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;
387 #if defined(AFS_DJGPP_ENV) && !defined(DEBUG)
388 __djgpp_set_quiet_socket(1);
395 if (rxinit_status == 0) {
396 tmp_status = rxinit_status;
398 return tmp_status; /* Already started; return previous error code. */
404 if (afs_winsockInit() < 0)
410 * Initialize anything necessary to provide a non-premptive threading
413 rxi_InitializeThreadSupport();
416 /* Allocate and initialize a socket for client and perhaps server
419 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
420 if (rx_socket == OSI_NULLSOCKET) {
424 #ifdef RX_ENABLE_LOCKS
427 #endif /* RX_LOCKS_DB */
428 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
429 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
430 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
431 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
432 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
434 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
436 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
438 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
440 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
442 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
444 #if defined(KERNEL) && defined(AFS_HPUX110_ENV)
446 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
447 #endif /* KERNEL && AFS_HPUX110_ENV */
448 #endif /* RX_ENABLE_LOCKS */
451 rx_connDeadTime = 12;
452 rx_tranquil = 0; /* reset flag */
453 memset((char *)&rx_stats, 0, sizeof(struct rx_stats));
455 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
456 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
457 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
458 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
459 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
460 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
462 /* Malloc up a bunch of packets & buffers */
464 queue_Init(&rx_freePacketQueue);
465 rxi_NeedMorePackets = FALSE;
466 #ifdef RX_ENABLE_TSFPQ
467 rx_nPackets = 0; /* in TSFPQ version, rx_nPackets is managed by rxi_MorePackets* */
468 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
469 #else /* RX_ENABLE_TSFPQ */
470 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
471 rxi_MorePackets(rx_nPackets);
472 #endif /* RX_ENABLE_TSFPQ */
479 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
480 tv.tv_sec = clock_now.sec;
481 tv.tv_usec = clock_now.usec;
482 srand((unsigned int)tv.tv_usec);
489 #if defined(KERNEL) && !defined(UKERNEL)
490 /* Really, this should never happen in a real kernel */
493 struct sockaddr_in addr;
494 int addrlen = sizeof(addr);
495 if (getsockname((int)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
499 rx_port = addr.sin_port;
502 rx_stats.minRtt.sec = 9999999;
504 rx_SetEpoch(tv.tv_sec | 0x80000000);
506 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
507 * will provide a randomer value. */
509 MUTEX_ENTER(&rx_stats_mutex);
510 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
511 MUTEX_EXIT(&rx_stats_mutex);
512 /* *Slightly* random start time for the cid. This is just to help
513 * out with the hashing function at the peer */
514 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
515 rx_connHashTable = (struct rx_connection **)htable;
516 rx_peerHashTable = (struct rx_peer **)ptable;
518 rx_lastAckDelay.sec = 0;
519 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
520 rx_hardAckDelay.sec = 0;
521 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
522 rx_softAckDelay.sec = 0;
523 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
525 rxevent_Init(20, rxi_ReScheduleEvents);
527 /* Initialize various global queues */
528 queue_Init(&rx_idleServerQueue);
529 queue_Init(&rx_incomingCallQueue);
530 queue_Init(&rx_freeCallQueue);
532 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
533 /* Initialize our list of usable IP addresses. */
537 /* Start listener process (exact function is dependent on the
538 * implementation environment--kernel or user space) */
542 tmp_status = rxinit_status = 0;
550 return rx_InitHost(htonl(INADDR_ANY), port);
553 /* called with unincremented nRequestsRunning to see if it is OK to start
554 * a new thread in this service. Could be "no" for two reasons: over the
555 * max quota, or would prevent others from reaching their min quota.
557 #ifdef RX_ENABLE_LOCKS
558 /* This verion of QuotaOK reserves quota if it's ok while the
559 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
562 QuotaOK(register struct rx_service *aservice)
564 /* check if over max quota */
565 if (aservice->nRequestsRunning >= aservice->maxProcs) {
569 /* under min quota, we're OK */
570 /* otherwise, can use only if there are enough to allow everyone
571 * to go to their min quota after this guy starts.
573 MUTEX_ENTER(&rx_stats_mutex);
574 if ((aservice->nRequestsRunning < aservice->minProcs)
575 || (rxi_availProcs > rxi_minDeficit)) {
576 aservice->nRequestsRunning++;
577 /* just started call in minProcs pool, need fewer to maintain
579 if (aservice->nRequestsRunning <= aservice->minProcs)
582 MUTEX_EXIT(&rx_stats_mutex);
585 MUTEX_EXIT(&rx_stats_mutex);
591 ReturnToServerPool(register struct rx_service *aservice)
593 aservice->nRequestsRunning--;
594 MUTEX_ENTER(&rx_stats_mutex);
595 if (aservice->nRequestsRunning < aservice->minProcs)
598 MUTEX_EXIT(&rx_stats_mutex);
601 #else /* RX_ENABLE_LOCKS */
603 QuotaOK(register struct rx_service *aservice)
606 /* under min quota, we're OK */
607 if (aservice->nRequestsRunning < aservice->minProcs)
610 /* check if over max quota */
611 if (aservice->nRequestsRunning >= aservice->maxProcs)
614 /* otherwise, can use only if there are enough to allow everyone
615 * to go to their min quota after this guy starts.
617 if (rxi_availProcs > rxi_minDeficit)
621 #endif /* RX_ENABLE_LOCKS */
624 /* Called by rx_StartServer to start up lwp's to service calls.
625 NExistingProcs gives the number of procs already existing, and which
626 therefore needn't be created. */
628 rxi_StartServerProcs(int nExistingProcs)
630 register struct rx_service *service;
635 /* For each service, reserve N processes, where N is the "minimum"
636 * number of processes that MUST be able to execute a request in parallel,
637 * at any time, for that process. Also compute the maximum difference
638 * between any service's maximum number of processes that can run
639 * (i.e. the maximum number that ever will be run, and a guarantee
640 * that this number will run if other services aren't running), and its
641 * minimum number. The result is the extra number of processes that
642 * we need in order to provide the latter guarantee */
643 for (i = 0; i < RX_MAX_SERVICES; i++) {
645 service = rx_services[i];
646 if (service == (struct rx_service *)0)
648 nProcs += service->minProcs;
649 diff = service->maxProcs - service->minProcs;
653 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
654 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
655 for (i = 0; i < nProcs; i++) {
656 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
662 /* This routine is only required on Windows */
664 rx_StartClientThread(void)
666 #ifdef AFS_PTHREAD_ENV
668 pid = (int) pthread_self();
669 #endif /* AFS_PTHREAD_ENV */
671 #endif /* AFS_NT40_ENV */
673 /* This routine must be called if any services are exported. If the
674 * donateMe flag is set, the calling process is donated to the server
677 rx_StartServer(int donateMe)
679 register struct rx_service *service;
685 /* Start server processes, if necessary (exact function is dependent
686 * on the implementation environment--kernel or user space). DonateMe
687 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
688 * case, one less new proc will be created rx_StartServerProcs.
690 rxi_StartServerProcs(donateMe);
692 /* count up the # of threads in minProcs, and add set the min deficit to
693 * be that value, too.
695 for (i = 0; i < RX_MAX_SERVICES; i++) {
696 service = rx_services[i];
697 if (service == (struct rx_service *)0)
699 MUTEX_ENTER(&rx_stats_mutex);
700 rxi_totalMin += service->minProcs;
701 /* below works even if a thread is running, since minDeficit would
702 * still have been decremented and later re-incremented.
704 rxi_minDeficit += service->minProcs;
705 MUTEX_EXIT(&rx_stats_mutex);
708 /* Turn on reaping of idle server connections */
709 rxi_ReapConnections();
718 #ifdef AFS_PTHREAD_ENV
720 pid = (pid_t) pthread_self();
721 #else /* AFS_PTHREAD_ENV */
723 LWP_CurrentProcess(&pid);
724 #endif /* AFS_PTHREAD_ENV */
726 sprintf(name, "srv_%d", ++nProcs);
728 (*registerProgram) (pid, name);
730 #endif /* AFS_NT40_ENV */
731 rx_ServerProc(); /* Never returns */
733 #ifdef RX_ENABLE_TSFPQ
734 /* no use leaving packets around in this thread's local queue if
735 * it isn't getting donated to the server thread pool.
737 rxi_FlushLocalPacketsTSFPQ();
738 #endif /* RX_ENABLE_TSFPQ */
742 /* Create a new client connection to the specified service, using the
743 * specified security object to implement the security model for this
745 struct rx_connection *
746 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
747 register struct rx_securityClass *securityObject,
748 int serviceSecurityIndex)
752 register struct rx_connection *conn;
757 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", ntohl(shost), ntohs(sport), sservice, securityObject, serviceSecurityIndex));
759 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
760 * the case of kmem_alloc? */
761 conn = rxi_AllocConnection();
762 #ifdef RX_ENABLE_LOCKS
763 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
764 MUTEX_INIT(&conn->conn_data_lock, "conn call lock", MUTEX_DEFAULT, 0);
765 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
768 MUTEX_ENTER(&rx_connHashTable_lock);
769 cid = (rx_nextCid += RX_MAXCALLS);
770 conn->type = RX_CLIENT_CONNECTION;
772 conn->epoch = rx_epoch;
773 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
774 conn->serviceId = sservice;
775 conn->securityObject = securityObject;
776 /* This doesn't work in all compilers with void (they're buggy), so fake it
778 conn->securityData = (VOID *) 0;
779 conn->securityIndex = serviceSecurityIndex;
780 rx_SetConnDeadTime(conn, rx_connDeadTime);
781 conn->ackRate = RX_FAST_ACK_RATE;
783 conn->specific = NULL;
784 conn->challengeEvent = NULL;
785 conn->delayedAbortEvent = NULL;
786 conn->abortCount = 0;
789 RXS_NewConnection(securityObject, conn);
791 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
793 conn->refCount++; /* no lock required since only this thread knows... */
794 conn->next = rx_connHashTable[hashindex];
795 rx_connHashTable[hashindex] = conn;
796 MUTEX_ENTER(&rx_stats_mutex);
797 rx_stats.nClientConns++;
798 MUTEX_EXIT(&rx_stats_mutex);
800 MUTEX_EXIT(&rx_connHashTable_lock);
806 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
808 /* The idea is to set the dead time to a value that allows several
809 * keepalives to be dropped without timing out the connection. */
810 conn->secondsUntilDead = MAX(seconds, 6);
811 conn->secondsUntilPing = conn->secondsUntilDead / 6;
814 int rxi_lowPeerRefCount = 0;
815 int rxi_lowConnRefCount = 0;
818 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
819 * NOTE: must not be called with rx_connHashTable_lock held.
822 rxi_CleanupConnection(struct rx_connection *conn)
824 /* Notify the service exporter, if requested, that this connection
825 * is being destroyed */
826 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
827 (*conn->service->destroyConnProc) (conn);
829 /* Notify the security module that this connection is being destroyed */
830 RXS_DestroyConnection(conn->securityObject, conn);
832 /* If this is the last connection using the rx_peer struct, set its
833 * idle time to now. rxi_ReapConnections will reap it if it's still
834 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
836 MUTEX_ENTER(&rx_peerHashTable_lock);
837 if (conn->peer->refCount < 2) {
838 conn->peer->idleWhen = clock_Sec();
839 if (conn->peer->refCount < 1) {
840 conn->peer->refCount = 1;
841 MUTEX_ENTER(&rx_stats_mutex);
842 rxi_lowPeerRefCount++;
843 MUTEX_EXIT(&rx_stats_mutex);
846 conn->peer->refCount--;
847 MUTEX_EXIT(&rx_peerHashTable_lock);
849 MUTEX_ENTER(&rx_stats_mutex);
850 if (conn->type == RX_SERVER_CONNECTION)
851 rx_stats.nServerConns--;
853 rx_stats.nClientConns--;
854 MUTEX_EXIT(&rx_stats_mutex);
857 if (conn->specific) {
859 for (i = 0; i < conn->nSpecific; i++) {
860 if (conn->specific[i] && rxi_keyCreate_destructor[i])
861 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
862 conn->specific[i] = NULL;
864 free(conn->specific);
866 conn->specific = NULL;
870 MUTEX_DESTROY(&conn->conn_call_lock);
871 MUTEX_DESTROY(&conn->conn_data_lock);
872 CV_DESTROY(&conn->conn_call_cv);
874 rxi_FreeConnection(conn);
877 /* Destroy the specified connection */
879 rxi_DestroyConnection(register struct rx_connection *conn)
881 MUTEX_ENTER(&rx_connHashTable_lock);
882 rxi_DestroyConnectionNoLock(conn);
883 /* conn should be at the head of the cleanup list */
884 if (conn == rx_connCleanup_list) {
885 rx_connCleanup_list = rx_connCleanup_list->next;
886 MUTEX_EXIT(&rx_connHashTable_lock);
887 rxi_CleanupConnection(conn);
889 #ifdef RX_ENABLE_LOCKS
891 MUTEX_EXIT(&rx_connHashTable_lock);
893 #endif /* RX_ENABLE_LOCKS */
897 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
899 register struct rx_connection **conn_ptr;
900 register int havecalls = 0;
901 struct rx_packet *packet;
908 MUTEX_ENTER(&conn->conn_data_lock);
909 if (conn->refCount > 0)
912 MUTEX_ENTER(&rx_stats_mutex);
913 rxi_lowConnRefCount++;
914 MUTEX_EXIT(&rx_stats_mutex);
917 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
918 /* Busy; wait till the last guy before proceeding */
919 MUTEX_EXIT(&conn->conn_data_lock);
924 /* If the client previously called rx_NewCall, but it is still
925 * waiting, treat this as a running call, and wait to destroy the
926 * connection later when the call completes. */
927 if ((conn->type == RX_CLIENT_CONNECTION)
928 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
929 conn->flags |= RX_CONN_DESTROY_ME;
930 MUTEX_EXIT(&conn->conn_data_lock);
934 MUTEX_EXIT(&conn->conn_data_lock);
936 /* Check for extant references to this connection */
937 for (i = 0; i < RX_MAXCALLS; i++) {
938 register struct rx_call *call = conn->call[i];
941 if (conn->type == RX_CLIENT_CONNECTION) {
942 MUTEX_ENTER(&call->lock);
943 if (call->delayedAckEvent) {
944 /* Push the final acknowledgment out now--there
945 * won't be a subsequent call to acknowledge the
946 * last reply packets */
947 rxevent_Cancel(call->delayedAckEvent, call,
948 RX_CALL_REFCOUNT_DELAY);
949 if (call->state == RX_STATE_PRECALL
950 || call->state == RX_STATE_ACTIVE) {
951 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
953 rxi_AckAll(NULL, call, 0);
956 MUTEX_EXIT(&call->lock);
960 #ifdef RX_ENABLE_LOCKS
962 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
963 MUTEX_EXIT(&conn->conn_data_lock);
965 /* Someone is accessing a packet right now. */
969 #endif /* RX_ENABLE_LOCKS */
972 /* Don't destroy the connection if there are any call
973 * structures still in use */
974 MUTEX_ENTER(&conn->conn_data_lock);
975 conn->flags |= RX_CONN_DESTROY_ME;
976 MUTEX_EXIT(&conn->conn_data_lock);
981 if (conn->delayedAbortEvent) {
982 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
983 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
985 MUTEX_ENTER(&conn->conn_data_lock);
986 rxi_SendConnectionAbort(conn, packet, 0, 1);
987 MUTEX_EXIT(&conn->conn_data_lock);
988 rxi_FreePacket(packet);
992 /* Remove from connection hash table before proceeding */
994 &rx_connHashTable[CONN_HASH
995 (peer->host, peer->port, conn->cid, conn->epoch,
997 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
998 if (*conn_ptr == conn) {
999 *conn_ptr = conn->next;
1003 /* if the conn that we are destroying was the last connection, then we
1004 * clear rxLastConn as well */
1005 if (rxLastConn == conn)
1008 /* Make sure the connection is completely reset before deleting it. */
1009 /* get rid of pending events that could zap us later */
1010 if (conn->challengeEvent)
1011 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1012 if (conn->checkReachEvent)
1013 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1015 /* Add the connection to the list of destroyed connections that
1016 * need to be cleaned up. This is necessary to avoid deadlocks
1017 * in the routines we call to inform others that this connection is
1018 * being destroyed. */
1019 conn->next = rx_connCleanup_list;
1020 rx_connCleanup_list = conn;
1023 /* Externally available version */
1025 rx_DestroyConnection(register struct rx_connection *conn)
1030 rxi_DestroyConnection(conn);
1035 rx_GetConnection(register struct rx_connection *conn)
1040 MUTEX_ENTER(&conn->conn_data_lock);
1042 MUTEX_EXIT(&conn->conn_data_lock);
1046 /* Start a new rx remote procedure call, on the specified connection.
1047 * If wait is set to 1, wait for a free call channel; otherwise return
1048 * 0. Maxtime gives the maximum number of seconds this call may take,
1049 * after rx_MakeCall returns. After this time interval, a call to any
1050 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1051 * For fine grain locking, we hold the conn_call_lock in order to
1052 * to ensure that we don't get signalle after we found a call in an active
1053 * state and before we go to sleep.
1056 rx_NewCall(register struct rx_connection *conn)
1059 register struct rx_call *call;
1060 struct clock queueTime;
1064 dpf(("rx_MakeCall(conn %x)\n", conn));
1067 clock_GetTime(&queueTime);
1068 MUTEX_ENTER(&conn->conn_call_lock);
1071 * Check if there are others waiting for a new call.
1072 * If so, let them go first to avoid starving them.
1073 * This is a fairly simple scheme, and might not be
1074 * a complete solution for large numbers of waiters.
1076 * makeCallWaiters keeps track of the number of
1077 * threads waiting to make calls and the
1078 * RX_CONN_MAKECALL_WAITING flag bit is used to
1079 * indicate that there are indeed calls waiting.
1080 * The flag is set when the waiter is incremented.
1081 * It is only cleared in rx_EndCall when
1082 * makeCallWaiters is 0. This prevents us from
1083 * accidently destroying the connection while it
1084 * is potentially about to be used.
1086 MUTEX_ENTER(&conn->conn_data_lock);
1087 if (conn->makeCallWaiters) {
1088 conn->flags |= RX_CONN_MAKECALL_WAITING;
1089 conn->makeCallWaiters++;
1090 MUTEX_EXIT(&conn->conn_data_lock);
1092 #ifdef RX_ENABLE_LOCKS
1093 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1097 MUTEX_ENTER(&conn->conn_data_lock);
1098 conn->makeCallWaiters--;
1100 MUTEX_EXIT(&conn->conn_data_lock);
1103 for (i = 0; i < RX_MAXCALLS; i++) {
1104 call = conn->call[i];
1106 MUTEX_ENTER(&call->lock);
1107 if (call->state == RX_STATE_DALLY) {
1108 rxi_ResetCall(call, 0);
1109 (*call->callNumber)++;
1112 MUTEX_EXIT(&call->lock);
1114 call = rxi_NewCall(conn, i);
1118 if (i < RX_MAXCALLS) {
1121 MUTEX_ENTER(&conn->conn_data_lock);
1122 conn->flags |= RX_CONN_MAKECALL_WAITING;
1123 conn->makeCallWaiters++;
1124 MUTEX_EXIT(&conn->conn_data_lock);
1126 #ifdef RX_ENABLE_LOCKS
1127 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1131 MUTEX_ENTER(&conn->conn_data_lock);
1132 conn->makeCallWaiters--;
1133 MUTEX_EXIT(&conn->conn_data_lock);
1136 * Wake up anyone else who might be giving us a chance to
1137 * run (see code above that avoids resource starvation).
1139 #ifdef RX_ENABLE_LOCKS
1140 CV_BROADCAST(&conn->conn_call_cv);
1145 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1147 /* Client is initially in send mode */
1148 call->state = RX_STATE_ACTIVE;
1149 call->error = conn->error;
1151 call->mode = RX_MODE_ERROR;
1153 call->mode = RX_MODE_SENDING;
1155 /* remember start time for call in case we have hard dead time limit */
1156 call->queueTime = queueTime;
1157 clock_GetTime(&call->startTime);
1158 hzero(call->bytesSent);
1159 hzero(call->bytesRcvd);
1161 /* Turn on busy protocol. */
1162 rxi_KeepAliveOn(call);
1164 MUTEX_EXIT(&call->lock);
1165 MUTEX_EXIT(&conn->conn_call_lock);
1168 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1169 /* Now, if TQ wasn't cleared earlier, do it now. */
1170 MUTEX_ENTER(&call->lock);
1171 while (call->flags & RX_CALL_TQ_BUSY) {
1172 call->flags |= RX_CALL_TQ_WAIT;
1174 #ifdef RX_ENABLE_LOCKS
1175 osirx_AssertMine(&call->lock, "rxi_Start lock4");
1176 CV_WAIT(&call->cv_tq, &call->lock);
1177 #else /* RX_ENABLE_LOCKS */
1178 osi_rxSleep(&call->tq);
1179 #endif /* RX_ENABLE_LOCKS */
1181 if (call->tqWaiters == 0) {
1182 call->flags &= ~RX_CALL_TQ_WAIT;
1185 if (call->flags & RX_CALL_TQ_CLEARME) {
1186 rxi_ClearTransmitQueue(call, 0);
1187 queue_Init(&call->tq);
1189 MUTEX_EXIT(&call->lock);
1190 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1196 rxi_HasActiveCalls(register struct rx_connection *aconn)
1199 register struct rx_call *tcall;
1203 for (i = 0; i < RX_MAXCALLS; i++) {
1204 if ((tcall = aconn->call[i])) {
1205 if ((tcall->state == RX_STATE_ACTIVE)
1206 || (tcall->state == RX_STATE_PRECALL)) {
1217 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1218 register afs_int32 * aint32s)
1221 register struct rx_call *tcall;
1225 for (i = 0; i < RX_MAXCALLS; i++) {
1226 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1227 aint32s[i] = aconn->callNumber[i] + 1;
1229 aint32s[i] = aconn->callNumber[i];
1236 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1237 register afs_int32 * aint32s)
1240 register struct rx_call *tcall;
1244 for (i = 0; i < RX_MAXCALLS; i++) {
1245 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1246 aconn->callNumber[i] = aint32s[i] - 1;
1248 aconn->callNumber[i] = aint32s[i];
1254 /* Advertise a new service. A service is named locally by a UDP port
1255 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1258 char *serviceName; Name for identification purposes (e.g. the
1259 service name might be used for probing for
1262 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1263 char *serviceName, struct rx_securityClass **securityObjects,
1264 int nSecurityObjects,
1265 afs_int32(*serviceProc) (struct rx_call * acall))
1267 osi_socket socket = OSI_NULLSOCKET;
1268 register struct rx_service *tservice;
1274 if (serviceId == 0) {
1276 "rx_NewService: service id for service %s is not non-zero.\n",
1283 "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",
1291 tservice = rxi_AllocService();
1293 for (i = 0; i < RX_MAX_SERVICES; i++) {
1294 register struct rx_service *service = rx_services[i];
1296 if (port == service->servicePort && host == service->serviceHost) {
1297 if (service->serviceId == serviceId) {
1298 /* The identical service has already been
1299 * installed; if the caller was intending to
1300 * change the security classes used by this
1301 * service, he/she loses. */
1303 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1304 serviceName, serviceId, service->serviceName);
1306 rxi_FreeService(tservice);
1309 /* Different service, same port: re-use the socket
1310 * which is bound to the same port */
1311 socket = service->socket;
1314 if (socket == OSI_NULLSOCKET) {
1315 /* If we don't already have a socket (from another
1316 * service on same port) get a new one */
1317 socket = rxi_GetHostUDPSocket(host, port);
1318 if (socket == OSI_NULLSOCKET) {
1320 rxi_FreeService(tservice);
1325 service->socket = socket;
1326 service->serviceHost = host;
1327 service->servicePort = port;
1328 service->serviceId = serviceId;
1329 service->serviceName = serviceName;
1330 service->nSecurityObjects = nSecurityObjects;
1331 service->securityObjects = securityObjects;
1332 service->minProcs = 0;
1333 service->maxProcs = 1;
1334 service->idleDeadTime = 60;
1335 service->connDeadTime = rx_connDeadTime;
1336 service->executeRequestProc = serviceProc;
1337 service->checkReach = 0;
1338 rx_services[i] = service; /* not visible until now */
1344 rxi_FreeService(tservice);
1345 (osi_Msg "rx_NewService: cannot support > %d services\n",
1351 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1352 struct rx_securityClass **securityObjects, int nSecurityObjects,
1353 afs_int32(*serviceProc) (struct rx_call * acall))
1355 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1358 /* Generic request processing loop. This routine should be called
1359 * by the implementation dependent rx_ServerProc. If socketp is
1360 * non-null, it will be set to the file descriptor that this thread
1361 * is now listening on. If socketp is null, this routine will never
1364 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1366 register struct rx_call *call;
1367 register afs_int32 code;
1368 register struct rx_service *tservice = NULL;
1375 call = rx_GetCall(threadID, tservice, socketp);
1376 if (socketp && *socketp != OSI_NULLSOCKET) {
1377 /* We are now a listener thread */
1382 /* if server is restarting( typically smooth shutdown) then do not
1383 * allow any new calls.
1386 if (rx_tranquil && (call != NULL)) {
1390 MUTEX_ENTER(&call->lock);
1392 rxi_CallError(call, RX_RESTARTING);
1393 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1395 MUTEX_EXIT(&call->lock);
1399 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1400 #ifdef RX_ENABLE_LOCKS
1402 #endif /* RX_ENABLE_LOCKS */
1403 afs_termState = AFSOP_STOP_AFS;
1404 afs_osi_Wakeup(&afs_termState);
1405 #ifdef RX_ENABLE_LOCKS
1407 #endif /* RX_ENABLE_LOCKS */
1412 tservice = call->conn->service;
1414 if (tservice->beforeProc)
1415 (*tservice->beforeProc) (call);
1417 code = call->conn->service->executeRequestProc(call);
1419 if (tservice->afterProc)
1420 (*tservice->afterProc) (call, code);
1422 rx_EndCall(call, code);
1423 MUTEX_ENTER(&rx_stats_mutex);
1425 MUTEX_EXIT(&rx_stats_mutex);
1431 rx_WakeupServerProcs(void)
1433 struct rx_serverQueueEntry *np, *tqp;
1437 MUTEX_ENTER(&rx_serverPool_lock);
1439 #ifdef RX_ENABLE_LOCKS
1440 if (rx_waitForPacket)
1441 CV_BROADCAST(&rx_waitForPacket->cv);
1442 #else /* RX_ENABLE_LOCKS */
1443 if (rx_waitForPacket)
1444 osi_rxWakeup(rx_waitForPacket);
1445 #endif /* RX_ENABLE_LOCKS */
1446 MUTEX_ENTER(&freeSQEList_lock);
1447 for (np = rx_FreeSQEList; np; np = tqp) {
1448 tqp = *(struct rx_serverQueueEntry **)np;
1449 #ifdef RX_ENABLE_LOCKS
1450 CV_BROADCAST(&np->cv);
1451 #else /* RX_ENABLE_LOCKS */
1453 #endif /* RX_ENABLE_LOCKS */
1455 MUTEX_EXIT(&freeSQEList_lock);
1456 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1457 #ifdef RX_ENABLE_LOCKS
1458 CV_BROADCAST(&np->cv);
1459 #else /* RX_ENABLE_LOCKS */
1461 #endif /* RX_ENABLE_LOCKS */
1463 MUTEX_EXIT(&rx_serverPool_lock);
1468 * One thing that seems to happen is that all the server threads get
1469 * tied up on some empty or slow call, and then a whole bunch of calls
1470 * arrive at once, using up the packet pool, so now there are more
1471 * empty calls. The most critical resources here are server threads
1472 * and the free packet pool. The "doreclaim" code seems to help in
1473 * general. I think that eventually we arrive in this state: there
1474 * are lots of pending calls which do have all their packets present,
1475 * so they won't be reclaimed, are multi-packet calls, so they won't
1476 * be scheduled until later, and thus are tying up most of the free
1477 * packet pool for a very long time.
1479 * 1. schedule multi-packet calls if all the packets are present.
1480 * Probably CPU-bound operation, useful to return packets to pool.
1481 * Do what if there is a full window, but the last packet isn't here?
1482 * 3. preserve one thread which *only* runs "best" calls, otherwise
1483 * it sleeps and waits for that type of call.
1484 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1485 * the current dataquota business is badly broken. The quota isn't adjusted
1486 * to reflect how many packets are presently queued for a running call.
1487 * So, when we schedule a queued call with a full window of packets queued
1488 * up for it, that *should* free up a window full of packets for other 2d-class
1489 * calls to be able to use from the packet pool. But it doesn't.
1491 * NB. Most of the time, this code doesn't run -- since idle server threads
1492 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1493 * as a new call arrives.
1495 /* Sleep until a call arrives. Returns a pointer to the call, ready
1496 * for an rx_Read. */
1497 #ifdef RX_ENABLE_LOCKS
1499 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1501 struct rx_serverQueueEntry *sq;
1502 register struct rx_call *call = (struct rx_call *)0;
1503 struct rx_service *service = NULL;
1506 MUTEX_ENTER(&freeSQEList_lock);
1508 if ((sq = rx_FreeSQEList)) {
1509 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1510 MUTEX_EXIT(&freeSQEList_lock);
1511 } else { /* otherwise allocate a new one and return that */
1512 MUTEX_EXIT(&freeSQEList_lock);
1513 sq = (struct rx_serverQueueEntry *)
1514 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1515 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1516 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1519 MUTEX_ENTER(&rx_serverPool_lock);
1520 if (cur_service != NULL) {
1521 ReturnToServerPool(cur_service);
1524 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1525 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1527 /* Scan for eligible incoming calls. A call is not eligible
1528 * if the maximum number of calls for its service type are
1529 * already executing */
1530 /* One thread will process calls FCFS (to prevent starvation),
1531 * while the other threads may run ahead looking for calls which
1532 * have all their input data available immediately. This helps
1533 * keep threads from blocking, waiting for data from the client. */
1534 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1535 service = tcall->conn->service;
1536 if (!QuotaOK(service)) {
1539 if (tno == rxi_fcfs_thread_num
1540 || !tcall->queue_item_header.next) {
1541 /* If we're the fcfs thread , then we'll just use
1542 * this call. If we haven't been able to find an optimal
1543 * choice, and we're at the end of the list, then use a
1544 * 2d choice if one has been identified. Otherwise... */
1545 call = (choice2 ? choice2 : tcall);
1546 service = call->conn->service;
1547 } else if (!queue_IsEmpty(&tcall->rq)) {
1548 struct rx_packet *rp;
1549 rp = queue_First(&tcall->rq, rx_packet);
1550 if (rp->header.seq == 1) {
1552 || (rp->header.flags & RX_LAST_PACKET)) {
1554 } else if (rxi_2dchoice && !choice2
1555 && !(tcall->flags & RX_CALL_CLEARED)
1556 && (tcall->rprev > rxi_HardAckRate)) {
1565 ReturnToServerPool(service);
1572 MUTEX_EXIT(&rx_serverPool_lock);
1573 MUTEX_ENTER(&call->lock);
1575 if (call->flags & RX_CALL_WAIT_PROC) {
1576 call->flags &= ~RX_CALL_WAIT_PROC;
1577 MUTEX_ENTER(&rx_stats_mutex);
1579 MUTEX_EXIT(&rx_stats_mutex);
1582 if (call->state != RX_STATE_PRECALL || call->error) {
1583 MUTEX_EXIT(&call->lock);
1584 MUTEX_ENTER(&rx_serverPool_lock);
1585 ReturnToServerPool(service);
1590 if (queue_IsEmpty(&call->rq)
1591 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1592 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1594 CLEAR_CALL_QUEUE_LOCK(call);
1597 /* If there are no eligible incoming calls, add this process
1598 * to the idle server queue, to wait for one */
1602 *socketp = OSI_NULLSOCKET;
1604 sq->socketp = socketp;
1605 queue_Append(&rx_idleServerQueue, sq);
1606 #ifndef AFS_AIX41_ENV
1607 rx_waitForPacket = sq;
1609 rx_waitingForPacket = sq;
1610 #endif /* AFS_AIX41_ENV */
1612 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1614 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1615 MUTEX_EXIT(&rx_serverPool_lock);
1616 return (struct rx_call *)0;
1619 } while (!(call = sq->newcall)
1620 && !(socketp && *socketp != OSI_NULLSOCKET));
1621 MUTEX_EXIT(&rx_serverPool_lock);
1623 MUTEX_ENTER(&call->lock);
1629 MUTEX_ENTER(&freeSQEList_lock);
1630 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1631 rx_FreeSQEList = sq;
1632 MUTEX_EXIT(&freeSQEList_lock);
1635 clock_GetTime(&call->startTime);
1636 call->state = RX_STATE_ACTIVE;
1637 call->mode = RX_MODE_RECEIVING;
1638 #ifdef RX_KERNEL_TRACE
1639 if (ICL_SETACTIVE(afs_iclSetp)) {
1640 int glockOwner = ISAFS_GLOCK();
1643 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1644 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1651 rxi_calltrace(RX_CALL_START, call);
1652 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1653 call->conn->service->servicePort, call->conn->service->serviceId,
1656 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1657 MUTEX_EXIT(&call->lock);
1659 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1664 #else /* RX_ENABLE_LOCKS */
1666 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1668 struct rx_serverQueueEntry *sq;
1669 register struct rx_call *call = (struct rx_call *)0, *choice2;
1670 struct rx_service *service = NULL;
1674 MUTEX_ENTER(&freeSQEList_lock);
1676 if ((sq = rx_FreeSQEList)) {
1677 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1678 MUTEX_EXIT(&freeSQEList_lock);
1679 } else { /* otherwise allocate a new one and return that */
1680 MUTEX_EXIT(&freeSQEList_lock);
1681 sq = (struct rx_serverQueueEntry *)
1682 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1683 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1684 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1686 MUTEX_ENTER(&sq->lock);
1688 if (cur_service != NULL) {
1689 cur_service->nRequestsRunning--;
1690 if (cur_service->nRequestsRunning < cur_service->minProcs)
1694 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1695 register struct rx_call *tcall, *ncall;
1696 /* Scan for eligible incoming calls. A call is not eligible
1697 * if the maximum number of calls for its service type are
1698 * already executing */
1699 /* One thread will process calls FCFS (to prevent starvation),
1700 * while the other threads may run ahead looking for calls which
1701 * have all their input data available immediately. This helps
1702 * keep threads from blocking, waiting for data from the client. */
1703 choice2 = (struct rx_call *)0;
1704 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1705 service = tcall->conn->service;
1706 if (QuotaOK(service)) {
1707 if (tno == rxi_fcfs_thread_num
1708 || !tcall->queue_item_header.next) {
1709 /* If we're the fcfs thread, then we'll just use
1710 * this call. If we haven't been able to find an optimal
1711 * choice, and we're at the end of the list, then use a
1712 * 2d choice if one has been identified. Otherwise... */
1713 call = (choice2 ? choice2 : tcall);
1714 service = call->conn->service;
1715 } else if (!queue_IsEmpty(&tcall->rq)) {
1716 struct rx_packet *rp;
1717 rp = queue_First(&tcall->rq, rx_packet);
1718 if (rp->header.seq == 1
1720 || (rp->header.flags & RX_LAST_PACKET))) {
1722 } else if (rxi_2dchoice && !choice2
1723 && !(tcall->flags & RX_CALL_CLEARED)
1724 && (tcall->rprev > rxi_HardAckRate)) {
1737 /* we can't schedule a call if there's no data!!! */
1738 /* send an ack if there's no data, if we're missing the
1739 * first packet, or we're missing something between first
1740 * and last -- there's a "hole" in the incoming data. */
1741 if (queue_IsEmpty(&call->rq)
1742 || queue_First(&call->rq, rx_packet)->header.seq != 1
1743 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1744 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1746 call->flags &= (~RX_CALL_WAIT_PROC);
1747 service->nRequestsRunning++;
1748 /* just started call in minProcs pool, need fewer to maintain
1750 if (service->nRequestsRunning <= service->minProcs)
1754 /* MUTEX_EXIT(&call->lock); */
1756 /* If there are no eligible incoming calls, add this process
1757 * to the idle server queue, to wait for one */
1760 *socketp = OSI_NULLSOCKET;
1762 sq->socketp = socketp;
1763 queue_Append(&rx_idleServerQueue, sq);
1767 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1769 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1770 return (struct rx_call *)0;
1773 } while (!(call = sq->newcall)
1774 && !(socketp && *socketp != OSI_NULLSOCKET));
1776 MUTEX_EXIT(&sq->lock);
1778 MUTEX_ENTER(&freeSQEList_lock);
1779 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1780 rx_FreeSQEList = sq;
1781 MUTEX_EXIT(&freeSQEList_lock);
1784 clock_GetTime(&call->startTime);
1785 call->state = RX_STATE_ACTIVE;
1786 call->mode = RX_MODE_RECEIVING;
1787 #ifdef RX_KERNEL_TRACE
1788 if (ICL_SETACTIVE(afs_iclSetp)) {
1789 int glockOwner = ISAFS_GLOCK();
1792 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1793 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1800 rxi_calltrace(RX_CALL_START, call);
1801 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1802 call->conn->service->servicePort, call->conn->service->serviceId,
1805 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1812 #endif /* RX_ENABLE_LOCKS */
1816 /* Establish a procedure to be called when a packet arrives for a
1817 * call. This routine will be called at most once after each call,
1818 * and will also be called if there is an error condition on the or
1819 * the call is complete. Used by multi rx to build a selection
1820 * function which determines which of several calls is likely to be a
1821 * good one to read from.
1822 * NOTE: the way this is currently implemented it is probably only a
1823 * good idea to (1) use it immediately after a newcall (clients only)
1824 * and (2) only use it once. Other uses currently void your warranty
1827 rx_SetArrivalProc(register struct rx_call *call,
1828 register void (*proc) (register struct rx_call * call,
1830 register int index),
1831 register VOID * handle, register int arg)
1833 call->arrivalProc = proc;
1834 call->arrivalProcHandle = handle;
1835 call->arrivalProcArg = arg;
1838 /* Call is finished (possibly prematurely). Return rc to the peer, if
1839 * appropriate, and return the final error code from the conversation
1843 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1845 register struct rx_connection *conn = call->conn;
1846 register struct rx_service *service;
1852 dpf(("rx_EndCall(call %x rc %d error %d abortCode %d)\n", call, rc, call->error, call->abortCode));
1855 MUTEX_ENTER(&call->lock);
1857 if (rc == 0 && call->error == 0) {
1858 call->abortCode = 0;
1859 call->abortCount = 0;
1862 call->arrivalProc = (void (*)())0;
1863 if (rc && call->error == 0) {
1864 rxi_CallError(call, rc);
1865 /* Send an abort message to the peer if this error code has
1866 * only just been set. If it was set previously, assume the
1867 * peer has already been sent the error code or will request it
1869 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1871 if (conn->type == RX_SERVER_CONNECTION) {
1872 /* Make sure reply or at least dummy reply is sent */
1873 if (call->mode == RX_MODE_RECEIVING) {
1874 rxi_WriteProc(call, 0, 0);
1876 if (call->mode == RX_MODE_SENDING) {
1877 rxi_FlushWrite(call);
1879 service = conn->service;
1880 rxi_calltrace(RX_CALL_END, call);
1881 /* Call goes to hold state until reply packets are acknowledged */
1882 if (call->tfirst + call->nSoftAcked < call->tnext) {
1883 call->state = RX_STATE_HOLD;
1885 call->state = RX_STATE_DALLY;
1886 rxi_ClearTransmitQueue(call, 0);
1887 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1888 rxevent_Cancel(call->keepAliveEvent, call,
1889 RX_CALL_REFCOUNT_ALIVE);
1891 } else { /* Client connection */
1893 /* Make sure server receives input packets, in the case where
1894 * no reply arguments are expected */
1895 if ((call->mode == RX_MODE_SENDING)
1896 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1897 (void)rxi_ReadProc(call, &dummy, 1);
1900 /* If we had an outstanding delayed ack, be nice to the server
1901 * and force-send it now.
1903 if (call->delayedAckEvent) {
1904 rxevent_Cancel(call->delayedAckEvent, call,
1905 RX_CALL_REFCOUNT_DELAY);
1906 call->delayedAckEvent = NULL;
1907 rxi_SendDelayedAck(NULL, call, NULL);
1910 /* We need to release the call lock since it's lower than the
1911 * conn_call_lock and we don't want to hold the conn_call_lock
1912 * over the rx_ReadProc call. The conn_call_lock needs to be held
1913 * here for the case where rx_NewCall is perusing the calls on
1914 * the connection structure. We don't want to signal until
1915 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1916 * have checked this call, found it active and by the time it
1917 * goes to sleep, will have missed the signal.
1919 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
1920 * there are threads waiting to use the conn object.
1922 MUTEX_EXIT(&call->lock);
1923 MUTEX_ENTER(&conn->conn_call_lock);
1924 MUTEX_ENTER(&call->lock);
1925 MUTEX_ENTER(&conn->conn_data_lock);
1926 conn->flags |= RX_CONN_BUSY;
1927 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1928 if (conn->makeCallWaiters == 0)
1929 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1930 MUTEX_EXIT(&conn->conn_data_lock);
1931 #ifdef RX_ENABLE_LOCKS
1932 CV_BROADCAST(&conn->conn_call_cv);
1937 #ifdef RX_ENABLE_LOCKS
1939 MUTEX_EXIT(&conn->conn_data_lock);
1941 #endif /* RX_ENABLE_LOCKS */
1942 call->state = RX_STATE_DALLY;
1944 error = call->error;
1946 /* currentPacket, nLeft, and NFree must be zeroed here, because
1947 * ResetCall cannot: ResetCall may be called at splnet(), in the
1948 * kernel version, and may interrupt the macros rx_Read or
1949 * rx_Write, which run at normal priority for efficiency. */
1950 if (call->currentPacket) {
1951 queue_Prepend(&call->iovq, call->currentPacket);
1952 call->currentPacket = (struct rx_packet *)0;
1955 call->nLeft = call->nFree = call->curlen = 0;
1957 /* Free any packets from the last call to ReadvProc/WritevProc */
1958 rxi_FreePackets(0, &call->iovq);
1960 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1961 MUTEX_EXIT(&call->lock);
1962 if (conn->type == RX_CLIENT_CONNECTION) {
1963 MUTEX_EXIT(&conn->conn_call_lock);
1964 conn->flags &= ~RX_CONN_BUSY;
1968 * Map errors to the local host's errno.h format.
1970 error = ntoh_syserr_conv(error);
1974 #if !defined(KERNEL)
1976 /* Call this routine when shutting down a server or client (especially
1977 * clients). This will allow Rx to gracefully garbage collect server
1978 * connections, and reduce the number of retries that a server might
1979 * make to a dead client.
1980 * This is not quite right, since some calls may still be ongoing and
1981 * we can't lock them to destroy them. */
1985 register struct rx_connection **conn_ptr, **conn_end;
1989 if (rxinit_status == 1) {
1991 return; /* Already shutdown. */
1993 rxi_DeleteCachedConnections();
1994 if (rx_connHashTable) {
1995 MUTEX_ENTER(&rx_connHashTable_lock);
1996 for (conn_ptr = &rx_connHashTable[0], conn_end =
1997 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
1999 struct rx_connection *conn, *next;
2000 for (conn = *conn_ptr; conn; conn = next) {
2002 if (conn->type == RX_CLIENT_CONNECTION) {
2003 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2005 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2006 #ifdef RX_ENABLE_LOCKS
2007 rxi_DestroyConnectionNoLock(conn);
2008 #else /* RX_ENABLE_LOCKS */
2009 rxi_DestroyConnection(conn);
2010 #endif /* RX_ENABLE_LOCKS */
2014 #ifdef RX_ENABLE_LOCKS
2015 while (rx_connCleanup_list) {
2016 struct rx_connection *conn;
2017 conn = rx_connCleanup_list;
2018 rx_connCleanup_list = rx_connCleanup_list->next;
2019 MUTEX_EXIT(&rx_connHashTable_lock);
2020 rxi_CleanupConnection(conn);
2021 MUTEX_ENTER(&rx_connHashTable_lock);
2023 MUTEX_EXIT(&rx_connHashTable_lock);
2024 #endif /* RX_ENABLE_LOCKS */
2029 afs_winsockCleanup();
2037 /* if we wakeup packet waiter too often, can get in loop with two
2038 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2040 rxi_PacketsUnWait(void)
2042 if (!rx_waitingForPackets) {
2046 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2047 return; /* still over quota */
2050 rx_waitingForPackets = 0;
2051 #ifdef RX_ENABLE_LOCKS
2052 CV_BROADCAST(&rx_waitingForPackets_cv);
2054 osi_rxWakeup(&rx_waitingForPackets);
2060 /* ------------------Internal interfaces------------------------- */
2062 /* Return this process's service structure for the
2063 * specified socket and service */
2065 rxi_FindService(register osi_socket socket, register u_short serviceId)
2067 register struct rx_service **sp;
2068 for (sp = &rx_services[0]; *sp; sp++) {
2069 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2075 /* Allocate a call structure, for the indicated channel of the
2076 * supplied connection. The mode and state of the call must be set by
2077 * the caller. Returns the call with mutex locked. */
2079 rxi_NewCall(register struct rx_connection *conn, register int channel)
2081 register struct rx_call *call;
2082 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2083 register struct rx_call *cp; /* Call pointer temp */
2084 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2085 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2087 /* Grab an existing call structure, or allocate a new one.
2088 * Existing call structures are assumed to have been left reset by
2090 MUTEX_ENTER(&rx_freeCallQueue_lock);
2092 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2094 * EXCEPT that the TQ might not yet be cleared out.
2095 * Skip over those with in-use TQs.
2098 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2099 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2105 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2106 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2107 call = queue_First(&rx_freeCallQueue, rx_call);
2108 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2110 MUTEX_ENTER(&rx_stats_mutex);
2111 rx_stats.nFreeCallStructs--;
2112 MUTEX_EXIT(&rx_stats_mutex);
2113 MUTEX_EXIT(&rx_freeCallQueue_lock);
2114 MUTEX_ENTER(&call->lock);
2115 CLEAR_CALL_QUEUE_LOCK(call);
2116 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2117 /* Now, if TQ wasn't cleared earlier, do it now. */
2118 if (call->flags & RX_CALL_TQ_CLEARME) {
2119 rxi_ClearTransmitQueue(call, 0);
2120 queue_Init(&call->tq);
2122 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2123 /* Bind the call to its connection structure */
2125 rxi_ResetCall(call, 1);
2127 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2129 MUTEX_EXIT(&rx_freeCallQueue_lock);
2130 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2131 MUTEX_ENTER(&call->lock);
2132 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2133 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2134 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2136 MUTEX_ENTER(&rx_stats_mutex);
2137 rx_stats.nCallStructs++;
2138 MUTEX_EXIT(&rx_stats_mutex);
2139 /* Initialize once-only items */
2140 queue_Init(&call->tq);
2141 queue_Init(&call->rq);
2142 queue_Init(&call->iovq);
2143 /* Bind the call to its connection structure (prereq for reset) */
2145 rxi_ResetCall(call, 1);
2147 call->channel = channel;
2148 call->callNumber = &conn->callNumber[channel];
2149 /* Note that the next expected call number is retained (in
2150 * conn->callNumber[i]), even if we reallocate the call structure
2152 conn->call[channel] = call;
2153 /* if the channel's never been used (== 0), we should start at 1, otherwise
2154 * the call number is valid from the last time this channel was used */
2155 if (*call->callNumber == 0)
2156 *call->callNumber = 1;
2161 /* A call has been inactive long enough that so we can throw away
2162 * state, including the call structure, which is placed on the call
2164 * Call is locked upon entry.
2165 * haveCTLock set if called from rxi_ReapConnections
2167 #ifdef RX_ENABLE_LOCKS
2169 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2170 #else /* RX_ENABLE_LOCKS */
2172 rxi_FreeCall(register struct rx_call *call)
2173 #endif /* RX_ENABLE_LOCKS */
2175 register int channel = call->channel;
2176 register struct rx_connection *conn = call->conn;
2179 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2180 (*call->callNumber)++;
2181 rxi_ResetCall(call, 0);
2182 call->conn->call[channel] = (struct rx_call *)0;
2184 MUTEX_ENTER(&rx_freeCallQueue_lock);
2185 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2186 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2187 /* A call may be free even though its transmit queue is still in use.
2188 * Since we search the call list from head to tail, put busy calls at
2189 * the head of the list, and idle calls at the tail.
2191 if (call->flags & RX_CALL_TQ_BUSY)
2192 queue_Prepend(&rx_freeCallQueue, call);
2194 queue_Append(&rx_freeCallQueue, call);
2195 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2196 queue_Append(&rx_freeCallQueue, call);
2197 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2198 MUTEX_ENTER(&rx_stats_mutex);
2199 rx_stats.nFreeCallStructs++;
2200 MUTEX_EXIT(&rx_stats_mutex);
2202 MUTEX_EXIT(&rx_freeCallQueue_lock);
2204 /* Destroy the connection if it was previously slated for
2205 * destruction, i.e. the Rx client code previously called
2206 * rx_DestroyConnection (client connections), or
2207 * rxi_ReapConnections called the same routine (server
2208 * connections). Only do this, however, if there are no
2209 * outstanding calls. Note that for fine grain locking, there appears
2210 * to be a deadlock in that rxi_FreeCall has a call locked and
2211 * DestroyConnectionNoLock locks each call in the conn. But note a
2212 * few lines up where we have removed this call from the conn.
2213 * If someone else destroys a connection, they either have no
2214 * call lock held or are going through this section of code.
2216 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2217 MUTEX_ENTER(&conn->conn_data_lock);
2219 MUTEX_EXIT(&conn->conn_data_lock);
2220 #ifdef RX_ENABLE_LOCKS
2222 rxi_DestroyConnectionNoLock(conn);
2224 rxi_DestroyConnection(conn);
2225 #else /* RX_ENABLE_LOCKS */
2226 rxi_DestroyConnection(conn);
2227 #endif /* RX_ENABLE_LOCKS */
2231 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2233 rxi_Alloc(register size_t size)
2237 MUTEX_ENTER(&rx_stats_mutex);
2239 rxi_Allocsize += (afs_int32)size;
2240 MUTEX_EXIT(&rx_stats_mutex);
2242 p = (char *)osi_Alloc(size);
2245 osi_Panic("rxi_Alloc error");
2251 rxi_Free(void *addr, register size_t size)
2253 MUTEX_ENTER(&rx_stats_mutex);
2255 rxi_Allocsize -= (afs_int32)size;
2256 MUTEX_EXIT(&rx_stats_mutex);
2258 osi_Free(addr, size);
2261 /* Find the peer process represented by the supplied (host,port)
2262 * combination. If there is no appropriate active peer structure, a
2263 * new one will be allocated and initialized
2264 * The origPeer, if set, is a pointer to a peer structure on which the
2265 * refcount will be be decremented. This is used to replace the peer
2266 * structure hanging off a connection structure */
2268 rxi_FindPeer(register afs_uint32 host, register u_short port,
2269 struct rx_peer *origPeer, int create)
2271 register struct rx_peer *pp;
2273 hashIndex = PEER_HASH(host, port);
2274 MUTEX_ENTER(&rx_peerHashTable_lock);
2275 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2276 if ((pp->host == host) && (pp->port == port))
2281 pp = rxi_AllocPeer(); /* This bzero's *pp */
2282 pp->host = host; /* set here or in InitPeerParams is zero */
2284 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2285 queue_Init(&pp->congestionQueue);
2286 queue_Init(&pp->rpcStats);
2287 pp->next = rx_peerHashTable[hashIndex];
2288 rx_peerHashTable[hashIndex] = pp;
2289 rxi_InitPeerParams(pp);
2290 MUTEX_ENTER(&rx_stats_mutex);
2291 rx_stats.nPeerStructs++;
2292 MUTEX_EXIT(&rx_stats_mutex);
2299 origPeer->refCount--;
2300 MUTEX_EXIT(&rx_peerHashTable_lock);
2305 /* Find the connection at (host, port) started at epoch, and with the
2306 * given connection id. Creates the server connection if necessary.
2307 * The type specifies whether a client connection or a server
2308 * connection is desired. In both cases, (host, port) specify the
2309 * peer's (host, pair) pair. Client connections are not made
2310 * automatically by this routine. The parameter socket gives the
2311 * socket descriptor on which the packet was received. This is used,
2312 * in the case of server connections, to check that *new* connections
2313 * come via a valid (port, serviceId). Finally, the securityIndex
2314 * parameter must match the existing index for the connection. If a
2315 * server connection is created, it will be created using the supplied
2316 * index, if the index is valid for this service */
2317 struct rx_connection *
2318 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2319 register u_short port, u_short serviceId, afs_uint32 cid,
2320 afs_uint32 epoch, int type, u_int securityIndex)
2322 int hashindex, flag;
2323 register struct rx_connection *conn;
2324 hashindex = CONN_HASH(host, port, cid, epoch, type);
2325 MUTEX_ENTER(&rx_connHashTable_lock);
2326 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2327 rx_connHashTable[hashindex],
2330 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2331 && (epoch == conn->epoch)) {
2332 register struct rx_peer *pp = conn->peer;
2333 if (securityIndex != conn->securityIndex) {
2334 /* this isn't supposed to happen, but someone could forge a packet
2335 * like this, and there seems to be some CM bug that makes this
2336 * happen from time to time -- in which case, the fileserver
2338 MUTEX_EXIT(&rx_connHashTable_lock);
2339 return (struct rx_connection *)0;
2341 if (pp->host == host && pp->port == port)
2343 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2345 /* So what happens when it's a callback connection? */
2346 if ( /*type == RX_CLIENT_CONNECTION && */
2347 (conn->epoch & 0x80000000))
2351 /* the connection rxLastConn that was used the last time is not the
2352 ** one we are looking for now. Hence, start searching in the hash */
2354 conn = rx_connHashTable[hashindex];
2359 struct rx_service *service;
2360 if (type == RX_CLIENT_CONNECTION) {
2361 MUTEX_EXIT(&rx_connHashTable_lock);
2362 return (struct rx_connection *)0;
2364 service = rxi_FindService(socket, serviceId);
2365 if (!service || (securityIndex >= service->nSecurityObjects)
2366 || (service->securityObjects[securityIndex] == 0)) {
2367 MUTEX_EXIT(&rx_connHashTable_lock);
2368 return (struct rx_connection *)0;
2370 conn = rxi_AllocConnection(); /* This bzero's the connection */
2371 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2372 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2373 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2374 conn->next = rx_connHashTable[hashindex];
2375 rx_connHashTable[hashindex] = conn;
2376 conn->peer = rxi_FindPeer(host, port, 0, 1);
2377 conn->type = RX_SERVER_CONNECTION;
2378 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2379 conn->epoch = epoch;
2380 conn->cid = cid & RX_CIDMASK;
2381 /* conn->serial = conn->lastSerial = 0; */
2382 /* conn->timeout = 0; */
2383 conn->ackRate = RX_FAST_ACK_RATE;
2384 conn->service = service;
2385 conn->serviceId = serviceId;
2386 conn->securityIndex = securityIndex;
2387 conn->securityObject = service->securityObjects[securityIndex];
2388 conn->nSpecific = 0;
2389 conn->specific = NULL;
2390 rx_SetConnDeadTime(conn, service->connDeadTime);
2391 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2392 /* Notify security object of the new connection */
2393 RXS_NewConnection(conn->securityObject, conn);
2394 /* XXXX Connection timeout? */
2395 if (service->newConnProc)
2396 (*service->newConnProc) (conn);
2397 MUTEX_ENTER(&rx_stats_mutex);
2398 rx_stats.nServerConns++;
2399 MUTEX_EXIT(&rx_stats_mutex);
2402 MUTEX_ENTER(&conn->conn_data_lock);
2404 MUTEX_EXIT(&conn->conn_data_lock);
2406 rxLastConn = conn; /* store this connection as the last conn used */
2407 MUTEX_EXIT(&rx_connHashTable_lock);
2411 /* There are two packet tracing routines available for testing and monitoring
2412 * Rx. One is called just after every packet is received and the other is
2413 * called just before every packet is sent. Received packets, have had their
2414 * headers decoded, and packets to be sent have not yet had their headers
2415 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2416 * containing the network address. Both can be modified. The return value, if
2417 * non-zero, indicates that the packet should be dropped. */
2419 int (*rx_justReceived) () = 0;
2420 int (*rx_almostSent) () = 0;
2422 /* A packet has been received off the interface. Np is the packet, socket is
2423 * the socket number it was received from (useful in determining which service
2424 * this packet corresponds to), and (host, port) reflect the host,port of the
2425 * sender. This call returns the packet to the caller if it is finished with
2426 * it, rather than de-allocating it, just as a small performance hack */
2429 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2430 afs_uint32 host, u_short port, int *tnop,
2431 struct rx_call **newcallp)
2433 register struct rx_call *call;
2434 register struct rx_connection *conn;
2436 afs_uint32 currentCallNumber;
2442 struct rx_packet *tnp;
2445 /* We don't print out the packet until now because (1) the time may not be
2446 * accurate enough until now in the lwp implementation (rx_Listener only gets
2447 * the time after the packet is read) and (2) from a protocol point of view,
2448 * this is the first time the packet has been seen */
2449 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2450 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2451 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2452 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2453 np->header.epoch, np->header.cid, np->header.callNumber,
2454 np->header.seq, np->header.flags, np));
2457 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2458 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2461 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2462 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2465 /* If an input tracer function is defined, call it with the packet and
2466 * network address. Note this function may modify its arguments. */
2467 if (rx_justReceived) {
2468 struct sockaddr_in addr;
2470 addr.sin_family = AF_INET;
2471 addr.sin_port = port;
2472 addr.sin_addr.s_addr = host;
2473 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2474 addr.sin_len = sizeof(addr);
2475 #endif /* AFS_OSF_ENV */
2476 drop = (*rx_justReceived) (np, &addr);
2477 /* drop packet if return value is non-zero */
2480 port = addr.sin_port; /* in case fcn changed addr */
2481 host = addr.sin_addr.s_addr;
2485 /* If packet was not sent by the client, then *we* must be the client */
2486 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2487 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2489 /* Find the connection (or fabricate one, if we're the server & if
2490 * necessary) associated with this packet */
2492 rxi_FindConnection(socket, host, port, np->header.serviceId,
2493 np->header.cid, np->header.epoch, type,
2494 np->header.securityIndex);
2497 /* If no connection found or fabricated, just ignore the packet.
2498 * (An argument could be made for sending an abort packet for
2503 MUTEX_ENTER(&conn->conn_data_lock);
2504 if (conn->maxSerial < np->header.serial)
2505 conn->maxSerial = np->header.serial;
2506 MUTEX_EXIT(&conn->conn_data_lock);
2508 /* If the connection is in an error state, send an abort packet and ignore
2509 * the incoming packet */
2511 /* Don't respond to an abort packet--we don't want loops! */
2512 MUTEX_ENTER(&conn->conn_data_lock);
2513 if (np->header.type != RX_PACKET_TYPE_ABORT)
2514 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2516 MUTEX_EXIT(&conn->conn_data_lock);
2520 /* Check for connection-only requests (i.e. not call specific). */
2521 if (np->header.callNumber == 0) {
2522 switch (np->header.type) {
2523 case RX_PACKET_TYPE_ABORT: {
2524 /* What if the supplied error is zero? */
2525 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2526 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2527 rxi_ConnectionError(conn, errcode);
2528 MUTEX_ENTER(&conn->conn_data_lock);
2530 MUTEX_EXIT(&conn->conn_data_lock);
2533 case RX_PACKET_TYPE_CHALLENGE:
2534 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2535 MUTEX_ENTER(&conn->conn_data_lock);
2537 MUTEX_EXIT(&conn->conn_data_lock);
2539 case RX_PACKET_TYPE_RESPONSE:
2540 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2541 MUTEX_ENTER(&conn->conn_data_lock);
2543 MUTEX_EXIT(&conn->conn_data_lock);
2545 case RX_PACKET_TYPE_PARAMS:
2546 case RX_PACKET_TYPE_PARAMS + 1:
2547 case RX_PACKET_TYPE_PARAMS + 2:
2548 /* ignore these packet types for now */
2549 MUTEX_ENTER(&conn->conn_data_lock);
2551 MUTEX_EXIT(&conn->conn_data_lock);
2556 /* Should not reach here, unless the peer is broken: send an
2558 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2559 MUTEX_ENTER(&conn->conn_data_lock);
2560 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2562 MUTEX_EXIT(&conn->conn_data_lock);
2567 channel = np->header.cid & RX_CHANNELMASK;
2568 call = conn->call[channel];
2569 #ifdef RX_ENABLE_LOCKS
2571 MUTEX_ENTER(&call->lock);
2572 /* Test to see if call struct is still attached to conn. */
2573 if (call != conn->call[channel]) {
2575 MUTEX_EXIT(&call->lock);
2576 if (type == RX_SERVER_CONNECTION) {
2577 call = conn->call[channel];
2578 /* If we started with no call attached and there is one now,
2579 * another thread is also running this routine and has gotten
2580 * the connection channel. We should drop this packet in the tests
2581 * below. If there was a call on this connection and it's now
2582 * gone, then we'll be making a new call below.
2583 * If there was previously a call and it's now different then
2584 * the old call was freed and another thread running this routine
2585 * has created a call on this channel. One of these two threads
2586 * has a packet for the old call and the code below handles those
2590 MUTEX_ENTER(&call->lock);
2592 /* This packet can't be for this call. If the new call address is
2593 * 0 then no call is running on this channel. If there is a call
2594 * then, since this is a client connection we're getting data for
2595 * it must be for the previous call.
2597 MUTEX_ENTER(&rx_stats_mutex);
2598 rx_stats.spuriousPacketsRead++;
2599 MUTEX_EXIT(&rx_stats_mutex);
2600 MUTEX_ENTER(&conn->conn_data_lock);
2602 MUTEX_EXIT(&conn->conn_data_lock);
2607 currentCallNumber = conn->callNumber[channel];
2609 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2610 if (np->header.callNumber < currentCallNumber) {
2611 MUTEX_ENTER(&rx_stats_mutex);
2612 rx_stats.spuriousPacketsRead++;
2613 MUTEX_EXIT(&rx_stats_mutex);
2614 #ifdef RX_ENABLE_LOCKS
2616 MUTEX_EXIT(&call->lock);
2618 MUTEX_ENTER(&conn->conn_data_lock);
2620 MUTEX_EXIT(&conn->conn_data_lock);
2624 MUTEX_ENTER(&conn->conn_call_lock);
2625 call = rxi_NewCall(conn, channel);
2626 MUTEX_EXIT(&conn->conn_call_lock);
2627 *call->callNumber = np->header.callNumber;
2628 if (np->header.callNumber == 0)
2629 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));
2631 call->state = RX_STATE_PRECALL;
2632 clock_GetTime(&call->queueTime);
2633 hzero(call->bytesSent);
2634 hzero(call->bytesRcvd);
2636 * If the number of queued calls exceeds the overload
2637 * threshold then abort this call.
2639 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2640 struct rx_packet *tp;
2642 rxi_CallError(call, rx_BusyError);
2643 tp = rxi_SendCallAbort(call, np, 1, 0);
2644 MUTEX_EXIT(&call->lock);
2645 MUTEX_ENTER(&conn->conn_data_lock);
2647 MUTEX_EXIT(&conn->conn_data_lock);
2648 MUTEX_ENTER(&rx_stats_mutex);
2650 MUTEX_EXIT(&rx_stats_mutex);
2653 rxi_KeepAliveOn(call);
2654 } else if (np->header.callNumber != currentCallNumber) {
2655 /* Wait until the transmit queue is idle before deciding
2656 * whether to reset the current call. Chances are that the
2657 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2660 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2661 while ((call->state == RX_STATE_ACTIVE)
2662 && (call->flags & RX_CALL_TQ_BUSY)) {
2663 call->flags |= RX_CALL_TQ_WAIT;
2665 #ifdef RX_ENABLE_LOCKS
2666 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2667 CV_WAIT(&call->cv_tq, &call->lock);
2668 #else /* RX_ENABLE_LOCKS */
2669 osi_rxSleep(&call->tq);
2670 #endif /* RX_ENABLE_LOCKS */
2672 if (call->tqWaiters == 0)
2673 call->flags &= ~RX_CALL_TQ_WAIT;
2675 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2676 /* If the new call cannot be taken right now send a busy and set
2677 * the error condition in this call, so that it terminates as
2678 * quickly as possible */
2679 if (call->state == RX_STATE_ACTIVE) {
2680 struct rx_packet *tp;
2682 rxi_CallError(call, RX_CALL_DEAD);
2683 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2685 MUTEX_EXIT(&call->lock);
2686 MUTEX_ENTER(&conn->conn_data_lock);
2688 MUTEX_EXIT(&conn->conn_data_lock);
2691 rxi_ResetCall(call, 0);
2692 *call->callNumber = np->header.callNumber;
2693 if (np->header.callNumber == 0)
2694 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));
2696 call->state = RX_STATE_PRECALL;
2697 clock_GetTime(&call->queueTime);
2698 hzero(call->bytesSent);
2699 hzero(call->bytesRcvd);
2701 * If the number of queued calls exceeds the overload
2702 * threshold then abort this call.
2704 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2705 struct rx_packet *tp;
2707 rxi_CallError(call, rx_BusyError);
2708 tp = rxi_SendCallAbort(call, np, 1, 0);
2709 MUTEX_EXIT(&call->lock);
2710 MUTEX_ENTER(&conn->conn_data_lock);
2712 MUTEX_EXIT(&conn->conn_data_lock);
2713 MUTEX_ENTER(&rx_stats_mutex);
2715 MUTEX_EXIT(&rx_stats_mutex);
2718 rxi_KeepAliveOn(call);
2720 /* Continuing call; do nothing here. */
2722 } else { /* we're the client */
2723 /* Ignore all incoming acknowledgements for calls in DALLY state */
2724 if (call && (call->state == RX_STATE_DALLY)
2725 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2726 MUTEX_ENTER(&rx_stats_mutex);
2727 rx_stats.ignorePacketDally++;
2728 MUTEX_EXIT(&rx_stats_mutex);
2729 #ifdef RX_ENABLE_LOCKS
2731 MUTEX_EXIT(&call->lock);
2734 MUTEX_ENTER(&conn->conn_data_lock);
2736 MUTEX_EXIT(&conn->conn_data_lock);
2740 /* Ignore anything that's not relevant to the current call. If there
2741 * isn't a current call, then no packet is relevant. */
2742 if (!call || (np->header.callNumber != currentCallNumber)) {
2743 MUTEX_ENTER(&rx_stats_mutex);
2744 rx_stats.spuriousPacketsRead++;
2745 MUTEX_EXIT(&rx_stats_mutex);
2746 #ifdef RX_ENABLE_LOCKS
2748 MUTEX_EXIT(&call->lock);
2751 MUTEX_ENTER(&conn->conn_data_lock);
2753 MUTEX_EXIT(&conn->conn_data_lock);
2756 /* If the service security object index stamped in the packet does not
2757 * match the connection's security index, ignore the packet */
2758 if (np->header.securityIndex != conn->securityIndex) {
2759 #ifdef RX_ENABLE_LOCKS
2760 MUTEX_EXIT(&call->lock);
2762 MUTEX_ENTER(&conn->conn_data_lock);
2764 MUTEX_EXIT(&conn->conn_data_lock);
2768 /* If we're receiving the response, then all transmit packets are
2769 * implicitly acknowledged. Get rid of them. */
2770 if (np->header.type == RX_PACKET_TYPE_DATA) {
2771 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2772 /* XXX Hack. Because we must release the global rx lock when
2773 * sending packets (osi_NetSend) we drop all acks while we're
2774 * traversing the tq in rxi_Start sending packets out because
2775 * packets may move to the freePacketQueue as result of being here!
2776 * So we drop these packets until we're safely out of the
2777 * traversing. Really ugly!
2778 * For fine grain RX locking, we set the acked field in the
2779 * packets and let rxi_Start remove them from the transmit queue.
2781 if (call->flags & RX_CALL_TQ_BUSY) {
2782 #ifdef RX_ENABLE_LOCKS
2783 rxi_SetAcksInTransmitQueue(call);
2786 return np; /* xmitting; drop packet */
2789 rxi_ClearTransmitQueue(call, 0);
2791 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2792 rxi_ClearTransmitQueue(call, 0);
2793 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2795 if (np->header.type == RX_PACKET_TYPE_ACK) {
2796 /* now check to see if this is an ack packet acknowledging that the
2797 * server actually *lost* some hard-acked data. If this happens we
2798 * ignore this packet, as it may indicate that the server restarted in
2799 * the middle of a call. It is also possible that this is an old ack
2800 * packet. We don't abort the connection in this case, because this
2801 * *might* just be an old ack packet. The right way to detect a server
2802 * restart in the midst of a call is to notice that the server epoch
2804 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2805 * XXX unacknowledged. I think that this is off-by-one, but
2806 * XXX I don't dare change it just yet, since it will
2807 * XXX interact badly with the server-restart detection
2808 * XXX code in receiveackpacket. */
2809 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2810 MUTEX_ENTER(&rx_stats_mutex);
2811 rx_stats.spuriousPacketsRead++;
2812 MUTEX_EXIT(&rx_stats_mutex);
2813 MUTEX_EXIT(&call->lock);
2814 MUTEX_ENTER(&conn->conn_data_lock);
2816 MUTEX_EXIT(&conn->conn_data_lock);
2820 } /* else not a data packet */
2823 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2824 /* Set remote user defined status from packet */
2825 call->remoteStatus = np->header.userStatus;
2827 /* Note the gap between the expected next packet and the actual
2828 * packet that arrived, when the new packet has a smaller serial number
2829 * than expected. Rioses frequently reorder packets all by themselves,
2830 * so this will be quite important with very large window sizes.
2831 * Skew is checked against 0 here to avoid any dependence on the type of
2832 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2834 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2835 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2836 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2838 MUTEX_ENTER(&conn->conn_data_lock);
2839 skew = conn->lastSerial - np->header.serial;
2840 conn->lastSerial = np->header.serial;
2841 MUTEX_EXIT(&conn->conn_data_lock);
2843 register struct rx_peer *peer;
2845 if (skew > peer->inPacketSkew) {
2846 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2848 peer->inPacketSkew = skew;
2852 /* Now do packet type-specific processing */
2853 switch (np->header.type) {
2854 case RX_PACKET_TYPE_DATA:
2855 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2858 case RX_PACKET_TYPE_ACK:
2859 /* Respond immediately to ack packets requesting acknowledgement
2861 if (np->header.flags & RX_REQUEST_ACK) {
2863 (void)rxi_SendCallAbort(call, 0, 1, 0);
2865 (void)rxi_SendAck(call, 0, np->header.serial,
2866 RX_ACK_PING_RESPONSE, 1);
2868 np = rxi_ReceiveAckPacket(call, np, 1);
2870 case RX_PACKET_TYPE_ABORT: {
2871 /* An abort packet: reset the call, passing the error up to the user. */
2872 /* What if error is zero? */
2873 /* What if the error is -1? the application will treat it as a timeout. */
2874 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
2875 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
2876 rxi_CallError(call, errdata);
2877 MUTEX_EXIT(&call->lock);
2878 MUTEX_ENTER(&conn->conn_data_lock);
2880 MUTEX_EXIT(&conn->conn_data_lock);
2881 return np; /* xmitting; drop packet */
2883 case RX_PACKET_TYPE_BUSY:
2886 case RX_PACKET_TYPE_ACKALL:
2887 /* All packets acknowledged, so we can drop all packets previously
2888 * readied for sending */
2889 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2890 /* XXX Hack. We because we can't release the global rx lock when
2891 * sending packets (osi_NetSend) we drop all ack pkts while we're
2892 * traversing the tq in rxi_Start sending packets out because
2893 * packets may move to the freePacketQueue as result of being
2894 * here! So we drop these packets until we're safely out of the
2895 * traversing. Really ugly!
2896 * For fine grain RX locking, we set the acked field in the packets
2897 * and let rxi_Start remove the packets from the transmit queue.
2899 if (call->flags & RX_CALL_TQ_BUSY) {
2900 #ifdef RX_ENABLE_LOCKS
2901 rxi_SetAcksInTransmitQueue(call);
2903 #else /* RX_ENABLE_LOCKS */
2904 MUTEX_EXIT(&call->lock);
2905 MUTEX_ENTER(&conn->conn_data_lock);
2907 MUTEX_EXIT(&conn->conn_data_lock);
2908 return np; /* xmitting; drop packet */
2909 #endif /* RX_ENABLE_LOCKS */
2911 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2912 rxi_ClearTransmitQueue(call, 0);
2915 /* Should not reach here, unless the peer is broken: send an abort
2917 rxi_CallError(call, RX_PROTOCOL_ERROR);
2918 np = rxi_SendCallAbort(call, np, 1, 0);
2921 /* Note when this last legitimate packet was received, for keep-alive
2922 * processing. Note, we delay getting the time until now in the hope that
2923 * the packet will be delivered to the user before any get time is required
2924 * (if not, then the time won't actually be re-evaluated here). */
2925 call->lastReceiveTime = clock_Sec();
2926 MUTEX_EXIT(&call->lock);
2927 MUTEX_ENTER(&conn->conn_data_lock);
2929 MUTEX_EXIT(&conn->conn_data_lock);
2933 /* return true if this is an "interesting" connection from the point of view
2934 of someone trying to debug the system */
2936 rxi_IsConnInteresting(struct rx_connection *aconn)
2939 register struct rx_call *tcall;
2941 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2943 for (i = 0; i < RX_MAXCALLS; i++) {
2944 tcall = aconn->call[i];
2946 if ((tcall->state == RX_STATE_PRECALL)
2947 || (tcall->state == RX_STATE_ACTIVE))
2949 if ((tcall->mode == RX_MODE_SENDING)
2950 || (tcall->mode == RX_MODE_RECEIVING))
2958 /* if this is one of the last few packets AND it wouldn't be used by the
2959 receiving call to immediately satisfy a read request, then drop it on
2960 the floor, since accepting it might prevent a lock-holding thread from
2961 making progress in its reading. If a call has been cleared while in
2962 the precall state then ignore all subsequent packets until the call
2963 is assigned to a thread. */
2966 TooLow(struct rx_packet *ap, struct rx_call *acall)
2969 MUTEX_ENTER(&rx_stats_mutex);
2970 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
2971 && (acall->state == RX_STATE_PRECALL))
2972 || ((rx_nFreePackets < rxi_dataQuota + 2)
2973 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
2974 && (acall->flags & RX_CALL_READER_WAIT)))) {
2977 MUTEX_EXIT(&rx_stats_mutex);
2983 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
2984 struct rx_call *acall)
2986 struct rx_call *call = acall;
2990 MUTEX_ENTER(&conn->conn_data_lock);
2991 conn->checkReachEvent = NULL;
2992 waiting = conn->flags & RX_CONN_ATTACHWAIT;
2995 MUTEX_EXIT(&conn->conn_data_lock);
2999 MUTEX_ENTER(&conn->conn_call_lock);
3000 MUTEX_ENTER(&conn->conn_data_lock);
3001 for (i = 0; i < RX_MAXCALLS; i++) {
3002 struct rx_call *tc = conn->call[i];
3003 if (tc && tc->state == RX_STATE_PRECALL) {
3009 /* Indicate that rxi_CheckReachEvent is no longer running by
3010 * clearing the flag. Must be atomic under conn_data_lock to
3011 * avoid a new call slipping by: rxi_CheckConnReach holds
3012 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3014 conn->flags &= ~RX_CONN_ATTACHWAIT;
3015 MUTEX_EXIT(&conn->conn_data_lock);
3016 MUTEX_EXIT(&conn->conn_call_lock);
3021 MUTEX_ENTER(&call->lock);
3022 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3024 MUTEX_EXIT(&call->lock);
3026 clock_GetTime(&when);
3027 when.sec += RX_CHECKREACH_TIMEOUT;
3028 MUTEX_ENTER(&conn->conn_data_lock);
3029 if (!conn->checkReachEvent) {
3031 conn->checkReachEvent =
3032 rxevent_Post(&when, rxi_CheckReachEvent, conn, NULL);
3034 MUTEX_EXIT(&conn->conn_data_lock);
3040 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3042 struct rx_service *service = conn->service;
3043 struct rx_peer *peer = conn->peer;
3044 afs_uint32 now, lastReach;
3046 if (service->checkReach == 0)
3050 MUTEX_ENTER(&peer->peer_lock);
3051 lastReach = peer->lastReachTime;
3052 MUTEX_EXIT(&peer->peer_lock);
3053 if (now - lastReach < RX_CHECKREACH_TTL)
3056 MUTEX_ENTER(&conn->conn_data_lock);
3057 if (conn->flags & RX_CONN_ATTACHWAIT) {
3058 MUTEX_EXIT(&conn->conn_data_lock);
3061 conn->flags |= RX_CONN_ATTACHWAIT;
3062 MUTEX_EXIT(&conn->conn_data_lock);
3063 if (!conn->checkReachEvent)
3064 rxi_CheckReachEvent(NULL, conn, call);
3069 /* try to attach call, if authentication is complete */
3071 TryAttach(register struct rx_call *acall, register osi_socket socket,
3072 register int *tnop, register struct rx_call **newcallp,
3075 struct rx_connection *conn = acall->conn;
3077 if (conn->type == RX_SERVER_CONNECTION
3078 && acall->state == RX_STATE_PRECALL) {
3079 /* Don't attach until we have any req'd. authentication. */
3080 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3081 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3082 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3083 /* Note: this does not necessarily succeed; there
3084 * may not any proc available
3087 rxi_ChallengeOn(acall->conn);
3092 /* A data packet has been received off the interface. This packet is
3093 * appropriate to the call (the call is in the right state, etc.). This
3094 * routine can return a packet to the caller, for re-use */
3097 rxi_ReceiveDataPacket(register struct rx_call *call,
3098 register struct rx_packet *np, int istack,
3099 osi_socket socket, afs_uint32 host, u_short port,
3100 int *tnop, struct rx_call **newcallp)
3102 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3106 afs_uint32 seq, serial, flags;
3108 struct rx_packet *tnp;
3110 MUTEX_ENTER(&rx_stats_mutex);
3111 rx_stats.dataPacketsRead++;
3112 MUTEX_EXIT(&rx_stats_mutex);
3115 /* If there are no packet buffers, drop this new packet, unless we can find
3116 * packet buffers from inactive calls */
3118 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3119 MUTEX_ENTER(&rx_freePktQ_lock);
3120 rxi_NeedMorePackets = TRUE;
3121 MUTEX_EXIT(&rx_freePktQ_lock);
3122 MUTEX_ENTER(&rx_stats_mutex);
3123 rx_stats.noPacketBuffersOnRead++;
3124 MUTEX_EXIT(&rx_stats_mutex);
3125 call->rprev = np->header.serial;
3126 rxi_calltrace(RX_TRACE_DROP, call);
3127 dpf(("packet %x dropped on receipt - quota problems", np));
3129 rxi_ClearReceiveQueue(call);
3130 clock_GetTime(&when);
3131 clock_Add(&when, &rx_softAckDelay);
3132 if (!call->delayedAckEvent
3133 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3134 rxevent_Cancel(call->delayedAckEvent, call,
3135 RX_CALL_REFCOUNT_DELAY);
3136 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3137 call->delayedAckEvent =
3138 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3140 /* we've damaged this call already, might as well do it in. */
3146 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3147 * packet is one of several packets transmitted as a single
3148 * datagram. Do not send any soft or hard acks until all packets
3149 * in a jumbogram have been processed. Send negative acks right away.
3151 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3152 /* tnp is non-null when there are more packets in the
3153 * current jumbo gram */
3160 seq = np->header.seq;
3161 serial = np->header.serial;
3162 flags = np->header.flags;
3164 /* If the call is in an error state, send an abort message */
3166 return rxi_SendCallAbort(call, np, istack, 0);
3168 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3169 * AFS 3.5 jumbogram. */
3170 if (flags & RX_JUMBO_PACKET) {
3171 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3176 if (np->header.spare != 0) {
3177 MUTEX_ENTER(&call->conn->conn_data_lock);
3178 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3179 MUTEX_EXIT(&call->conn->conn_data_lock);
3182 /* The usual case is that this is the expected next packet */
3183 if (seq == call->rnext) {
3185 /* Check to make sure it is not a duplicate of one already queued */
3186 if (queue_IsNotEmpty(&call->rq)
3187 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3188 MUTEX_ENTER(&rx_stats_mutex);
3189 rx_stats.dupPacketsRead++;
3190 MUTEX_EXIT(&rx_stats_mutex);
3191 dpf(("packet %x dropped on receipt - duplicate", np));
3192 rxevent_Cancel(call->delayedAckEvent, call,
3193 RX_CALL_REFCOUNT_DELAY);
3194 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3200 /* It's the next packet. Stick it on the receive queue
3201 * for this call. Set newPackets to make sure we wake
3202 * the reader once all packets have been processed */
3203 queue_Prepend(&call->rq, np);
3205 np = NULL; /* We can't use this anymore */
3208 /* If an ack is requested then set a flag to make sure we
3209 * send an acknowledgement for this packet */
3210 if (flags & RX_REQUEST_ACK) {
3211 ackNeeded = RX_ACK_REQUESTED;
3214 /* Keep track of whether we have received the last packet */
3215 if (flags & RX_LAST_PACKET) {
3216 call->flags |= RX_CALL_HAVE_LAST;
3220 /* Check whether we have all of the packets for this call */
3221 if (call->flags & RX_CALL_HAVE_LAST) {
3222 afs_uint32 tseq; /* temporary sequence number */
3223 struct rx_packet *tp; /* Temporary packet pointer */
3224 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3226 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3227 if (tseq != tp->header.seq)
3229 if (tp->header.flags & RX_LAST_PACKET) {
3230 call->flags |= RX_CALL_RECEIVE_DONE;
3237 /* Provide asynchronous notification for those who want it
3238 * (e.g. multi rx) */
3239 if (call->arrivalProc) {
3240 (*call->arrivalProc) (call, call->arrivalProcHandle,
3241 call->arrivalProcArg);
3242 call->arrivalProc = (void (*)())0;
3245 /* Update last packet received */
3248 /* If there is no server process serving this call, grab
3249 * one, if available. We only need to do this once. If a
3250 * server thread is available, this thread becomes a server
3251 * thread and the server thread becomes a listener thread. */
3253 TryAttach(call, socket, tnop, newcallp, 0);
3256 /* This is not the expected next packet. */
3258 /* Determine whether this is a new or old packet, and if it's
3259 * a new one, whether it fits into the current receive window.
3260 * Also figure out whether the packet was delivered in sequence.
3261 * We use the prev variable to determine whether the new packet
3262 * is the successor of its immediate predecessor in the
3263 * receive queue, and the missing flag to determine whether
3264 * any of this packets predecessors are missing. */
3266 afs_uint32 prev; /* "Previous packet" sequence number */
3267 struct rx_packet *tp; /* Temporary packet pointer */
3268 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3269 int missing; /* Are any predecessors missing? */
3271 /* If the new packet's sequence number has been sent to the
3272 * application already, then this is a duplicate */
3273 if (seq < call->rnext) {
3274 MUTEX_ENTER(&rx_stats_mutex);
3275 rx_stats.dupPacketsRead++;
3276 MUTEX_EXIT(&rx_stats_mutex);
3277 rxevent_Cancel(call->delayedAckEvent, call,
3278 RX_CALL_REFCOUNT_DELAY);
3279 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3285 /* If the sequence number is greater than what can be
3286 * accomodated by the current window, then send a negative
3287 * acknowledge and drop the packet */
3288 if ((call->rnext + call->rwind) <= seq) {
3289 rxevent_Cancel(call->delayedAckEvent, call,
3290 RX_CALL_REFCOUNT_DELAY);
3291 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3298 /* Look for the packet in the queue of old received packets */
3299 for (prev = call->rnext - 1, missing =
3300 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3301 /*Check for duplicate packet */
3302 if (seq == tp->header.seq) {
3303 MUTEX_ENTER(&rx_stats_mutex);
3304 rx_stats.dupPacketsRead++;
3305 MUTEX_EXIT(&rx_stats_mutex);
3306 rxevent_Cancel(call->delayedAckEvent, call,
3307 RX_CALL_REFCOUNT_DELAY);
3308 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3314 /* If we find a higher sequence packet, break out and
3315 * insert the new packet here. */
3316 if (seq < tp->header.seq)
3318 /* Check for missing packet */
3319 if (tp->header.seq != prev + 1) {
3323 prev = tp->header.seq;
3326 /* Keep track of whether we have received the last packet. */
3327 if (flags & RX_LAST_PACKET) {
3328 call->flags |= RX_CALL_HAVE_LAST;
3331 /* It's within the window: add it to the the receive queue.
3332 * tp is left by the previous loop either pointing at the
3333 * packet before which to insert the new packet, or at the
3334 * queue head if the queue is empty or the packet should be
3336 queue_InsertBefore(tp, np);
3340 /* Check whether we have all of the packets for this call */
3341 if ((call->flags & RX_CALL_HAVE_LAST)
3342 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3343 afs_uint32 tseq; /* temporary sequence number */
3346 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3347 if (tseq != tp->header.seq)
3349 if (tp->header.flags & RX_LAST_PACKET) {
3350 call->flags |= RX_CALL_RECEIVE_DONE;
3357 /* We need to send an ack of the packet is out of sequence,
3358 * or if an ack was requested by the peer. */
3359 if (seq != prev + 1 || missing || (flags & RX_REQUEST_ACK)) {
3360 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3363 /* Acknowledge the last packet for each call */
3364 if (flags & RX_LAST_PACKET) {
3375 * If the receiver is waiting for an iovec, fill the iovec
3376 * using the data from the receive queue */
3377 if (call->flags & RX_CALL_IOVEC_WAIT) {
3378 didHardAck = rxi_FillReadVec(call, serial);
3379 /* the call may have been aborted */
3388 /* Wakeup the reader if any */
3389 if ((call->flags & RX_CALL_READER_WAIT)
3390 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3391 || (call->iovNext >= call->iovMax)
3392 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3393 call->flags &= ~RX_CALL_READER_WAIT;
3394 #ifdef RX_ENABLE_LOCKS
3395 CV_BROADCAST(&call->cv_rq);
3397 osi_rxWakeup(&call->rq);
3403 * Send an ack when requested by the peer, or once every
3404 * rxi_SoftAckRate packets until the last packet has been
3405 * received. Always send a soft ack for the last packet in
3406 * the server's reply. */
3408 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3409 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3410 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3411 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3412 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3413 } else if (call->nSoftAcks) {
3414 clock_GetTime(&when);
3415 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3416 clock_Add(&when, &rx_lastAckDelay);
3418 clock_Add(&when, &rx_softAckDelay);
3420 if (!call->delayedAckEvent
3421 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3422 rxevent_Cancel(call->delayedAckEvent, call,
3423 RX_CALL_REFCOUNT_DELAY);
3424 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3425 call->delayedAckEvent =
3426 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3428 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3429 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3436 static void rxi_ComputeRate();
3440 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3442 struct rx_peer *peer = conn->peer;
3444 MUTEX_ENTER(&peer->peer_lock);
3445 peer->lastReachTime = clock_Sec();
3446 MUTEX_EXIT(&peer->peer_lock);
3448 MUTEX_ENTER(&conn->conn_data_lock);
3449 if (conn->flags & RX_CONN_ATTACHWAIT) {
3452 conn->flags &= ~RX_CONN_ATTACHWAIT;
3453 MUTEX_EXIT(&conn->conn_data_lock);
3455 for (i = 0; i < RX_MAXCALLS; i++) {
3456 struct rx_call *call = conn->call[i];
3459 MUTEX_ENTER(&call->lock);
3460 /* tnop can be null if newcallp is null */
3461 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3463 MUTEX_EXIT(&call->lock);
3467 MUTEX_EXIT(&conn->conn_data_lock);
3471 rx_ack_reason(int reason)
3474 case RX_ACK_REQUESTED:
3476 case RX_ACK_DUPLICATE:
3478 case RX_ACK_OUT_OF_SEQUENCE:
3480 case RX_ACK_EXCEEDS_WINDOW:
3482 case RX_ACK_NOSPACE:
3486 case RX_ACK_PING_RESPONSE:
3498 /* rxi_ComputePeerNetStats
3500 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3501 * estimates (like RTT and throughput) based on ack packets. Caller
3502 * must ensure that the packet in question is the right one (i.e.
3503 * serial number matches).
3506 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3507 struct rx_ackPacket *ap, struct rx_packet *np)
3509 struct rx_peer *peer = call->conn->peer;
3511 /* Use RTT if not delayed by client. */
3512 if (ap->reason != RX_ACK_DELAY)
3513 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3515 rxi_ComputeRate(peer, call, p, np, ap->reason);
3519 /* The real smarts of the whole thing. */
3521 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3524 struct rx_ackPacket *ap;
3526 register struct rx_packet *tp;
3527 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3528 register struct rx_connection *conn = call->conn;
3529 struct rx_peer *peer = conn->peer;
3532 /* because there are CM's that are bogus, sending weird values for this. */
3533 afs_uint32 skew = 0;
3538 int newAckCount = 0;
3539 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3540 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3542 MUTEX_ENTER(&rx_stats_mutex);
3543 rx_stats.ackPacketsRead++;
3544 MUTEX_EXIT(&rx_stats_mutex);
3545 ap = (struct rx_ackPacket *)rx_DataOf(np);
3546 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3548 return np; /* truncated ack packet */
3550 /* depends on ack packet struct */
3551 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3552 first = ntohl(ap->firstPacket);
3553 serial = ntohl(ap->serial);
3554 /* temporarily disabled -- needs to degrade over time
3555 * skew = ntohs(ap->maxSkew); */
3557 /* Ignore ack packets received out of order */
3558 if (first < call->tfirst) {
3562 if (np->header.flags & RX_SLOW_START_OK) {
3563 call->flags |= RX_CALL_SLOW_START_OK;
3566 if (ap->reason == RX_ACK_PING_RESPONSE)
3567 rxi_UpdatePeerReach(conn, call);
3571 if (rxdebug_active) {
3575 len = _snprintf(msg, sizeof(msg),
3576 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3577 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3578 ntohl(ap->serial), ntohl(ap->previousPacket),
3579 (unsigned int)np->header.seq, (unsigned int)skew,
3580 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3584 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3585 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3589 OutputDebugString(msg);
3591 #else /* AFS_NT40_ENV */
3594 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3595 ap->reason, ntohl(ap->previousPacket),
3596 (unsigned int)np->header.seq, (unsigned int)serial,
3597 (unsigned int)skew, ntohl(ap->firstPacket));
3600 for (offset = 0; offset < nAcks; offset++)
3601 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3606 #endif /* AFS_NT40_ENV */
3609 /* Update the outgoing packet skew value to the latest value of
3610 * the peer's incoming packet skew value. The ack packet, of
3611 * course, could arrive out of order, but that won't affect things
3613 MUTEX_ENTER(&peer->peer_lock);
3614 peer->outPacketSkew = skew;
3616 /* Check for packets that no longer need to be transmitted, and
3617 * discard them. This only applies to packets positively
3618 * acknowledged as having been sent to the peer's upper level.
3619 * All other packets must be retained. So only packets with
3620 * sequence numbers < ap->firstPacket are candidates. */
3621 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3622 if (tp->header.seq >= first)
3624 call->tfirst = tp->header.seq + 1;
3626 && (tp->header.serial == serial || tp->firstSerial == serial))
3627 rxi_ComputePeerNetStats(call, tp, ap, np);
3628 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3631 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3632 /* XXX Hack. Because we have to release the global rx lock when sending
3633 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3634 * in rxi_Start sending packets out because packets may move to the
3635 * freePacketQueue as result of being here! So we drop these packets until
3636 * we're safely out of the traversing. Really ugly!
3637 * To make it even uglier, if we're using fine grain locking, we can
3638 * set the ack bits in the packets and have rxi_Start remove the packets
3639 * when it's done transmitting.
3641 if (call->flags & RX_CALL_TQ_BUSY) {
3642 #ifdef RX_ENABLE_LOCKS
3643 tp->flags |= RX_PKTFLAG_ACKED;
3644 call->flags |= RX_CALL_TQ_SOME_ACKED;
3645 #else /* RX_ENABLE_LOCKS */
3647 #endif /* RX_ENABLE_LOCKS */
3649 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3652 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3657 /* Give rate detector a chance to respond to ping requests */
3658 if (ap->reason == RX_ACK_PING_RESPONSE) {
3659 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3663 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3665 /* Now go through explicit acks/nacks and record the results in
3666 * the waiting packets. These are packets that can't be released
3667 * yet, even with a positive acknowledge. This positive
3668 * acknowledge only means the packet has been received by the
3669 * peer, not that it will be retained long enough to be sent to
3670 * the peer's upper level. In addition, reset the transmit timers
3671 * of any missing packets (those packets that must be missing
3672 * because this packet was out of sequence) */
3674 call->nSoftAcked = 0;
3675 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3676 /* Update round trip time if the ack was stimulated on receipt
3678 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3679 #ifdef RX_ENABLE_LOCKS
3680 if (tp->header.seq >= first)
3681 #endif /* RX_ENABLE_LOCKS */
3682 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3684 && (tp->header.serial == serial || tp->firstSerial == serial))
3685 rxi_ComputePeerNetStats(call, tp, ap, np);
3687 /* Set the acknowledge flag per packet based on the
3688 * information in the ack packet. An acknowlegded packet can
3689 * be downgraded when the server has discarded a packet it
3690 * soacked previously, or when an ack packet is received
3691 * out of sequence. */
3692 if (tp->header.seq < first) {
3693 /* Implicit ack information */
3694 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3697 tp->flags |= RX_PKTFLAG_ACKED;
3698 } else if (tp->header.seq < first + nAcks) {
3699 /* Explicit ack information: set it in the packet appropriately */
3700 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3701 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3703 tp->flags |= RX_PKTFLAG_ACKED;
3710 } else /* RX_ACK_TYPE_NACK */ {
3711 tp->flags &= ~RX_PKTFLAG_ACKED;
3715 tp->flags &= ~RX_PKTFLAG_ACKED;
3719 /* If packet isn't yet acked, and it has been transmitted at least
3720 * once, reset retransmit time using latest timeout
3721 * ie, this should readjust the retransmit timer for all outstanding
3722 * packets... So we don't just retransmit when we should know better*/
3724 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3725 tp->retryTime = tp->timeSent;
3726 clock_Add(&tp->retryTime, &peer->timeout);
3727 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3728 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3732 /* If the window has been extended by this acknowledge packet,
3733 * then wakeup a sender waiting in alloc for window space, or try
3734 * sending packets now, if he's been sitting on packets due to
3735 * lack of window space */
3736 if (call->tnext < (call->tfirst + call->twind)) {
3737 #ifdef RX_ENABLE_LOCKS
3738 CV_SIGNAL(&call->cv_twind);
3740 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3741 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3742 osi_rxWakeup(&call->twind);
3745 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3746 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3750 /* if the ack packet has a receivelen field hanging off it,
3751 * update our state */
3752 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3755 /* If the ack packet has a "recommended" size that is less than
3756 * what I am using now, reduce my size to match */
3757 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3758 (int)sizeof(afs_int32), &tSize);
3759 tSize = (afs_uint32) ntohl(tSize);
3760 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3762 /* Get the maximum packet size to send to this peer */
3763 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
3765 tSize = (afs_uint32) ntohl(tSize);
3766 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3767 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3769 /* sanity check - peer might have restarted with different params.
3770 * If peer says "send less", dammit, send less... Peer should never
3771 * be unable to accept packets of the size that prior AFS versions would
3772 * send without asking. */
3773 if (peer->maxMTU != tSize) {
3774 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
3776 peer->maxMTU = tSize;
3777 peer->MTU = MIN(tSize, peer->MTU);
3778 call->MTU = MIN(call->MTU, tSize);
3781 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3784 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3785 (int)sizeof(afs_int32), &tSize);
3786 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3787 if (tSize < call->twind) { /* smaller than our send */
3788 call->twind = tSize; /* window, we must send less... */
3789 call->ssthresh = MIN(call->twind, call->ssthresh);
3792 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3793 * network MTU confused with the loopback MTU. Calculate the
3794 * maximum MTU here for use in the slow start code below.
3796 maxMTU = peer->maxMTU;
3797 /* Did peer restart with older RX version? */
3798 if (peer->maxDgramPackets > 1) {
3799 peer->maxDgramPackets = 1;
3801 } else if (np->length >=
3802 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3805 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3806 sizeof(afs_int32), &tSize);
3807 tSize = (afs_uint32) ntohl(tSize);
3809 * As of AFS 3.5 we set the send window to match the receive window.
3811 if (tSize < call->twind) {
3812 call->twind = tSize;
3813 call->ssthresh = MIN(call->twind, call->ssthresh);
3814 } else if (tSize > call->twind) {
3815 call->twind = tSize;
3819 * As of AFS 3.5, a jumbogram is more than one fixed size
3820 * packet transmitted in a single UDP datagram. If the remote
3821 * MTU is smaller than our local MTU then never send a datagram
3822 * larger than the natural MTU.
3825 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3826 sizeof(afs_int32), &tSize);
3827 maxDgramPackets = (afs_uint32) ntohl(tSize);
3828 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3830 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
3831 maxDgramPackets = MIN(maxDgramPackets, tSize);
3832 if (maxDgramPackets > 1) {
3833 peer->maxDgramPackets = maxDgramPackets;
3834 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3836 peer->maxDgramPackets = 1;
3837 call->MTU = peer->natMTU;
3839 } else if (peer->maxDgramPackets > 1) {
3840 /* Restarted with lower version of RX */
3841 peer->maxDgramPackets = 1;
3843 } else if (peer->maxDgramPackets > 1
3844 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3845 /* Restarted with lower version of RX */
3846 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3847 peer->natMTU = OLD_MAX_PACKET_SIZE;
3848 peer->MTU = OLD_MAX_PACKET_SIZE;
3849 peer->maxDgramPackets = 1;
3850 peer->nDgramPackets = 1;
3852 call->MTU = OLD_MAX_PACKET_SIZE;
3857 * Calculate how many datagrams were successfully received after
3858 * the first missing packet and adjust the negative ack counter
3863 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3864 if (call->nNacks < nNacked) {
3865 call->nNacks = nNacked;
3874 if (call->flags & RX_CALL_FAST_RECOVER) {
3876 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3878 call->flags &= ~RX_CALL_FAST_RECOVER;
3879 call->cwind = call->nextCwind;
3880 call->nextCwind = 0;
3883 call->nCwindAcks = 0;
3884 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3885 /* Three negative acks in a row trigger congestion recovery */
3886 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3887 MUTEX_EXIT(&peer->peer_lock);
3888 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3889 /* someone else is waiting to start recovery */
3892 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3893 while (call->flags & RX_CALL_TQ_BUSY) {
3894 call->flags |= RX_CALL_TQ_WAIT;
3896 #ifdef RX_ENABLE_LOCKS
3897 osirx_AssertMine(&call->lock, "rxi_Start lock2");
3898 CV_WAIT(&call->cv_tq, &call->lock);
3899 #else /* RX_ENABLE_LOCKS */
3900 osi_rxSleep(&call->tq);
3901 #endif /* RX_ENABLE_LOCKS */
3903 if (call->tqWaiters == 0)
3904 call->flags &= ~RX_CALL_TQ_WAIT;
3906 MUTEX_ENTER(&peer->peer_lock);
3907 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3908 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3909 call->flags |= RX_CALL_FAST_RECOVER;
3910 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3912 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3913 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3914 call->nextCwind = call->ssthresh;
3917 peer->MTU = call->MTU;
3918 peer->cwind = call->nextCwind;
3919 peer->nDgramPackets = call->nDgramPackets;
3921 call->congestSeq = peer->congestSeq;
3922 /* Reset the resend times on the packets that were nacked
3923 * so we will retransmit as soon as the window permits*/
3924 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3926 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3927 clock_Zero(&tp->retryTime);
3929 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3934 /* If cwind is smaller than ssthresh, then increase
3935 * the window one packet for each ack we receive (exponential
3937 * If cwind is greater than or equal to ssthresh then increase
3938 * the congestion window by one packet for each cwind acks we
3939 * receive (linear growth). */
3940 if (call->cwind < call->ssthresh) {
3942 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3943 call->nCwindAcks = 0;
3945 call->nCwindAcks += newAckCount;
3946 if (call->nCwindAcks >= call->cwind) {
3947 call->nCwindAcks = 0;
3948 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3952 * If we have received several acknowledgements in a row then
3953 * it is time to increase the size of our datagrams
3955 if ((int)call->nAcks > rx_nDgramThreshold) {
3956 if (peer->maxDgramPackets > 1) {
3957 if (call->nDgramPackets < peer->maxDgramPackets) {
3958 call->nDgramPackets++;
3960 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3961 } else if (call->MTU < peer->maxMTU) {
3962 call->MTU += peer->natMTU;
3963 call->MTU = MIN(call->MTU, peer->maxMTU);
3969 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3971 /* Servers need to hold the call until all response packets have
3972 * been acknowledged. Soft acks are good enough since clients
3973 * are not allowed to clear their receive queues. */
3974 if (call->state == RX_STATE_HOLD
3975 && call->tfirst + call->nSoftAcked >= call->tnext) {
3976 call->state = RX_STATE_DALLY;
3977 rxi_ClearTransmitQueue(call, 0);
3978 } else if (!queue_IsEmpty(&call->tq)) {
3979 rxi_Start(0, call, 0, istack);
3984 /* Received a response to a challenge packet */
3986 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
3987 register struct rx_packet *np, int istack)
3991 /* Ignore the packet if we're the client */
3992 if (conn->type == RX_CLIENT_CONNECTION)
3995 /* If already authenticated, ignore the packet (it's probably a retry) */
3996 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
3999 /* Otherwise, have the security object evaluate the response packet */
4000 error = RXS_CheckResponse(conn->securityObject, conn, np);
4002 /* If the response is invalid, reset the connection, sending
4003 * an abort to the peer */
4007 rxi_ConnectionError(conn, error);
4008 MUTEX_ENTER(&conn->conn_data_lock);
4009 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4010 MUTEX_EXIT(&conn->conn_data_lock);
4013 /* If the response is valid, any calls waiting to attach
4014 * servers can now do so */
4017 for (i = 0; i < RX_MAXCALLS; i++) {
4018 struct rx_call *call = conn->call[i];
4020 MUTEX_ENTER(&call->lock);
4021 if (call->state == RX_STATE_PRECALL)
4022 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4023 /* tnop can be null if newcallp is null */
4024 MUTEX_EXIT(&call->lock);
4028 /* Update the peer reachability information, just in case
4029 * some calls went into attach-wait while we were waiting
4030 * for authentication..
4032 rxi_UpdatePeerReach(conn, NULL);
4037 /* A client has received an authentication challenge: the security
4038 * object is asked to cough up a respectable response packet to send
4039 * back to the server. The server is responsible for retrying the
4040 * challenge if it fails to get a response. */
4043 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
4044 register struct rx_packet *np, int istack)
4048 /* Ignore the challenge if we're the server */
4049 if (conn->type == RX_SERVER_CONNECTION)
4052 /* Ignore the challenge if the connection is otherwise idle; someone's
4053 * trying to use us as an oracle. */
4054 if (!rxi_HasActiveCalls(conn))
4057 /* Send the security object the challenge packet. It is expected to fill
4058 * in the response. */
4059 error = RXS_GetResponse(conn->securityObject, conn, np);
4061 /* If the security object is unable to return a valid response, reset the
4062 * connection and send an abort to the peer. Otherwise send the response
4063 * packet to the peer connection. */
4065 rxi_ConnectionError(conn, error);
4066 MUTEX_ENTER(&conn->conn_data_lock);
4067 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4068 MUTEX_EXIT(&conn->conn_data_lock);
4070 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4071 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4077 /* Find an available server process to service the current request in
4078 * the given call structure. If one isn't available, queue up this
4079 * call so it eventually gets one */
4081 rxi_AttachServerProc(register struct rx_call *call,
4082 register osi_socket socket, register int *tnop,
4083 register struct rx_call **newcallp)
4085 register struct rx_serverQueueEntry *sq;
4086 register struct rx_service *service = call->conn->service;
4087 register int haveQuota = 0;
4089 /* May already be attached */
4090 if (call->state == RX_STATE_ACTIVE)
4093 MUTEX_ENTER(&rx_serverPool_lock);
4095 haveQuota = QuotaOK(service);
4096 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4097 /* If there are no processes available to service this call,
4098 * put the call on the incoming call queue (unless it's
4099 * already on the queue).
4101 #ifdef RX_ENABLE_LOCKS
4103 ReturnToServerPool(service);
4104 #endif /* RX_ENABLE_LOCKS */
4106 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4107 call->flags |= RX_CALL_WAIT_PROC;
4108 MUTEX_ENTER(&rx_stats_mutex);
4111 MUTEX_EXIT(&rx_stats_mutex);
4112 rxi_calltrace(RX_CALL_ARRIVAL, call);
4113 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4114 queue_Append(&rx_incomingCallQueue, call);
4117 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4119 /* If hot threads are enabled, and both newcallp and sq->socketp
4120 * are non-null, then this thread will process the call, and the
4121 * idle server thread will start listening on this threads socket.
4124 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4127 *sq->socketp = socket;
4128 clock_GetTime(&call->startTime);
4129 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4133 if (call->flags & RX_CALL_WAIT_PROC) {
4134 /* Conservative: I don't think this should happen */
4135 call->flags &= ~RX_CALL_WAIT_PROC;
4136 if (queue_IsOnQueue(call)) {
4138 MUTEX_ENTER(&rx_stats_mutex);
4140 MUTEX_EXIT(&rx_stats_mutex);
4143 call->state = RX_STATE_ACTIVE;
4144 call->mode = RX_MODE_RECEIVING;
4145 #ifdef RX_KERNEL_TRACE
4147 int glockOwner = ISAFS_GLOCK();
4150 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4151 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4157 if (call->flags & RX_CALL_CLEARED) {
4158 /* send an ack now to start the packet flow up again */
4159 call->flags &= ~RX_CALL_CLEARED;
4160 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4162 #ifdef RX_ENABLE_LOCKS
4165 service->nRequestsRunning++;
4166 if (service->nRequestsRunning <= service->minProcs)
4172 MUTEX_EXIT(&rx_serverPool_lock);
4175 /* Delay the sending of an acknowledge event for a short while, while
4176 * a new call is being prepared (in the case of a client) or a reply
4177 * is being prepared (in the case of a server). Rather than sending
4178 * an ack packet, an ACKALL packet is sent. */
4180 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4182 #ifdef RX_ENABLE_LOCKS
4184 MUTEX_ENTER(&call->lock);
4185 call->delayedAckEvent = NULL;
4186 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4188 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4189 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4191 MUTEX_EXIT(&call->lock);
4192 #else /* RX_ENABLE_LOCKS */
4194 call->delayedAckEvent = NULL;
4195 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4196 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4197 #endif /* RX_ENABLE_LOCKS */
4201 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4204 #ifdef RX_ENABLE_LOCKS
4206 MUTEX_ENTER(&call->lock);
4207 if (event == call->delayedAckEvent)
4208 call->delayedAckEvent = NULL;
4209 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4211 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4213 MUTEX_EXIT(&call->lock);
4214 #else /* RX_ENABLE_LOCKS */
4216 call->delayedAckEvent = NULL;
4217 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4218 #endif /* RX_ENABLE_LOCKS */
4222 #ifdef RX_ENABLE_LOCKS
4223 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4224 * clearing them out.
4227 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4229 register struct rx_packet *p, *tp;
4232 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4233 p->flags |= RX_PKTFLAG_ACKED;
4237 call->flags |= RX_CALL_TQ_CLEARME;
4238 call->flags |= RX_CALL_TQ_SOME_ACKED;
4241 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4242 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4243 call->tfirst = call->tnext;
4244 call->nSoftAcked = 0;
4246 if (call->flags & RX_CALL_FAST_RECOVER) {
4247 call->flags &= ~RX_CALL_FAST_RECOVER;
4248 call->cwind = call->nextCwind;
4249 call->nextCwind = 0;
4252 CV_SIGNAL(&call->cv_twind);
4254 #endif /* RX_ENABLE_LOCKS */
4256 /* Clear out the transmit queue for the current call (all packets have
4257 * been received by peer) */
4259 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4261 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4262 register struct rx_packet *p, *tp;
4264 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4266 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4267 p->flags |= RX_PKTFLAG_ACKED;
4271 call->flags |= RX_CALL_TQ_CLEARME;
4272 call->flags |= RX_CALL_TQ_SOME_ACKED;
4275 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4276 rxi_FreePackets(0, &call->tq);
4277 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4278 call->flags &= ~RX_CALL_TQ_CLEARME;
4280 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4282 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4283 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4284 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4285 call->nSoftAcked = 0;
4287 if (call->flags & RX_CALL_FAST_RECOVER) {
4288 call->flags &= ~RX_CALL_FAST_RECOVER;
4289 call->cwind = call->nextCwind;
4291 #ifdef RX_ENABLE_LOCKS
4292 CV_SIGNAL(&call->cv_twind);
4294 osi_rxWakeup(&call->twind);
4299 rxi_ClearReceiveQueue(register struct rx_call *call)
4301 if (queue_IsNotEmpty(&call->rq)) {
4302 rx_packetReclaims += rxi_FreePackets(0, &call->rq);
4303 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4305 if (call->state == RX_STATE_PRECALL) {
4306 call->flags |= RX_CALL_CLEARED;
4310 /* Send an abort packet for the specified call */
4312 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4313 int istack, int force)
4321 /* Clients should never delay abort messages */
4322 if (rx_IsClientConn(call->conn))
4325 if (call->abortCode != call->error) {
4326 call->abortCode = call->error;
4327 call->abortCount = 0;
4330 if (force || rxi_callAbortThreshhold == 0
4331 || call->abortCount < rxi_callAbortThreshhold) {
4332 if (call->delayedAbortEvent) {
4333 rxevent_Cancel(call->delayedAbortEvent, call,
4334 RX_CALL_REFCOUNT_ABORT);
4336 error = htonl(call->error);
4339 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4340 (char *)&error, sizeof(error), istack);
4341 } else if (!call->delayedAbortEvent) {
4342 clock_GetTime(&when);
4343 clock_Addmsec(&when, rxi_callAbortDelay);
4344 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4345 call->delayedAbortEvent =
4346 rxevent_Post(&when, rxi_SendDelayedCallAbort, call, 0);
4351 /* Send an abort packet for the specified connection. Packet is an
4352 * optional pointer to a packet that can be used to send the abort.
4353 * Once the number of abort messages reaches the threshhold, an
4354 * event is scheduled to send the abort. Setting the force flag
4355 * overrides sending delayed abort messages.
4357 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4358 * to send the abort packet.
4361 rxi_SendConnectionAbort(register struct rx_connection *conn,
4362 struct rx_packet *packet, int istack, int force)
4370 /* Clients should never delay abort messages */
4371 if (rx_IsClientConn(conn))
4374 if (force || rxi_connAbortThreshhold == 0
4375 || conn->abortCount < rxi_connAbortThreshhold) {
4376 if (conn->delayedAbortEvent) {
4377 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4379 error = htonl(conn->error);
4381 MUTEX_EXIT(&conn->conn_data_lock);
4383 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4384 RX_PACKET_TYPE_ABORT, (char *)&error,
4385 sizeof(error), istack);
4386 MUTEX_ENTER(&conn->conn_data_lock);
4387 } else if (!conn->delayedAbortEvent) {
4388 clock_GetTime(&when);
4389 clock_Addmsec(&when, rxi_connAbortDelay);
4390 conn->delayedAbortEvent =
4391 rxevent_Post(&when, rxi_SendDelayedConnAbort, conn, 0);
4396 /* Associate an error all of the calls owned by a connection. Called
4397 * with error non-zero. This is only for really fatal things, like
4398 * bad authentication responses. The connection itself is set in
4399 * error at this point, so that future packets received will be
4402 rxi_ConnectionError(register struct rx_connection *conn,
4403 register afs_int32 error)
4408 dpf(("rxi_ConnectionError conn %x error %d", conn, error));
4410 MUTEX_ENTER(&conn->conn_data_lock);
4411 if (conn->challengeEvent)
4412 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4413 if (conn->checkReachEvent) {
4414 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4415 conn->checkReachEvent = 0;
4416 conn->flags &= ~RX_CONN_ATTACHWAIT;
4419 MUTEX_EXIT(&conn->conn_data_lock);
4420 for (i = 0; i < RX_MAXCALLS; i++) {
4421 struct rx_call *call = conn->call[i];
4423 MUTEX_ENTER(&call->lock);
4424 rxi_CallError(call, error);
4425 MUTEX_EXIT(&call->lock);
4428 conn->error = error;
4429 MUTEX_ENTER(&rx_stats_mutex);
4430 rx_stats.fatalErrors++;
4431 MUTEX_EXIT(&rx_stats_mutex);
4436 rxi_CallError(register struct rx_call *call, afs_int32 error)
4438 dpf(("rxi_CallError call %x error %d call->error %d", call, error, call->error));
4440 error = call->error;
4442 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4443 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4444 rxi_ResetCall(call, 0);
4447 rxi_ResetCall(call, 0);
4449 call->error = error;
4450 call->mode = RX_MODE_ERROR;
4453 /* Reset various fields in a call structure, and wakeup waiting
4454 * processes. Some fields aren't changed: state & mode are not
4455 * touched (these must be set by the caller), and bufptr, nLeft, and
4456 * nFree are not reset, since these fields are manipulated by
4457 * unprotected macros, and may only be reset by non-interrupting code.
4460 /* this code requires that call->conn be set properly as a pre-condition. */
4461 #endif /* ADAPT_WINDOW */
4464 rxi_ResetCall(register struct rx_call *call, register int newcall)
4467 register struct rx_peer *peer;
4468 struct rx_packet *packet;
4470 /* Notify anyone who is waiting for asynchronous packet arrival */
4471 if (call->arrivalProc) {
4472 (*call->arrivalProc) (call, call->arrivalProcHandle,
4473 call->arrivalProcArg);
4474 call->arrivalProc = (void (*)())0;
4477 if (call->delayedAbortEvent) {
4478 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4479 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4481 rxi_SendCallAbort(call, packet, 0, 1);
4482 rxi_FreePacket(packet);
4487 * Update the peer with the congestion information in this call
4488 * so other calls on this connection can pick up where this call
4489 * left off. If the congestion sequence numbers don't match then
4490 * another call experienced a retransmission.
4492 peer = call->conn->peer;
4493 MUTEX_ENTER(&peer->peer_lock);
4495 if (call->congestSeq == peer->congestSeq) {
4496 peer->cwind = MAX(peer->cwind, call->cwind);
4497 peer->MTU = MAX(peer->MTU, call->MTU);
4498 peer->nDgramPackets =
4499 MAX(peer->nDgramPackets, call->nDgramPackets);
4502 call->abortCode = 0;
4503 call->abortCount = 0;
4505 if (peer->maxDgramPackets > 1) {
4506 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4508 call->MTU = peer->MTU;
4510 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4511 call->ssthresh = rx_maxSendWindow;
4512 call->nDgramPackets = peer->nDgramPackets;
4513 call->congestSeq = peer->congestSeq;
4514 MUTEX_EXIT(&peer->peer_lock);
4516 flags = call->flags;
4517 rxi_ClearReceiveQueue(call);
4518 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4519 if (flags & RX_CALL_TQ_BUSY) {
4520 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4521 call->flags |= (flags & RX_CALL_TQ_WAIT);
4523 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4525 rxi_ClearTransmitQueue(call, 0);
4526 queue_Init(&call->tq);
4527 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4528 dpf(("rcall %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4531 while (call->tqWaiters) {
4532 #ifdef RX_ENABLE_LOCKS
4533 CV_BROADCAST(&call->cv_tq);
4534 #else /* RX_ENABLE_LOCKS */
4535 osi_rxWakeup(&call->tq);
4536 #endif /* RX_ENABLE_LOCKS */
4540 queue_Init(&call->rq);
4542 call->rwind = rx_initReceiveWindow;
4543 call->twind = rx_initSendWindow;
4544 call->nSoftAcked = 0;
4545 call->nextCwind = 0;
4548 call->nCwindAcks = 0;
4549 call->nSoftAcks = 0;
4550 call->nHardAcks = 0;
4552 call->tfirst = call->rnext = call->tnext = 1;
4554 call->lastAcked = 0;
4555 call->localStatus = call->remoteStatus = 0;
4557 if (flags & RX_CALL_READER_WAIT) {
4558 #ifdef RX_ENABLE_LOCKS
4559 CV_BROADCAST(&call->cv_rq);
4561 osi_rxWakeup(&call->rq);
4564 if (flags & RX_CALL_WAIT_PACKETS) {
4565 MUTEX_ENTER(&rx_freePktQ_lock);
4566 rxi_PacketsUnWait(); /* XXX */
4567 MUTEX_EXIT(&rx_freePktQ_lock);
4569 #ifdef RX_ENABLE_LOCKS
4570 CV_SIGNAL(&call->cv_twind);
4572 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4573 osi_rxWakeup(&call->twind);
4576 #ifdef RX_ENABLE_LOCKS
4577 /* The following ensures that we don't mess with any queue while some
4578 * other thread might also be doing so. The call_queue_lock field is
4579 * is only modified under the call lock. If the call is in the process
4580 * of being removed from a queue, the call is not locked until the
4581 * the queue lock is dropped and only then is the call_queue_lock field
4582 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4583 * Note that any other routine which removes a call from a queue has to
4584 * obtain the queue lock before examing the queue and removing the call.
4586 if (call->call_queue_lock) {
4587 MUTEX_ENTER(call->call_queue_lock);
4588 if (queue_IsOnQueue(call)) {
4590 if (flags & RX_CALL_WAIT_PROC) {
4591 MUTEX_ENTER(&rx_stats_mutex);
4593 MUTEX_EXIT(&rx_stats_mutex);
4596 MUTEX_EXIT(call->call_queue_lock);
4597 CLEAR_CALL_QUEUE_LOCK(call);
4599 #else /* RX_ENABLE_LOCKS */
4600 if (queue_IsOnQueue(call)) {
4602 if (flags & RX_CALL_WAIT_PROC)
4605 #endif /* RX_ENABLE_LOCKS */
4607 rxi_KeepAliveOff(call);
4608 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4611 /* Send an acknowledge for the indicated packet (seq,serial) of the
4612 * indicated call, for the indicated reason (reason). This
4613 * acknowledge will specifically acknowledge receiving the packet, and
4614 * will also specify which other packets for this call have been
4615 * received. This routine returns the packet that was used to the
4616 * caller. The caller is responsible for freeing it or re-using it.
4617 * This acknowledgement also returns the highest sequence number
4618 * actually read out by the higher level to the sender; the sender
4619 * promises to keep around packets that have not been read by the
4620 * higher level yet (unless, of course, the sender decides to abort
4621 * the call altogether). Any of p, seq, serial, pflags, or reason may
4622 * be set to zero without ill effect. That is, if they are zero, they
4623 * will not convey any information.
4624 * NOW there is a trailer field, after the ack where it will safely be
4625 * ignored by mundanes, which indicates the maximum size packet this
4626 * host can swallow. */
4628 register struct rx_packet *optionalPacket; use to send ack (or null)
4629 int seq; Sequence number of the packet we are acking
4630 int serial; Serial number of the packet
4631 int pflags; Flags field from packet header
4632 int reason; Reason an acknowledge was prompted
4636 rxi_SendAck(register struct rx_call *call,
4637 register struct rx_packet *optionalPacket, int serial, int reason,
4640 struct rx_ackPacket *ap;
4641 register struct rx_packet *rqp;
4642 register struct rx_packet *nxp; /* For queue_Scan */
4643 register struct rx_packet *p;
4646 #ifdef RX_ENABLE_TSFPQ
4647 struct rx_ts_info_t * rx_ts_info;
4651 * Open the receive window once a thread starts reading packets
4653 if (call->rnext > 1) {
4654 call->rwind = rx_maxReceiveWindow;
4657 call->nHardAcks = 0;
4658 call->nSoftAcks = 0;
4659 if (call->rnext > call->lastAcked)
4660 call->lastAcked = call->rnext;
4664 rx_computelen(p, p->length); /* reset length, you never know */
4665 } /* where that's been... */
4666 #ifdef RX_ENABLE_TSFPQ
4668 RX_TS_INFO_GET(rx_ts_info);
4669 if ((p = rx_ts_info->local_special_packet)) {
4670 rx_computelen(p, p->length);
4671 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4672 rx_ts_info->local_special_packet = p;
4673 } else { /* We won't send the ack, but don't panic. */
4674 return optionalPacket;
4678 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4679 /* We won't send the ack, but don't panic. */
4680 return optionalPacket;
4685 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4688 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4689 #ifndef RX_ENABLE_TSFPQ
4690 if (!optionalPacket)
4693 return optionalPacket;
4695 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4696 if (rx_Contiguous(p) < templ) {
4697 #ifndef RX_ENABLE_TSFPQ
4698 if (!optionalPacket)
4701 return optionalPacket;
4706 /* MTUXXX failing to send an ack is very serious. We should */
4707 /* try as hard as possible to send even a partial ack; it's */
4708 /* better than nothing. */
4709 ap = (struct rx_ackPacket *)rx_DataOf(p);
4710 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4711 ap->reason = reason;
4713 /* The skew computation used to be bogus, I think it's better now. */
4714 /* We should start paying attention to skew. XXX */
4715 ap->serial = htonl(serial);
4716 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4718 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4719 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4721 /* No fear of running out of ack packet here because there can only be at most
4722 * one window full of unacknowledged packets. The window size must be constrained
4723 * to be less than the maximum ack size, of course. Also, an ack should always
4724 * fit into a single packet -- it should not ever be fragmented. */
4725 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4726 if (!rqp || !call->rq.next
4727 || (rqp->header.seq > (call->rnext + call->rwind))) {
4728 #ifndef RX_ENABLE_TSFPQ
4729 if (!optionalPacket)
4732 rxi_CallError(call, RX_CALL_DEAD);
4733 return optionalPacket;
4736 while (rqp->header.seq > call->rnext + offset)
4737 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4738 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4740 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4741 #ifndef RX_ENABLE_TSFPQ
4742 if (!optionalPacket)
4745 rxi_CallError(call, RX_CALL_DEAD);
4746 return optionalPacket;
4751 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4753 /* these are new for AFS 3.3 */
4754 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4755 templ = htonl(templ);
4756 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4757 templ = htonl(call->conn->peer->ifMTU);
4758 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4759 sizeof(afs_int32), &templ);
4761 /* new for AFS 3.4 */
4762 templ = htonl(call->rwind);
4763 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4764 sizeof(afs_int32), &templ);
4766 /* new for AFS 3.5 */
4767 templ = htonl(call->conn->peer->ifDgramPackets);
4768 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4769 sizeof(afs_int32), &templ);
4771 p->header.serviceId = call->conn->serviceId;
4772 p->header.cid = (call->conn->cid | call->channel);
4773 p->header.callNumber = *call->callNumber;
4775 p->header.securityIndex = call->conn->securityIndex;
4776 p->header.epoch = call->conn->epoch;
4777 p->header.type = RX_PACKET_TYPE_ACK;
4778 p->header.flags = RX_SLOW_START_OK;
4779 if (reason == RX_ACK_PING) {
4780 p->header.flags |= RX_REQUEST_ACK;
4782 clock_GetTime(&call->pingRequestTime);
4785 if (call->conn->type == RX_CLIENT_CONNECTION)
4786 p->header.flags |= RX_CLIENT_INITIATED;
4790 if (rxdebug_active) {
4794 len = _snprintf(msg, sizeof(msg),
4795 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4796 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4797 ntohl(ap->serial), ntohl(ap->previousPacket),
4798 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
4799 ap->nAcks, ntohs(ap->bufferSpace) );
4803 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
4804 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4808 OutputDebugString(msg);
4810 #else /* AFS_NT40_ENV */
4812 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
4813 ap->reason, ntohl(ap->previousPacket),
4814 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4816 for (offset = 0; offset < ap->nAcks; offset++)
4817 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4822 #endif /* AFS_NT40_ENV */
4825 register int i, nbytes = p->length;
4827 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4828 if (nbytes <= p->wirevec[i].iov_len) {
4829 register int savelen, saven;
4831 savelen = p->wirevec[i].iov_len;
4833 p->wirevec[i].iov_len = nbytes;
4835 rxi_Send(call, p, istack);
4836 p->wirevec[i].iov_len = savelen;
4840 nbytes -= p->wirevec[i].iov_len;
4843 MUTEX_ENTER(&rx_stats_mutex);
4844 rx_stats.ackPacketsSent++;
4845 MUTEX_EXIT(&rx_stats_mutex);
4846 #ifndef RX_ENABLE_TSFPQ
4847 if (!optionalPacket)
4850 return optionalPacket; /* Return packet for re-use by caller */
4853 /* Send all of the packets in the list in single datagram */
4855 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4856 int istack, int moreFlag, struct clock *now,
4857 struct clock *retryTime, int resending)
4862 struct rx_connection *conn = call->conn;
4863 struct rx_peer *peer = conn->peer;
4865 MUTEX_ENTER(&peer->peer_lock);
4868 peer->reSends += len;
4869 MUTEX_ENTER(&rx_stats_mutex);
4870 rx_stats.dataPacketsSent += len;
4871 MUTEX_EXIT(&rx_stats_mutex);
4872 MUTEX_EXIT(&peer->peer_lock);
4874 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4878 /* Set the packet flags and schedule the resend events */
4879 /* Only request an ack for the last packet in the list */
4880 for (i = 0; i < len; i++) {
4881 list[i]->retryTime = *retryTime;
4882 if (list[i]->header.serial) {
4883 /* Exponentially backoff retry times */
4884 if (list[i]->backoff < MAXBACKOFF) {
4885 /* so it can't stay == 0 */
4886 list[i]->backoff = (list[i]->backoff << 1) + 1;
4889 clock_Addmsec(&(list[i]->retryTime),
4890 ((afs_uint32) list[i]->backoff) << 8);
4893 /* Wait a little extra for the ack on the last packet */
4894 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4895 clock_Addmsec(&(list[i]->retryTime), 400);
4898 /* Record the time sent */
4899 list[i]->timeSent = *now;
4901 /* Ask for an ack on retransmitted packets, on every other packet
4902 * if the peer doesn't support slow start. Ask for an ack on every
4903 * packet until the congestion window reaches the ack rate. */
4904 if (list[i]->header.serial) {
4906 MUTEX_ENTER(&rx_stats_mutex);
4907 rx_stats.dataPacketsReSent++;
4908 MUTEX_EXIT(&rx_stats_mutex);
4910 /* improved RTO calculation- not Karn */
4911 list[i]->firstSent = *now;
4912 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4913 || (!(call->flags & RX_CALL_SLOW_START_OK)
4914 && (list[i]->header.seq & 1)))) {
4919 MUTEX_ENTER(&peer->peer_lock);
4923 MUTEX_ENTER(&rx_stats_mutex);
4924 rx_stats.dataPacketsSent++;
4925 MUTEX_EXIT(&rx_stats_mutex);
4926 MUTEX_EXIT(&peer->peer_lock);
4928 /* Tag this packet as not being the last in this group,
4929 * for the receiver's benefit */
4930 if (i < len - 1 || moreFlag) {
4931 list[i]->header.flags |= RX_MORE_PACKETS;
4934 /* Install the new retransmit time for the packet, and
4935 * record the time sent */
4936 list[i]->timeSent = *now;
4940 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4943 /* Since we're about to send a data packet to the peer, it's
4944 * safe to nuke any scheduled end-of-packets ack */
4945 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4947 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4948 MUTEX_EXIT(&call->lock);
4950 rxi_SendPacketList(call, conn, list, len, istack);
4952 rxi_SendPacket(call, conn, list[0], istack);
4954 MUTEX_ENTER(&call->lock);
4955 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4957 /* Update last send time for this call (for keep-alive
4958 * processing), and for the connection (so that we can discover
4959 * idle connections) */
4960 conn->lastSendTime = call->lastSendTime = clock_Sec();
4963 /* When sending packets we need to follow these rules:
4964 * 1. Never send more than maxDgramPackets in a jumbogram.
4965 * 2. Never send a packet with more than two iovecs in a jumbogram.
4966 * 3. Never send a retransmitted packet in a jumbogram.
4967 * 4. Never send more than cwind/4 packets in a jumbogram
4968 * We always keep the last list we should have sent so we
4969 * can set the RX_MORE_PACKETS flags correctly.
4972 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4973 int istack, struct clock *now, struct clock *retryTime,
4976 int i, cnt, lastCnt = 0;
4977 struct rx_packet **listP, **lastP = 0;
4978 struct rx_peer *peer = call->conn->peer;
4979 int morePackets = 0;
4981 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4982 /* Does the current packet force us to flush the current list? */
4984 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4985 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4987 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
4989 /* If the call enters an error state stop sending, or if
4990 * we entered congestion recovery mode, stop sending */
4991 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4999 /* Add the current packet to the list if it hasn't been acked.
5000 * Otherwise adjust the list pointer to skip the current packet. */
5001 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5003 /* Do we need to flush the list? */
5004 if (cnt >= (int)peer->maxDgramPackets
5005 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5006 || list[i]->header.serial
5007 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5009 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5010 retryTime, resending);
5011 /* If the call enters an error state stop sending, or if
5012 * we entered congestion recovery mode, stop sending */
5014 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5019 listP = &list[i + 1];
5024 osi_Panic("rxi_SendList error");
5026 listP = &list[i + 1];
5030 /* Send the whole list when the call is in receive mode, when
5031 * the call is in eof mode, when we are in fast recovery mode,
5032 * and when we have the last packet */
5033 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5034 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5035 || (call->flags & RX_CALL_FAST_RECOVER)) {
5036 /* Check for the case where the current list contains
5037 * an acked packet. Since we always send retransmissions
5038 * in a separate packet, we only need to check the first
5039 * packet in the list */
5040 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5044 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5045 retryTime, resending);
5046 /* If the call enters an error state stop sending, or if
5047 * we entered congestion recovery mode, stop sending */
5048 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5052 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5055 } else if (lastCnt > 0) {
5056 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5061 #ifdef RX_ENABLE_LOCKS
5062 /* Call rxi_Start, below, but with the call lock held. */
5064 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
5065 void *arg1, int istack)
5067 MUTEX_ENTER(&call->lock);
5068 rxi_Start(event, call, arg1, istack);
5069 MUTEX_EXIT(&call->lock);
5071 #endif /* RX_ENABLE_LOCKS */
5073 /* This routine is called when new packets are readied for
5074 * transmission and when retransmission may be necessary, or when the
5075 * transmission window or burst count are favourable. This should be
5076 * better optimized for new packets, the usual case, now that we've
5077 * got rid of queues of send packets. XXXXXXXXXXX */
5079 rxi_Start(struct rxevent *event, register struct rx_call *call,
5080 void *arg1, int istack)
5082 struct rx_packet *p;
5083 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
5084 struct rx_peer *peer = call->conn->peer;
5085 struct clock now, retryTime;
5089 struct rx_packet **xmitList;
5092 /* If rxi_Start is being called as a result of a resend event,
5093 * then make sure that the event pointer is removed from the call
5094 * structure, since there is no longer a per-call retransmission
5096 if (event && event == call->resendEvent) {
5097 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5098 call->resendEvent = NULL;
5100 if (queue_IsEmpty(&call->tq)) {
5104 /* Timeouts trigger congestion recovery */
5105 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5106 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5107 /* someone else is waiting to start recovery */
5110 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5111 while (call->flags & RX_CALL_TQ_BUSY) {
5112 call->flags |= RX_CALL_TQ_WAIT;
5114 #ifdef RX_ENABLE_LOCKS
5115 osirx_AssertMine(&call->lock, "rxi_Start lock1");
5116 CV_WAIT(&call->cv_tq, &call->lock);
5117 #else /* RX_ENABLE_LOCKS */
5118 osi_rxSleep(&call->tq);
5119 #endif /* RX_ENABLE_LOCKS */
5121 if (call->tqWaiters == 0)
5122 call->flags &= ~RX_CALL_TQ_WAIT;
5124 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5125 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5126 call->flags |= RX_CALL_FAST_RECOVER;
5127 if (peer->maxDgramPackets > 1) {
5128 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5130 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5132 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5133 call->nDgramPackets = 1;
5135 call->nextCwind = 1;
5138 MUTEX_ENTER(&peer->peer_lock);
5139 peer->MTU = call->MTU;
5140 peer->cwind = call->cwind;
5141 peer->nDgramPackets = 1;
5143 call->congestSeq = peer->congestSeq;
5144 MUTEX_EXIT(&peer->peer_lock);
5145 /* Clear retry times on packets. Otherwise, it's possible for
5146 * some packets in the queue to force resends at rates faster
5147 * than recovery rates.
5149 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5150 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5151 clock_Zero(&p->retryTime);
5156 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5157 MUTEX_ENTER(&rx_stats_mutex);
5158 rx_tq_debug.rxi_start_in_error++;
5159 MUTEX_EXIT(&rx_stats_mutex);
5164 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5165 /* Get clock to compute the re-transmit time for any packets
5166 * in this burst. Note, if we back off, it's reasonable to
5167 * back off all of the packets in the same manner, even if
5168 * some of them have been retransmitted more times than more
5169 * recent additions */
5170 clock_GetTime(&now);
5171 retryTime = now; /* initialize before use */
5172 MUTEX_ENTER(&peer->peer_lock);
5173 clock_Add(&retryTime, &peer->timeout);
5174 MUTEX_EXIT(&peer->peer_lock);
5176 /* Send (or resend) any packets that need it, subject to
5177 * window restrictions and congestion burst control
5178 * restrictions. Ask for an ack on the last packet sent in
5179 * this burst. For now, we're relying upon the window being
5180 * considerably bigger than the largest number of packets that
5181 * are typically sent at once by one initial call to
5182 * rxi_Start. This is probably bogus (perhaps we should ask
5183 * for an ack when we're half way through the current
5184 * window?). Also, for non file transfer applications, this
5185 * may end up asking for an ack for every packet. Bogus. XXXX
5188 * But check whether we're here recursively, and let the other guy
5191 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5192 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5193 call->flags |= RX_CALL_TQ_BUSY;
5195 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5197 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5198 call->flags &= ~RX_CALL_NEED_START;
5199 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5201 maxXmitPackets = MIN(call->twind, call->cwind);
5202 xmitList = (struct rx_packet **)
5203 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5204 if (xmitList == NULL)
5205 osi_Panic("rxi_Start, failed to allocate xmit list");
5206 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5207 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5208 /* We shouldn't be sending packets if a thread is waiting
5209 * to initiate congestion recovery */
5213 && (call->flags & RX_CALL_FAST_RECOVER)) {
5214 /* Only send one packet during fast recovery */
5217 if ((p->flags & RX_PKTFLAG_FREE)
5218 || (!queue_IsEnd(&call->tq, nxp)
5219 && (nxp->flags & RX_PKTFLAG_FREE))
5220 || (p == (struct rx_packet *)&rx_freePacketQueue)
5221 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5222 osi_Panic("rxi_Start: xmit queue clobbered");
5224 if (p->flags & RX_PKTFLAG_ACKED) {
5225 MUTEX_ENTER(&rx_stats_mutex);
5226 rx_stats.ignoreAckedPacket++;
5227 MUTEX_EXIT(&rx_stats_mutex);
5228 continue; /* Ignore this packet if it has been acknowledged */
5231 /* Turn off all flags except these ones, which are the same
5232 * on each transmission */
5233 p->header.flags &= RX_PRESET_FLAGS;
5235 if (p->header.seq >=
5236 call->tfirst + MIN((int)call->twind,
5237 (int)(call->nSoftAcked +
5239 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5240 /* Note: if we're waiting for more window space, we can
5241 * still send retransmits; hence we don't return here, but
5242 * break out to schedule a retransmit event */
5243 dpf(("call %d waiting for window",
5244 *(call->callNumber)));
5248 /* Transmit the packet if it needs to be sent. */
5249 if (!clock_Lt(&now, &p->retryTime)) {
5250 if (nXmitPackets == maxXmitPackets) {
5251 rxi_SendXmitList(call, xmitList, nXmitPackets,
5252 istack, &now, &retryTime,
5254 osi_Free(xmitList, maxXmitPackets *
5255 sizeof(struct rx_packet *));
5258 xmitList[nXmitPackets++] = p;
5262 /* xmitList now hold pointers to all of the packets that are
5263 * ready to send. Now we loop to send the packets */
5264 if (nXmitPackets > 0) {
5265 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5266 &now, &retryTime, resending);
5269 maxXmitPackets * sizeof(struct rx_packet *));
5271 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5273 * TQ references no longer protected by this flag; they must remain
5274 * protected by the global lock.
5276 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5277 call->flags &= ~RX_CALL_TQ_BUSY;
5278 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5279 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5280 #ifdef RX_ENABLE_LOCKS
5281 osirx_AssertMine(&call->lock, "rxi_Start start");
5282 CV_BROADCAST(&call->cv_tq);
5283 #else /* RX_ENABLE_LOCKS */
5284 osi_rxWakeup(&call->tq);
5285 #endif /* RX_ENABLE_LOCKS */
5290 /* We went into the error state while sending packets. Now is
5291 * the time to reset the call. This will also inform the using
5292 * process that the call is in an error state.
5294 MUTEX_ENTER(&rx_stats_mutex);
5295 rx_tq_debug.rxi_start_aborted++;
5296 MUTEX_EXIT(&rx_stats_mutex);
5297 call->flags &= ~RX_CALL_TQ_BUSY;
5298 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5299 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5300 #ifdef RX_ENABLE_LOCKS
5301 osirx_AssertMine(&call->lock, "rxi_Start middle");
5302 CV_BROADCAST(&call->cv_tq);
5303 #else /* RX_ENABLE_LOCKS */
5304 osi_rxWakeup(&call->tq);
5305 #endif /* RX_ENABLE_LOCKS */
5307 rxi_CallError(call, call->error);
5310 #ifdef RX_ENABLE_LOCKS
5311 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5312 register int missing;
5313 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5314 /* Some packets have received acks. If they all have, we can clear
5315 * the transmit queue.
5318 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5319 if (p->header.seq < call->tfirst
5320 && (p->flags & RX_PKTFLAG_ACKED)) {
5327 call->flags |= RX_CALL_TQ_CLEARME;
5329 #endif /* RX_ENABLE_LOCKS */
5330 /* Don't bother doing retransmits if the TQ is cleared. */
5331 if (call->flags & RX_CALL_TQ_CLEARME) {
5332 rxi_ClearTransmitQueue(call, 1);
5334 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5337 /* Always post a resend event, if there is anything in the
5338 * queue, and resend is possible. There should be at least
5339 * one unacknowledged packet in the queue ... otherwise none
5340 * of these packets should be on the queue in the first place.
5342 if (call->resendEvent) {
5343 /* Cancel the existing event and post a new one */
5344 rxevent_Cancel(call->resendEvent, call,
5345 RX_CALL_REFCOUNT_RESEND);
5348 /* The retry time is the retry time on the first unacknowledged
5349 * packet inside the current window */
5351 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5352 /* Don't set timers for packets outside the window */
5353 if (p->header.seq >= call->tfirst + call->twind) {
5357 if (!(p->flags & RX_PKTFLAG_ACKED)
5358 && !clock_IsZero(&p->retryTime)) {
5360 retryTime = p->retryTime;
5365 /* Post a new event to re-run rxi_Start when retries may be needed */
5366 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5367 #ifdef RX_ENABLE_LOCKS
5368 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5370 rxevent_Post2(&retryTime, rxi_StartUnlocked,
5371 (void *)call, 0, istack);
5372 #else /* RX_ENABLE_LOCKS */
5374 rxevent_Post2(&retryTime, rxi_Start, (void *)call,
5376 #endif /* RX_ENABLE_LOCKS */
5379 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5380 } while (call->flags & RX_CALL_NEED_START);
5382 * TQ references no longer protected by this flag; they must remain
5383 * protected by the global lock.
5385 call->flags &= ~RX_CALL_TQ_BUSY;
5386 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5387 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5388 #ifdef RX_ENABLE_LOCKS
5389 osirx_AssertMine(&call->lock, "rxi_Start end");
5390 CV_BROADCAST(&call->cv_tq);
5391 #else /* RX_ENABLE_LOCKS */
5392 osi_rxWakeup(&call->tq);
5393 #endif /* RX_ENABLE_LOCKS */
5396 call->flags |= RX_CALL_NEED_START;
5398 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5400 if (call->resendEvent) {
5401 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5406 /* Also adjusts the keep alive parameters for the call, to reflect
5407 * that we have just sent a packet (so keep alives aren't sent
5410 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5413 register struct rx_connection *conn = call->conn;
5415 /* Stamp each packet with the user supplied status */
5416 p->header.userStatus = call->localStatus;
5418 /* Allow the security object controlling this call's security to
5419 * make any last-minute changes to the packet */
5420 RXS_SendPacket(conn->securityObject, call, p);
5422 /* Since we're about to send SOME sort of packet to the peer, it's
5423 * safe to nuke any scheduled end-of-packets ack */
5424 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5426 /* Actually send the packet, filling in more connection-specific fields */
5427 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5428 MUTEX_EXIT(&call->lock);
5429 rxi_SendPacket(call, conn, p, istack);
5430 MUTEX_ENTER(&call->lock);
5431 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5433 /* Update last send time for this call (for keep-alive
5434 * processing), and for the connection (so that we can discover
5435 * idle connections) */
5436 conn->lastSendTime = call->lastSendTime = clock_Sec();
5440 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5441 * that things are fine. Also called periodically to guarantee that nothing
5442 * falls through the cracks (e.g. (error + dally) connections have keepalive
5443 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5445 * haveCTLock Set if calling from rxi_ReapConnections
5447 #ifdef RX_ENABLE_LOCKS
5449 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5450 #else /* RX_ENABLE_LOCKS */
5452 rxi_CheckCall(register struct rx_call *call)
5453 #endif /* RX_ENABLE_LOCKS */
5455 register struct rx_connection *conn = call->conn;
5457 afs_uint32 deadTime;
5459 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5460 if (call->flags & RX_CALL_TQ_BUSY) {
5461 /* Call is active and will be reset by rxi_Start if it's
5462 * in an error state.
5467 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5469 (((afs_uint32) conn->secondsUntilDead << 10) +
5470 ((afs_uint32) conn->peer->rtt >> 3) +
5471 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5473 /* These are computed to the second (+- 1 second). But that's
5474 * good enough for these values, which should be a significant
5475 * number of seconds. */
5476 if (now > (call->lastReceiveTime + deadTime)) {
5477 if (call->state == RX_STATE_ACTIVE) {
5478 rxi_CallError(call, RX_CALL_DEAD);
5481 #ifdef RX_ENABLE_LOCKS
5482 /* Cancel pending events */
5483 rxevent_Cancel(call->delayedAckEvent, call,
5484 RX_CALL_REFCOUNT_DELAY);
5485 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5486 rxevent_Cancel(call->keepAliveEvent, call,
5487 RX_CALL_REFCOUNT_ALIVE);
5488 if (call->refCount == 0) {
5489 rxi_FreeCall(call, haveCTLock);
5493 #else /* RX_ENABLE_LOCKS */
5496 #endif /* RX_ENABLE_LOCKS */
5498 /* Non-active calls are destroyed if they are not responding
5499 * to pings; active calls are simply flagged in error, so the
5500 * attached process can die reasonably gracefully. */
5502 /* see if we have a non-activity timeout */
5503 if (call->startWait && conn->idleDeadTime
5504 && ((call->startWait + conn->idleDeadTime) < now)) {
5505 if (call->state == RX_STATE_ACTIVE) {
5506 rxi_CallError(call, RX_CALL_TIMEOUT);
5510 /* see if we have a hard timeout */
5511 if (conn->hardDeadTime
5512 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5513 if (call->state == RX_STATE_ACTIVE)
5514 rxi_CallError(call, RX_CALL_TIMEOUT);
5521 /* When a call is in progress, this routine is called occasionally to
5522 * make sure that some traffic has arrived (or been sent to) the peer.
5523 * If nothing has arrived in a reasonable amount of time, the call is
5524 * declared dead; if nothing has been sent for a while, we send a
5525 * keep-alive packet (if we're actually trying to keep the call alive)
5528 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5531 struct rx_connection *conn;
5534 MUTEX_ENTER(&call->lock);
5535 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5536 if (event == call->keepAliveEvent)
5537 call->keepAliveEvent = NULL;
5540 #ifdef RX_ENABLE_LOCKS
5541 if (rxi_CheckCall(call, 0)) {
5542 MUTEX_EXIT(&call->lock);
5545 #else /* RX_ENABLE_LOCKS */
5546 if (rxi_CheckCall(call))
5548 #endif /* RX_ENABLE_LOCKS */
5550 /* Don't try to keep alive dallying calls */
5551 if (call->state == RX_STATE_DALLY) {
5552 MUTEX_EXIT(&call->lock);
5557 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5558 /* Don't try to send keepalives if there is unacknowledged data */
5559 /* the rexmit code should be good enough, this little hack
5560 * doesn't quite work XXX */
5561 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5563 rxi_ScheduleKeepAliveEvent(call);
5564 MUTEX_EXIT(&call->lock);
5569 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5571 if (!call->keepAliveEvent) {
5573 clock_GetTime(&when);
5574 when.sec += call->conn->secondsUntilPing;
5575 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5576 call->keepAliveEvent =
5577 rxevent_Post(&when, rxi_KeepAliveEvent, call, 0);
5581 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5583 rxi_KeepAliveOn(register struct rx_call *call)
5585 /* Pretend last packet received was received now--i.e. if another
5586 * packet isn't received within the keep alive time, then the call
5587 * will die; Initialize last send time to the current time--even
5588 * if a packet hasn't been sent yet. This will guarantee that a
5589 * keep-alive is sent within the ping time */
5590 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5591 rxi_ScheduleKeepAliveEvent(call);
5594 /* This routine is called to send connection abort messages
5595 * that have been delayed to throttle looping clients. */
5597 rxi_SendDelayedConnAbort(struct rxevent *event,
5598 register struct rx_connection *conn, char *dummy)
5601 struct rx_packet *packet;
5603 MUTEX_ENTER(&conn->conn_data_lock);
5604 conn->delayedAbortEvent = NULL;
5605 error = htonl(conn->error);
5607 MUTEX_EXIT(&conn->conn_data_lock);
5608 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5611 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5612 RX_PACKET_TYPE_ABORT, (char *)&error,
5614 rxi_FreePacket(packet);
5618 /* This routine is called to send call abort messages
5619 * that have been delayed to throttle looping clients. */
5621 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5625 struct rx_packet *packet;
5627 MUTEX_ENTER(&call->lock);
5628 call->delayedAbortEvent = NULL;
5629 error = htonl(call->error);
5631 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5634 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5635 (char *)&error, sizeof(error), 0);
5636 rxi_FreePacket(packet);
5638 MUTEX_EXIT(&call->lock);
5641 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5642 * seconds) to ask the client to authenticate itself. The routine
5643 * issues a challenge to the client, which is obtained from the
5644 * security object associated with the connection */
5646 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5647 void *arg1, int tries)
5649 conn->challengeEvent = NULL;
5650 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5651 register struct rx_packet *packet;
5655 /* We've failed to authenticate for too long.
5656 * Reset any calls waiting for authentication;
5657 * they are all in RX_STATE_PRECALL.
5661 MUTEX_ENTER(&conn->conn_call_lock);
5662 for (i = 0; i < RX_MAXCALLS; i++) {
5663 struct rx_call *call = conn->call[i];
5665 MUTEX_ENTER(&call->lock);
5666 if (call->state == RX_STATE_PRECALL) {
5667 rxi_CallError(call, RX_CALL_DEAD);
5668 rxi_SendCallAbort(call, NULL, 0, 0);
5670 MUTEX_EXIT(&call->lock);
5673 MUTEX_EXIT(&conn->conn_call_lock);
5677 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5679 /* If there's no packet available, do this later. */
5680 RXS_GetChallenge(conn->securityObject, conn, packet);
5681 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5682 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5683 rxi_FreePacket(packet);
5685 clock_GetTime(&when);
5686 when.sec += RX_CHALLENGE_TIMEOUT;
5687 conn->challengeEvent =
5688 rxevent_Post2(&when, rxi_ChallengeEvent, conn, 0,
5693 /* Call this routine to start requesting the client to authenticate
5694 * itself. This will continue until authentication is established,
5695 * the call times out, or an invalid response is returned. The
5696 * security object associated with the connection is asked to create
5697 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5698 * defined earlier. */
5700 rxi_ChallengeOn(register struct rx_connection *conn)
5702 if (!conn->challengeEvent) {
5703 RXS_CreateChallenge(conn->securityObject, conn);
5704 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5709 /* Compute round trip time of the packet provided, in *rttp.
5712 /* rxi_ComputeRoundTripTime is called with peer locked. */
5713 /* sentp and/or peer may be null */
5715 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5716 register struct clock *sentp,
5717 register struct rx_peer *peer)
5719 struct clock thisRtt, *rttp = &thisRtt;
5721 register int rtt_timeout;
5723 clock_GetTime(rttp);
5725 if (clock_Lt(rttp, sentp)) {
5727 return; /* somebody set the clock back, don't count this time. */
5729 clock_Sub(rttp, sentp);
5730 MUTEX_ENTER(&rx_stats_mutex);
5731 if (clock_Lt(rttp, &rx_stats.minRtt))
5732 rx_stats.minRtt = *rttp;
5733 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5734 if (rttp->sec > 60) {
5735 MUTEX_EXIT(&rx_stats_mutex);
5736 return; /* somebody set the clock ahead */
5738 rx_stats.maxRtt = *rttp;
5740 clock_Add(&rx_stats.totalRtt, rttp);
5741 rx_stats.nRttSamples++;
5742 MUTEX_EXIT(&rx_stats_mutex);
5744 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5746 /* Apply VanJacobson round-trip estimations */
5751 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5752 * srtt is stored as fixed point with 3 bits after the binary
5753 * point (i.e., scaled by 8). The following magic is
5754 * equivalent to the smoothing algorithm in rfc793 with an
5755 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5756 * srtt*8 = srtt*8 + rtt - srtt
5757 * srtt = srtt + rtt/8 - srtt/8
5760 delta = MSEC(rttp) - (peer->rtt >> 3);
5764 * We accumulate a smoothed rtt variance (actually, a smoothed
5765 * mean difference), then set the retransmit timer to smoothed
5766 * rtt + 4 times the smoothed variance (was 2x in van's original
5767 * paper, but 4x works better for me, and apparently for him as
5769 * rttvar is stored as
5770 * fixed point with 2 bits after the binary point (scaled by
5771 * 4). The following is equivalent to rfc793 smoothing with
5772 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5773 * replaces rfc793's wired-in beta.
5774 * dev*4 = dev*4 + (|actual - expected| - dev)
5780 delta -= (peer->rtt_dev >> 2);
5781 peer->rtt_dev += delta;
5783 /* I don't have a stored RTT so I start with this value. Since I'm
5784 * probably just starting a call, and will be pushing more data down
5785 * this, I expect congestion to increase rapidly. So I fudge a
5786 * little, and I set deviance to half the rtt. In practice,
5787 * deviance tends to approach something a little less than
5788 * half the smoothed rtt. */
5789 peer->rtt = (MSEC(rttp) << 3) + 8;
5790 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5792 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5793 * the other of these connections is usually in a user process, and can
5794 * be switched and/or swapped out. So on fast, reliable networks, the
5795 * timeout would otherwise be too short.
5797 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5798 clock_Zero(&(peer->timeout));
5799 clock_Addmsec(&(peer->timeout), rtt_timeout);
5801 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)));
5805 /* Find all server connections that have not been active for a long time, and
5808 rxi_ReapConnections(void)
5811 clock_GetTime(&now);
5813 /* Find server connection structures that haven't been used for
5814 * greater than rx_idleConnectionTime */
5816 struct rx_connection **conn_ptr, **conn_end;
5817 int i, havecalls = 0;
5818 MUTEX_ENTER(&rx_connHashTable_lock);
5819 for (conn_ptr = &rx_connHashTable[0], conn_end =
5820 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5822 struct rx_connection *conn, *next;
5823 struct rx_call *call;
5827 for (conn = *conn_ptr; conn; conn = next) {
5828 /* XXX -- Shouldn't the connection be locked? */
5831 for (i = 0; i < RX_MAXCALLS; i++) {
5832 call = conn->call[i];
5835 MUTEX_ENTER(&call->lock);
5836 #ifdef RX_ENABLE_LOCKS
5837 result = rxi_CheckCall(call, 1);
5838 #else /* RX_ENABLE_LOCKS */
5839 result = rxi_CheckCall(call);
5840 #endif /* RX_ENABLE_LOCKS */
5841 MUTEX_EXIT(&call->lock);
5843 /* If CheckCall freed the call, it might
5844 * have destroyed the connection as well,
5845 * which screws up the linked lists.
5851 if (conn->type == RX_SERVER_CONNECTION) {
5852 /* This only actually destroys the connection if
5853 * there are no outstanding calls */
5854 MUTEX_ENTER(&conn->conn_data_lock);
5855 if (!havecalls && !conn->refCount
5856 && ((conn->lastSendTime + rx_idleConnectionTime) <
5858 conn->refCount++; /* it will be decr in rx_DestroyConn */
5859 MUTEX_EXIT(&conn->conn_data_lock);
5860 #ifdef RX_ENABLE_LOCKS
5861 rxi_DestroyConnectionNoLock(conn);
5862 #else /* RX_ENABLE_LOCKS */
5863 rxi_DestroyConnection(conn);
5864 #endif /* RX_ENABLE_LOCKS */
5866 #ifdef RX_ENABLE_LOCKS
5868 MUTEX_EXIT(&conn->conn_data_lock);
5870 #endif /* RX_ENABLE_LOCKS */
5874 #ifdef RX_ENABLE_LOCKS
5875 while (rx_connCleanup_list) {
5876 struct rx_connection *conn;
5877 conn = rx_connCleanup_list;
5878 rx_connCleanup_list = rx_connCleanup_list->next;
5879 MUTEX_EXIT(&rx_connHashTable_lock);
5880 rxi_CleanupConnection(conn);
5881 MUTEX_ENTER(&rx_connHashTable_lock);
5883 MUTEX_EXIT(&rx_connHashTable_lock);
5884 #endif /* RX_ENABLE_LOCKS */
5887 /* Find any peer structures that haven't been used (haven't had an
5888 * associated connection) for greater than rx_idlePeerTime */
5890 struct rx_peer **peer_ptr, **peer_end;
5892 MUTEX_ENTER(&rx_rpc_stats);
5893 MUTEX_ENTER(&rx_peerHashTable_lock);
5894 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5895 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5897 struct rx_peer *peer, *next, *prev;
5898 for (prev = peer = *peer_ptr; peer; peer = next) {
5900 code = MUTEX_TRYENTER(&peer->peer_lock);
5901 if ((code) && (peer->refCount == 0)
5902 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5903 rx_interface_stat_p rpc_stat, nrpc_stat;
5905 MUTEX_EXIT(&peer->peer_lock);
5906 MUTEX_DESTROY(&peer->peer_lock);
5908 (&peer->rpcStats, rpc_stat, nrpc_stat,
5909 rx_interface_stat)) {
5910 unsigned int num_funcs;
5913 queue_Remove(&rpc_stat->queue_header);
5914 queue_Remove(&rpc_stat->all_peers);
5915 num_funcs = rpc_stat->stats[0].func_total;
5917 sizeof(rx_interface_stat_t) +
5918 rpc_stat->stats[0].func_total *
5919 sizeof(rx_function_entry_v1_t);
5921 rxi_Free(rpc_stat, space);
5922 rxi_rpc_peer_stat_cnt -= num_funcs;
5925 MUTEX_ENTER(&rx_stats_mutex);
5926 rx_stats.nPeerStructs--;
5927 MUTEX_EXIT(&rx_stats_mutex);
5928 if (peer == *peer_ptr) {
5935 MUTEX_EXIT(&peer->peer_lock);
5941 MUTEX_EXIT(&rx_peerHashTable_lock);
5942 MUTEX_EXIT(&rx_rpc_stats);
5945 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5946 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5947 * GC, just below. Really, we shouldn't have to keep moving packets from
5948 * one place to another, but instead ought to always know if we can
5949 * afford to hold onto a packet in its particular use. */
5950 MUTEX_ENTER(&rx_freePktQ_lock);
5951 if (rx_waitingForPackets) {
5952 rx_waitingForPackets = 0;
5953 #ifdef RX_ENABLE_LOCKS
5954 CV_BROADCAST(&rx_waitingForPackets_cv);
5956 osi_rxWakeup(&rx_waitingForPackets);
5959 MUTEX_EXIT(&rx_freePktQ_lock);
5961 now.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5962 rxevent_Post(&now, rxi_ReapConnections, 0, 0);
5966 /* rxs_Release - This isn't strictly necessary but, since the macro name from
5967 * rx.h is sort of strange this is better. This is called with a security
5968 * object before it is discarded. Each connection using a security object has
5969 * its own refcount to the object so it won't actually be freed until the last
5970 * connection is destroyed.
5972 * This is the only rxs module call. A hold could also be written but no one
5976 rxs_Release(struct rx_securityClass *aobj)
5978 return RXS_Close(aobj);
5982 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
5983 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
5984 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
5985 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
5987 /* Adjust our estimate of the transmission rate to this peer, given
5988 * that the packet p was just acked. We can adjust peer->timeout and
5989 * call->twind. Pragmatically, this is called
5990 * only with packets of maximal length.
5991 * Called with peer and call locked.
5995 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
5996 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
5998 afs_int32 xferSize, xferMs;
5999 register afs_int32 minTime;
6002 /* Count down packets */
6003 if (peer->rateFlag > 0)
6005 /* Do nothing until we're enabled */
6006 if (peer->rateFlag != 0)
6011 /* Count only when the ack seems legitimate */
6012 switch (ackReason) {
6013 case RX_ACK_REQUESTED:
6015 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6019 case RX_ACK_PING_RESPONSE:
6020 if (p) /* want the response to ping-request, not data send */
6022 clock_GetTime(&newTO);
6023 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6024 clock_Sub(&newTO, &call->pingRequestTime);
6025 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6029 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6036 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));
6038 /* Track only packets that are big enough. */
6039 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6043 /* absorb RTT data (in milliseconds) for these big packets */
6044 if (peer->smRtt == 0) {
6045 peer->smRtt = xferMs;
6047 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6052 if (peer->countDown) {
6056 peer->countDown = 10; /* recalculate only every so often */
6058 /* In practice, we can measure only the RTT for full packets,
6059 * because of the way Rx acks the data that it receives. (If it's
6060 * smaller than a full packet, it often gets implicitly acked
6061 * either by the call response (from a server) or by the next call
6062 * (from a client), and either case confuses transmission times
6063 * with processing times.) Therefore, replace the above
6064 * more-sophisticated processing with a simpler version, where the
6065 * smoothed RTT is kept for full-size packets, and the time to
6066 * transmit a windowful of full-size packets is simply RTT *
6067 * windowSize. Again, we take two steps:
6068 - ensure the timeout is large enough for a single packet's RTT;
6069 - ensure that the window is small enough to fit in the desired timeout.*/
6071 /* First, the timeout check. */
6072 minTime = peer->smRtt;
6073 /* Get a reasonable estimate for a timeout period */
6075 newTO.sec = minTime / 1000;
6076 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6078 /* Increase the timeout period so that we can always do at least
6079 * one packet exchange */
6080 if (clock_Gt(&newTO, &peer->timeout)) {
6082 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));
6084 peer->timeout = newTO;
6087 /* Now, get an estimate for the transmit window size. */
6088 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6089 /* Now, convert to the number of full packets that could fit in a
6090 * reasonable fraction of that interval */
6091 minTime /= (peer->smRtt << 1);
6092 xferSize = minTime; /* (make a copy) */
6094 /* Now clamp the size to reasonable bounds. */
6097 else if (minTime > rx_Window)
6098 minTime = rx_Window;
6099 /* if (minTime != peer->maxWindow) {
6100 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6101 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6102 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6104 peer->maxWindow = minTime;
6105 elide... call->twind = minTime;
6109 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6110 * Discern this by calculating the timeout necessary for rx_Window
6112 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6113 /* calculate estimate for transmission interval in milliseconds */
6114 minTime = rx_Window * peer->smRtt;
6115 if (minTime < 1000) {
6116 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
6117 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6118 peer->timeout.usec, peer->smRtt, peer->packetSize));
6120 newTO.sec = 0; /* cut back on timeout by half a second */
6121 newTO.usec = 500000;
6122 clock_Sub(&peer->timeout, &newTO);
6127 } /* end of rxi_ComputeRate */
6128 #endif /* ADAPT_WINDOW */
6136 #define TRACE_OPTION_DEBUGLOG 4
6144 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6145 0, KEY_QUERY_VALUE, &parmKey);
6146 if (code != ERROR_SUCCESS)
6149 dummyLen = sizeof(TraceOption);
6150 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6151 (BYTE *) &TraceOption, &dummyLen);
6152 if (code == ERROR_SUCCESS) {
6153 rxdebug_active = (TraceOption & TRACE_OPTION_DEBUGLOG) ? 1 : 0;
6155 RegCloseKey (parmKey);
6156 #endif /* AFS_NT40_ENV */
6161 rx_DebugOnOff(int on)
6163 rxdebug_active = on;
6165 #endif /* AFS_NT40_ENV */
6168 /* Don't call this debugging routine directly; use dpf */
6170 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
6171 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
6179 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6182 len = _snprintf(msg, sizeof(msg)-2,
6183 tformat, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10,
6184 a11, a12, a13, a14, a15);
6186 if (msg[len-1] != '\n') {
6190 OutputDebugString(msg);
6195 clock_GetTime(&now);
6196 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
6197 (unsigned int)now.usec / 1000);
6198 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
6205 * This function is used to process the rx_stats structure that is local
6206 * to a process as well as an rx_stats structure received from a remote
6207 * process (via rxdebug). Therefore, it needs to do minimal version
6211 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
6212 afs_int32 freePackets, char version)
6216 if (size != sizeof(struct rx_stats)) {
6218 "Unexpected size of stats structure: was %d, expected %d\n",
6219 size, sizeof(struct rx_stats));
6222 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6225 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6226 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6227 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6228 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6229 s->specialPktAllocFailures);
6231 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6232 s->receivePktAllocFailures, s->sendPktAllocFailures,
6233 s->specialPktAllocFailures);
6237 " greedy %d, " "bogusReads %d (last from host %x), "
6238 "noPackets %d, " "noBuffers %d, " "selects %d, "
6239 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6240 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6241 s->selects, s->sendSelects);
6243 fprintf(file, " packets read: ");
6244 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6245 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6247 fprintf(file, "\n");
6250 " other read counters: data %d, " "ack %d, " "dup %d "
6251 "spurious %d " "dally %d\n", s->dataPacketsRead,
6252 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6253 s->ignorePacketDally);
6255 fprintf(file, " packets sent: ");
6256 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6257 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6259 fprintf(file, "\n");
6262 " other send counters: ack %d, " "data %d (not resends), "
6263 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6264 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6265 s->dataPacketsPushed, s->ignoreAckedPacket);
6268 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6269 s->netSendFailures, (int)s->fatalErrors);
6271 if (s->nRttSamples) {
6272 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6273 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6275 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6276 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6280 " %d server connections, " "%d client connections, "
6281 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6282 s->nServerConns, s->nClientConns, s->nPeerStructs,
6283 s->nCallStructs, s->nFreeCallStructs);
6285 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6286 fprintf(file, " %d clock updates\n", clock_nUpdates);
6291 /* for backward compatibility */
6293 rx_PrintStats(FILE * file)
6295 MUTEX_ENTER(&rx_stats_mutex);
6296 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6298 MUTEX_EXIT(&rx_stats_mutex);
6302 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6304 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6305 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6306 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6309 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6310 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6311 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6314 " Packet size %d, " "max in packet skew %d, "
6315 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6316 (int)peer->outPacketSkew);
6319 #ifdef AFS_PTHREAD_ENV
6321 * This mutex protects the following static variables:
6325 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6326 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6328 #define LOCK_RX_DEBUG
6329 #define UNLOCK_RX_DEBUG
6330 #endif /* AFS_PTHREAD_ENV */
6333 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6334 u_char type, void *inputData, size_t inputLength,
6335 void *outputData, size_t outputLength)
6337 static afs_int32 counter = 100;
6339 struct rx_header theader;
6341 register afs_int32 code;
6343 struct sockaddr_in taddr, faddr;
6348 endTime = time(0) + 20; /* try for 20 seconds */
6352 tp = &tbuffer[sizeof(struct rx_header)];
6353 taddr.sin_family = AF_INET;
6354 taddr.sin_port = remotePort;
6355 taddr.sin_addr.s_addr = remoteAddr;
6356 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6357 taddr.sin_len = sizeof(struct sockaddr_in);
6360 memset(&theader, 0, sizeof(theader));
6361 theader.epoch = htonl(999);
6363 theader.callNumber = htonl(counter);
6366 theader.type = type;
6367 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6368 theader.serviceId = 0;
6370 memcpy(tbuffer, &theader, sizeof(theader));
6371 memcpy(tp, inputData, inputLength);
6373 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6374 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6376 /* see if there's a packet available */
6378 FD_SET(socket, &imask);
6381 code = select((int)(socket + 1), &imask, 0, 0, &tv);
6382 if (code == 1 && FD_ISSET(socket, &imask)) {
6383 /* now receive a packet */
6384 faddrLen = sizeof(struct sockaddr_in);
6386 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6387 (struct sockaddr *)&faddr, &faddrLen);
6390 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6391 if (counter == ntohl(theader.callNumber))
6396 /* see if we've timed out */
6397 if (endTime < time(0))
6400 code -= sizeof(struct rx_header);
6401 if (code > outputLength)
6402 code = outputLength;
6403 memcpy(outputData, tp, code);
6408 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6409 afs_uint16 remotePort, struct rx_debugStats * stat,
6410 afs_uint32 * supportedValues)
6412 struct rx_debugIn in;
6415 *supportedValues = 0;
6416 in.type = htonl(RX_DEBUGI_GETSTATS);
6419 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6420 &in, sizeof(in), stat, sizeof(*stat));
6423 * If the call was successful, fixup the version and indicate
6424 * what contents of the stat structure are valid.
6425 * Also do net to host conversion of fields here.
6429 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6430 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6432 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6433 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6435 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6436 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6438 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6439 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6441 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6442 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6444 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6445 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6447 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6448 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6450 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6451 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6454 stat->nFreePackets = ntohl(stat->nFreePackets);
6455 stat->packetReclaims = ntohl(stat->packetReclaims);
6456 stat->callsExecuted = ntohl(stat->callsExecuted);
6457 stat->nWaiting = ntohl(stat->nWaiting);
6458 stat->idleThreads = ntohl(stat->idleThreads);
6465 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6466 afs_uint16 remotePort, struct rx_stats * stat,
6467 afs_uint32 * supportedValues)
6469 struct rx_debugIn in;
6470 afs_int32 *lp = (afs_int32 *) stat;
6475 * supportedValues is currently unused, but added to allow future
6476 * versioning of this function.
6479 *supportedValues = 0;
6480 in.type = htonl(RX_DEBUGI_RXSTATS);
6482 memset(stat, 0, sizeof(*stat));
6484 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6485 &in, sizeof(in), stat, sizeof(*stat));
6490 * Do net to host conversion here
6493 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6502 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6503 afs_uint16 remotePort, size_t version_length,
6507 return MakeDebugCall(socket, remoteAddr, remotePort,
6508 RX_PACKET_TYPE_VERSION, a, 1, version,
6513 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6514 afs_uint16 remotePort, afs_int32 * nextConnection,
6515 int allConnections, afs_uint32 debugSupportedValues,
6516 struct rx_debugConn * conn,
6517 afs_uint32 * supportedValues)
6519 struct rx_debugIn in;
6524 * supportedValues is currently unused, but added to allow future
6525 * versioning of this function.
6528 *supportedValues = 0;
6529 if (allConnections) {
6530 in.type = htonl(RX_DEBUGI_GETALLCONN);
6532 in.type = htonl(RX_DEBUGI_GETCONN);
6534 in.index = htonl(*nextConnection);
6535 memset(conn, 0, sizeof(*conn));
6537 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6538 &in, sizeof(in), conn, sizeof(*conn));
6541 *nextConnection += 1;
6544 * Convert old connection format to new structure.
6547 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6548 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6549 #define MOVEvL(a) (conn->a = vL->a)
6551 /* any old or unrecognized version... */
6552 for (i = 0; i < RX_MAXCALLS; i++) {
6553 MOVEvL(callState[i]);
6554 MOVEvL(callMode[i]);
6555 MOVEvL(callFlags[i]);
6556 MOVEvL(callOther[i]);
6558 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6559 MOVEvL(secStats.type);
6560 MOVEvL(secStats.level);
6561 MOVEvL(secStats.flags);
6562 MOVEvL(secStats.expires);
6563 MOVEvL(secStats.packetsReceived);
6564 MOVEvL(secStats.packetsSent);
6565 MOVEvL(secStats.bytesReceived);
6566 MOVEvL(secStats.bytesSent);
6571 * Do net to host conversion here
6573 * I don't convert host or port since we are most likely
6574 * going to want these in NBO.
6576 conn->cid = ntohl(conn->cid);
6577 conn->serial = ntohl(conn->serial);
6578 for (i = 0; i < RX_MAXCALLS; i++) {
6579 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6581 conn->error = ntohl(conn->error);
6582 conn->secStats.flags = ntohl(conn->secStats.flags);
6583 conn->secStats.expires = ntohl(conn->secStats.expires);
6584 conn->secStats.packetsReceived =
6585 ntohl(conn->secStats.packetsReceived);
6586 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6587 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6588 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6589 conn->epoch = ntohl(conn->epoch);
6590 conn->natMTU = ntohl(conn->natMTU);
6597 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6598 afs_uint16 remotePort, afs_int32 * nextPeer,
6599 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6600 afs_uint32 * supportedValues)
6602 struct rx_debugIn in;
6606 * supportedValues is currently unused, but added to allow future
6607 * versioning of this function.
6610 *supportedValues = 0;
6611 in.type = htonl(RX_DEBUGI_GETPEER);
6612 in.index = htonl(*nextPeer);
6613 memset(peer, 0, sizeof(*peer));
6615 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6616 &in, sizeof(in), peer, sizeof(*peer));
6622 * Do net to host conversion here
6624 * I don't convert host or port since we are most likely
6625 * going to want these in NBO.
6627 peer->ifMTU = ntohs(peer->ifMTU);
6628 peer->idleWhen = ntohl(peer->idleWhen);
6629 peer->refCount = ntohs(peer->refCount);
6630 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6631 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6632 peer->rtt = ntohl(peer->rtt);
6633 peer->rtt_dev = ntohl(peer->rtt_dev);
6634 peer->timeout.sec = ntohl(peer->timeout.sec);
6635 peer->timeout.usec = ntohl(peer->timeout.usec);
6636 peer->nSent = ntohl(peer->nSent);
6637 peer->reSends = ntohl(peer->reSends);
6638 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6639 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6640 peer->rateFlag = ntohl(peer->rateFlag);
6641 peer->natMTU = ntohs(peer->natMTU);
6642 peer->maxMTU = ntohs(peer->maxMTU);
6643 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6644 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6645 peer->MTU = ntohs(peer->MTU);
6646 peer->cwind = ntohs(peer->cwind);
6647 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6648 peer->congestSeq = ntohs(peer->congestSeq);
6649 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6650 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6651 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6652 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6657 #endif /* RXDEBUG */
6662 struct rx_serverQueueEntry *np;
6665 register struct rx_call *call;
6666 register struct rx_serverQueueEntry *sq;
6670 if (rxinit_status == 1) {
6672 return; /* Already shutdown. */
6676 #ifndef AFS_PTHREAD_ENV
6677 FD_ZERO(&rx_selectMask);
6678 #endif /* AFS_PTHREAD_ENV */
6679 rxi_dataQuota = RX_MAX_QUOTA;
6680 #ifndef AFS_PTHREAD_ENV
6682 #endif /* AFS_PTHREAD_ENV */
6685 #ifndef AFS_PTHREAD_ENV
6686 #ifndef AFS_USE_GETTIMEOFDAY
6688 #endif /* AFS_USE_GETTIMEOFDAY */
6689 #endif /* AFS_PTHREAD_ENV */
6691 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6692 call = queue_First(&rx_freeCallQueue, rx_call);
6694 rxi_Free(call, sizeof(struct rx_call));
6697 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6698 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6704 struct rx_peer **peer_ptr, **peer_end;
6705 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6706 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6708 struct rx_peer *peer, *next;
6709 for (peer = *peer_ptr; peer; peer = next) {
6710 rx_interface_stat_p rpc_stat, nrpc_stat;
6713 (&peer->rpcStats, rpc_stat, nrpc_stat,
6714 rx_interface_stat)) {
6715 unsigned int num_funcs;
6718 queue_Remove(&rpc_stat->queue_header);
6719 queue_Remove(&rpc_stat->all_peers);
6720 num_funcs = rpc_stat->stats[0].func_total;
6722 sizeof(rx_interface_stat_t) +
6723 rpc_stat->stats[0].func_total *
6724 sizeof(rx_function_entry_v1_t);
6726 rxi_Free(rpc_stat, space);
6727 MUTEX_ENTER(&rx_rpc_stats);
6728 rxi_rpc_peer_stat_cnt -= num_funcs;
6729 MUTEX_EXIT(&rx_rpc_stats);
6733 MUTEX_ENTER(&rx_stats_mutex);
6734 rx_stats.nPeerStructs--;
6735 MUTEX_EXIT(&rx_stats_mutex);
6739 for (i = 0; i < RX_MAX_SERVICES; i++) {
6741 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6743 for (i = 0; i < rx_hashTableSize; i++) {
6744 register struct rx_connection *tc, *ntc;
6745 MUTEX_ENTER(&rx_connHashTable_lock);
6746 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6748 for (j = 0; j < RX_MAXCALLS; j++) {
6750 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6753 rxi_Free(tc, sizeof(*tc));
6755 MUTEX_EXIT(&rx_connHashTable_lock);
6758 MUTEX_ENTER(&freeSQEList_lock);
6760 while ((np = rx_FreeSQEList)) {
6761 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6762 MUTEX_DESTROY(&np->lock);
6763 rxi_Free(np, sizeof(*np));
6766 MUTEX_EXIT(&freeSQEList_lock);
6767 MUTEX_DESTROY(&freeSQEList_lock);
6768 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6769 MUTEX_DESTROY(&rx_connHashTable_lock);
6770 MUTEX_DESTROY(&rx_peerHashTable_lock);
6771 MUTEX_DESTROY(&rx_serverPool_lock);
6773 osi_Free(rx_connHashTable,
6774 rx_hashTableSize * sizeof(struct rx_connection *));
6775 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6777 UNPIN(rx_connHashTable,
6778 rx_hashTableSize * sizeof(struct rx_connection *));
6779 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6781 rxi_FreeAllPackets();
6783 MUTEX_ENTER(&rx_stats_mutex);
6784 rxi_dataQuota = RX_MAX_QUOTA;
6785 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6786 MUTEX_EXIT(&rx_stats_mutex);
6792 #ifdef RX_ENABLE_LOCKS
6794 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6796 if (!MUTEX_ISMINE(lockaddr))
6797 osi_Panic("Lock not held: %s", msg);
6799 #endif /* RX_ENABLE_LOCKS */
6804 * Routines to implement connection specific data.
6808 rx_KeyCreate(rx_destructor_t rtn)
6811 MUTEX_ENTER(&rxi_keyCreate_lock);
6812 key = rxi_keyCreate_counter++;
6813 rxi_keyCreate_destructor = (rx_destructor_t *)
6814 realloc((void *)rxi_keyCreate_destructor,
6815 (key + 1) * sizeof(rx_destructor_t));
6816 rxi_keyCreate_destructor[key] = rtn;
6817 MUTEX_EXIT(&rxi_keyCreate_lock);
6822 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6825 MUTEX_ENTER(&conn->conn_data_lock);
6826 if (!conn->specific) {
6827 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6828 for (i = 0; i < key; i++)
6829 conn->specific[i] = NULL;
6830 conn->nSpecific = key + 1;
6831 conn->specific[key] = ptr;
6832 } else if (key >= conn->nSpecific) {
6833 conn->specific = (void **)
6834 realloc(conn->specific, (key + 1) * sizeof(void *));
6835 for (i = conn->nSpecific; i < key; i++)
6836 conn->specific[i] = NULL;
6837 conn->nSpecific = key + 1;
6838 conn->specific[key] = ptr;
6840 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6841 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6842 conn->specific[key] = ptr;
6844 MUTEX_EXIT(&conn->conn_data_lock);
6848 rx_GetSpecific(struct rx_connection *conn, int key)
6851 MUTEX_ENTER(&conn->conn_data_lock);
6852 if (key >= conn->nSpecific)
6855 ptr = conn->specific[key];
6856 MUTEX_EXIT(&conn->conn_data_lock);
6860 #endif /* !KERNEL */
6863 * processStats is a queue used to store the statistics for the local
6864 * process. Its contents are similar to the contents of the rpcStats
6865 * queue on a rx_peer structure, but the actual data stored within
6866 * this queue contains totals across the lifetime of the process (assuming
6867 * the stats have not been reset) - unlike the per peer structures
6868 * which can come and go based upon the peer lifetime.
6871 static struct rx_queue processStats = { &processStats, &processStats };
6874 * peerStats is a queue used to store the statistics for all peer structs.
6875 * Its contents are the union of all the peer rpcStats queues.
6878 static struct rx_queue peerStats = { &peerStats, &peerStats };
6881 * rxi_monitor_processStats is used to turn process wide stat collection
6885 static int rxi_monitor_processStats = 0;
6888 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6891 static int rxi_monitor_peerStats = 0;
6894 * rxi_AddRpcStat - given all of the information for a particular rpc
6895 * call, create (if needed) and update the stat totals for the rpc.
6899 * IN stats - the queue of stats that will be updated with the new value
6901 * IN rxInterface - a unique number that identifies the rpc interface
6903 * IN currentFunc - the index of the function being invoked
6905 * IN totalFunc - the total number of functions in this interface
6907 * IN queueTime - the amount of time this function waited for a thread
6909 * IN execTime - the amount of time this function invocation took to execute
6911 * IN bytesSent - the number bytes sent by this invocation
6913 * IN bytesRcvd - the number bytes received by this invocation
6915 * IN isServer - if true, this invocation was made to a server
6917 * IN remoteHost - the ip address of the remote host
6919 * IN remotePort - the port of the remote host
6921 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6923 * INOUT counter - if a new stats structure is allocated, the counter will
6924 * be updated with the new number of allocated stat structures
6932 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6933 afs_uint32 currentFunc, afs_uint32 totalFunc,
6934 struct clock *queueTime, struct clock *execTime,
6935 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6936 afs_uint32 remoteHost, afs_uint32 remotePort,
6937 int addToPeerList, unsigned int *counter)
6940 rx_interface_stat_p rpc_stat, nrpc_stat;
6943 * See if there's already a structure for this interface
6946 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6947 if ((rpc_stat->stats[0].interfaceId == rxInterface)
6948 && (rpc_stat->stats[0].remote_is_server == isServer))
6953 * Didn't find a match so allocate a new structure and add it to the
6957 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
6958 || (rpc_stat->stats[0].interfaceId != rxInterface)
6959 || (rpc_stat->stats[0].remote_is_server != isServer)) {
6964 sizeof(rx_interface_stat_t) +
6965 totalFunc * sizeof(rx_function_entry_v1_t);
6967 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
6968 if (rpc_stat == NULL) {
6972 *counter += totalFunc;
6973 for (i = 0; i < totalFunc; i++) {
6974 rpc_stat->stats[i].remote_peer = remoteHost;
6975 rpc_stat->stats[i].remote_port = remotePort;
6976 rpc_stat->stats[i].remote_is_server = isServer;
6977 rpc_stat->stats[i].interfaceId = rxInterface;
6978 rpc_stat->stats[i].func_total = totalFunc;
6979 rpc_stat->stats[i].func_index = i;
6980 hzero(rpc_stat->stats[i].invocations);
6981 hzero(rpc_stat->stats[i].bytes_sent);
6982 hzero(rpc_stat->stats[i].bytes_rcvd);
6983 rpc_stat->stats[i].queue_time_sum.sec = 0;
6984 rpc_stat->stats[i].queue_time_sum.usec = 0;
6985 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
6986 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
6987 rpc_stat->stats[i].queue_time_min.sec = 9999999;
6988 rpc_stat->stats[i].queue_time_min.usec = 9999999;
6989 rpc_stat->stats[i].queue_time_max.sec = 0;
6990 rpc_stat->stats[i].queue_time_max.usec = 0;
6991 rpc_stat->stats[i].execution_time_sum.sec = 0;
6992 rpc_stat->stats[i].execution_time_sum.usec = 0;
6993 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
6994 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
6995 rpc_stat->stats[i].execution_time_min.sec = 9999999;
6996 rpc_stat->stats[i].execution_time_min.usec = 9999999;
6997 rpc_stat->stats[i].execution_time_max.sec = 0;
6998 rpc_stat->stats[i].execution_time_max.usec = 0;
7000 queue_Prepend(stats, rpc_stat);
7001 if (addToPeerList) {
7002 queue_Prepend(&peerStats, &rpc_stat->all_peers);
7007 * Increment the stats for this function
7010 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
7011 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
7012 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
7013 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
7014 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7015 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7016 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7018 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7019 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7021 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7022 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7024 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7025 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7027 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7028 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7036 * rx_IncrementTimeAndCount - increment the times and count for a particular
7041 * IN peer - the peer who invoked the rpc
7043 * IN rxInterface - a unique number that identifies the rpc interface
7045 * IN currentFunc - the index of the function being invoked
7047 * IN totalFunc - the total number of functions in this interface
7049 * IN queueTime - the amount of time this function waited for a thread
7051 * IN execTime - the amount of time this function invocation took to execute
7053 * IN bytesSent - the number bytes sent by this invocation
7055 * IN bytesRcvd - the number bytes received by this invocation
7057 * IN isServer - if true, this invocation was made to a server
7065 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7066 afs_uint32 currentFunc, afs_uint32 totalFunc,
7067 struct clock *queueTime, struct clock *execTime,
7068 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7072 MUTEX_ENTER(&rx_rpc_stats);
7073 MUTEX_ENTER(&peer->peer_lock);
7075 if (rxi_monitor_peerStats) {
7076 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7077 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7078 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7081 if (rxi_monitor_processStats) {
7082 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7083 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7084 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7087 MUTEX_EXIT(&peer->peer_lock);
7088 MUTEX_EXIT(&rx_rpc_stats);
7093 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7097 * IN callerVersion - the rpc stat version of the caller.
7099 * IN count - the number of entries to marshall.
7101 * IN stats - pointer to stats to be marshalled.
7103 * OUT ptr - Where to store the marshalled data.
7110 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7111 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7117 * We only support the first version
7119 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7120 *(ptr++) = stats->remote_peer;
7121 *(ptr++) = stats->remote_port;
7122 *(ptr++) = stats->remote_is_server;
7123 *(ptr++) = stats->interfaceId;
7124 *(ptr++) = stats->func_total;
7125 *(ptr++) = stats->func_index;
7126 *(ptr++) = hgethi(stats->invocations);
7127 *(ptr++) = hgetlo(stats->invocations);
7128 *(ptr++) = hgethi(stats->bytes_sent);
7129 *(ptr++) = hgetlo(stats->bytes_sent);
7130 *(ptr++) = hgethi(stats->bytes_rcvd);
7131 *(ptr++) = hgetlo(stats->bytes_rcvd);
7132 *(ptr++) = stats->queue_time_sum.sec;
7133 *(ptr++) = stats->queue_time_sum.usec;
7134 *(ptr++) = stats->queue_time_sum_sqr.sec;
7135 *(ptr++) = stats->queue_time_sum_sqr.usec;
7136 *(ptr++) = stats->queue_time_min.sec;
7137 *(ptr++) = stats->queue_time_min.usec;
7138 *(ptr++) = stats->queue_time_max.sec;
7139 *(ptr++) = stats->queue_time_max.usec;
7140 *(ptr++) = stats->execution_time_sum.sec;
7141 *(ptr++) = stats->execution_time_sum.usec;
7142 *(ptr++) = stats->execution_time_sum_sqr.sec;
7143 *(ptr++) = stats->execution_time_sum_sqr.usec;
7144 *(ptr++) = stats->execution_time_min.sec;
7145 *(ptr++) = stats->execution_time_min.usec;
7146 *(ptr++) = stats->execution_time_max.sec;
7147 *(ptr++) = stats->execution_time_max.usec;
7153 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7158 * IN callerVersion - the rpc stat version of the caller
7160 * OUT myVersion - the rpc stat version of this function
7162 * OUT clock_sec - local time seconds
7164 * OUT clock_usec - local time microseconds
7166 * OUT allocSize - the number of bytes allocated to contain stats
7168 * OUT statCount - the number stats retrieved from this process.
7170 * OUT stats - the actual stats retrieved from this process.
7174 * Returns void. If successful, stats will != NULL.
7178 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7179 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7180 size_t * allocSize, afs_uint32 * statCount,
7181 afs_uint32 ** stats)
7191 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7194 * Check to see if stats are enabled
7197 MUTEX_ENTER(&rx_rpc_stats);
7198 if (!rxi_monitor_processStats) {
7199 MUTEX_EXIT(&rx_rpc_stats);
7203 clock_GetTime(&now);
7204 *clock_sec = now.sec;
7205 *clock_usec = now.usec;
7208 * Allocate the space based upon the caller version
7210 * If the client is at an older version than we are,
7211 * we return the statistic data in the older data format, but
7212 * we still return our version number so the client knows we
7213 * are maintaining more data than it can retrieve.
7216 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7217 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7218 *statCount = rxi_rpc_process_stat_cnt;
7221 * This can't happen yet, but in the future version changes
7222 * can be handled by adding additional code here
7226 if (space > (size_t) 0) {
7228 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7231 rx_interface_stat_p rpc_stat, nrpc_stat;
7235 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7237 * Copy the data based upon the caller version
7239 rx_MarshallProcessRPCStats(callerVersion,
7240 rpc_stat->stats[0].func_total,
7241 rpc_stat->stats, &ptr);
7247 MUTEX_EXIT(&rx_rpc_stats);
7252 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7256 * IN callerVersion - the rpc stat version of the caller
7258 * OUT myVersion - the rpc stat version of this function
7260 * OUT clock_sec - local time seconds
7262 * OUT clock_usec - local time microseconds
7264 * OUT allocSize - the number of bytes allocated to contain stats
7266 * OUT statCount - the number of stats retrieved from the individual
7269 * OUT stats - the actual stats retrieved from the individual peer structures.
7273 * Returns void. If successful, stats will != NULL.
7277 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7278 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7279 size_t * allocSize, afs_uint32 * statCount,
7280 afs_uint32 ** stats)
7290 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7293 * Check to see if stats are enabled
7296 MUTEX_ENTER(&rx_rpc_stats);
7297 if (!rxi_monitor_peerStats) {
7298 MUTEX_EXIT(&rx_rpc_stats);
7302 clock_GetTime(&now);
7303 *clock_sec = now.sec;
7304 *clock_usec = now.usec;
7307 * Allocate the space based upon the caller version
7309 * If the client is at an older version than we are,
7310 * we return the statistic data in the older data format, but
7311 * we still return our version number so the client knows we
7312 * are maintaining more data than it can retrieve.
7315 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7316 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7317 *statCount = rxi_rpc_peer_stat_cnt;
7320 * This can't happen yet, but in the future version changes
7321 * can be handled by adding additional code here
7325 if (space > (size_t) 0) {
7327 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7330 rx_interface_stat_p rpc_stat, nrpc_stat;
7334 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7336 * We have to fix the offset of rpc_stat since we are
7337 * keeping this structure on two rx_queues. The rx_queue
7338 * package assumes that the rx_queue member is the first
7339 * member of the structure. That is, rx_queue assumes that
7340 * any one item is only on one queue at a time. We are
7341 * breaking that assumption and so we have to do a little
7342 * math to fix our pointers.
7345 fix_offset = (char *)rpc_stat;
7346 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7347 rpc_stat = (rx_interface_stat_p) fix_offset;
7350 * Copy the data based upon the caller version
7352 rx_MarshallProcessRPCStats(callerVersion,
7353 rpc_stat->stats[0].func_total,
7354 rpc_stat->stats, &ptr);
7360 MUTEX_EXIT(&rx_rpc_stats);
7365 * rx_FreeRPCStats - free memory allocated by
7366 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7370 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7371 * rx_RetrievePeerRPCStats
7373 * IN allocSize - the number of bytes in stats.
7381 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7383 rxi_Free(stats, allocSize);
7387 * rx_queryProcessRPCStats - see if process rpc stat collection is
7388 * currently enabled.
7394 * Returns 0 if stats are not enabled != 0 otherwise
7398 rx_queryProcessRPCStats(void)
7401 MUTEX_ENTER(&rx_rpc_stats);
7402 rc = rxi_monitor_processStats;
7403 MUTEX_EXIT(&rx_rpc_stats);
7408 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7414 * Returns 0 if stats are not enabled != 0 otherwise
7418 rx_queryPeerRPCStats(void)
7421 MUTEX_ENTER(&rx_rpc_stats);
7422 rc = rxi_monitor_peerStats;
7423 MUTEX_EXIT(&rx_rpc_stats);
7428 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7438 rx_enableProcessRPCStats(void)
7440 MUTEX_ENTER(&rx_rpc_stats);
7441 rx_enable_stats = 1;
7442 rxi_monitor_processStats = 1;
7443 MUTEX_EXIT(&rx_rpc_stats);
7447 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7457 rx_enablePeerRPCStats(void)
7459 MUTEX_ENTER(&rx_rpc_stats);
7460 rx_enable_stats = 1;
7461 rxi_monitor_peerStats = 1;
7462 MUTEX_EXIT(&rx_rpc_stats);
7466 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7476 rx_disableProcessRPCStats(void)
7478 rx_interface_stat_p rpc_stat, nrpc_stat;
7481 MUTEX_ENTER(&rx_rpc_stats);
7484 * Turn off process statistics and if peer stats is also off, turn
7488 rxi_monitor_processStats = 0;
7489 if (rxi_monitor_peerStats == 0) {
7490 rx_enable_stats = 0;
7493 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7494 unsigned int num_funcs = 0;
7497 queue_Remove(rpc_stat);
7498 num_funcs = rpc_stat->stats[0].func_total;
7500 sizeof(rx_interface_stat_t) +
7501 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7503 rxi_Free(rpc_stat, space);
7504 rxi_rpc_process_stat_cnt -= num_funcs;
7506 MUTEX_EXIT(&rx_rpc_stats);
7510 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7520 rx_disablePeerRPCStats(void)
7522 struct rx_peer **peer_ptr, **peer_end;
7525 MUTEX_ENTER(&rx_rpc_stats);
7528 * Turn off peer statistics and if process stats is also off, turn
7532 rxi_monitor_peerStats = 0;
7533 if (rxi_monitor_processStats == 0) {
7534 rx_enable_stats = 0;
7537 MUTEX_ENTER(&rx_peerHashTable_lock);
7538 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7539 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7541 struct rx_peer *peer, *next, *prev;
7542 for (prev = peer = *peer_ptr; peer; peer = next) {
7544 code = MUTEX_TRYENTER(&peer->peer_lock);
7546 rx_interface_stat_p rpc_stat, nrpc_stat;
7549 (&peer->rpcStats, rpc_stat, nrpc_stat,
7550 rx_interface_stat)) {
7551 unsigned int num_funcs = 0;
7554 queue_Remove(&rpc_stat->queue_header);
7555 queue_Remove(&rpc_stat->all_peers);
7556 num_funcs = rpc_stat->stats[0].func_total;
7558 sizeof(rx_interface_stat_t) +
7559 rpc_stat->stats[0].func_total *
7560 sizeof(rx_function_entry_v1_t);
7562 rxi_Free(rpc_stat, space);
7563 rxi_rpc_peer_stat_cnt -= num_funcs;
7565 MUTEX_EXIT(&peer->peer_lock);
7566 if (prev == *peer_ptr) {
7576 MUTEX_EXIT(&rx_peerHashTable_lock);
7577 MUTEX_EXIT(&rx_rpc_stats);
7581 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7586 * IN clearFlag - flag indicating which stats to clear
7594 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7596 rx_interface_stat_p rpc_stat, nrpc_stat;
7598 MUTEX_ENTER(&rx_rpc_stats);
7600 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7601 unsigned int num_funcs = 0, i;
7602 num_funcs = rpc_stat->stats[0].func_total;
7603 for (i = 0; i < num_funcs; i++) {
7604 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7605 hzero(rpc_stat->stats[i].invocations);
7607 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7608 hzero(rpc_stat->stats[i].bytes_sent);
7610 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7611 hzero(rpc_stat->stats[i].bytes_rcvd);
7613 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7614 rpc_stat->stats[i].queue_time_sum.sec = 0;
7615 rpc_stat->stats[i].queue_time_sum.usec = 0;
7617 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7618 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7619 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7621 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7622 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7623 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7625 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7626 rpc_stat->stats[i].queue_time_max.sec = 0;
7627 rpc_stat->stats[i].queue_time_max.usec = 0;
7629 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7630 rpc_stat->stats[i].execution_time_sum.sec = 0;
7631 rpc_stat->stats[i].execution_time_sum.usec = 0;
7633 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7634 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7635 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7637 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7638 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7639 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7641 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7642 rpc_stat->stats[i].execution_time_max.sec = 0;
7643 rpc_stat->stats[i].execution_time_max.usec = 0;
7648 MUTEX_EXIT(&rx_rpc_stats);
7652 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7657 * IN clearFlag - flag indicating which stats to clear
7665 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7667 rx_interface_stat_p rpc_stat, nrpc_stat;
7669 MUTEX_ENTER(&rx_rpc_stats);
7671 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7672 unsigned int num_funcs = 0, i;
7675 * We have to fix the offset of rpc_stat since we are
7676 * keeping this structure on two rx_queues. The rx_queue
7677 * package assumes that the rx_queue member is the first
7678 * member of the structure. That is, rx_queue assumes that
7679 * any one item is only on one queue at a time. We are
7680 * breaking that assumption and so we have to do a little
7681 * math to fix our pointers.
7684 fix_offset = (char *)rpc_stat;
7685 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7686 rpc_stat = (rx_interface_stat_p) fix_offset;
7688 num_funcs = rpc_stat->stats[0].func_total;
7689 for (i = 0; i < num_funcs; i++) {
7690 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7691 hzero(rpc_stat->stats[i].invocations);
7693 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7694 hzero(rpc_stat->stats[i].bytes_sent);
7696 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7697 hzero(rpc_stat->stats[i].bytes_rcvd);
7699 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7700 rpc_stat->stats[i].queue_time_sum.sec = 0;
7701 rpc_stat->stats[i].queue_time_sum.usec = 0;
7703 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7704 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7705 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7707 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7708 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7709 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7711 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7712 rpc_stat->stats[i].queue_time_max.sec = 0;
7713 rpc_stat->stats[i].queue_time_max.usec = 0;
7715 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7716 rpc_stat->stats[i].execution_time_sum.sec = 0;
7717 rpc_stat->stats[i].execution_time_sum.usec = 0;
7719 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7720 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7721 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7723 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7724 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7725 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7727 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7728 rpc_stat->stats[i].execution_time_max.sec = 0;
7729 rpc_stat->stats[i].execution_time_max.usec = 0;
7734 MUTEX_EXIT(&rx_rpc_stats);
7738 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7739 * is authorized to enable/disable/clear RX statistics.
7741 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7744 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7746 rxi_rxstat_userok = proc;
7750 rx_RxStatUserOk(struct rx_call *call)
7752 if (!rxi_rxstat_userok)
7754 return rxi_rxstat_userok(call);
7759 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
7760 * function in the MSVC runtime DLL (msvcrt.dll).
7762 * Note: the system serializes calls to this function.
7765 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
7766 DWORD reason, /* reason function is being called */
7767 LPVOID reserved) /* reserved for future use */
7770 case DLL_PROCESS_ATTACH:
7771 /* library is being attached to a process */
7775 case DLL_PROCESS_DETACH: