are not enabled != 0 otherwise
7414 rx_queryPeerRPCStats(void)
7417 MUTEX_ENTER(&rx_rpc_stats);
7418 rc = rxi_monitor_peerStats;
7419 MUTEX_EXIT(&rx_rpc_stats);
7424 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7434 rx_enableProcessRPCStats(void)
7436 MUTEX_ENTER(&rx_rpc_stats);
7437 rx_enable_stats = 1;
7438 rxi_monitor_processStats = 1;
7439 MUTEX_EXIT(&rx_rpc_stats);
7443 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7453 rx_enablePeerRPCStats(void)
7455 MUTEX_ENTER(&rx_rpc_stats);
7456 rx_enable_stats = 1;
7457 rxi_monitor_peerStats = 1;
7458 MUTEX_EXIT(&rx_rpc_stats);
7462 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7472 rx_disableProcessRPCStats(void)
7474 rx_interface_stat_p rpc_stat, nrpc_stat;
7477 MUTEX_ENTER(&rx_rpc_stats);
7480 * Turn off process statistics and if peer stats is also off, turn
7484 rxi_monitor_processStats = 0;
7485 if (rxi_monitor_peerStats == 0) {
7486 rx_enable_stats = 0;
7489 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7490 unsigned int num_funcs = 0;
7493 queue_Remove(rpc_stat);
7494 num_funcs = rpc_stat->stats[0].func_total;
7496 sizeof(rx_interface_stat_t) +
7497 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7499 rxi_Free(rpc_stat, space);
7500 rxi_rpc_process_stat_cnt -= num_funcs;
7502 MUTEX_EXIT(&rx_rpc_stats);
7506 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7516 rx_disablePeerRPCStats(void)
7518 struct rx_peer **peer_ptr, **peer_end;
7521 MUTEX_ENTER(&rx_rpc_stats);
7524 * Turn off peer statistics and if process stats is also off, turn
7528 rxi_monitor_peerStats = 0;
7529 if (rxi_monitor_processStats == 0) {
7530 rx_enable_stats = 0;
7533 MUTEX_ENTER(&rx_peerHashTable_lock);
7534 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7535 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7537 struct rx_peer *peer, *next, *prev;
7538 for (prev = peer = *peer_ptr; peer; peer = next) {
7540 code = MUTEX_TRYENTER(&peer->peer_lock);
7542 rx_interface_stat_p rpc_stat, nrpc_stat;
7545 (&peer->rpcStats, rpc_stat, nrpc_stat,
7546 rx_interface_stat)) {
7547 unsigned int num_funcs = 0;
7550 queue_Remove(&rpc_stat->queue_header);
7551 queue_Remove(&rpc_stat->all_peers);
7552 num_funcs = rpc_stat->stats[0].func_total;
7554 sizeof(rx_interface_stat_t) +
7555 rpc_stat->stats[0].func_total *
7556 sizeof(rx_function_entry_v1_t);
7558 rxi_Free(rpc_stat, space);
7559 rxi_rpc_peer_stat_cnt -= num_funcs;
7561 MUTEX_EXIT(&peer->peer_lock);
7562 if (prev == *peer_ptr) {
7572 MUTEX_EXIT(&rx_peerHashTable_lock);
7573 MUTEX_EXIT(&rx_rpc_stats);
7577 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7582 * IN clearFlag - flag indicating which stats to clear
7590 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7592 rx_interface_stat_p rpc_stat, nrpc_stat;
7594 MUTEX_ENTER(&rx_rpc_stats);
7596 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7597 unsigned int num_funcs = 0, i;
7598 num_funcs = rpc_stat->stats[0].func_total;
7599 for (i = 0; i < num_funcs; i++) {
7600 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7601 hzero(rpc_stat->stats[i].invocations);
7603 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7604 hzero(rpc_stat->stats[i].bytes_sent);
7606 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7607 hzero(rpc_stat->stats[i].bytes_rcvd);
7609 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7610 rpc_stat->stats[i].queue_time_sum.sec = 0;
7611 rpc_stat->stats[i].queue_time_sum.usec = 0;
7613 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7614 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7615 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7617 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7618 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7619 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7621 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7622 rpc_stat->stats[i].queue_time_max.sec = 0;
7623 rpc_stat->stats[i].queue_time_max.usec = 0;
7625 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7626 rpc_stat->stats[i].execution_time_sum.sec = 0;
7627 rpc_stat->stats[i].execution_time_sum.usec = 0;
7629 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7630 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7631 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7633 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7634 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7635 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7637 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7638 rpc_stat->stats[i].execution_time_max.sec = 0;
7639 rpc_stat->stats[i].execution_time_max.usec = 0;
7644 MUTEX_EXIT(&rx_rpc_stats);
7648 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7653 * IN clearFlag - flag indicating which stats to clear
7661 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7663 rx_interface_stat_p rpc_stat, nrpc_stat;
7665 MUTEX_ENTER(&rx_rpc_stats);
7667 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7668 unsigned int num_funcs = 0, i;
7671 * We have to fix the offset of rpc_stat since we are
7672 * keeping this structure on two rx_queues. The rx_queue
7673 * package assumes that the rx_queue member is the first
7674 * member of the structure. That is, rx_queue assumes that
7675 * any one item is only on one queue at a time. We are
7676 * breaking that assumption and so we have to do a little
7677 * math to fix our pointers.
7680 fix_offset = (char *)rpc_stat;
7681 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7682 rpc_stat = (rx_interface_stat_p) fix_offset;
7684 num_funcs = rpc_stat->stats[0].func_total;
7685 for (i = 0; i < num_funcs; i++) {
7686 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7687 hzero(rpc_stat->stats[i].invocations);
7689 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7690 hzero(rpc_stat->stats[i].bytes_sent);
7692 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7693 hzero(rpc_stat->stats[i].bytes_rcvd);
7695 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7696 rpc_stat->stats[i].queue_time_sum.sec = 0;
7697 rpc_stat->stats[i].queue_time_sum.usec = 0;
7699 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7700 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7701 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7703 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7704 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7705 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7707 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7708 rpc_stat->stats[i].queue_time_max.sec = 0;
7709 rpc_stat->stats[i].queue_time_max.usec = 0;
7711 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7712 rpc_stat->stats[i].execution_time_sum.sec = 0;
7713 rpc_stat->stats[i].execution_time_sum.usec = 0;
7715 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7716 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7717 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7719 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7720 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7721 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7723 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7724 rpc_stat->stats[i].execution_time_max.sec = 0;
7725 rpc_stat->stats[i].execution_time_max.usec = 0;
7730 MUTEX_EXIT(&rx_rpc_stats);
7734 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7735 * is authorized to enable/disable/clear RX statistics.
7737 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7740 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7742 rxi_rxstat_userok = proc;
7746 rx_RxStatUserOk(struct rx_call *call)
7748 if (!rxi_rxstat_userok)
7750 return rxi_rxstat_userok(call);
7755 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
7756 * function in the MSVC runtime DLL (msvcrt.dll).
7758 * Note: the system serializes calls to this function.
7761 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
7762 DWORD reason, /* reason function is being called */
7763 LPVOID reserved) /* reserved for future use */
7766 case DLL_PROCESS_ATTACH:
7767 /* library is being attached to a process */
7771 case DLL_PROCESS_DETACH: