2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
14 #include "afs/param.h"
16 #include <afs/param.h>
23 #include "afs/sysincludes.h"
24 #include "afsincludes.h"
30 #include <net/net_globals.h>
31 #endif /* AFS_OSF_ENV */
32 #ifdef AFS_LINUX20_ENV
35 #include "netinet/in.h"
36 #include "afs/afs_args.h"
37 #include "afs/afs_osi.h"
38 #ifdef RX_KERNEL_TRACE
39 #include "rx_kcommon.h"
41 #if (defined(AFS_AUX_ENV) || defined(AFS_AIX_ENV))
45 #undef RXDEBUG /* turn off debugging */
47 #if defined(AFS_SGI_ENV)
48 #include "sys/debug.h"
57 #endif /* AFS_OSF_ENV */
59 #include "afs/sysincludes.h"
60 #include "afsincludes.h"
63 #include "rx_kmutex.h"
64 #include "rx_kernel.h"
68 #include "rx_globals.h"
70 #define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
71 #define AFSOP_STOP_AFS 211 /* Stop AFS process */
72 #define AFSOP_STOP_BKG 212 /* Stop BKG process */
74 extern afs_int32 afs_termState;
76 #include "sys/lockl.h"
77 #include "sys/lock_def.h"
78 #endif /* AFS_AIX41_ENV */
79 # include "rxgen_consts.h"
81 # include <sys/types.h>
86 # include <afs/afsutil.h>
87 # include <WINNT\afsreg.h>
89 # include <sys/socket.h>
90 # include <sys/file.h>
92 # include <sys/stat.h>
93 # include <netinet/in.h>
94 # include <sys/time.h>
104 # include "rx_user.h"
105 # include "rx_clock.h"
106 # include "rx_queue.h"
107 # include "rx_globals.h"
108 # include "rx_trace.h"
109 # include <afs/rxgen_consts.h>
112 int (*registerProgram) () = 0;
113 int (*swapNameProgram) () = 0;
115 /* Local static routines */
116 static void rxi_DestroyConnectionNoLock(register struct rx_connection *conn);
117 #ifdef RX_ENABLE_LOCKS
118 static void rxi_SetAcksInTransmitQueue(register struct rx_call *call);
121 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
123 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
124 afs_int32 rxi_start_in_error;
126 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
129 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
130 * currently allocated within rx. This number is used to allocate the
131 * memory required to return the statistics when queried.
134 static unsigned int rxi_rpc_peer_stat_cnt;
137 * rxi_rpc_process_stat_cnt counts the total number of local process stat
138 * structures currently allocated within rx. The number is used to allocate
139 * the memory required to return the statistics when queried.
142 static unsigned int rxi_rpc_process_stat_cnt;
144 #if !defined(offsetof)
145 #include <stddef.h> /* for definition of offsetof() */
148 #ifdef AFS_PTHREAD_ENV
152 * Use procedural initialization of mutexes/condition variables
156 extern pthread_mutex_t rx_stats_mutex;
157 extern pthread_mutex_t des_init_mutex;
158 extern pthread_mutex_t des_random_mutex;
159 extern pthread_mutex_t rx_clock_mutex;
160 extern pthread_mutex_t rxi_connCacheMutex;
161 extern pthread_mutex_t rx_event_mutex;
162 extern pthread_mutex_t osi_malloc_mutex;
163 extern pthread_mutex_t event_handler_mutex;
164 extern pthread_mutex_t listener_mutex;
165 extern pthread_mutex_t rx_if_init_mutex;
166 extern pthread_mutex_t rx_if_mutex;
167 extern pthread_mutex_t rxkad_client_uid_mutex;
168 extern pthread_mutex_t rxkad_random_mutex;
170 extern pthread_cond_t rx_event_handler_cond;
171 extern pthread_cond_t rx_listener_cond;
173 static pthread_mutex_t epoch_mutex;
174 static pthread_mutex_t rx_init_mutex;
175 static pthread_mutex_t rx_debug_mutex;
178 rxi_InitPthread(void)
180 assert(pthread_mutex_init(&rx_clock_mutex, (const pthread_mutexattr_t *)0)
182 assert(pthread_mutex_init(&rx_stats_mutex, (const pthread_mutexattr_t *)0)
184 assert(pthread_mutex_init
185 (&rxi_connCacheMutex, (const pthread_mutexattr_t *)0) == 0);
186 assert(pthread_mutex_init(&rx_init_mutex, (const pthread_mutexattr_t *)0)
188 assert(pthread_mutex_init(&epoch_mutex, (const pthread_mutexattr_t *)0) ==
190 assert(pthread_mutex_init(&rx_event_mutex, (const pthread_mutexattr_t *)0)
192 assert(pthread_mutex_init(&des_init_mutex, (const pthread_mutexattr_t *)0)
194 assert(pthread_mutex_init
195 (&des_random_mutex, (const pthread_mutexattr_t *)0) == 0);
196 assert(pthread_mutex_init
197 (&osi_malloc_mutex, (const pthread_mutexattr_t *)0) == 0);
198 assert(pthread_mutex_init
199 (&event_handler_mutex, (const pthread_mutexattr_t *)0) == 0);
200 assert(pthread_mutex_init(&listener_mutex, (const pthread_mutexattr_t *)0)
202 assert(pthread_mutex_init
203 (&rx_if_init_mutex, (const pthread_mutexattr_t *)0) == 0);
204 assert(pthread_mutex_init(&rx_if_mutex, (const pthread_mutexattr_t *)0) ==
206 assert(pthread_mutex_init
207 (&rxkad_client_uid_mutex, (const pthread_mutexattr_t *)0) == 0);
208 assert(pthread_mutex_init
209 (&rxkad_random_mutex, (const pthread_mutexattr_t *)0) == 0);
210 assert(pthread_mutex_init(&rx_debug_mutex, (const pthread_mutexattr_t *)0)
213 assert(pthread_cond_init
214 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
215 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
217 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
218 assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
220 rxkad_global_stats_init();
223 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
224 #define INIT_PTHREAD_LOCKS \
225 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
227 * The rx_stats_mutex mutex protects the following global variables:
232 * rxi_lowConnRefCount
233 * rxi_lowPeerRefCount
242 #define INIT_PTHREAD_LOCKS
246 /* Variables for handling the minProcs implementation. availProcs gives the
247 * number of threads available in the pool at this moment (not counting dudes
248 * executing right now). totalMin gives the total number of procs required
249 * for handling all minProcs requests. minDeficit is a dynamic variable
250 * tracking the # of procs required to satisfy all of the remaining minProcs
252 * For fine grain locking to work, the quota check and the reservation of
253 * a server thread has to come while rxi_availProcs and rxi_minDeficit
254 * are locked. To this end, the code has been modified under #ifdef
255 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
256 * same time. A new function, ReturnToServerPool() returns the allocation.
258 * A call can be on several queue's (but only one at a time). When
259 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
260 * that no one else is touching the queue. To this end, we store the address
261 * of the queue lock in the call structure (under the call lock) when we
262 * put the call on a queue, and we clear the call_queue_lock when the
263 * call is removed from a queue (once the call lock has been obtained).
264 * This allows rxi_ResetCall to safely synchronize with others wishing
265 * to manipulate the queue.
268 #ifdef RX_ENABLE_LOCKS
269 static afs_kmutex_t rx_rpc_stats;
270 void rxi_StartUnlocked();
273 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
274 ** pretty good that the next packet coming in is from the same connection
275 ** as the last packet, since we're send multiple packets in a transmit window.
277 struct rx_connection *rxLastConn = 0;
279 #ifdef RX_ENABLE_LOCKS
280 /* The locking hierarchy for rx fine grain locking is composed of these
283 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
284 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
285 * call->lock - locks call data fields.
286 * These are independent of each other:
287 * rx_freeCallQueue_lock
292 * serverQueueEntry->lock
294 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
295 * peer->lock - locks peer data fields.
296 * conn_data_lock - that more than one thread is not updating a conn data
297 * field at the same time.
305 * Do we need a lock to protect the peer field in the conn structure?
306 * conn->peer was previously a constant for all intents and so has no
307 * lock protecting this field. The multihomed client delta introduced
308 * a RX code change : change the peer field in the connection structure
309 * to that remote inetrface from which the last packet for this
310 * connection was sent out. This may become an issue if further changes
313 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
314 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
316 /* rxdb_fileID is used to identify the lock location, along with line#. */
317 static int rxdb_fileID = RXDB_FILE_RX;
318 #endif /* RX_LOCKS_DB */
319 #else /* RX_ENABLE_LOCKS */
320 #define SET_CALL_QUEUE_LOCK(C, L)
321 #define CLEAR_CALL_QUEUE_LOCK(C)
322 #endif /* RX_ENABLE_LOCKS */
323 struct rx_serverQueueEntry *rx_waitForPacket = 0;
324 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
326 /* ------------Exported Interfaces------------- */
328 /* This function allows rxkad to set the epoch to a suitably random number
329 * which rx_NewConnection will use in the future. The principle purpose is to
330 * get rxnull connections to use the same epoch as the rxkad connections do, at
331 * least once the first rxkad connection is established. This is important now
332 * that the host/port addresses aren't used in FindConnection: the uniqueness
333 * of epoch/cid matters and the start time won't do. */
335 #ifdef AFS_PTHREAD_ENV
337 * This mutex protects the following global variables:
341 #define LOCK_EPOCH assert(pthread_mutex_lock(&epoch_mutex)==0)
342 #define UNLOCK_EPOCH assert(pthread_mutex_unlock(&epoch_mutex)==0)
346 #endif /* AFS_PTHREAD_ENV */
349 rx_SetEpoch(afs_uint32 epoch)
356 /* Initialize rx. A port number may be mentioned, in which case this
357 * becomes the default port number for any service installed later.
358 * If 0 is provided for the port number, a random port will be chosen
359 * by the kernel. Whether this will ever overlap anything in
360 * /etc/services is anybody's guess... Returns 0 on success, -1 on
362 static int rxinit_status = 1;
363 #ifdef AFS_PTHREAD_ENV
365 * This mutex protects the following global variables:
369 #define LOCK_RX_INIT assert(pthread_mutex_lock(&rx_init_mutex)==0)
370 #define UNLOCK_RX_INIT assert(pthread_mutex_unlock(&rx_init_mutex)==0)
373 #define UNLOCK_RX_INIT
377 rx_InitHost(u_int host, u_int port)
384 char *htable, *ptable;
391 if (rxinit_status == 0) {
392 tmp_status = rxinit_status;
394 return tmp_status; /* Already started; return previous error code. */
400 if (afs_winsockInit() < 0)
406 * Initialize anything necessary to provide a non-premptive threading
409 rxi_InitializeThreadSupport();
412 /* Allocate and initialize a socket for client and perhaps server
415 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
416 if (rx_socket == OSI_NULLSOCKET) {
420 #ifdef RX_ENABLE_LOCKS
423 #endif /* RX_LOCKS_DB */
424 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
425 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
426 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
427 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
428 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
430 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
432 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
434 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
436 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
438 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
440 #if defined(KERNEL) && defined(AFS_HPUX110_ENV)
442 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
443 #endif /* KERNEL && AFS_HPUX110_ENV */
444 #endif /* RX_ENABLE_LOCKS */
447 rx_connDeadTime = 12;
448 rx_tranquil = 0; /* reset flag */
449 memset((char *)&rx_stats, 0, sizeof(struct rx_stats));
451 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
452 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
453 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
454 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
455 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
456 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
458 /* Malloc up a bunch of packets & buffers */
460 queue_Init(&rx_freePacketQueue);
461 rxi_NeedMorePackets = FALSE;
462 #ifdef RX_ENABLE_TSFPQ
463 rx_nPackets = 0; /* in TSFPQ version, rx_nPackets is managed by rxi_MorePackets* */
464 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
465 #else /* RX_ENABLE_TSFPQ */
466 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
467 rxi_MorePackets(rx_nPackets);
468 #endif /* RX_ENABLE_TSFPQ */
475 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
476 tv.tv_sec = clock_now.sec;
477 tv.tv_usec = clock_now.usec;
478 srand((unsigned int)tv.tv_usec);
485 #if defined(KERNEL) && !defined(UKERNEL)
486 /* Really, this should never happen in a real kernel */
489 struct sockaddr_in addr;
490 int addrlen = sizeof(addr);
491 if (getsockname((int)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
495 rx_port = addr.sin_port;
498 rx_stats.minRtt.sec = 9999999;
500 rx_SetEpoch(tv.tv_sec | 0x80000000);
502 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
503 * will provide a randomer value. */
505 MUTEX_ENTER(&rx_stats_mutex);
506 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
507 MUTEX_EXIT(&rx_stats_mutex);
508 /* *Slightly* random start time for the cid. This is just to help
509 * out with the hashing function at the peer */
510 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
511 rx_connHashTable = (struct rx_connection **)htable;
512 rx_peerHashTable = (struct rx_peer **)ptable;
514 rx_lastAckDelay.sec = 0;
515 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
516 rx_hardAckDelay.sec = 0;
517 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
518 rx_softAckDelay.sec = 0;
519 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
521 rxevent_Init(20, rxi_ReScheduleEvents);
523 /* Initialize various global queues */
524 queue_Init(&rx_idleServerQueue);
525 queue_Init(&rx_incomingCallQueue);
526 queue_Init(&rx_freeCallQueue);
528 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
529 /* Initialize our list of usable IP addresses. */
533 /* Start listener process (exact function is dependent on the
534 * implementation environment--kernel or user space) */
538 tmp_status = rxinit_status = 0;
546 return rx_InitHost(htonl(INADDR_ANY), port);
549 /* called with unincremented nRequestsRunning to see if it is OK to start
550 * a new thread in this service. Could be "no" for two reasons: over the
551 * max quota, or would prevent others from reaching their min quota.
553 #ifdef RX_ENABLE_LOCKS
554 /* This verion of QuotaOK reserves quota if it's ok while the
555 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
558 QuotaOK(register struct rx_service *aservice)
560 /* check if over max quota */
561 if (aservice->nRequestsRunning >= aservice->maxProcs) {
565 /* under min quota, we're OK */
566 /* otherwise, can use only if there are enough to allow everyone
567 * to go to their min quota after this guy starts.
569 MUTEX_ENTER(&rx_stats_mutex);
570 if ((aservice->nRequestsRunning < aservice->minProcs)
571 || (rxi_availProcs > rxi_minDeficit)) {
572 aservice->nRequestsRunning++;
573 /* just started call in minProcs pool, need fewer to maintain
575 if (aservice->nRequestsRunning <= aservice->minProcs)
578 MUTEX_EXIT(&rx_stats_mutex);
581 MUTEX_EXIT(&rx_stats_mutex);
587 ReturnToServerPool(register struct rx_service *aservice)
589 aservice->nRequestsRunning--;
590 MUTEX_ENTER(&rx_stats_mutex);
591 if (aservice->nRequestsRunning < aservice->minProcs)
594 MUTEX_EXIT(&rx_stats_mutex);
597 #else /* RX_ENABLE_LOCKS */
599 QuotaOK(register struct rx_service *aservice)
602 /* under min quota, we're OK */
603 if (aservice->nRequestsRunning < aservice->minProcs)
606 /* check if over max quota */
607 if (aservice->nRequestsRunning >= aservice->maxProcs)
610 /* otherwise, can use only if there are enough to allow everyone
611 * to go to their min quota after this guy starts.
613 if (rxi_availProcs > rxi_minDeficit)
617 #endif /* RX_ENABLE_LOCKS */
620 /* Called by rx_StartServer to start up lwp's to service calls.
621 NExistingProcs gives the number of procs already existing, and which
622 therefore needn't be created. */
624 rxi_StartServerProcs(int nExistingProcs)
626 register struct rx_service *service;
631 /* For each service, reserve N processes, where N is the "minimum"
632 * number of processes that MUST be able to execute a request in parallel,
633 * at any time, for that process. Also compute the maximum difference
634 * between any service's maximum number of processes that can run
635 * (i.e. the maximum number that ever will be run, and a guarantee
636 * that this number will run if other services aren't running), and its
637 * minimum number. The result is the extra number of processes that
638 * we need in order to provide the latter guarantee */
639 for (i = 0; i < RX_MAX_SERVICES; i++) {
641 service = rx_services[i];
642 if (service == (struct rx_service *)0)
644 nProcs += service->minProcs;
645 diff = service->maxProcs - service->minProcs;
649 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
650 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
651 for (i = 0; i < nProcs; i++) {
652 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
658 /* This routine is only required on Windows */
660 rx_StartClientThread(void)
662 #ifdef AFS_PTHREAD_ENV
664 pid = (int) pthread_self();
665 #endif /* AFS_PTHREAD_ENV */
667 #endif /* AFS_NT40_ENV */
669 /* This routine must be called if any services are exported. If the
670 * donateMe flag is set, the calling process is donated to the server
673 rx_StartServer(int donateMe)
675 register struct rx_service *service;
681 /* Start server processes, if necessary (exact function is dependent
682 * on the implementation environment--kernel or user space). DonateMe
683 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
684 * case, one less new proc will be created rx_StartServerProcs.
686 rxi_StartServerProcs(donateMe);
688 /* count up the # of threads in minProcs, and add set the min deficit to
689 * be that value, too.
691 for (i = 0; i < RX_MAX_SERVICES; i++) {
692 service = rx_services[i];
693 if (service == (struct rx_service *)0)
695 MUTEX_ENTER(&rx_stats_mutex);
696 rxi_totalMin += service->minProcs;
697 /* below works even if a thread is running, since minDeficit would
698 * still have been decremented and later re-incremented.
700 rxi_minDeficit += service->minProcs;
701 MUTEX_EXIT(&rx_stats_mutex);
704 /* Turn on reaping of idle server connections */
705 rxi_ReapConnections();
714 #ifdef AFS_PTHREAD_ENV
716 pid = (pid_t) pthread_self();
717 #else /* AFS_PTHREAD_ENV */
719 LWP_CurrentProcess(&pid);
720 #endif /* AFS_PTHREAD_ENV */
722 sprintf(name, "srv_%d", ++nProcs);
724 (*registerProgram) (pid, name);
726 #endif /* AFS_NT40_ENV */
727 rx_ServerProc(); /* Never returns */
729 #ifdef RX_ENABLE_TSFPQ
730 /* no use leaving packets around in this thread's local queue if
731 * it isn't getting donated to the server thread pool.
733 rxi_FlushLocalPacketsTSFPQ();
734 #endif /* RX_ENABLE_TSFPQ */
738 /* Create a new client connection to the specified service, using the
739 * specified security object to implement the security model for this
741 struct rx_connection *
742 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
743 register struct rx_securityClass *securityObject,
744 int serviceSecurityIndex)
748 register struct rx_connection *conn;
753 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", ntohl(shost), ntohs(sport), sservice, securityObject, serviceSecurityIndex));
755 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
756 * the case of kmem_alloc? */
757 conn = rxi_AllocConnection();
758 #ifdef RX_ENABLE_LOCKS
759 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
760 MUTEX_INIT(&conn->conn_data_lock, "conn call lock", MUTEX_DEFAULT, 0);
761 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
764 MUTEX_ENTER(&rx_connHashTable_lock);
765 cid = (rx_nextCid += RX_MAXCALLS);
766 conn->type = RX_CLIENT_CONNECTION;
768 conn->epoch = rx_epoch;
769 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
770 conn->serviceId = sservice;
771 conn->securityObject = securityObject;
772 /* This doesn't work in all compilers with void (they're buggy), so fake it
774 conn->securityData = (VOID *) 0;
775 conn->securityIndex = serviceSecurityIndex;
776 rx_SetConnDeadTime(conn, rx_connDeadTime);
777 conn->ackRate = RX_FAST_ACK_RATE;
779 conn->specific = NULL;
780 conn->challengeEvent = NULL;
781 conn->delayedAbortEvent = NULL;
782 conn->abortCount = 0;
785 RXS_NewConnection(securityObject, conn);
787 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
789 conn->refCount++; /* no lock required since only this thread knows... */
790 conn->next = rx_connHashTable[hashindex];
791 rx_connHashTable[hashindex] = conn;
792 MUTEX_ENTER(&rx_stats_mutex);
793 rx_stats.nClientConns++;
794 MUTEX_EXIT(&rx_stats_mutex);
796 MUTEX_EXIT(&rx_connHashTable_lock);
802 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
804 /* The idea is to set the dead time to a value that allows several
805 * keepalives to be dropped without timing out the connection. */
806 conn->secondsUntilDead = MAX(seconds, 6);
807 conn->secondsUntilPing = conn->secondsUntilDead / 6;
810 int rxi_lowPeerRefCount = 0;
811 int rxi_lowConnRefCount = 0;
814 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
815 * NOTE: must not be called with rx_connHashTable_lock held.
818 rxi_CleanupConnection(struct rx_connection *conn)
820 /* Notify the service exporter, if requested, that this connection
821 * is being destroyed */
822 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
823 (*conn->service->destroyConnProc) (conn);
825 /* Notify the security module that this connection is being destroyed */
826 RXS_DestroyConnection(conn->securityObject, conn);
828 /* If this is the last connection using the rx_peer struct, set its
829 * idle time to now. rxi_ReapConnections will reap it if it's still
830 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
832 MUTEX_ENTER(&rx_peerHashTable_lock);
833 if (conn->peer->refCount < 2) {
834 conn->peer->idleWhen = clock_Sec();
835 if (conn->peer->refCount < 1) {
836 conn->peer->refCount = 1;
837 MUTEX_ENTER(&rx_stats_mutex);
838 rxi_lowPeerRefCount++;
839 MUTEX_EXIT(&rx_stats_mutex);
842 conn->peer->refCount--;
843 MUTEX_EXIT(&rx_peerHashTable_lock);
845 MUTEX_ENTER(&rx_stats_mutex);
846 if (conn->type == RX_SERVER_CONNECTION)
847 rx_stats.nServerConns--;
849 rx_stats.nClientConns--;
850 MUTEX_EXIT(&rx_stats_mutex);
853 if (conn->specific) {
855 for (i = 0; i < conn->nSpecific; i++) {
856 if (conn->specific[i] && rxi_keyCreate_destructor[i])
857 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
858 conn->specific[i] = NULL;
860 free(conn->specific);
862 conn->specific = NULL;
866 MUTEX_DESTROY(&conn->conn_call_lock);
867 MUTEX_DESTROY(&conn->conn_data_lock);
868 CV_DESTROY(&conn->conn_call_cv);
870 rxi_FreeConnection(conn);
873 /* Destroy the specified connection */
875 rxi_DestroyConnection(register struct rx_connection *conn)
877 MUTEX_ENTER(&rx_connHashTable_lock);
878 rxi_DestroyConnectionNoLock(conn);
879 /* conn should be at the head of the cleanup list */
880 if (conn == rx_connCleanup_list) {
881 rx_connCleanup_list = rx_connCleanup_list->next;
882 MUTEX_EXIT(&rx_connHashTable_lock);
883 rxi_CleanupConnection(conn);
885 #ifdef RX_ENABLE_LOCKS
887 MUTEX_EXIT(&rx_connHashTable_lock);
889 #endif /* RX_ENABLE_LOCKS */
893 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
895 register struct rx_connection **conn_ptr;
896 register int havecalls = 0;
897 struct rx_packet *packet;
904 MUTEX_ENTER(&conn->conn_data_lock);
905 if (conn->refCount > 0)
908 MUTEX_ENTER(&rx_stats_mutex);
909 rxi_lowConnRefCount++;
910 MUTEX_EXIT(&rx_stats_mutex);
913 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
914 /* Busy; wait till the last guy before proceeding */
915 MUTEX_EXIT(&conn->conn_data_lock);
920 /* If the client previously called rx_NewCall, but it is still
921 * waiting, treat this as a running call, and wait to destroy the
922 * connection later when the call completes. */
923 if ((conn->type == RX_CLIENT_CONNECTION)
924 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
925 conn->flags |= RX_CONN_DESTROY_ME;
926 MUTEX_EXIT(&conn->conn_data_lock);
930 MUTEX_EXIT(&conn->conn_data_lock);
932 /* Check for extant references to this connection */
933 for (i = 0; i < RX_MAXCALLS; i++) {
934 register struct rx_call *call = conn->call[i];
937 if (conn->type == RX_CLIENT_CONNECTION) {
938 MUTEX_ENTER(&call->lock);
939 if (call->delayedAckEvent) {
940 /* Push the final acknowledgment out now--there
941 * won't be a subsequent call to acknowledge the
942 * last reply packets */
943 rxevent_Cancel(call->delayedAckEvent, call,
944 RX_CALL_REFCOUNT_DELAY);
945 if (call->state == RX_STATE_PRECALL
946 || call->state == RX_STATE_ACTIVE) {
947 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
949 rxi_AckAll(NULL, call, 0);
952 MUTEX_EXIT(&call->lock);
956 #ifdef RX_ENABLE_LOCKS
958 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
959 MUTEX_EXIT(&conn->conn_data_lock);
961 /* Someone is accessing a packet right now. */
965 #endif /* RX_ENABLE_LOCKS */
968 /* Don't destroy the connection if there are any call
969 * structures still in use */
970 MUTEX_ENTER(&conn->conn_data_lock);
971 conn->flags |= RX_CONN_DESTROY_ME;
972 MUTEX_EXIT(&conn->conn_data_lock);
977 if (conn->delayedAbortEvent) {
978 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
979 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
981 MUTEX_ENTER(&conn->conn_data_lock);
982 rxi_SendConnectionAbort(conn, packet, 0, 1);
983 MUTEX_EXIT(&conn->conn_data_lock);
984 rxi_FreePacket(packet);
988 /* Remove from connection hash table before proceeding */
990 &rx_connHashTable[CONN_HASH
991 (peer->host, peer->port, conn->cid, conn->epoch,
993 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
994 if (*conn_ptr == conn) {
995 *conn_ptr = conn->next;
999 /* if the conn that we are destroying was the last connection, then we
1000 * clear rxLastConn as well */
1001 if (rxLastConn == conn)
1004 /* Make sure the connection is completely reset before deleting it. */
1005 /* get rid of pending events that could zap us later */
1006 if (conn->challengeEvent)
1007 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1008 if (conn->checkReachEvent)
1009 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1011 /* Add the connection to the list of destroyed connections that
1012 * need to be cleaned up. This is necessary to avoid deadlocks
1013 * in the routines we call to inform others that this connection is
1014 * being destroyed. */
1015 conn->next = rx_connCleanup_list;
1016 rx_connCleanup_list = conn;
1019 /* Externally available version */
1021 rx_DestroyConnection(register struct rx_connection *conn)
1026 rxi_DestroyConnection(conn);
1031 rx_GetConnection(register struct rx_connection *conn)
1036 MUTEX_ENTER(&conn->conn_data_lock);
1038 MUTEX_EXIT(&conn->conn_data_lock);
1042 /* Wait for the transmit queue to no longer be busy.
1043 * requires the call->lock to be held */
1044 static void rxi_WaitforTQBusy(struct rx_call *call) {
1045 while (call->flags & RX_CALL_TQ_BUSY) {
1046 call->flags |= RX_CALL_TQ_WAIT;
1048 #ifdef RX_ENABLE_LOCKS
1049 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1050 CV_WAIT(&call->cv_tq, &call->lock);
1051 #else /* RX_ENABLE_LOCKS */
1052 osi_rxSleep(&call->tq);
1053 #endif /* RX_ENABLE_LOCKS */
1055 if (call->tqWaiters == 0) {
1056 call->flags &= ~RX_CALL_TQ_WAIT;
1060 /* Start a new rx remote procedure call, on the specified connection.
1061 * If wait is set to 1, wait for a free call channel; otherwise return
1062 * 0. Maxtime gives the maximum number of seconds this call may take,
1063 * after rx_NewCall returns. After this time interval, a call to any
1064 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1065 * For fine grain locking, we hold the conn_call_lock in order to
1066 * to ensure that we don't get signalle after we found a call in an active
1067 * state and before we go to sleep.
1070 rx_NewCall(register struct rx_connection *conn)
1073 register struct rx_call *call;
1074 struct clock queueTime;
1078 dpf(("rx_NewCall(conn %x)\n", conn));
1081 clock_GetTime(&queueTime);
1082 MUTEX_ENTER(&conn->conn_call_lock);
1085 * Check if there are others waiting for a new call.
1086 * If so, let them go first to avoid starving them.
1087 * This is a fairly simple scheme, and might not be
1088 * a complete solution for large numbers of waiters.
1090 * makeCallWaiters keeps track of the number of
1091 * threads waiting to make calls and the
1092 * RX_CONN_MAKECALL_WAITING flag bit is used to
1093 * indicate that there are indeed calls waiting.
1094 * The flag is set when the waiter is incremented.
1095 * It is only cleared in rx_EndCall when
1096 * makeCallWaiters is 0. This prevents us from
1097 * accidently destroying the connection while it
1098 * is potentially about to be used.
1100 MUTEX_ENTER(&conn->conn_data_lock);
1101 if (conn->makeCallWaiters) {
1102 conn->flags |= RX_CONN_MAKECALL_WAITING;
1103 conn->makeCallWaiters++;
1104 MUTEX_EXIT(&conn->conn_data_lock);
1106 #ifdef RX_ENABLE_LOCKS
1107 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1111 MUTEX_ENTER(&conn->conn_data_lock);
1112 conn->makeCallWaiters--;
1114 MUTEX_EXIT(&conn->conn_data_lock);
1117 for (i = 0; i < RX_MAXCALLS; i++) {
1118 call = conn->call[i];
1120 MUTEX_ENTER(&call->lock);
1121 if (call->state == RX_STATE_DALLY) {
1122 rxi_ResetCall(call, 0);
1123 (*call->callNumber)++;
1126 MUTEX_EXIT(&call->lock);
1128 call = rxi_NewCall(conn, i);
1132 if (i < RX_MAXCALLS) {
1135 MUTEX_ENTER(&conn->conn_data_lock);
1136 conn->flags |= RX_CONN_MAKECALL_WAITING;
1137 conn->makeCallWaiters++;
1138 MUTEX_EXIT(&conn->conn_data_lock);
1140 #ifdef RX_ENABLE_LOCKS
1141 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1145 MUTEX_ENTER(&conn->conn_data_lock);
1146 conn->makeCallWaiters--;
1147 MUTEX_EXIT(&conn->conn_data_lock);
1150 * Wake up anyone else who might be giving us a chance to
1151 * run (see code above that avoids resource starvation).
1153 #ifdef RX_ENABLE_LOCKS
1154 CV_BROADCAST(&conn->conn_call_cv);
1159 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1161 /* Client is initially in send mode */
1162 call->state = RX_STATE_ACTIVE;
1163 call->error = conn->error;
1165 call->mode = RX_MODE_ERROR;
1167 call->mode = RX_MODE_SENDING;
1169 /* remember start time for call in case we have hard dead time limit */
1170 call->queueTime = queueTime;
1171 clock_GetTime(&call->startTime);
1172 hzero(call->bytesSent);
1173 hzero(call->bytesRcvd);
1175 /* Turn on busy protocol. */
1176 rxi_KeepAliveOn(call);
1178 MUTEX_EXIT(&call->lock);
1179 MUTEX_EXIT(&conn->conn_call_lock);
1182 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1183 /* Now, if TQ wasn't cleared earlier, do it now. */
1184 MUTEX_ENTER(&call->lock);
1185 rxi_WaitforTQBusy(call);
1186 if (call->flags & RX_CALL_TQ_CLEARME) {
1187 rxi_ClearTransmitQueue(call, 0);
1188 queue_Init(&call->tq);
1190 MUTEX_EXIT(&call->lock);
1191 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1193 dpf(("rx_NewCall(call %x)\n", call));
1198 rxi_HasActiveCalls(register struct rx_connection *aconn)
1201 register struct rx_call *tcall;
1205 for (i = 0; i < RX_MAXCALLS; i++) {
1206 if ((tcall = aconn->call[i])) {
1207 if ((tcall->state == RX_STATE_ACTIVE)
1208 || (tcall->state == RX_STATE_PRECALL)) {
1219 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1220 register afs_int32 * aint32s)
1223 register struct rx_call *tcall;
1227 for (i = 0; i < RX_MAXCALLS; i++) {
1228 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1229 aint32s[i] = aconn->callNumber[i] + 1;
1231 aint32s[i] = aconn->callNumber[i];
1238 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1239 register afs_int32 * aint32s)
1242 register struct rx_call *tcall;
1246 for (i = 0; i < RX_MAXCALLS; i++) {
1247 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1248 aconn->callNumber[i] = aint32s[i] - 1;
1250 aconn->callNumber[i] = aint32s[i];
1256 /* Advertise a new service. A service is named locally by a UDP port
1257 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1260 char *serviceName; Name for identification purposes (e.g. the
1261 service name might be used for probing for
1264 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1265 char *serviceName, struct rx_securityClass **securityObjects,
1266 int nSecurityObjects,
1267 afs_int32(*serviceProc) (struct rx_call * acall))
1269 osi_socket socket = OSI_NULLSOCKET;
1270 register struct rx_service *tservice;
1276 if (serviceId == 0) {
1278 "rx_NewService: service id for service %s is not non-zero.\n",
1285 "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",
1293 tservice = rxi_AllocService();
1295 for (i = 0; i < RX_MAX_SERVICES; i++) {
1296 register struct rx_service *service = rx_services[i];
1298 if (port == service->servicePort && host == service->serviceHost) {
1299 if (service->serviceId == serviceId) {
1300 /* The identical service has already been
1301 * installed; if the caller was intending to
1302 * change the security classes used by this
1303 * service, he/she loses. */
1305 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1306 serviceName, serviceId, service->serviceName);
1308 rxi_FreeService(tservice);
1311 /* Different service, same port: re-use the socket
1312 * which is bound to the same port */
1313 socket = service->socket;
1316 if (socket == OSI_NULLSOCKET) {
1317 /* If we don't already have a socket (from another
1318 * service on same port) get a new one */
1319 socket = rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
1320 if (socket == OSI_NULLSOCKET) {
1322 rxi_FreeService(tservice);
1327 service->socket = socket;
1328 service->serviceHost = host;
1329 service->servicePort = port;
1330 service->serviceId = serviceId;
1331 service->serviceName = serviceName;
1332 service->nSecurityObjects = nSecurityObjects;
1333 service->securityObjects = securityObjects;
1334 service->minProcs = 0;
1335 service->maxProcs = 1;
1336 service->idleDeadTime = 60;
1337 service->connDeadTime = rx_connDeadTime;
1338 service->executeRequestProc = serviceProc;
1339 service->checkReach = 0;
1340 rx_services[i] = service; /* not visible until now */
1346 rxi_FreeService(tservice);
1347 (osi_Msg "rx_NewService: cannot support > %d services\n",
1353 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1354 struct rx_securityClass **securityObjects, int nSecurityObjects,
1355 afs_int32(*serviceProc) (struct rx_call * acall))
1357 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1360 /* Generic request processing loop. This routine should be called
1361 * by the implementation dependent rx_ServerProc. If socketp is
1362 * non-null, it will be set to the file descriptor that this thread
1363 * is now listening on. If socketp is null, this routine will never
1366 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1368 register struct rx_call *call;
1369 register afs_int32 code;
1370 register struct rx_service *tservice = NULL;
1377 call = rx_GetCall(threadID, tservice, socketp);
1378 if (socketp && *socketp != OSI_NULLSOCKET) {
1379 /* We are now a listener thread */
1384 /* if server is restarting( typically smooth shutdown) then do not
1385 * allow any new calls.
1388 if (rx_tranquil && (call != NULL)) {
1392 MUTEX_ENTER(&call->lock);
1394 rxi_CallError(call, RX_RESTARTING);
1395 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1397 MUTEX_EXIT(&call->lock);
1401 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1402 #ifdef RX_ENABLE_LOCKS
1404 #endif /* RX_ENABLE_LOCKS */
1405 afs_termState = AFSOP_STOP_AFS;
1406 afs_osi_Wakeup(&afs_termState);
1407 #ifdef RX_ENABLE_LOCKS
1409 #endif /* RX_ENABLE_LOCKS */
1414 tservice = call->conn->service;
1416 if (tservice->beforeProc)
1417 (*tservice->beforeProc) (call);
1419 code = call->conn->service->executeRequestProc(call);
1421 if (tservice->afterProc)
1422 (*tservice->afterProc) (call, code);
1424 rx_EndCall(call, code);
1425 MUTEX_ENTER(&rx_stats_mutex);
1427 MUTEX_EXIT(&rx_stats_mutex);
1433 rx_WakeupServerProcs(void)
1435 struct rx_serverQueueEntry *np, *tqp;
1439 MUTEX_ENTER(&rx_serverPool_lock);
1441 #ifdef RX_ENABLE_LOCKS
1442 if (rx_waitForPacket)
1443 CV_BROADCAST(&rx_waitForPacket->cv);
1444 #else /* RX_ENABLE_LOCKS */
1445 if (rx_waitForPacket)
1446 osi_rxWakeup(rx_waitForPacket);
1447 #endif /* RX_ENABLE_LOCKS */
1448 MUTEX_ENTER(&freeSQEList_lock);
1449 for (np = rx_FreeSQEList; np; np = tqp) {
1450 tqp = *(struct rx_serverQueueEntry **)np;
1451 #ifdef RX_ENABLE_LOCKS
1452 CV_BROADCAST(&np->cv);
1453 #else /* RX_ENABLE_LOCKS */
1455 #endif /* RX_ENABLE_LOCKS */
1457 MUTEX_EXIT(&freeSQEList_lock);
1458 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1459 #ifdef RX_ENABLE_LOCKS
1460 CV_BROADCAST(&np->cv);
1461 #else /* RX_ENABLE_LOCKS */
1463 #endif /* RX_ENABLE_LOCKS */
1465 MUTEX_EXIT(&rx_serverPool_lock);
1470 * One thing that seems to happen is that all the server threads get
1471 * tied up on some empty or slow call, and then a whole bunch of calls
1472 * arrive at once, using up the packet pool, so now there are more
1473 * empty calls. The most critical resources here are server threads
1474 * and the free packet pool. The "doreclaim" code seems to help in
1475 * general. I think that eventually we arrive in this state: there
1476 * are lots of pending calls which do have all their packets present,
1477 * so they won't be reclaimed, are multi-packet calls, so they won't
1478 * be scheduled until later, and thus are tying up most of the free
1479 * packet pool for a very long time.
1481 * 1. schedule multi-packet calls if all the packets are present.
1482 * Probably CPU-bound operation, useful to return packets to pool.
1483 * Do what if there is a full window, but the last packet isn't here?
1484 * 3. preserve one thread which *only* runs "best" calls, otherwise
1485 * it sleeps and waits for that type of call.
1486 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1487 * the current dataquota business is badly broken. The quota isn't adjusted
1488 * to reflect how many packets are presently queued for a running call.
1489 * So, when we schedule a queued call with a full window of packets queued
1490 * up for it, that *should* free up a window full of packets for other 2d-class
1491 * calls to be able to use from the packet pool. But it doesn't.
1493 * NB. Most of the time, this code doesn't run -- since idle server threads
1494 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1495 * as a new call arrives.
1497 /* Sleep until a call arrives. Returns a pointer to the call, ready
1498 * for an rx_Read. */
1499 #ifdef RX_ENABLE_LOCKS
1501 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1503 struct rx_serverQueueEntry *sq;
1504 register struct rx_call *call = (struct rx_call *)0;
1505 struct rx_service *service = NULL;
1508 MUTEX_ENTER(&freeSQEList_lock);
1510 if ((sq = rx_FreeSQEList)) {
1511 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1512 MUTEX_EXIT(&freeSQEList_lock);
1513 } else { /* otherwise allocate a new one and return that */
1514 MUTEX_EXIT(&freeSQEList_lock);
1515 sq = (struct rx_serverQueueEntry *)
1516 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1517 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1518 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1521 MUTEX_ENTER(&rx_serverPool_lock);
1522 if (cur_service != NULL) {
1523 ReturnToServerPool(cur_service);
1526 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1527 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1529 /* Scan for eligible incoming calls. A call is not eligible
1530 * if the maximum number of calls for its service type are
1531 * already executing */
1532 /* One thread will process calls FCFS (to prevent starvation),
1533 * while the other threads may run ahead looking for calls which
1534 * have all their input data available immediately. This helps
1535 * keep threads from blocking, waiting for data from the client. */
1536 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1537 service = tcall->conn->service;
1538 if (!QuotaOK(service)) {
1541 if (tno == rxi_fcfs_thread_num
1542 || !tcall->queue_item_header.next) {
1543 /* If we're the fcfs thread , then we'll just use
1544 * this call. If we haven't been able to find an optimal
1545 * choice, and we're at the end of the list, then use a
1546 * 2d choice if one has been identified. Otherwise... */
1547 call = (choice2 ? choice2 : tcall);
1548 service = call->conn->service;
1549 } else if (!queue_IsEmpty(&tcall->rq)) {
1550 struct rx_packet *rp;
1551 rp = queue_First(&tcall->rq, rx_packet);
1552 if (rp->header.seq == 1) {
1554 || (rp->header.flags & RX_LAST_PACKET)) {
1556 } else if (rxi_2dchoice && !choice2
1557 && !(tcall->flags & RX_CALL_CLEARED)
1558 && (tcall->rprev > rxi_HardAckRate)) {
1567 ReturnToServerPool(service);
1574 MUTEX_EXIT(&rx_serverPool_lock);
1575 MUTEX_ENTER(&call->lock);
1577 if (call->flags & RX_CALL_WAIT_PROC) {
1578 call->flags &= ~RX_CALL_WAIT_PROC;
1579 MUTEX_ENTER(&rx_stats_mutex);
1581 MUTEX_EXIT(&rx_stats_mutex);
1584 if (call->state != RX_STATE_PRECALL || call->error) {
1585 MUTEX_EXIT(&call->lock);
1586 MUTEX_ENTER(&rx_serverPool_lock);
1587 ReturnToServerPool(service);
1592 if (queue_IsEmpty(&call->rq)
1593 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1594 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1596 CLEAR_CALL_QUEUE_LOCK(call);
1599 /* If there are no eligible incoming calls, add this process
1600 * to the idle server queue, to wait for one */
1604 *socketp = OSI_NULLSOCKET;
1606 sq->socketp = socketp;
1607 queue_Append(&rx_idleServerQueue, sq);
1608 #ifndef AFS_AIX41_ENV
1609 rx_waitForPacket = sq;
1611 rx_waitingForPacket = sq;
1612 #endif /* AFS_AIX41_ENV */
1614 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1616 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1617 MUTEX_EXIT(&rx_serverPool_lock);
1618 return (struct rx_call *)0;
1621 } while (!(call = sq->newcall)
1622 && !(socketp && *socketp != OSI_NULLSOCKET));
1623 MUTEX_EXIT(&rx_serverPool_lock);
1625 MUTEX_ENTER(&call->lock);
1631 MUTEX_ENTER(&freeSQEList_lock);
1632 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1633 rx_FreeSQEList = sq;
1634 MUTEX_EXIT(&freeSQEList_lock);
1637 clock_GetTime(&call->startTime);
1638 call->state = RX_STATE_ACTIVE;
1639 call->mode = RX_MODE_RECEIVING;
1640 #ifdef RX_KERNEL_TRACE
1641 if (ICL_SETACTIVE(afs_iclSetp)) {
1642 int glockOwner = ISAFS_GLOCK();
1645 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1646 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1653 rxi_calltrace(RX_CALL_START, call);
1654 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1655 call->conn->service->servicePort, call->conn->service->serviceId,
1658 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1659 MUTEX_EXIT(&call->lock);
1661 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1666 #else /* RX_ENABLE_LOCKS */
1668 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1670 struct rx_serverQueueEntry *sq;
1671 register struct rx_call *call = (struct rx_call *)0, *choice2;
1672 struct rx_service *service = NULL;
1676 MUTEX_ENTER(&freeSQEList_lock);
1678 if ((sq = rx_FreeSQEList)) {
1679 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1680 MUTEX_EXIT(&freeSQEList_lock);
1681 } else { /* otherwise allocate a new one and return that */
1682 MUTEX_EXIT(&freeSQEList_lock);
1683 sq = (struct rx_serverQueueEntry *)
1684 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1685 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1686 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1688 MUTEX_ENTER(&sq->lock);
1690 if (cur_service != NULL) {
1691 cur_service->nRequestsRunning--;
1692 if (cur_service->nRequestsRunning < cur_service->minProcs)
1696 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1697 register struct rx_call *tcall, *ncall;
1698 /* Scan for eligible incoming calls. A call is not eligible
1699 * if the maximum number of calls for its service type are
1700 * already executing */
1701 /* One thread will process calls FCFS (to prevent starvation),
1702 * while the other threads may run ahead looking for calls which
1703 * have all their input data available immediately. This helps
1704 * keep threads from blocking, waiting for data from the client. */
1705 choice2 = (struct rx_call *)0;
1706 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1707 service = tcall->conn->service;
1708 if (QuotaOK(service)) {
1709 if (tno == rxi_fcfs_thread_num
1710 || !tcall->queue_item_header.next) {
1711 /* If we're the fcfs thread, then we'll just use
1712 * this call. If we haven't been able to find an optimal
1713 * choice, and we're at the end of the list, then use a
1714 * 2d choice if one has been identified. Otherwise... */
1715 call = (choice2 ? choice2 : tcall);
1716 service = call->conn->service;
1717 } else if (!queue_IsEmpty(&tcall->rq)) {
1718 struct rx_packet *rp;
1719 rp = queue_First(&tcall->rq, rx_packet);
1720 if (rp->header.seq == 1
1722 || (rp->header.flags & RX_LAST_PACKET))) {
1724 } else if (rxi_2dchoice && !choice2
1725 && !(tcall->flags & RX_CALL_CLEARED)
1726 && (tcall->rprev > rxi_HardAckRate)) {
1739 /* we can't schedule a call if there's no data!!! */
1740 /* send an ack if there's no data, if we're missing the
1741 * first packet, or we're missing something between first
1742 * and last -- there's a "hole" in the incoming data. */
1743 if (queue_IsEmpty(&call->rq)
1744 || queue_First(&call->rq, rx_packet)->header.seq != 1
1745 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1746 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1748 call->flags &= (~RX_CALL_WAIT_PROC);
1749 service->nRequestsRunning++;
1750 /* just started call in minProcs pool, need fewer to maintain
1752 if (service->nRequestsRunning <= service->minProcs)
1756 /* MUTEX_EXIT(&call->lock); */
1758 /* If there are no eligible incoming calls, add this process
1759 * to the idle server queue, to wait for one */
1762 *socketp = OSI_NULLSOCKET;
1764 sq->socketp = socketp;
1765 queue_Append(&rx_idleServerQueue, sq);
1769 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1771 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1772 return (struct rx_call *)0;
1775 } while (!(call = sq->newcall)
1776 && !(socketp && *socketp != OSI_NULLSOCKET));
1778 MUTEX_EXIT(&sq->lock);
1780 MUTEX_ENTER(&freeSQEList_lock);
1781 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1782 rx_FreeSQEList = sq;
1783 MUTEX_EXIT(&freeSQEList_lock);
1786 clock_GetTime(&call->startTime);
1787 call->state = RX_STATE_ACTIVE;
1788 call->mode = RX_MODE_RECEIVING;
1789 #ifdef RX_KERNEL_TRACE
1790 if (ICL_SETACTIVE(afs_iclSetp)) {
1791 int glockOwner = ISAFS_GLOCK();
1794 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1795 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1802 rxi_calltrace(RX_CALL_START, call);
1803 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1804 call->conn->service->servicePort, call->conn->service->serviceId,
1807 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1814 #endif /* RX_ENABLE_LOCKS */
1818 /* Establish a procedure to be called when a packet arrives for a
1819 * call. This routine will be called at most once after each call,
1820 * and will also be called if there is an error condition on the or
1821 * the call is complete. Used by multi rx to build a selection
1822 * function which determines which of several calls is likely to be a
1823 * good one to read from.
1824 * NOTE: the way this is currently implemented it is probably only a
1825 * good idea to (1) use it immediately after a newcall (clients only)
1826 * and (2) only use it once. Other uses currently void your warranty
1829 rx_SetArrivalProc(register struct rx_call *call,
1830 register void (*proc) (register struct rx_call * call,
1832 register int index),
1833 register VOID * handle, register int arg)
1835 call->arrivalProc = proc;
1836 call->arrivalProcHandle = handle;
1837 call->arrivalProcArg = arg;
1840 /* Call is finished (possibly prematurely). Return rc to the peer, if
1841 * appropriate, and return the final error code from the conversation
1845 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1847 register struct rx_connection *conn = call->conn;
1848 register struct rx_service *service;
1854 dpf(("rx_EndCall(call %x rc %d error %d abortCode %d)\n", call, rc, call->error, call->abortCode));
1857 MUTEX_ENTER(&call->lock);
1859 if (rc == 0 && call->error == 0) {
1860 call->abortCode = 0;
1861 call->abortCount = 0;
1864 call->arrivalProc = (void (*)())0;
1865 if (rc && call->error == 0) {
1866 rxi_CallError(call, rc);
1867 /* Send an abort message to the peer if this error code has
1868 * only just been set. If it was set previously, assume the
1869 * peer has already been sent the error code or will request it
1871 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1873 if (conn->type == RX_SERVER_CONNECTION) {
1874 /* Make sure reply or at least dummy reply is sent */
1875 if (call->mode == RX_MODE_RECEIVING) {
1876 rxi_WriteProc(call, 0, 0);
1878 if (call->mode == RX_MODE_SENDING) {
1879 rxi_FlushWrite(call);
1881 service = conn->service;
1882 rxi_calltrace(RX_CALL_END, call);
1883 /* Call goes to hold state until reply packets are acknowledged */
1884 if (call->tfirst + call->nSoftAcked < call->tnext) {
1885 call->state = RX_STATE_HOLD;
1887 call->state = RX_STATE_DALLY;
1888 rxi_ClearTransmitQueue(call, 0);
1889 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1890 rxevent_Cancel(call->keepAliveEvent, call,
1891 RX_CALL_REFCOUNT_ALIVE);
1893 } else { /* Client connection */
1895 /* Make sure server receives input packets, in the case where
1896 * no reply arguments are expected */
1897 if ((call->mode == RX_MODE_SENDING)
1898 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1899 (void)rxi_ReadProc(call, &dummy, 1);
1902 /* If we had an outstanding delayed ack, be nice to the server
1903 * and force-send it now.
1905 if (call->delayedAckEvent) {
1906 rxevent_Cancel(call->delayedAckEvent, call,
1907 RX_CALL_REFCOUNT_DELAY);
1908 call->delayedAckEvent = NULL;
1909 rxi_SendDelayedAck(NULL, call, NULL);
1912 /* We need to release the call lock since it's lower than the
1913 * conn_call_lock and we don't want to hold the conn_call_lock
1914 * over the rx_ReadProc call. The conn_call_lock needs to be held
1915 * here for the case where rx_NewCall is perusing the calls on
1916 * the connection structure. We don't want to signal until
1917 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1918 * have checked this call, found it active and by the time it
1919 * goes to sleep, will have missed the signal.
1921 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
1922 * there are threads waiting to use the conn object.
1924 MUTEX_EXIT(&call->lock);
1925 MUTEX_ENTER(&conn->conn_call_lock);
1926 MUTEX_ENTER(&call->lock);
1927 MUTEX_ENTER(&conn->conn_data_lock);
1928 conn->flags |= RX_CONN_BUSY;
1929 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1930 if (conn->makeCallWaiters == 0)
1931 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1932 MUTEX_EXIT(&conn->conn_data_lock);
1933 #ifdef RX_ENABLE_LOCKS
1934 CV_BROADCAST(&conn->conn_call_cv);
1939 #ifdef RX_ENABLE_LOCKS
1941 MUTEX_EXIT(&conn->conn_data_lock);
1943 #endif /* RX_ENABLE_LOCKS */
1944 call->state = RX_STATE_DALLY;
1946 error = call->error;
1948 /* currentPacket, nLeft, and NFree must be zeroed here, because
1949 * ResetCall cannot: ResetCall may be called at splnet(), in the
1950 * kernel version, and may interrupt the macros rx_Read or
1951 * rx_Write, which run at normal priority for efficiency. */
1952 if (call->currentPacket) {
1953 queue_Prepend(&call->iovq, call->currentPacket);
1954 call->currentPacket = (struct rx_packet *)0;
1957 call->nLeft = call->nFree = call->curlen = 0;
1959 /* Free any packets from the last call to ReadvProc/WritevProc */
1960 rxi_FreePackets(0, &call->iovq);
1962 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1963 MUTEX_EXIT(&call->lock);
1964 if (conn->type == RX_CLIENT_CONNECTION) {
1965 MUTEX_EXIT(&conn->conn_call_lock);
1966 conn->flags &= ~RX_CONN_BUSY;
1970 * Map errors to the local host's errno.h format.
1972 error = ntoh_syserr_conv(error);
1976 #if !defined(KERNEL)
1978 /* Call this routine when shutting down a server or client (especially
1979 * clients). This will allow Rx to gracefully garbage collect server
1980 * connections, and reduce the number of retries that a server might
1981 * make to a dead client.
1982 * This is not quite right, since some calls may still be ongoing and
1983 * we can't lock them to destroy them. */
1987 register struct rx_connection **conn_ptr, **conn_end;
1991 if (rxinit_status == 1) {
1993 return; /* Already shutdown. */
1995 rxi_DeleteCachedConnections();
1996 if (rx_connHashTable) {
1997 MUTEX_ENTER(&rx_connHashTable_lock);
1998 for (conn_ptr = &rx_connHashTable[0], conn_end =
1999 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2001 struct rx_connection *conn, *next;
2002 for (conn = *conn_ptr; conn; conn = next) {
2004 if (conn->type == RX_CLIENT_CONNECTION) {
2005 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2007 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2008 #ifdef RX_ENABLE_LOCKS
2009 rxi_DestroyConnectionNoLock(conn);
2010 #else /* RX_ENABLE_LOCKS */
2011 rxi_DestroyConnection(conn);
2012 #endif /* RX_ENABLE_LOCKS */
2016 #ifdef RX_ENABLE_LOCKS
2017 while (rx_connCleanup_list) {
2018 struct rx_connection *conn;
2019 conn = rx_connCleanup_list;
2020 rx_connCleanup_list = rx_connCleanup_list->next;
2021 MUTEX_EXIT(&rx_connHashTable_lock);
2022 rxi_CleanupConnection(conn);
2023 MUTEX_ENTER(&rx_connHashTable_lock);
2025 MUTEX_EXIT(&rx_connHashTable_lock);
2026 #endif /* RX_ENABLE_LOCKS */
2031 afs_winsockCleanup();
2039 /* if we wakeup packet waiter too often, can get in loop with two
2040 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2042 rxi_PacketsUnWait(void)
2044 if (!rx_waitingForPackets) {
2048 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2049 return; /* still over quota */
2052 rx_waitingForPackets = 0;
2053 #ifdef RX_ENABLE_LOCKS
2054 CV_BROADCAST(&rx_waitingForPackets_cv);
2056 osi_rxWakeup(&rx_waitingForPackets);
2062 /* ------------------Internal interfaces------------------------- */
2064 /* Return this process's service structure for the
2065 * specified socket and service */
2067 rxi_FindService(register osi_socket socket, register u_short serviceId)
2069 register struct rx_service **sp;
2070 for (sp = &rx_services[0]; *sp; sp++) {
2071 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2077 /* Allocate a call structure, for the indicated channel of the
2078 * supplied connection. The mode and state of the call must be set by
2079 * the caller. Returns the call with mutex locked. */
2081 rxi_NewCall(register struct rx_connection *conn, register int channel)
2083 register struct rx_call *call;
2084 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2085 register struct rx_call *cp; /* Call pointer temp */
2086 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2087 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2089 dpf(("rxi_NewCall(conn %x, channel %d)\n", conn, channel));
2091 /* Grab an existing call structure, or allocate a new one.
2092 * Existing call structures are assumed to have been left reset by
2094 MUTEX_ENTER(&rx_freeCallQueue_lock);
2096 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2098 * EXCEPT that the TQ might not yet be cleared out.
2099 * Skip over those with in-use TQs.
2102 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2103 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2109 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2110 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2111 call = queue_First(&rx_freeCallQueue, rx_call);
2112 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2114 MUTEX_ENTER(&rx_stats_mutex);
2115 rx_stats.nFreeCallStructs--;
2116 MUTEX_EXIT(&rx_stats_mutex);
2117 MUTEX_EXIT(&rx_freeCallQueue_lock);
2118 MUTEX_ENTER(&call->lock);
2119 CLEAR_CALL_QUEUE_LOCK(call);
2120 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2121 /* Now, if TQ wasn't cleared earlier, do it now. */
2122 if (call->flags & RX_CALL_TQ_CLEARME) {
2123 rxi_ClearTransmitQueue(call, 0);
2124 queue_Init(&call->tq);
2126 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2127 /* Bind the call to its connection structure */
2129 rxi_ResetCall(call, 1);
2131 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2133 MUTEX_EXIT(&rx_freeCallQueue_lock);
2134 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2135 MUTEX_ENTER(&call->lock);
2136 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2137 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2138 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2140 MUTEX_ENTER(&rx_stats_mutex);
2141 rx_stats.nCallStructs++;
2142 MUTEX_EXIT(&rx_stats_mutex);
2143 /* Initialize once-only items */
2144 queue_Init(&call->tq);
2145 queue_Init(&call->rq);
2146 queue_Init(&call->iovq);
2147 /* Bind the call to its connection structure (prereq for reset) */
2149 rxi_ResetCall(call, 1);
2151 call->channel = channel;
2152 call->callNumber = &conn->callNumber[channel];
2153 /* Note that the next expected call number is retained (in
2154 * conn->callNumber[i]), even if we reallocate the call structure
2156 conn->call[channel] = call;
2157 /* if the channel's never been used (== 0), we should start at 1, otherwise
2158 * the call number is valid from the last time this channel was used */
2159 if (*call->callNumber == 0)
2160 *call->callNumber = 1;
2165 /* A call has been inactive long enough that so we can throw away
2166 * state, including the call structure, which is placed on the call
2168 * Call is locked upon entry.
2169 * haveCTLock set if called from rxi_ReapConnections
2171 #ifdef RX_ENABLE_LOCKS
2173 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2174 #else /* RX_ENABLE_LOCKS */
2176 rxi_FreeCall(register struct rx_call *call)
2177 #endif /* RX_ENABLE_LOCKS */
2179 register int channel = call->channel;
2180 register struct rx_connection *conn = call->conn;
2183 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2184 (*call->callNumber)++;
2185 rxi_ResetCall(call, 0);
2186 call->conn->call[channel] = (struct rx_call *)0;
2188 MUTEX_ENTER(&rx_freeCallQueue_lock);
2189 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2190 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2191 /* A call may be free even though its transmit queue is still in use.
2192 * Since we search the call list from head to tail, put busy calls at
2193 * the head of the list, and idle calls at the tail.
2195 if (call->flags & RX_CALL_TQ_BUSY)
2196 queue_Prepend(&rx_freeCallQueue, call);
2198 queue_Append(&rx_freeCallQueue, call);
2199 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2200 queue_Append(&rx_freeCallQueue, call);
2201 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2202 MUTEX_ENTER(&rx_stats_mutex);
2203 rx_stats.nFreeCallStructs++;
2204 MUTEX_EXIT(&rx_stats_mutex);
2206 MUTEX_EXIT(&rx_freeCallQueue_lock);
2208 /* Destroy the connection if it was previously slated for
2209 * destruction, i.e. the Rx client code previously called
2210 * rx_DestroyConnection (client connections), or
2211 * rxi_ReapConnections called the same routine (server
2212 * connections). Only do this, however, if there are no
2213 * outstanding calls. Note that for fine grain locking, there appears
2214 * to be a deadlock in that rxi_FreeCall has a call locked and
2215 * DestroyConnectionNoLock locks each call in the conn. But note a
2216 * few lines up where we have removed this call from the conn.
2217 * If someone else destroys a connection, they either have no
2218 * call lock held or are going through this section of code.
2220 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2221 MUTEX_ENTER(&conn->conn_data_lock);
2223 MUTEX_EXIT(&conn->conn_data_lock);
2224 #ifdef RX_ENABLE_LOCKS
2226 rxi_DestroyConnectionNoLock(conn);
2228 rxi_DestroyConnection(conn);
2229 #else /* RX_ENABLE_LOCKS */
2230 rxi_DestroyConnection(conn);
2231 #endif /* RX_ENABLE_LOCKS */
2235 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2237 rxi_Alloc(register size_t size)
2241 MUTEX_ENTER(&rx_stats_mutex);
2243 rxi_Allocsize += (afs_int32)size;
2244 MUTEX_EXIT(&rx_stats_mutex);
2246 p = (char *)osi_Alloc(size);
2249 osi_Panic("rxi_Alloc error");
2255 rxi_Free(void *addr, register size_t size)
2257 MUTEX_ENTER(&rx_stats_mutex);
2259 rxi_Allocsize -= (afs_int32)size;
2260 MUTEX_EXIT(&rx_stats_mutex);
2262 osi_Free(addr, size);
2265 /* Find the peer process represented by the supplied (host,port)
2266 * combination. If there is no appropriate active peer structure, a
2267 * new one will be allocated and initialized
2268 * The origPeer, if set, is a pointer to a peer structure on which the
2269 * refcount will be be decremented. This is used to replace the peer
2270 * structure hanging off a connection structure */
2272 rxi_FindPeer(register afs_uint32 host, register u_short port,
2273 struct rx_peer *origPeer, int create)
2275 register struct rx_peer *pp;
2277 hashIndex = PEER_HASH(host, port);
2278 MUTEX_ENTER(&rx_peerHashTable_lock);
2279 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2280 if ((pp->host == host) && (pp->port == port))
2285 pp = rxi_AllocPeer(); /* This bzero's *pp */
2286 pp->host = host; /* set here or in InitPeerParams is zero */
2288 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2289 queue_Init(&pp->congestionQueue);
2290 queue_Init(&pp->rpcStats);
2291 pp->next = rx_peerHashTable[hashIndex];
2292 rx_peerHashTable[hashIndex] = pp;
2293 rxi_InitPeerParams(pp);
2294 MUTEX_ENTER(&rx_stats_mutex);
2295 rx_stats.nPeerStructs++;
2296 MUTEX_EXIT(&rx_stats_mutex);
2303 origPeer->refCount--;
2304 MUTEX_EXIT(&rx_peerHashTable_lock);
2309 /* Find the connection at (host, port) started at epoch, and with the
2310 * given connection id. Creates the server connection if necessary.
2311 * The type specifies whether a client connection or a server
2312 * connection is desired. In both cases, (host, port) specify the
2313 * peer's (host, pair) pair. Client connections are not made
2314 * automatically by this routine. The parameter socket gives the
2315 * socket descriptor on which the packet was received. This is used,
2316 * in the case of server connections, to check that *new* connections
2317 * come via a valid (port, serviceId). Finally, the securityIndex
2318 * parameter must match the existing index for the connection. If a
2319 * server connection is created, it will be created using the supplied
2320 * index, if the index is valid for this service */
2321 struct rx_connection *
2322 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2323 register u_short port, u_short serviceId, afs_uint32 cid,
2324 afs_uint32 epoch, int type, u_int securityIndex)
2326 int hashindex, flag;
2327 register struct rx_connection *conn;
2328 hashindex = CONN_HASH(host, port, cid, epoch, type);
2329 MUTEX_ENTER(&rx_connHashTable_lock);
2330 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2331 rx_connHashTable[hashindex],
2334 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2335 && (epoch == conn->epoch)) {
2336 register struct rx_peer *pp = conn->peer;
2337 if (securityIndex != conn->securityIndex) {
2338 /* this isn't supposed to happen, but someone could forge a packet
2339 * like this, and there seems to be some CM bug that makes this
2340 * happen from time to time -- in which case, the fileserver
2342 MUTEX_EXIT(&rx_connHashTable_lock);
2343 return (struct rx_connection *)0;
2345 if (pp->host == host && pp->port == port)
2347 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2349 /* So what happens when it's a callback connection? */
2350 if ( /*type == RX_CLIENT_CONNECTION && */
2351 (conn->epoch & 0x80000000))
2355 /* the connection rxLastConn that was used the last time is not the
2356 ** one we are looking for now. Hence, start searching in the hash */
2358 conn = rx_connHashTable[hashindex];
2363 struct rx_service *service;
2364 if (type == RX_CLIENT_CONNECTION) {
2365 MUTEX_EXIT(&rx_connHashTable_lock);
2366 return (struct rx_connection *)0;
2368 service = rxi_FindService(socket, serviceId);
2369 if (!service || (securityIndex >= service->nSecurityObjects)
2370 || (service->securityObjects[securityIndex] == 0)) {
2371 MUTEX_EXIT(&rx_connHashTable_lock);
2372 return (struct rx_connection *)0;
2374 conn = rxi_AllocConnection(); /* This bzero's the connection */
2375 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2376 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2377 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2378 conn->next = rx_connHashTable[hashindex];
2379 rx_connHashTable[hashindex] = conn;
2380 conn->peer = rxi_FindPeer(host, port, 0, 1);
2381 conn->type = RX_SERVER_CONNECTION;
2382 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2383 conn->epoch = epoch;
2384 conn->cid = cid & RX_CIDMASK;
2385 /* conn->serial = conn->lastSerial = 0; */
2386 /* conn->timeout = 0; */
2387 conn->ackRate = RX_FAST_ACK_RATE;
2388 conn->service = service;
2389 conn->serviceId = serviceId;
2390 conn->securityIndex = securityIndex;
2391 conn->securityObject = service->securityObjects[securityIndex];
2392 conn->nSpecific = 0;
2393 conn->specific = NULL;
2394 rx_SetConnDeadTime(conn, service->connDeadTime);
2395 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2396 /* Notify security object of the new connection */
2397 RXS_NewConnection(conn->securityObject, conn);
2398 /* XXXX Connection timeout? */
2399 if (service->newConnProc)
2400 (*service->newConnProc) (conn);
2401 MUTEX_ENTER(&rx_stats_mutex);
2402 rx_stats.nServerConns++;
2403 MUTEX_EXIT(&rx_stats_mutex);
2406 MUTEX_ENTER(&conn->conn_data_lock);
2408 MUTEX_EXIT(&conn->conn_data_lock);
2410 rxLastConn = conn; /* store this connection as the last conn used */
2411 MUTEX_EXIT(&rx_connHashTable_lock);
2415 /* There are two packet tracing routines available for testing and monitoring
2416 * Rx. One is called just after every packet is received and the other is
2417 * called just before every packet is sent. Received packets, have had their
2418 * headers decoded, and packets to be sent have not yet had their headers
2419 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2420 * containing the network address. Both can be modified. The return value, if
2421 * non-zero, indicates that the packet should be dropped. */
2423 int (*rx_justReceived) () = 0;
2424 int (*rx_almostSent) () = 0;
2426 /* A packet has been received off the interface. Np is the packet, socket is
2427 * the socket number it was received from (useful in determining which service
2428 * this packet corresponds to), and (host, port) reflect the host,port of the
2429 * sender. This call returns the packet to the caller if it is finished with
2430 * it, rather than de-allocating it, just as a small performance hack */
2433 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2434 afs_uint32 host, u_short port, int *tnop,
2435 struct rx_call **newcallp)
2437 register struct rx_call *call;
2438 register struct rx_connection *conn;
2440 afs_uint32 currentCallNumber;
2446 struct rx_packet *tnp;
2449 /* We don't print out the packet until now because (1) the time may not be
2450 * accurate enough until now in the lwp implementation (rx_Listener only gets
2451 * the time after the packet is read) and (2) from a protocol point of view,
2452 * this is the first time the packet has been seen */
2453 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2454 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2455 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2456 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2457 np->header.epoch, np->header.cid, np->header.callNumber,
2458 np->header.seq, np->header.flags, np));
2461 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2462 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2465 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2466 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2469 /* If an input tracer function is defined, call it with the packet and
2470 * network address. Note this function may modify its arguments. */
2471 if (rx_justReceived) {
2472 struct sockaddr_in addr;
2474 addr.sin_family = AF_INET;
2475 addr.sin_port = port;
2476 addr.sin_addr.s_addr = host;
2477 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2478 addr.sin_len = sizeof(addr);
2479 #endif /* AFS_OSF_ENV */
2480 drop = (*rx_justReceived) (np, &addr);
2481 /* drop packet if return value is non-zero */
2484 port = addr.sin_port; /* in case fcn changed addr */
2485 host = addr.sin_addr.s_addr;
2489 /* If packet was not sent by the client, then *we* must be the client */
2490 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2491 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2493 /* Find the connection (or fabricate one, if we're the server & if
2494 * necessary) associated with this packet */
2496 rxi_FindConnection(socket, host, port, np->header.serviceId,
2497 np->header.cid, np->header.epoch, type,
2498 np->header.securityIndex);
2501 /* If no connection found or fabricated, just ignore the packet.
2502 * (An argument could be made for sending an abort packet for
2507 MUTEX_ENTER(&conn->conn_data_lock);
2508 if (conn->maxSerial < np->header.serial)
2509 conn->maxSerial = np->header.serial;
2510 MUTEX_EXIT(&conn->conn_data_lock);
2512 /* If the connection is in an error state, send an abort packet and ignore
2513 * the incoming packet */
2515 /* Don't respond to an abort packet--we don't want loops! */
2516 MUTEX_ENTER(&conn->conn_data_lock);
2517 if (np->header.type != RX_PACKET_TYPE_ABORT)
2518 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2520 MUTEX_EXIT(&conn->conn_data_lock);
2524 /* Check for connection-only requests (i.e. not call specific). */
2525 if (np->header.callNumber == 0) {
2526 switch (np->header.type) {
2527 case RX_PACKET_TYPE_ABORT: {
2528 /* What if the supplied error is zero? */
2529 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2530 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2531 rxi_ConnectionError(conn, errcode);
2532 MUTEX_ENTER(&conn->conn_data_lock);
2534 MUTEX_EXIT(&conn->conn_data_lock);
2537 case RX_PACKET_TYPE_CHALLENGE:
2538 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2539 MUTEX_ENTER(&conn->conn_data_lock);
2541 MUTEX_EXIT(&conn->conn_data_lock);
2543 case RX_PACKET_TYPE_RESPONSE:
2544 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2545 MUTEX_ENTER(&conn->conn_data_lock);
2547 MUTEX_EXIT(&conn->conn_data_lock);
2549 case RX_PACKET_TYPE_PARAMS:
2550 case RX_PACKET_TYPE_PARAMS + 1:
2551 case RX_PACKET_TYPE_PARAMS + 2:
2552 /* ignore these packet types for now */
2553 MUTEX_ENTER(&conn->conn_data_lock);
2555 MUTEX_EXIT(&conn->conn_data_lock);
2560 /* Should not reach here, unless the peer is broken: send an
2562 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2563 MUTEX_ENTER(&conn->conn_data_lock);
2564 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2566 MUTEX_EXIT(&conn->conn_data_lock);
2571 channel = np->header.cid & RX_CHANNELMASK;
2572 call = conn->call[channel];
2573 #ifdef RX_ENABLE_LOCKS
2575 MUTEX_ENTER(&call->lock);
2576 /* Test to see if call struct is still attached to conn. */
2577 if (call != conn->call[channel]) {
2579 MUTEX_EXIT(&call->lock);
2580 if (type == RX_SERVER_CONNECTION) {
2581 call = conn->call[channel];
2582 /* If we started with no call attached and there is one now,
2583 * another thread is also running this routine and has gotten
2584 * the connection channel. We should drop this packet in the tests
2585 * below. If there was a call on this connection and it's now
2586 * gone, then we'll be making a new call below.
2587 * If there was previously a call and it's now different then
2588 * the old call was freed and another thread running this routine
2589 * has created a call on this channel. One of these two threads
2590 * has a packet for the old call and the code below handles those
2594 MUTEX_ENTER(&call->lock);
2596 /* This packet can't be for this call. If the new call address is
2597 * 0 then no call is running on this channel. If there is a call
2598 * then, since this is a client connection we're getting data for
2599 * it must be for the previous call.
2601 MUTEX_ENTER(&rx_stats_mutex);
2602 rx_stats.spuriousPacketsRead++;
2603 MUTEX_EXIT(&rx_stats_mutex);
2604 MUTEX_ENTER(&conn->conn_data_lock);
2606 MUTEX_EXIT(&conn->conn_data_lock);
2611 currentCallNumber = conn->callNumber[channel];
2613 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2614 if (np->header.callNumber < currentCallNumber) {
2615 MUTEX_ENTER(&rx_stats_mutex);
2616 rx_stats.spuriousPacketsRead++;
2617 MUTEX_EXIT(&rx_stats_mutex);
2618 #ifdef RX_ENABLE_LOCKS
2620 MUTEX_EXIT(&call->lock);
2622 MUTEX_ENTER(&conn->conn_data_lock);
2624 MUTEX_EXIT(&conn->conn_data_lock);
2628 MUTEX_ENTER(&conn->conn_call_lock);
2629 call = rxi_NewCall(conn, channel);
2630 MUTEX_EXIT(&conn->conn_call_lock);
2631 *call->callNumber = np->header.callNumber;
2632 if (np->header.callNumber == 0)
2633 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));
2635 call->state = RX_STATE_PRECALL;
2636 clock_GetTime(&call->queueTime);
2637 hzero(call->bytesSent);
2638 hzero(call->bytesRcvd);
2640 * If the number of queued calls exceeds the overload
2641 * threshold then abort this call.
2643 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2644 struct rx_packet *tp;
2646 rxi_CallError(call, rx_BusyError);
2647 tp = rxi_SendCallAbort(call, np, 1, 0);
2648 MUTEX_EXIT(&call->lock);
2649 MUTEX_ENTER(&conn->conn_data_lock);
2651 MUTEX_EXIT(&conn->conn_data_lock);
2652 MUTEX_ENTER(&rx_stats_mutex);
2654 MUTEX_EXIT(&rx_stats_mutex);
2657 rxi_KeepAliveOn(call);
2658 } else if (np->header.callNumber != currentCallNumber) {
2659 /* Wait until the transmit queue is idle before deciding
2660 * whether to reset the current call. Chances are that the
2661 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2664 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2665 while ((call->state == RX_STATE_ACTIVE)
2666 && (call->flags & RX_CALL_TQ_BUSY)) {
2667 call->flags |= RX_CALL_TQ_WAIT;
2669 #ifdef RX_ENABLE_LOCKS
2670 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2671 CV_WAIT(&call->cv_tq, &call->lock);
2672 #else /* RX_ENABLE_LOCKS */
2673 osi_rxSleep(&call->tq);
2674 #endif /* RX_ENABLE_LOCKS */
2676 if (call->tqWaiters == 0)
2677 call->flags &= ~RX_CALL_TQ_WAIT;
2679 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2680 /* If the new call cannot be taken right now send a busy and set
2681 * the error condition in this call, so that it terminates as
2682 * quickly as possible */
2683 if (call->state == RX_STATE_ACTIVE) {
2684 struct rx_packet *tp;
2686 rxi_CallError(call, RX_CALL_DEAD);
2687 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2689 MUTEX_EXIT(&call->lock);
2690 MUTEX_ENTER(&conn->conn_data_lock);
2692 MUTEX_EXIT(&conn->conn_data_lock);
2695 rxi_ResetCall(call, 0);
2696 *call->callNumber = np->header.callNumber;
2697 if (np->header.callNumber == 0)
2698 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));
2700 call->state = RX_STATE_PRECALL;
2701 clock_GetTime(&call->queueTime);
2702 hzero(call->bytesSent);
2703 hzero(call->bytesRcvd);
2705 * If the number of queued calls exceeds the overload
2706 * threshold then abort this call.
2708 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2709 struct rx_packet *tp;
2711 rxi_CallError(call, rx_BusyError);
2712 tp = rxi_SendCallAbort(call, np, 1, 0);
2713 MUTEX_EXIT(&call->lock);
2714 MUTEX_ENTER(&conn->conn_data_lock);
2716 MUTEX_EXIT(&conn->conn_data_lock);
2717 MUTEX_ENTER(&rx_stats_mutex);
2719 MUTEX_EXIT(&rx_stats_mutex);
2722 rxi_KeepAliveOn(call);
2724 /* Continuing call; do nothing here. */
2726 } else { /* we're the client */
2727 /* Ignore all incoming acknowledgements for calls in DALLY state */
2728 if (call && (call->state == RX_STATE_DALLY)
2729 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2730 MUTEX_ENTER(&rx_stats_mutex);
2731 rx_stats.ignorePacketDally++;
2732 MUTEX_EXIT(&rx_stats_mutex);
2733 #ifdef RX_ENABLE_LOCKS
2735 MUTEX_EXIT(&call->lock);
2738 MUTEX_ENTER(&conn->conn_data_lock);
2740 MUTEX_EXIT(&conn->conn_data_lock);
2744 /* Ignore anything that's not relevant to the current call. If there
2745 * isn't a current call, then no packet is relevant. */
2746 if (!call || (np->header.callNumber != currentCallNumber)) {
2747 MUTEX_ENTER(&rx_stats_mutex);
2748 rx_stats.spuriousPacketsRead++;
2749 MUTEX_EXIT(&rx_stats_mutex);
2750 #ifdef RX_ENABLE_LOCKS
2752 MUTEX_EXIT(&call->lock);
2755 MUTEX_ENTER(&conn->conn_data_lock);
2757 MUTEX_EXIT(&conn->conn_data_lock);
2760 /* If the service security object index stamped in the packet does not
2761 * match the connection's security index, ignore the packet */
2762 if (np->header.securityIndex != conn->securityIndex) {
2763 #ifdef RX_ENABLE_LOCKS
2764 MUTEX_EXIT(&call->lock);
2766 MUTEX_ENTER(&conn->conn_data_lock);
2768 MUTEX_EXIT(&conn->conn_data_lock);
2772 /* If we're receiving the response, then all transmit packets are
2773 * implicitly acknowledged. Get rid of them. */
2774 if (np->header.type == RX_PACKET_TYPE_DATA) {
2775 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2776 /* XXX Hack. Because we must release the global rx lock when
2777 * sending packets (osi_NetSend) we drop all acks while we're
2778 * traversing the tq in rxi_Start sending packets out because
2779 * packets may move to the freePacketQueue as result of being here!
2780 * So we drop these packets until we're safely out of the
2781 * traversing. Really ugly!
2782 * For fine grain RX locking, we set the acked field in the
2783 * packets and let rxi_Start remove them from the transmit queue.
2785 if (call->flags & RX_CALL_TQ_BUSY) {
2786 #ifdef RX_ENABLE_LOCKS
2787 rxi_SetAcksInTransmitQueue(call);
2790 return np; /* xmitting; drop packet */
2793 rxi_ClearTransmitQueue(call, 0);
2795 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2796 rxi_ClearTransmitQueue(call, 0);
2797 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2799 if (np->header.type == RX_PACKET_TYPE_ACK) {
2800 /* now check to see if this is an ack packet acknowledging that the
2801 * server actually *lost* some hard-acked data. If this happens we
2802 * ignore this packet, as it may indicate that the server restarted in
2803 * the middle of a call. It is also possible that this is an old ack
2804 * packet. We don't abort the connection in this case, because this
2805 * *might* just be an old ack packet. The right way to detect a server
2806 * restart in the midst of a call is to notice that the server epoch
2808 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2809 * XXX unacknowledged. I think that this is off-by-one, but
2810 * XXX I don't dare change it just yet, since it will
2811 * XXX interact badly with the server-restart detection
2812 * XXX code in receiveackpacket. */
2813 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2814 MUTEX_ENTER(&rx_stats_mutex);
2815 rx_stats.spuriousPacketsRead++;
2816 MUTEX_EXIT(&rx_stats_mutex);
2817 MUTEX_EXIT(&call->lock);
2818 MUTEX_ENTER(&conn->conn_data_lock);
2820 MUTEX_EXIT(&conn->conn_data_lock);
2824 } /* else not a data packet */
2827 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2828 /* Set remote user defined status from packet */
2829 call->remoteStatus = np->header.userStatus;
2831 /* Note the gap between the expected next packet and the actual
2832 * packet that arrived, when the new packet has a smaller serial number
2833 * than expected. Rioses frequently reorder packets all by themselves,
2834 * so this will be quite important with very large window sizes.
2835 * Skew is checked against 0 here to avoid any dependence on the type of
2836 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2838 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2839 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2840 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2842 MUTEX_ENTER(&conn->conn_data_lock);
2843 skew = conn->lastSerial - np->header.serial;
2844 conn->lastSerial = np->header.serial;
2845 MUTEX_EXIT(&conn->conn_data_lock);
2847 register struct rx_peer *peer;
2849 if (skew > peer->inPacketSkew) {
2850 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2852 peer->inPacketSkew = skew;
2856 /* Now do packet type-specific processing */
2857 switch (np->header.type) {
2858 case RX_PACKET_TYPE_DATA:
2859 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2862 case RX_PACKET_TYPE_ACK:
2863 /* Respond immediately to ack packets requesting acknowledgement
2865 if (np->header.flags & RX_REQUEST_ACK) {
2867 (void)rxi_SendCallAbort(call, 0, 1, 0);
2869 (void)rxi_SendAck(call, 0, np->header.serial,
2870 RX_ACK_PING_RESPONSE, 1);
2872 np = rxi_ReceiveAckPacket(call, np, 1);
2874 case RX_PACKET_TYPE_ABORT: {
2875 /* An abort packet: reset the call, passing the error up to the user. */
2876 /* What if error is zero? */
2877 /* What if the error is -1? the application will treat it as a timeout. */
2878 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
2879 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
2880 rxi_CallError(call, errdata);
2881 MUTEX_EXIT(&call->lock);
2882 MUTEX_ENTER(&conn->conn_data_lock);
2884 MUTEX_EXIT(&conn->conn_data_lock);
2885 return np; /* xmitting; drop packet */
2887 case RX_PACKET_TYPE_BUSY:
2890 case RX_PACKET_TYPE_ACKALL:
2891 /* All packets acknowledged, so we can drop all packets previously
2892 * readied for sending */
2893 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2894 /* XXX Hack. We because we can't release the global rx lock when
2895 * sending packets (osi_NetSend) we drop all ack pkts while we're
2896 * traversing the tq in rxi_Start sending packets out because
2897 * packets may move to the freePacketQueue as result of being
2898 * here! So we drop these packets until we're safely out of the
2899 * traversing. Really ugly!
2900 * For fine grain RX locking, we set the acked field in the packets
2901 * and let rxi_Start remove the packets from the transmit queue.
2903 if (call->flags & RX_CALL_TQ_BUSY) {
2904 #ifdef RX_ENABLE_LOCKS
2905 rxi_SetAcksInTransmitQueue(call);
2907 #else /* RX_ENABLE_LOCKS */
2908 MUTEX_EXIT(&call->lock);
2909 MUTEX_ENTER(&conn->conn_data_lock);
2911 MUTEX_EXIT(&conn->conn_data_lock);
2912 return np; /* xmitting; drop packet */
2913 #endif /* RX_ENABLE_LOCKS */
2915 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2916 rxi_ClearTransmitQueue(call, 0);
2919 /* Should not reach here, unless the peer is broken: send an abort
2921 rxi_CallError(call, RX_PROTOCOL_ERROR);
2922 np = rxi_SendCallAbort(call, np, 1, 0);
2925 /* Note when this last legitimate packet was received, for keep-alive
2926 * processing. Note, we delay getting the time until now in the hope that
2927 * the packet will be delivered to the user before any get time is required
2928 * (if not, then the time won't actually be re-evaluated here). */
2929 call->lastReceiveTime = clock_Sec();
2930 MUTEX_EXIT(&call->lock);
2931 MUTEX_ENTER(&conn->conn_data_lock);
2933 MUTEX_EXIT(&conn->conn_data_lock);
2937 /* return true if this is an "interesting" connection from the point of view
2938 of someone trying to debug the system */
2940 rxi_IsConnInteresting(struct rx_connection *aconn)
2943 register struct rx_call *tcall;
2945 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2947 for (i = 0; i < RX_MAXCALLS; i++) {
2948 tcall = aconn->call[i];
2950 if ((tcall->state == RX_STATE_PRECALL)
2951 || (tcall->state == RX_STATE_ACTIVE))
2953 if ((tcall->mode == RX_MODE_SENDING)
2954 || (tcall->mode == RX_MODE_RECEIVING))
2962 /* if this is one of the last few packets AND it wouldn't be used by the
2963 receiving call to immediately satisfy a read request, then drop it on
2964 the floor, since accepting it might prevent a lock-holding thread from
2965 making progress in its reading. If a call has been cleared while in
2966 the precall state then ignore all subsequent packets until the call
2967 is assigned to a thread. */
2970 TooLow(struct rx_packet *ap, struct rx_call *acall)
2973 MUTEX_ENTER(&rx_stats_mutex);
2974 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
2975 && (acall->state == RX_STATE_PRECALL))
2976 || ((rx_nFreePackets < rxi_dataQuota + 2)
2977 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
2978 && (acall->flags & RX_CALL_READER_WAIT)))) {
2981 MUTEX_EXIT(&rx_stats_mutex);
2987 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
2988 struct rx_call *acall)
2990 struct rx_call *call = acall;
2994 MUTEX_ENTER(&conn->conn_data_lock);
2995 conn->checkReachEvent = NULL;
2996 waiting = conn->flags & RX_CONN_ATTACHWAIT;
2999 MUTEX_EXIT(&conn->conn_data_lock);
3003 MUTEX_ENTER(&conn->conn_call_lock);
3004 MUTEX_ENTER(&conn->conn_data_lock);
3005 for (i = 0; i < RX_MAXCALLS; i++) {
3006 struct rx_call *tc = conn->call[i];
3007 if (tc && tc->state == RX_STATE_PRECALL) {
3013 /* Indicate that rxi_CheckReachEvent is no longer running by
3014 * clearing the flag. Must be atomic under conn_data_lock to
3015 * avoid a new call slipping by: rxi_CheckConnReach holds
3016 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3018 conn->flags &= ~RX_CONN_ATTACHWAIT;
3019 MUTEX_EXIT(&conn->conn_data_lock);
3020 MUTEX_EXIT(&conn->conn_call_lock);
3025 MUTEX_ENTER(&call->lock);
3026 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3028 MUTEX_EXIT(&call->lock);
3030 clock_GetTime(&when);
3031 when.sec += RX_CHECKREACH_TIMEOUT;
3032 MUTEX_ENTER(&conn->conn_data_lock);
3033 if (!conn->checkReachEvent) {
3035 conn->checkReachEvent =
3036 rxevent_Post(&when, rxi_CheckReachEvent, conn, NULL);
3038 MUTEX_EXIT(&conn->conn_data_lock);
3044 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3046 struct rx_service *service = conn->service;
3047 struct rx_peer *peer = conn->peer;
3048 afs_uint32 now, lastReach;
3050 if (service->checkReach == 0)
3054 MUTEX_ENTER(&peer->peer_lock);
3055 lastReach = peer->lastReachTime;
3056 MUTEX_EXIT(&peer->peer_lock);
3057 if (now - lastReach < RX_CHECKREACH_TTL)
3060 MUTEX_ENTER(&conn->conn_data_lock);
3061 if (conn->flags & RX_CONN_ATTACHWAIT) {
3062 MUTEX_EXIT(&conn->conn_data_lock);
3065 conn->flags |= RX_CONN_ATTACHWAIT;
3066 MUTEX_EXIT(&conn->conn_data_lock);
3067 if (!conn->checkReachEvent)
3068 rxi_CheckReachEvent(NULL, conn, call);
3073 /* try to attach call, if authentication is complete */
3075 TryAttach(register struct rx_call *acall, register osi_socket socket,
3076 register int *tnop, register struct rx_call **newcallp,
3079 struct rx_connection *conn = acall->conn;
3081 if (conn->type == RX_SERVER_CONNECTION
3082 && acall->state == RX_STATE_PRECALL) {
3083 /* Don't attach until we have any req'd. authentication. */
3084 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3085 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3086 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3087 /* Note: this does not necessarily succeed; there
3088 * may not any proc available
3091 rxi_ChallengeOn(acall->conn);
3096 /* A data packet has been received off the interface. This packet is
3097 * appropriate to the call (the call is in the right state, etc.). This
3098 * routine can return a packet to the caller, for re-use */
3101 rxi_ReceiveDataPacket(register struct rx_call *call,
3102 register struct rx_packet *np, int istack,
3103 osi_socket socket, afs_uint32 host, u_short port,
3104 int *tnop, struct rx_call **newcallp)
3106 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3110 afs_uint32 seq, serial, flags;
3112 struct rx_packet *tnp;
3114 MUTEX_ENTER(&rx_stats_mutex);
3115 rx_stats.dataPacketsRead++;
3116 MUTEX_EXIT(&rx_stats_mutex);
3119 /* If there are no packet buffers, drop this new packet, unless we can find
3120 * packet buffers from inactive calls */
3122 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3123 MUTEX_ENTER(&rx_freePktQ_lock);
3124 rxi_NeedMorePackets = TRUE;
3125 MUTEX_EXIT(&rx_freePktQ_lock);
3126 MUTEX_ENTER(&rx_stats_mutex);
3127 rx_stats.noPacketBuffersOnRead++;
3128 MUTEX_EXIT(&rx_stats_mutex);
3129 call->rprev = np->header.serial;
3130 rxi_calltrace(RX_TRACE_DROP, call);
3131 dpf(("packet %x dropped on receipt - quota problems", np));
3133 rxi_ClearReceiveQueue(call);
3134 clock_GetTime(&when);
3135 clock_Add(&when, &rx_softAckDelay);
3136 if (!call->delayedAckEvent
3137 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3138 rxevent_Cancel(call->delayedAckEvent, call,
3139 RX_CALL_REFCOUNT_DELAY);
3140 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3141 call->delayedAckEvent =
3142 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3144 /* we've damaged this call already, might as well do it in. */
3150 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3151 * packet is one of several packets transmitted as a single
3152 * datagram. Do not send any soft or hard acks until all packets
3153 * in a jumbogram have been processed. Send negative acks right away.
3155 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3156 /* tnp is non-null when there are more packets in the
3157 * current jumbo gram */
3164 seq = np->header.seq;
3165 serial = np->header.serial;
3166 flags = np->header.flags;
3168 /* If the call is in an error state, send an abort message */
3170 return rxi_SendCallAbort(call, np, istack, 0);
3172 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3173 * AFS 3.5 jumbogram. */
3174 if (flags & RX_JUMBO_PACKET) {
3175 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3180 if (np->header.spare != 0) {
3181 MUTEX_ENTER(&call->conn->conn_data_lock);
3182 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3183 MUTEX_EXIT(&call->conn->conn_data_lock);
3186 /* The usual case is that this is the expected next packet */
3187 if (seq == call->rnext) {
3189 /* Check to make sure it is not a duplicate of one already queued */
3190 if (queue_IsNotEmpty(&call->rq)
3191 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3192 MUTEX_ENTER(&rx_stats_mutex);
3193 rx_stats.dupPacketsRead++;
3194 MUTEX_EXIT(&rx_stats_mutex);
3195 dpf(("packet %x dropped on receipt - duplicate", np));
3196 rxevent_Cancel(call->delayedAckEvent, call,
3197 RX_CALL_REFCOUNT_DELAY);
3198 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3204 /* It's the next packet. Stick it on the receive queue
3205 * for this call. Set newPackets to make sure we wake
3206 * the reader once all packets have been processed */
3207 queue_Prepend(&call->rq, np);
3209 np = NULL; /* We can't use this anymore */
3212 /* If an ack is requested then set a flag to make sure we
3213 * send an acknowledgement for this packet */
3214 if (flags & RX_REQUEST_ACK) {
3215 ackNeeded = RX_ACK_REQUESTED;
3218 /* Keep track of whether we have received the last packet */
3219 if (flags & RX_LAST_PACKET) {
3220 call->flags |= RX_CALL_HAVE_LAST;
3224 /* Check whether we have all of the packets for this call */
3225 if (call->flags & RX_CALL_HAVE_LAST) {
3226 afs_uint32 tseq; /* temporary sequence number */
3227 struct rx_packet *tp; /* Temporary packet pointer */
3228 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3230 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3231 if (tseq != tp->header.seq)
3233 if (tp->header.flags & RX_LAST_PACKET) {
3234 call->flags |= RX_CALL_RECEIVE_DONE;
3241 /* Provide asynchronous notification for those who want it
3242 * (e.g. multi rx) */
3243 if (call->arrivalProc) {
3244 (*call->arrivalProc) (call, call->arrivalProcHandle,
3245 call->arrivalProcArg);
3246 call->arrivalProc = (void (*)())0;
3249 /* Update last packet received */
3252 /* If there is no server process serving this call, grab
3253 * one, if available. We only need to do this once. If a
3254 * server thread is available, this thread becomes a server
3255 * thread and the server thread becomes a listener thread. */
3257 TryAttach(call, socket, tnop, newcallp, 0);
3260 /* This is not the expected next packet. */
3262 /* Determine whether this is a new or old packet, and if it's
3263 * a new one, whether it fits into the current receive window.
3264 * Also figure out whether the packet was delivered in sequence.
3265 * We use the prev variable to determine whether the new packet
3266 * is the successor of its immediate predecessor in the
3267 * receive queue, and the missing flag to determine whether
3268 * any of this packets predecessors are missing. */
3270 afs_uint32 prev; /* "Previous packet" sequence number */
3271 struct rx_packet *tp; /* Temporary packet pointer */
3272 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3273 int missing; /* Are any predecessors missing? */
3275 /* If the new packet's sequence number has been sent to the
3276 * application already, then this is a duplicate */
3277 if (seq < call->rnext) {
3278 MUTEX_ENTER(&rx_stats_mutex);
3279 rx_stats.dupPacketsRead++;
3280 MUTEX_EXIT(&rx_stats_mutex);
3281 rxevent_Cancel(call->delayedAckEvent, call,
3282 RX_CALL_REFCOUNT_DELAY);
3283 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3289 /* If the sequence number is greater than what can be
3290 * accomodated by the current window, then send a negative
3291 * acknowledge and drop the packet */
3292 if ((call->rnext + call->rwind) <= seq) {
3293 rxevent_Cancel(call->delayedAckEvent, call,
3294 RX_CALL_REFCOUNT_DELAY);
3295 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3302 /* Look for the packet in the queue of old received packets */
3303 for (prev = call->rnext - 1, missing =
3304 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3305 /*Check for duplicate packet */
3306 if (seq == tp->header.seq) {
3307 MUTEX_ENTER(&rx_stats_mutex);
3308 rx_stats.dupPacketsRead++;
3309 MUTEX_EXIT(&rx_stats_mutex);
3310 rxevent_Cancel(call->delayedAckEvent, call,
3311 RX_CALL_REFCOUNT_DELAY);
3312 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3318 /* If we find a higher sequence packet, break out and
3319 * insert the new packet here. */
3320 if (seq < tp->header.seq)
3322 /* Check for missing packet */
3323 if (tp->header.seq != prev + 1) {
3327 prev = tp->header.seq;
3330 /* Keep track of whether we have received the last packet. */
3331 if (flags & RX_LAST_PACKET) {
3332 call->flags |= RX_CALL_HAVE_LAST;
3335 /* It's within the window: add it to the the receive queue.
3336 * tp is left by the previous loop either pointing at the
3337 * packet before which to insert the new packet, or at the
3338 * queue head if the queue is empty or the packet should be
3340 queue_InsertBefore(tp, np);
3344 /* Check whether we have all of the packets for this call */
3345 if ((call->flags & RX_CALL_HAVE_LAST)
3346 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3347 afs_uint32 tseq; /* temporary sequence number */
3350 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3351 if (tseq != tp->header.seq)
3353 if (tp->header.flags & RX_LAST_PACKET) {
3354 call->flags |= RX_CALL_RECEIVE_DONE;
3361 /* We need to send an ack of the packet is out of sequence,
3362 * or if an ack was requested by the peer. */
3363 if (seq != prev + 1 || missing) {
3364 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3365 } else if (flags & RX_REQUEST_ACK) {
3366 ackNeeded = RX_ACK_REQUESTED;
3369 /* Acknowledge the last packet for each call */
3370 if (flags & RX_LAST_PACKET) {
3381 * If the receiver is waiting for an iovec, fill the iovec
3382 * using the data from the receive queue */
3383 if (call->flags & RX_CALL_IOVEC_WAIT) {
3384 didHardAck = rxi_FillReadVec(call, serial);
3385 /* the call may have been aborted */
3394 /* Wakeup the reader if any */
3395 if ((call->flags & RX_CALL_READER_WAIT)
3396 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3397 || (call->iovNext >= call->iovMax)
3398 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3399 call->flags &= ~RX_CALL_READER_WAIT;
3400 #ifdef RX_ENABLE_LOCKS
3401 CV_BROADCAST(&call->cv_rq);
3403 osi_rxWakeup(&call->rq);
3409 * Send an ack when requested by the peer, or once every
3410 * rxi_SoftAckRate packets until the last packet has been
3411 * received. Always send a soft ack for the last packet in
3412 * the server's reply. */
3414 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3415 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3416 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3417 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3418 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3419 } else if (call->nSoftAcks) {
3420 clock_GetTime(&when);
3421 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3422 clock_Add(&when, &rx_lastAckDelay);
3424 clock_Add(&when, &rx_softAckDelay);
3426 if (!call->delayedAckEvent
3427 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3428 rxevent_Cancel(call->delayedAckEvent, call,
3429 RX_CALL_REFCOUNT_DELAY);
3430 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3431 call->delayedAckEvent =
3432 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3434 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3435 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3442 static void rxi_ComputeRate();
3446 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3448 struct rx_peer *peer = conn->peer;
3450 MUTEX_ENTER(&peer->peer_lock);
3451 peer->lastReachTime = clock_Sec();
3452 MUTEX_EXIT(&peer->peer_lock);
3454 MUTEX_ENTER(&conn->conn_data_lock);
3455 if (conn->flags & RX_CONN_ATTACHWAIT) {
3458 conn->flags &= ~RX_CONN_ATTACHWAIT;
3459 MUTEX_EXIT(&conn->conn_data_lock);
3461 for (i = 0; i < RX_MAXCALLS; i++) {
3462 struct rx_call *call = conn->call[i];
3465 MUTEX_ENTER(&call->lock);
3466 /* tnop can be null if newcallp is null */
3467 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3469 MUTEX_EXIT(&call->lock);
3473 MUTEX_EXIT(&conn->conn_data_lock);
3477 rx_ack_reason(int reason)
3480 case RX_ACK_REQUESTED:
3482 case RX_ACK_DUPLICATE:
3484 case RX_ACK_OUT_OF_SEQUENCE:
3486 case RX_ACK_EXCEEDS_WINDOW:
3488 case RX_ACK_NOSPACE:
3492 case RX_ACK_PING_RESPONSE:
3504 /* rxi_ComputePeerNetStats
3506 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3507 * estimates (like RTT and throughput) based on ack packets. Caller
3508 * must ensure that the packet in question is the right one (i.e.
3509 * serial number matches).
3512 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3513 struct rx_ackPacket *ap, struct rx_packet *np)
3515 struct rx_peer *peer = call->conn->peer;
3517 /* Use RTT if not delayed by client. */
3518 if (ap->reason != RX_ACK_DELAY)
3519 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3521 rxi_ComputeRate(peer, call, p, np, ap->reason);
3525 /* The real smarts of the whole thing. */
3527 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3530 struct rx_ackPacket *ap;
3532 register struct rx_packet *tp;
3533 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3534 register struct rx_connection *conn = call->conn;
3535 struct rx_peer *peer = conn->peer;
3538 /* because there are CM's that are bogus, sending weird values for this. */
3539 afs_uint32 skew = 0;
3544 int newAckCount = 0;
3545 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3546 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3548 MUTEX_ENTER(&rx_stats_mutex);
3549 rx_stats.ackPacketsRead++;
3550 MUTEX_EXIT(&rx_stats_mutex);
3551 ap = (struct rx_ackPacket *)rx_DataOf(np);
3552 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3554 return np; /* truncated ack packet */
3556 /* depends on ack packet struct */
3557 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3558 first = ntohl(ap->firstPacket);
3559 serial = ntohl(ap->serial);
3560 /* temporarily disabled -- needs to degrade over time
3561 * skew = ntohs(ap->maxSkew); */
3563 /* Ignore ack packets received out of order */
3564 if (first < call->tfirst) {
3568 if (np->header.flags & RX_SLOW_START_OK) {
3569 call->flags |= RX_CALL_SLOW_START_OK;
3572 if (ap->reason == RX_ACK_PING_RESPONSE)
3573 rxi_UpdatePeerReach(conn, call);
3577 if (rxdebug_active) {
3581 len = _snprintf(msg, sizeof(msg),
3582 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3583 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3584 ntohl(ap->serial), ntohl(ap->previousPacket),
3585 (unsigned int)np->header.seq, (unsigned int)skew,
3586 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3590 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3591 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3595 OutputDebugString(msg);
3597 #else /* AFS_NT40_ENV */
3600 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3601 ap->reason, ntohl(ap->previousPacket),
3602 (unsigned int)np->header.seq, (unsigned int)serial,
3603 (unsigned int)skew, ntohl(ap->firstPacket));
3606 for (offset = 0; offset < nAcks; offset++)
3607 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3612 #endif /* AFS_NT40_ENV */
3615 /* Update the outgoing packet skew value to the latest value of
3616 * the peer's incoming packet skew value. The ack packet, of
3617 * course, could arrive out of order, but that won't affect things
3619 MUTEX_ENTER(&peer->peer_lock);
3620 peer->outPacketSkew = skew;
3622 /* Check for packets that no longer need to be transmitted, and
3623 * discard them. This only applies to packets positively
3624 * acknowledged as having been sent to the peer's upper level.
3625 * All other packets must be retained. So only packets with
3626 * sequence numbers < ap->firstPacket are candidates. */
3627 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3628 if (tp->header.seq >= first)
3630 call->tfirst = tp->header.seq + 1;
3632 && (tp->header.serial == serial || tp->firstSerial == serial))
3633 rxi_ComputePeerNetStats(call, tp, ap, np);
3634 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3637 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3638 /* XXX Hack. Because we have to release the global rx lock when sending
3639 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3640 * in rxi_Start sending packets out because packets may move to the
3641 * freePacketQueue as result of being here! So we drop these packets until
3642 * we're safely out of the traversing. Really ugly!
3643 * To make it even uglier, if we're using fine grain locking, we can
3644 * set the ack bits in the packets and have rxi_Start remove the packets
3645 * when it's done transmitting.
3647 if (call->flags & RX_CALL_TQ_BUSY) {
3648 #ifdef RX_ENABLE_LOCKS
3649 tp->flags |= RX_PKTFLAG_ACKED;
3650 call->flags |= RX_CALL_TQ_SOME_ACKED;
3651 #else /* RX_ENABLE_LOCKS */
3653 #endif /* RX_ENABLE_LOCKS */
3655 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3658 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3663 /* Give rate detector a chance to respond to ping requests */
3664 if (ap->reason == RX_ACK_PING_RESPONSE) {
3665 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3669 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3671 /* Now go through explicit acks/nacks and record the results in
3672 * the waiting packets. These are packets that can't be released
3673 * yet, even with a positive acknowledge. This positive
3674 * acknowledge only means the packet has been received by the
3675 * peer, not that it will be retained long enough to be sent to
3676 * the peer's upper level. In addition, reset the transmit timers
3677 * of any missing packets (those packets that must be missing
3678 * because this packet was out of sequence) */
3680 call->nSoftAcked = 0;
3681 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3682 /* Update round trip time if the ack was stimulated on receipt
3684 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3685 #ifdef RX_ENABLE_LOCKS
3686 if (tp->header.seq >= first)
3687 #endif /* RX_ENABLE_LOCKS */
3688 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3690 && (tp->header.serial == serial || tp->firstSerial == serial))
3691 rxi_ComputePeerNetStats(call, tp, ap, np);
3693 /* Set the acknowledge flag per packet based on the
3694 * information in the ack packet. An acknowlegded packet can
3695 * be downgraded when the server has discarded a packet it
3696 * soacked previously, or when an ack packet is received
3697 * out of sequence. */
3698 if (tp->header.seq < first) {
3699 /* Implicit ack information */
3700 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3703 tp->flags |= RX_PKTFLAG_ACKED;
3704 } else if (tp->header.seq < first + nAcks) {
3705 /* Explicit ack information: set it in the packet appropriately */
3706 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3707 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3709 tp->flags |= RX_PKTFLAG_ACKED;
3716 } else /* RX_ACK_TYPE_NACK */ {
3717 tp->flags &= ~RX_PKTFLAG_ACKED;
3721 tp->flags &= ~RX_PKTFLAG_ACKED;
3725 /* If packet isn't yet acked, and it has been transmitted at least
3726 * once, reset retransmit time using latest timeout
3727 * ie, this should readjust the retransmit timer for all outstanding
3728 * packets... So we don't just retransmit when we should know better*/
3730 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3731 tp->retryTime = tp->timeSent;
3732 clock_Add(&tp->retryTime, &peer->timeout);
3733 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3734 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3738 /* If the window has been extended by this acknowledge packet,
3739 * then wakeup a sender waiting in alloc for window space, or try
3740 * sending packets now, if he's been sitting on packets due to
3741 * lack of window space */
3742 if (call->tnext < (call->tfirst + call->twind)) {
3743 #ifdef RX_ENABLE_LOCKS
3744 CV_SIGNAL(&call->cv_twind);
3746 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3747 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3748 osi_rxWakeup(&call->twind);
3751 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3752 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3756 /* if the ack packet has a receivelen field hanging off it,
3757 * update our state */
3758 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3761 /* If the ack packet has a "recommended" size that is less than
3762 * what I am using now, reduce my size to match */
3763 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3764 (int)sizeof(afs_int32), &tSize);
3765 tSize = (afs_uint32) ntohl(tSize);
3766 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3768 /* Get the maximum packet size to send to this peer */
3769 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
3771 tSize = (afs_uint32) ntohl(tSize);
3772 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3773 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3775 /* sanity check - peer might have restarted with different params.
3776 * If peer says "send less", dammit, send less... Peer should never
3777 * be unable to accept packets of the size that prior AFS versions would
3778 * send without asking. */
3779 if (peer->maxMTU != tSize) {
3780 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
3782 peer->maxMTU = tSize;
3783 peer->MTU = MIN(tSize, peer->MTU);
3784 call->MTU = MIN(call->MTU, tSize);
3787 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3790 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3791 (int)sizeof(afs_int32), &tSize);
3792 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3793 if (tSize < call->twind) { /* smaller than our send */
3794 call->twind = tSize; /* window, we must send less... */
3795 call->ssthresh = MIN(call->twind, call->ssthresh);
3798 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3799 * network MTU confused with the loopback MTU. Calculate the
3800 * maximum MTU here for use in the slow start code below.
3802 maxMTU = peer->maxMTU;
3803 /* Did peer restart with older RX version? */
3804 if (peer->maxDgramPackets > 1) {
3805 peer->maxDgramPackets = 1;
3807 } else if (np->length >=
3808 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3811 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3812 sizeof(afs_int32), &tSize);
3813 tSize = (afs_uint32) ntohl(tSize);
3815 * As of AFS 3.5 we set the send window to match the receive window.
3817 if (tSize < call->twind) {
3818 call->twind = tSize;
3819 call->ssthresh = MIN(call->twind, call->ssthresh);
3820 } else if (tSize > call->twind) {
3821 call->twind = tSize;
3825 * As of AFS 3.5, a jumbogram is more than one fixed size
3826 * packet transmitted in a single UDP datagram. If the remote
3827 * MTU is smaller than our local MTU then never send a datagram
3828 * larger than the natural MTU.
3831 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3832 sizeof(afs_int32), &tSize);
3833 maxDgramPackets = (afs_uint32) ntohl(tSize);
3834 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3835 maxDgramPackets = MIN(maxDgramPackets, peer->ifDgramPackets);
3836 if (peer->natMTU < peer->ifMTU)
3837 maxDgramPackets = MIN(maxDgramPackets, rxi_AdjustDgramPackets(1, peer->natMTU));
3838 if (maxDgramPackets > 1) {
3839 peer->maxDgramPackets = maxDgramPackets;
3840 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3842 peer->maxDgramPackets = 1;
3843 call->MTU = peer->natMTU;
3845 } else if (peer->maxDgramPackets > 1) {
3846 /* Restarted with lower version of RX */
3847 peer->maxDgramPackets = 1;
3849 } else if (peer->maxDgramPackets > 1
3850 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3851 /* Restarted with lower version of RX */
3852 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3853 peer->natMTU = OLD_MAX_PACKET_SIZE;
3854 peer->MTU = OLD_MAX_PACKET_SIZE;
3855 peer->maxDgramPackets = 1;
3856 peer->nDgramPackets = 1;
3858 call->MTU = OLD_MAX_PACKET_SIZE;
3863 * Calculate how many datagrams were successfully received after
3864 * the first missing packet and adjust the negative ack counter
3869 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3870 if (call->nNacks < nNacked) {
3871 call->nNacks = nNacked;
3880 if (call->flags & RX_CALL_FAST_RECOVER) {
3882 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3884 call->flags &= ~RX_CALL_FAST_RECOVER;
3885 call->cwind = call->nextCwind;
3886 call->nextCwind = 0;
3889 call->nCwindAcks = 0;
3890 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3891 /* Three negative acks in a row trigger congestion recovery */
3892 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3893 MUTEX_EXIT(&peer->peer_lock);
3894 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3895 /* someone else is waiting to start recovery */
3898 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3899 rxi_WaitforTQBusy(call);
3900 MUTEX_ENTER(&peer->peer_lock);
3901 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3902 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3903 call->flags |= RX_CALL_FAST_RECOVER;
3904 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3906 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3907 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3908 call->nextCwind = call->ssthresh;
3911 peer->MTU = call->MTU;
3912 peer->cwind = call->nextCwind;
3913 peer->nDgramPackets = call->nDgramPackets;
3915 call->congestSeq = peer->congestSeq;
3916 /* Reset the resend times on the packets that were nacked
3917 * so we will retransmit as soon as the window permits*/
3918 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3920 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3921 clock_Zero(&tp->retryTime);
3923 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3928 /* If cwind is smaller than ssthresh, then increase
3929 * the window one packet for each ack we receive (exponential
3931 * If cwind is greater than or equal to ssthresh then increase
3932 * the congestion window by one packet for each cwind acks we
3933 * receive (linear growth). */
3934 if (call->cwind < call->ssthresh) {
3936 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3937 call->nCwindAcks = 0;
3939 call->nCwindAcks += newAckCount;
3940 if (call->nCwindAcks >= call->cwind) {
3941 call->nCwindAcks = 0;
3942 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3946 * If we have received several acknowledgements in a row then
3947 * it is time to increase the size of our datagrams
3949 if ((int)call->nAcks > rx_nDgramThreshold) {
3950 if (peer->maxDgramPackets > 1) {
3951 if (call->nDgramPackets < peer->maxDgramPackets) {
3952 call->nDgramPackets++;
3954 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3955 } else if (call->MTU < peer->maxMTU) {
3956 call->MTU += peer->natMTU;
3957 call->MTU = MIN(call->MTU, peer->maxMTU);
3963 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3965 /* Servers need to hold the call until all response packets have
3966 * been acknowledged. Soft acks are good enough since clients
3967 * are not allowed to clear their receive queues. */
3968 if (call->state == RX_STATE_HOLD
3969 && call->tfirst + call->nSoftAcked >= call->tnext) {
3970 call->state = RX_STATE_DALLY;
3971 rxi_ClearTransmitQueue(call, 0);
3972 } else if (!queue_IsEmpty(&call->tq)) {
3973 rxi_Start(0, call, 0, istack);
3978 /* Received a response to a challenge packet */
3980 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
3981 register struct rx_packet *np, int istack)
3985 /* Ignore the packet if we're the client */
3986 if (conn->type == RX_CLIENT_CONNECTION)
3989 /* If already authenticated, ignore the packet (it's probably a retry) */
3990 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
3993 /* Otherwise, have the security object evaluate the response packet */
3994 error = RXS_CheckResponse(conn->securityObject, conn, np);
3996 /* If the response is invalid, reset the connection, sending
3997 * an abort to the peer */
4001 rxi_ConnectionError(conn, error);
4002 MUTEX_ENTER(&conn->conn_data_lock);
4003 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4004 MUTEX_EXIT(&conn->conn_data_lock);
4007 /* If the response is valid, any calls waiting to attach
4008 * servers can now do so */
4011 for (i = 0; i < RX_MAXCALLS; i++) {
4012 struct rx_call *call = conn->call[i];
4014 MUTEX_ENTER(&call->lock);
4015 if (call->state == RX_STATE_PRECALL)
4016 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4017 /* tnop can be null if newcallp is null */
4018 MUTEX_EXIT(&call->lock);
4022 /* Update the peer reachability information, just in case
4023 * some calls went into attach-wait while we were waiting
4024 * for authentication..
4026 rxi_UpdatePeerReach(conn, NULL);
4031 /* A client has received an authentication challenge: the security
4032 * object is asked to cough up a respectable response packet to send
4033 * back to the server. The server is responsible for retrying the
4034 * challenge if it fails to get a response. */
4037 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
4038 register struct rx_packet *np, int istack)
4042 /* Ignore the challenge if we're the server */
4043 if (conn->type == RX_SERVER_CONNECTION)
4046 /* Ignore the challenge if the connection is otherwise idle; someone's
4047 * trying to use us as an oracle. */
4048 if (!rxi_HasActiveCalls(conn))
4051 /* Send the security object the challenge packet. It is expected to fill
4052 * in the response. */
4053 error = RXS_GetResponse(conn->securityObject, conn, np);
4055 /* If the security object is unable to return a valid response, reset the
4056 * connection and send an abort to the peer. Otherwise send the response
4057 * packet to the peer connection. */
4059 rxi_ConnectionError(conn, error);
4060 MUTEX_ENTER(&conn->conn_data_lock);
4061 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4062 MUTEX_EXIT(&conn->conn_data_lock);
4064 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4065 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4071 /* Find an available server process to service the current request in
4072 * the given call structure. If one isn't available, queue up this
4073 * call so it eventually gets one */
4075 rxi_AttachServerProc(register struct rx_call *call,
4076 register osi_socket socket, register int *tnop,
4077 register struct rx_call **newcallp)
4079 register struct rx_serverQueueEntry *sq;
4080 register struct rx_service *service = call->conn->service;
4081 register int haveQuota = 0;
4083 /* May already be attached */
4084 if (call->state == RX_STATE_ACTIVE)
4087 MUTEX_ENTER(&rx_serverPool_lock);
4089 haveQuota = QuotaOK(service);
4090 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4091 /* If there are no processes available to service this call,
4092 * put the call on the incoming call queue (unless it's
4093 * already on the queue).
4095 #ifdef RX_ENABLE_LOCKS
4097 ReturnToServerPool(service);
4098 #endif /* RX_ENABLE_LOCKS */
4100 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4101 call->flags |= RX_CALL_WAIT_PROC;
4102 MUTEX_ENTER(&rx_stats_mutex);
4105 MUTEX_EXIT(&rx_stats_mutex);
4106 rxi_calltrace(RX_CALL_ARRIVAL, call);
4107 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4108 queue_Append(&rx_incomingCallQueue, call);
4111 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4113 /* If hot threads are enabled, and both newcallp and sq->socketp
4114 * are non-null, then this thread will process the call, and the
4115 * idle server thread will start listening on this threads socket.
4118 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4121 *sq->socketp = socket;
4122 clock_GetTime(&call->startTime);
4123 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4127 if (call->flags & RX_CALL_WAIT_PROC) {
4128 /* Conservative: I don't think this should happen */
4129 call->flags &= ~RX_CALL_WAIT_PROC;
4130 if (queue_IsOnQueue(call)) {
4132 MUTEX_ENTER(&rx_stats_mutex);
4134 MUTEX_EXIT(&rx_stats_mutex);
4137 call->state = RX_STATE_ACTIVE;
4138 call->mode = RX_MODE_RECEIVING;
4139 #ifdef RX_KERNEL_TRACE
4141 int glockOwner = ISAFS_GLOCK();
4144 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4145 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4151 if (call->flags & RX_CALL_CLEARED) {
4152 /* send an ack now to start the packet flow up again */
4153 call->flags &= ~RX_CALL_CLEARED;
4154 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4156 #ifdef RX_ENABLE_LOCKS
4159 service->nRequestsRunning++;
4160 if (service->nRequestsRunning <= service->minProcs)
4166 MUTEX_EXIT(&rx_serverPool_lock);
4169 /* Delay the sending of an acknowledge event for a short while, while
4170 * a new call is being prepared (in the case of a client) or a reply
4171 * is being prepared (in the case of a server). Rather than sending
4172 * an ack packet, an ACKALL packet is sent. */
4174 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4176 #ifdef RX_ENABLE_LOCKS
4178 MUTEX_ENTER(&call->lock);
4179 call->delayedAckEvent = NULL;
4180 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4182 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4183 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4185 MUTEX_EXIT(&call->lock);
4186 #else /* RX_ENABLE_LOCKS */
4188 call->delayedAckEvent = NULL;
4189 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4190 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4191 #endif /* RX_ENABLE_LOCKS */
4195 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4198 #ifdef RX_ENABLE_LOCKS
4200 MUTEX_ENTER(&call->lock);
4201 if (event == call->delayedAckEvent)
4202 call->delayedAckEvent = NULL;
4203 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4205 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4207 MUTEX_EXIT(&call->lock);
4208 #else /* RX_ENABLE_LOCKS */
4210 call->delayedAckEvent = NULL;
4211 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4212 #endif /* RX_ENABLE_LOCKS */
4216 #ifdef RX_ENABLE_LOCKS
4217 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4218 * clearing them out.
4221 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4223 register struct rx_packet *p, *tp;
4226 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4227 p->flags |= RX_PKTFLAG_ACKED;
4231 call->flags |= RX_CALL_TQ_CLEARME;
4232 call->flags |= RX_CALL_TQ_SOME_ACKED;
4235 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4236 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4237 call->tfirst = call->tnext;
4238 call->nSoftAcked = 0;
4240 if (call->flags & RX_CALL_FAST_RECOVER) {
4241 call->flags &= ~RX_CALL_FAST_RECOVER;
4242 call->cwind = call->nextCwind;
4243 call->nextCwind = 0;
4246 CV_SIGNAL(&call->cv_twind);
4248 #endif /* RX_ENABLE_LOCKS */
4250 /* Clear out the transmit queue for the current call (all packets have
4251 * been received by peer) */
4253 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4255 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4256 register struct rx_packet *p, *tp;
4258 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4260 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4261 p->flags |= RX_PKTFLAG_ACKED;
4265 call->flags |= RX_CALL_TQ_CLEARME;
4266 call->flags |= RX_CALL_TQ_SOME_ACKED;
4269 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4270 rxi_FreePackets(0, &call->tq);
4271 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4272 call->flags &= ~RX_CALL_TQ_CLEARME;
4274 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4276 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4277 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4278 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4279 call->nSoftAcked = 0;
4281 if (call->flags & RX_CALL_FAST_RECOVER) {
4282 call->flags &= ~RX_CALL_FAST_RECOVER;
4283 call->cwind = call->nextCwind;
4285 #ifdef RX_ENABLE_LOCKS
4286 CV_SIGNAL(&call->cv_twind);
4288 osi_rxWakeup(&call->twind);
4293 rxi_ClearReceiveQueue(register struct rx_call *call)
4295 if (queue_IsNotEmpty(&call->rq)) {
4296 rx_packetReclaims += rxi_FreePackets(0, &call->rq);
4297 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4299 if (call->state == RX_STATE_PRECALL) {
4300 call->flags |= RX_CALL_CLEARED;
4304 /* Send an abort packet for the specified call */
4306 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4307 int istack, int force)
4315 /* Clients should never delay abort messages */
4316 if (rx_IsClientConn(call->conn))
4319 if (call->abortCode != call->error) {
4320 call->abortCode = call->error;
4321 call->abortCount = 0;
4324 if (force || rxi_callAbortThreshhold == 0
4325 || call->abortCount < rxi_callAbortThreshhold) {
4326 if (call->delayedAbortEvent) {
4327 rxevent_Cancel(call->delayedAbortEvent, call,
4328 RX_CALL_REFCOUNT_ABORT);
4330 error = htonl(call->error);
4333 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4334 (char *)&error, sizeof(error), istack);
4335 } else if (!call->delayedAbortEvent) {
4336 clock_GetTime(&when);
4337 clock_Addmsec(&when, rxi_callAbortDelay);
4338 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4339 call->delayedAbortEvent =
4340 rxevent_Post(&when, rxi_SendDelayedCallAbort, call, 0);
4345 /* Send an abort packet for the specified connection. Packet is an
4346 * optional pointer to a packet that can be used to send the abort.
4347 * Once the number of abort messages reaches the threshhold, an
4348 * event is scheduled to send the abort. Setting the force flag
4349 * overrides sending delayed abort messages.
4351 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4352 * to send the abort packet.
4355 rxi_SendConnectionAbort(register struct rx_connection *conn,
4356 struct rx_packet *packet, int istack, int force)
4364 /* Clients should never delay abort messages */
4365 if (rx_IsClientConn(conn))
4368 if (force || rxi_connAbortThreshhold == 0
4369 || conn->abortCount < rxi_connAbortThreshhold) {
4370 if (conn->delayedAbortEvent) {
4371 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4373 error = htonl(conn->error);
4375 MUTEX_EXIT(&conn->conn_data_lock);
4377 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4378 RX_PACKET_TYPE_ABORT, (char *)&error,
4379 sizeof(error), istack);
4380 MUTEX_ENTER(&conn->conn_data_lock);
4381 } else if (!conn->delayedAbortEvent) {
4382 clock_GetTime(&when);
4383 clock_Addmsec(&when, rxi_connAbortDelay);
4384 conn->delayedAbortEvent =
4385 rxevent_Post(&when, rxi_SendDelayedConnAbort, conn, 0);
4390 /* Associate an error all of the calls owned by a connection. Called
4391 * with error non-zero. This is only for really fatal things, like
4392 * bad authentication responses. The connection itself is set in
4393 * error at this point, so that future packets received will be
4396 rxi_ConnectionError(register struct rx_connection *conn,
4397 register afs_int32 error)
4402 dpf(("rxi_ConnectionError conn %x error %d", conn, error));
4404 MUTEX_ENTER(&conn->conn_data_lock);
4405 if (conn->challengeEvent)
4406 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4407 if (conn->checkReachEvent) {
4408 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4409 conn->checkReachEvent = 0;
4410 conn->flags &= ~RX_CONN_ATTACHWAIT;
4413 MUTEX_EXIT(&conn->conn_data_lock);
4414 for (i = 0; i < RX_MAXCALLS; i++) {
4415 struct rx_call *call = conn->call[i];
4417 MUTEX_ENTER(&call->lock);
4418 rxi_CallError(call, error);
4419 MUTEX_EXIT(&call->lock);
4422 conn->error = error;
4423 MUTEX_ENTER(&rx_stats_mutex);
4424 rx_stats.fatalErrors++;
4425 MUTEX_EXIT(&rx_stats_mutex);
4430 rxi_CallError(register struct rx_call *call, afs_int32 error)
4432 dpf(("rxi_CallError call %x error %d call->error %d", call, error, call->error));
4434 error = call->error;
4436 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4437 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4438 rxi_ResetCall(call, 0);
4441 rxi_ResetCall(call, 0);
4443 call->error = error;
4444 call->mode = RX_MODE_ERROR;
4447 /* Reset various fields in a call structure, and wakeup waiting
4448 * processes. Some fields aren't changed: state & mode are not
4449 * touched (these must be set by the caller), and bufptr, nLeft, and
4450 * nFree are not reset, since these fields are manipulated by
4451 * unprotected macros, and may only be reset by non-interrupting code.
4454 /* this code requires that call->conn be set properly as a pre-condition. */
4455 #endif /* ADAPT_WINDOW */
4458 rxi_ResetCall(register struct rx_call *call, register int newcall)
4461 register struct rx_peer *peer;
4462 struct rx_packet *packet;
4464 dpf(("rxi_ResetCall(call %x, newcall %d)\n", call, newcall));
4466 /* Notify anyone who is waiting for asynchronous packet arrival */
4467 if (call->arrivalProc) {
4468 (*call->arrivalProc) (call, call->arrivalProcHandle,
4469 call->arrivalProcArg);
4470 call->arrivalProc = (void (*)())0;
4473 if (call->delayedAbortEvent) {
4474 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4475 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4477 rxi_SendCallAbort(call, packet, 0, 1);
4478 rxi_FreePacket(packet);
4483 * Update the peer with the congestion information in this call
4484 * so other calls on this connection can pick up where this call
4485 * left off. If the congestion sequence numbers don't match then
4486 * another call experienced a retransmission.
4488 peer = call->conn->peer;
4489 MUTEX_ENTER(&peer->peer_lock);
4491 if (call->congestSeq == peer->congestSeq) {
4492 peer->cwind = MAX(peer->cwind, call->cwind);
4493 peer->MTU = MAX(peer->MTU, call->MTU);
4494 peer->nDgramPackets =
4495 MAX(peer->nDgramPackets, call->nDgramPackets);
4498 call->abortCode = 0;
4499 call->abortCount = 0;
4501 if (peer->maxDgramPackets > 1) {
4502 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4504 call->MTU = peer->MTU;
4506 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4507 call->ssthresh = rx_maxSendWindow;
4508 call->nDgramPackets = peer->nDgramPackets;
4509 call->congestSeq = peer->congestSeq;
4510 MUTEX_EXIT(&peer->peer_lock);
4512 flags = call->flags;
4513 rxi_ClearReceiveQueue(call);
4514 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4515 if (flags & RX_CALL_TQ_BUSY) {
4516 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4517 call->flags |= (flags & RX_CALL_TQ_WAIT);
4519 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4521 rxi_ClearTransmitQueue(call, 0);
4522 queue_Init(&call->tq);
4523 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4524 dpf(("rcall %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4527 while (call->tqWaiters) {
4528 #ifdef RX_ENABLE_LOCKS
4529 CV_BROADCAST(&call->cv_tq);
4530 #else /* RX_ENABLE_LOCKS */
4531 osi_rxWakeup(&call->tq);
4532 #endif /* RX_ENABLE_LOCKS */
4536 queue_Init(&call->rq);
4538 call->rwind = rx_initReceiveWindow;
4539 call->twind = rx_initSendWindow;
4540 call->nSoftAcked = 0;
4541 call->nextCwind = 0;
4544 call->nCwindAcks = 0;
4545 call->nSoftAcks = 0;
4546 call->nHardAcks = 0;
4548 call->tfirst = call->rnext = call->tnext = 1;
4550 call->lastAcked = 0;
4551 call->localStatus = call->remoteStatus = 0;
4553 if (flags & RX_CALL_READER_WAIT) {
4554 #ifdef RX_ENABLE_LOCKS
4555 CV_BROADCAST(&call->cv_rq);
4557 osi_rxWakeup(&call->rq);
4560 if (flags & RX_CALL_WAIT_PACKETS) {
4561 MUTEX_ENTER(&rx_freePktQ_lock);
4562 rxi_PacketsUnWait(); /* XXX */
4563 MUTEX_EXIT(&rx_freePktQ_lock);
4565 #ifdef RX_ENABLE_LOCKS
4566 CV_SIGNAL(&call->cv_twind);
4568 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4569 osi_rxWakeup(&call->twind);
4572 #ifdef RX_ENABLE_LOCKS
4573 /* The following ensures that we don't mess with any queue while some
4574 * other thread might also be doing so. The call_queue_lock field is
4575 * is only modified under the call lock. If the call is in the process
4576 * of being removed from a queue, the call is not locked until the
4577 * the queue lock is dropped and only then is the call_queue_lock field
4578 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4579 * Note that any other routine which removes a call from a queue has to
4580 * obtain the queue lock before examing the queue and removing the call.
4582 if (call->call_queue_lock) {
4583 MUTEX_ENTER(call->call_queue_lock);
4584 if (queue_IsOnQueue(call)) {
4586 if (flags & RX_CALL_WAIT_PROC) {
4587 MUTEX_ENTER(&rx_stats_mutex);
4589 MUTEX_EXIT(&rx_stats_mutex);
4592 MUTEX_EXIT(call->call_queue_lock);
4593 CLEAR_CALL_QUEUE_LOCK(call);
4595 #else /* RX_ENABLE_LOCKS */
4596 if (queue_IsOnQueue(call)) {
4598 if (flags & RX_CALL_WAIT_PROC)
4601 #endif /* RX_ENABLE_LOCKS */
4603 rxi_KeepAliveOff(call);
4604 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4607 /* Send an acknowledge for the indicated packet (seq,serial) of the
4608 * indicated call, for the indicated reason (reason). This
4609 * acknowledge will specifically acknowledge receiving the packet, and
4610 * will also specify which other packets for this call have been
4611 * received. This routine returns the packet that was used to the
4612 * caller. The caller is responsible for freeing it or re-using it.
4613 * This acknowledgement also returns the highest sequence number
4614 * actually read out by the higher level to the sender; the sender
4615 * promises to keep around packets that have not been read by the
4616 * higher level yet (unless, of course, the sender decides to abort
4617 * the call altogether). Any of p, seq, serial, pflags, or reason may
4618 * be set to zero without ill effect. That is, if they are zero, they
4619 * will not convey any information.
4620 * NOW there is a trailer field, after the ack where it will safely be
4621 * ignored by mundanes, which indicates the maximum size packet this
4622 * host can swallow. */
4624 register struct rx_packet *optionalPacket; use to send ack (or null)
4625 int seq; Sequence number of the packet we are acking
4626 int serial; Serial number of the packet
4627 int pflags; Flags field from packet header
4628 int reason; Reason an acknowledge was prompted
4632 rxi_SendAck(register struct rx_call *call,
4633 register struct rx_packet *optionalPacket, int serial, int reason,
4636 struct rx_ackPacket *ap;
4637 register struct rx_packet *rqp;
4638 register struct rx_packet *nxp; /* For queue_Scan */
4639 register struct rx_packet *p;
4642 #ifdef RX_ENABLE_TSFPQ
4643 struct rx_ts_info_t * rx_ts_info;
4647 * Open the receive window once a thread starts reading packets
4649 if (call->rnext > 1) {
4650 call->rwind = rx_maxReceiveWindow;
4653 call->nHardAcks = 0;
4654 call->nSoftAcks = 0;
4655 if (call->rnext > call->lastAcked)
4656 call->lastAcked = call->rnext;
4660 rx_computelen(p, p->length); /* reset length, you never know */
4661 } /* where that's been... */
4662 #ifdef RX_ENABLE_TSFPQ
4664 RX_TS_INFO_GET(rx_ts_info);
4665 if ((p = rx_ts_info->local_special_packet)) {
4666 rx_computelen(p, p->length);
4667 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4668 rx_ts_info->local_special_packet = p;
4669 } else { /* We won't send the ack, but don't panic. */
4670 return optionalPacket;
4674 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4675 /* We won't send the ack, but don't panic. */
4676 return optionalPacket;
4681 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4684 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4685 #ifndef RX_ENABLE_TSFPQ
4686 if (!optionalPacket)
4689 return optionalPacket;
4691 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4692 if (rx_Contiguous(p) < templ) {
4693 #ifndef RX_ENABLE_TSFPQ
4694 if (!optionalPacket)
4697 return optionalPacket;
4702 /* MTUXXX failing to send an ack is very serious. We should */
4703 /* try as hard as possible to send even a partial ack; it's */
4704 /* better than nothing. */
4705 ap = (struct rx_ackPacket *)rx_DataOf(p);
4706 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4707 ap->reason = reason;
4709 /* The skew computation used to be bogus, I think it's better now. */
4710 /* We should start paying attention to skew. XXX */
4711 ap->serial = htonl(serial);
4712 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4714 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4715 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4717 /* No fear of running out of ack packet here because there can only be at most
4718 * one window full of unacknowledged packets. The window size must be constrained
4719 * to be less than the maximum ack size, of course. Also, an ack should always
4720 * fit into a single packet -- it should not ever be fragmented. */
4721 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4722 if (!rqp || !call->rq.next
4723 || (rqp->header.seq > (call->rnext + call->rwind))) {
4724 #ifndef RX_ENABLE_TSFPQ
4725 if (!optionalPacket)
4728 rxi_CallError(call, RX_CALL_DEAD);
4729 return optionalPacket;
4732 while (rqp->header.seq > call->rnext + offset)
4733 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4734 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4736 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4737 #ifndef RX_ENABLE_TSFPQ
4738 if (!optionalPacket)
4741 rxi_CallError(call, RX_CALL_DEAD);
4742 return optionalPacket;
4747 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4749 /* these are new for AFS 3.3 */
4750 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4751 templ = htonl(templ);
4752 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4753 templ = htonl(call->conn->peer->ifMTU);
4754 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4755 sizeof(afs_int32), &templ);
4757 /* new for AFS 3.4 */
4758 templ = htonl(call->rwind);
4759 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4760 sizeof(afs_int32), &templ);
4762 /* new for AFS 3.5 */
4763 templ = htonl(call->conn->peer->ifDgramPackets);
4764 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4765 sizeof(afs_int32), &templ);
4767 p->header.serviceId = call->conn->serviceId;
4768 p->header.cid = (call->conn->cid | call->channel);
4769 p->header.callNumber = *call->callNumber;
4771 p->header.securityIndex = call->conn->securityIndex;
4772 p->header.epoch = call->conn->epoch;
4773 p->header.type = RX_PACKET_TYPE_ACK;
4774 p->header.flags = RX_SLOW_START_OK;
4775 if (reason == RX_ACK_PING) {
4776 p->header.flags |= RX_REQUEST_ACK;
4778 clock_GetTime(&call->pingRequestTime);
4781 if (call->conn->type == RX_CLIENT_CONNECTION)
4782 p->header.flags |= RX_CLIENT_INITIATED;
4786 if (rxdebug_active) {
4790 len = _snprintf(msg, sizeof(msg),
4791 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4792 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4793 ntohl(ap->serial), ntohl(ap->previousPacket),
4794 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
4795 ap->nAcks, ntohs(ap->bufferSpace) );
4799 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
4800 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4804 OutputDebugString(msg);
4806 #else /* AFS_NT40_ENV */
4808 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
4809 ap->reason, ntohl(ap->previousPacket),
4810 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4812 for (offset = 0; offset < ap->nAcks; offset++)
4813 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4818 #endif /* AFS_NT40_ENV */
4821 register int i, nbytes = p->length;
4823 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4824 if (nbytes <= p->wirevec[i].iov_len) {
4825 register int savelen, saven;
4827 savelen = p->wirevec[i].iov_len;
4829 p->wirevec[i].iov_len = nbytes;
4831 rxi_Send(call, p, istack);
4832 p->wirevec[i].iov_len = savelen;
4836 nbytes -= p->wirevec[i].iov_len;
4839 MUTEX_ENTER(&rx_stats_mutex);
4840 rx_stats.ackPacketsSent++;
4841 MUTEX_EXIT(&rx_stats_mutex);
4842 #ifndef RX_ENABLE_TSFPQ
4843 if (!optionalPacket)
4846 return optionalPacket; /* Return packet for re-use by caller */
4849 /* Send all of the packets in the list in single datagram */
4851 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4852 int istack, int moreFlag, struct clock *now,
4853 struct clock *retryTime, int resending)
4858 struct rx_connection *conn = call->conn;
4859 struct rx_peer *peer = conn->peer;
4861 MUTEX_ENTER(&peer->peer_lock);
4864 peer->reSends += len;
4865 MUTEX_ENTER(&rx_stats_mutex);
4866 rx_stats.dataPacketsSent += len;
4867 MUTEX_EXIT(&rx_stats_mutex);
4868 MUTEX_EXIT(&peer->peer_lock);
4870 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4874 /* Set the packet flags and schedule the resend events */
4875 /* Only request an ack for the last packet in the list */
4876 for (i = 0; i < len; i++) {
4877 list[i]->retryTime = *retryTime;
4878 if (list[i]->header.serial) {
4879 /* Exponentially backoff retry times */
4880 if (list[i]->backoff < MAXBACKOFF) {
4881 /* so it can't stay == 0 */
4882 list[i]->backoff = (list[i]->backoff << 1) + 1;
4885 clock_Addmsec(&(list[i]->retryTime),
4886 ((afs_uint32) list[i]->backoff) << 8);
4889 /* Wait a little extra for the ack on the last packet */
4890 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4891 clock_Addmsec(&(list[i]->retryTime), 400);
4894 /* Record the time sent */
4895 list[i]->timeSent = *now;
4897 /* Ask for an ack on retransmitted packets, on every other packet
4898 * if the peer doesn't support slow start. Ask for an ack on every
4899 * packet until the congestion window reaches the ack rate. */
4900 if (list[i]->header.serial) {
4902 MUTEX_ENTER(&rx_stats_mutex);
4903 rx_stats.dataPacketsReSent++;
4904 MUTEX_EXIT(&rx_stats_mutex);
4906 /* improved RTO calculation- not Karn */
4907 list[i]->firstSent = *now;
4908 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4909 || (!(call->flags & RX_CALL_SLOW_START_OK)
4910 && (list[i]->header.seq & 1)))) {
4915 MUTEX_ENTER(&peer->peer_lock);
4919 MUTEX_ENTER(&rx_stats_mutex);
4920 rx_stats.dataPacketsSent++;
4921 MUTEX_EXIT(&rx_stats_mutex);
4922 MUTEX_EXIT(&peer->peer_lock);
4924 /* Tag this packet as not being the last in this group,
4925 * for the receiver's benefit */
4926 if (i < len - 1 || moreFlag) {
4927 list[i]->header.flags |= RX_MORE_PACKETS;
4930 /* Install the new retransmit time for the packet, and
4931 * record the time sent */
4932 list[i]->timeSent = *now;
4936 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4939 /* Since we're about to send a data packet to the peer, it's
4940 * safe to nuke any scheduled end-of-packets ack */
4941 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4943 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4944 MUTEX_EXIT(&call->lock);
4946 rxi_SendPacketList(call, conn, list, len, istack);
4948 rxi_SendPacket(call, conn, list[0], istack);
4950 MUTEX_ENTER(&call->lock);
4951 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4953 /* Update last send time for this call (for keep-alive
4954 * processing), and for the connection (so that we can discover
4955 * idle connections) */
4956 conn->lastSendTime = call->lastSendTime = clock_Sec();
4959 /* When sending packets we need to follow these rules:
4960 * 1. Never send more than maxDgramPackets in a jumbogram.
4961 * 2. Never send a packet with more than two iovecs in a jumbogram.
4962 * 3. Never send a retransmitted packet in a jumbogram.
4963 * 4. Never send more than cwind/4 packets in a jumbogram
4964 * We always keep the last list we should have sent so we
4965 * can set the RX_MORE_PACKETS flags correctly.
4968 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4969 int istack, struct clock *now, struct clock *retryTime,
4972 int i, cnt, lastCnt = 0;
4973 struct rx_packet **listP, **lastP = 0;
4974 struct rx_peer *peer = call->conn->peer;
4975 int morePackets = 0;
4977 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4978 /* Does the current packet force us to flush the current list? */
4980 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4981 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4983 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
4985 /* If the call enters an error state stop sending, or if
4986 * we entered congestion recovery mode, stop sending */
4987 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4995 /* Add the current packet to the list if it hasn't been acked.
4996 * Otherwise adjust the list pointer to skip the current packet. */
4997 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
4999 /* Do we need to flush the list? */
5000 if (cnt >= (int)peer->maxDgramPackets
5001 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5002 || list[i]->header.serial
5003 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5005 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5006 retryTime, resending);
5007 /* If the call enters an error state stop sending, or if
5008 * we entered congestion recovery mode, stop sending */
5010 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5015 listP = &list[i + 1];
5020 osi_Panic("rxi_SendList error");
5022 listP = &list[i + 1];
5026 /* Send the whole list when the call is in receive mode, when
5027 * the call is in eof mode, when we are in fast recovery mode,
5028 * and when we have the last packet */
5029 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5030 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5031 || (call->flags & RX_CALL_FAST_RECOVER)) {
5032 /* Check for the case where the current list contains
5033 * an acked packet. Since we always send retransmissions
5034 * in a separate packet, we only need to check the first
5035 * packet in the list */
5036 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5040 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5041 retryTime, resending);
5042 /* If the call enters an error state stop sending, or if
5043 * we entered congestion recovery mode, stop sending */
5044 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5048 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5051 } else if (lastCnt > 0) {
5052 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5057 #ifdef RX_ENABLE_LOCKS
5058 /* Call rxi_Start, below, but with the call lock held. */
5060 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
5061 void *arg1, int istack)
5063 MUTEX_ENTER(&call->lock);
5064 rxi_Start(event, call, arg1, istack);
5065 MUTEX_EXIT(&call->lock);
5067 #endif /* RX_ENABLE_LOCKS */
5069 /* This routine is called when new packets are readied for
5070 * transmission and when retransmission may be necessary, or when the
5071 * transmission window or burst count are favourable. This should be
5072 * better optimized for new packets, the usual case, now that we've
5073 * got rid of queues of send packets. XXXXXXXXXXX */
5075 rxi_Start(struct rxevent *event, register struct rx_call *call,
5076 void *arg1, int istack)
5078 struct rx_packet *p;
5079 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
5080 struct rx_peer *peer = call->conn->peer;
5081 struct clock now, retryTime;
5085 struct rx_packet **xmitList;
5088 /* If rxi_Start is being called as a result of a resend event,
5089 * then make sure that the event pointer is removed from the call
5090 * structure, since there is no longer a per-call retransmission
5092 if (event && event == call->resendEvent) {
5093 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5094 call->resendEvent = NULL;
5096 if (queue_IsEmpty(&call->tq)) {
5100 /* Timeouts trigger congestion recovery */
5101 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5102 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5103 /* someone else is waiting to start recovery */
5106 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5107 rxi_WaitforTQBusy(call);
5108 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5109 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5110 call->flags |= RX_CALL_FAST_RECOVER;
5111 if (peer->maxDgramPackets > 1) {
5112 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5114 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5116 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5117 call->nDgramPackets = 1;
5119 call->nextCwind = 1;
5122 MUTEX_ENTER(&peer->peer_lock);
5123 peer->MTU = call->MTU;
5124 peer->cwind = call->cwind;
5125 peer->nDgramPackets = 1;
5127 call->congestSeq = peer->congestSeq;
5128 MUTEX_EXIT(&peer->peer_lock);
5129 /* Clear retry times on packets. Otherwise, it's possible for
5130 * some packets in the queue to force resends at rates faster
5131 * than recovery rates.
5133 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5134 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5135 clock_Zero(&p->retryTime);
5140 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5141 MUTEX_ENTER(&rx_stats_mutex);
5142 rx_tq_debug.rxi_start_in_error++;
5143 MUTEX_EXIT(&rx_stats_mutex);
5148 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5149 /* Get clock to compute the re-transmit time for any packets
5150 * in this burst. Note, if we back off, it's reasonable to
5151 * back off all of the packets in the same manner, even if
5152 * some of them have been retransmitted more times than more
5153 * recent additions */
5154 clock_GetTime(&now);
5155 retryTime = now; /* initialize before use */
5156 MUTEX_ENTER(&peer->peer_lock);
5157 clock_Add(&retryTime, &peer->timeout);
5158 MUTEX_EXIT(&peer->peer_lock);
5160 /* Send (or resend) any packets that need it, subject to
5161 * window restrictions and congestion burst control
5162 * restrictions. Ask for an ack on the last packet sent in
5163 * this burst. For now, we're relying upon the window being
5164 * considerably bigger than the largest number of packets that
5165 * are typically sent at once by one initial call to
5166 * rxi_Start. This is probably bogus (perhaps we should ask
5167 * for an ack when we're half way through the current
5168 * window?). Also, for non file transfer applications, this
5169 * may end up asking for an ack for every packet. Bogus. XXXX
5172 * But check whether we're here recursively, and let the other guy
5175 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5176 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5177 call->flags |= RX_CALL_TQ_BUSY;
5179 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5181 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5182 call->flags &= ~RX_CALL_NEED_START;
5183 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5185 maxXmitPackets = MIN(call->twind, call->cwind);
5186 xmitList = (struct rx_packet **)
5187 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5188 if (xmitList == NULL)
5189 osi_Panic("rxi_Start, failed to allocate xmit list");
5190 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5191 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5192 /* We shouldn't be sending packets if a thread is waiting
5193 * to initiate congestion recovery */
5197 && (call->flags & RX_CALL_FAST_RECOVER)) {
5198 /* Only send one packet during fast recovery */
5201 if ((p->flags & RX_PKTFLAG_FREE)
5202 || (!queue_IsEnd(&call->tq, nxp)
5203 && (nxp->flags & RX_PKTFLAG_FREE))
5204 || (p == (struct rx_packet *)&rx_freePacketQueue)
5205 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5206 osi_Panic("rxi_Start: xmit queue clobbered");
5208 if (p->flags & RX_PKTFLAG_ACKED) {
5209 MUTEX_ENTER(&rx_stats_mutex);
5210 rx_stats.ignoreAckedPacket++;
5211 MUTEX_EXIT(&rx_stats_mutex);
5212 continue; /* Ignore this packet if it has been acknowledged */
5215 /* Turn off all flags except these ones, which are the same
5216 * on each transmission */
5217 p->header.flags &= RX_PRESET_FLAGS;
5219 if (p->header.seq >=
5220 call->tfirst + MIN((int)call->twind,
5221 (int)(call->nSoftAcked +
5223 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5224 /* Note: if we're waiting for more window space, we can
5225 * still send retransmits; hence we don't return here, but
5226 * break out to schedule a retransmit event */
5227 dpf(("call %d waiting for window",
5228 *(call->callNumber)));
5232 /* Transmit the packet if it needs to be sent. */
5233 if (!clock_Lt(&now, &p->retryTime)) {
5234 if (nXmitPackets == maxXmitPackets) {
5235 rxi_SendXmitList(call, xmitList, nXmitPackets,
5236 istack, &now, &retryTime,
5238 osi_Free(xmitList, maxXmitPackets *
5239 sizeof(struct rx_packet *));
5242 xmitList[nXmitPackets++] = p;
5246 /* xmitList now hold pointers to all of the packets that are
5247 * ready to send. Now we loop to send the packets */
5248 if (nXmitPackets > 0) {
5249 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5250 &now, &retryTime, resending);
5253 maxXmitPackets * sizeof(struct rx_packet *));
5255 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5257 * TQ references no longer protected by this flag; they must remain
5258 * protected by the global lock.
5260 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5261 call->flags &= ~RX_CALL_TQ_BUSY;
5262 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5263 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5264 #ifdef RX_ENABLE_LOCKS
5265 osirx_AssertMine(&call->lock, "rxi_Start start");
5266 CV_BROADCAST(&call->cv_tq);
5267 #else /* RX_ENABLE_LOCKS */
5268 osi_rxWakeup(&call->tq);
5269 #endif /* RX_ENABLE_LOCKS */
5274 /* We went into the error state while sending packets. Now is
5275 * the time to reset the call. This will also inform the using
5276 * process that the call is in an error state.
5278 MUTEX_ENTER(&rx_stats_mutex);
5279 rx_tq_debug.rxi_start_aborted++;
5280 MUTEX_EXIT(&rx_stats_mutex);
5281 call->flags &= ~RX_CALL_TQ_BUSY;
5282 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5283 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5284 #ifdef RX_ENABLE_LOCKS
5285 osirx_AssertMine(&call->lock, "rxi_Start middle");
5286 CV_BROADCAST(&call->cv_tq);
5287 #else /* RX_ENABLE_LOCKS */
5288 osi_rxWakeup(&call->tq);
5289 #endif /* RX_ENABLE_LOCKS */
5291 rxi_CallError(call, call->error);
5294 #ifdef RX_ENABLE_LOCKS
5295 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5296 register int missing;
5297 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5298 /* Some packets have received acks. If they all have, we can clear
5299 * the transmit queue.
5302 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5303 if (p->header.seq < call->tfirst
5304 && (p->flags & RX_PKTFLAG_ACKED)) {
5311 call->flags |= RX_CALL_TQ_CLEARME;
5313 #endif /* RX_ENABLE_LOCKS */
5314 /* Don't bother doing retransmits if the TQ is cleared. */
5315 if (call->flags & RX_CALL_TQ_CLEARME) {
5316 rxi_ClearTransmitQueue(call, 1);
5318 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5321 /* Always post a resend event, if there is anything in the
5322 * queue, and resend is possible. There should be at least
5323 * one unacknowledged packet in the queue ... otherwise none
5324 * of these packets should be on the queue in the first place.
5326 if (call->resendEvent) {
5327 /* Cancel the existing event and post a new one */
5328 rxevent_Cancel(call->resendEvent, call,
5329 RX_CALL_REFCOUNT_RESEND);
5332 /* The retry time is the retry time on the first unacknowledged
5333 * packet inside the current window */
5335 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5336 /* Don't set timers for packets outside the window */
5337 if (p->header.seq >= call->tfirst + call->twind) {
5341 if (!(p->flags & RX_PKTFLAG_ACKED)
5342 && !clock_IsZero(&p->retryTime)) {
5344 retryTime = p->retryTime;
5349 /* Post a new event to re-run rxi_Start when retries may be needed */
5350 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5351 #ifdef RX_ENABLE_LOCKS
5352 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5354 rxevent_Post2(&retryTime, rxi_StartUnlocked,
5355 (void *)call, 0, istack);
5356 #else /* RX_ENABLE_LOCKS */
5358 rxevent_Post2(&retryTime, rxi_Start, (void *)call,
5360 #endif /* RX_ENABLE_LOCKS */
5363 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5364 } while (call->flags & RX_CALL_NEED_START);
5366 * TQ references no longer protected by this flag; they must remain
5367 * protected by the global lock.
5369 call->flags &= ~RX_CALL_TQ_BUSY;
5370 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5371 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5372 #ifdef RX_ENABLE_LOCKS
5373 osirx_AssertMine(&call->lock, "rxi_Start end");
5374 CV_BROADCAST(&call->cv_tq);
5375 #else /* RX_ENABLE_LOCKS */
5376 osi_rxWakeup(&call->tq);
5377 #endif /* RX_ENABLE_LOCKS */
5380 call->flags |= RX_CALL_NEED_START;
5382 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5384 if (call->resendEvent) {
5385 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5390 /* Also adjusts the keep alive parameters for the call, to reflect
5391 * that we have just sent a packet (so keep alives aren't sent
5394 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5397 register struct rx_connection *conn = call->conn;
5399 /* Stamp each packet with the user supplied status */
5400 p->header.userStatus = call->localStatus;
5402 /* Allow the security object controlling this call's security to
5403 * make any last-minute changes to the packet */
5404 RXS_SendPacket(conn->securityObject, call, p);
5406 /* Since we're about to send SOME sort of packet to the peer, it's
5407 * safe to nuke any scheduled end-of-packets ack */
5408 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5410 /* Actually send the packet, filling in more connection-specific fields */
5411 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5412 MUTEX_EXIT(&call->lock);
5413 rxi_SendPacket(call, conn, p, istack);
5414 MUTEX_ENTER(&call->lock);
5415 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5417 /* Update last send time for this call (for keep-alive
5418 * processing), and for the connection (so that we can discover
5419 * idle connections) */
5420 conn->lastSendTime = call->lastSendTime = clock_Sec();
5424 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5425 * that things are fine. Also called periodically to guarantee that nothing
5426 * falls through the cracks (e.g. (error + dally) connections have keepalive
5427 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5429 * haveCTLock Set if calling from rxi_ReapConnections
5431 #ifdef RX_ENABLE_LOCKS
5433 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5434 #else /* RX_ENABLE_LOCKS */
5436 rxi_CheckCall(register struct rx_call *call)
5437 #endif /* RX_ENABLE_LOCKS */
5439 register struct rx_connection *conn = call->conn;
5441 afs_uint32 deadTime;
5443 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5444 if (call->flags & RX_CALL_TQ_BUSY) {
5445 /* Call is active and will be reset by rxi_Start if it's
5446 * in an error state.
5451 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5453 (((afs_uint32) conn->secondsUntilDead << 10) +
5454 ((afs_uint32) conn->peer->rtt >> 3) +
5455 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5457 /* These are computed to the second (+- 1 second). But that's
5458 * good enough for these values, which should be a significant
5459 * number of seconds. */
5460 if (now > (call->lastReceiveTime + deadTime)) {
5461 if (call->state == RX_STATE_ACTIVE) {
5462 rxi_CallError(call, RX_CALL_DEAD);
5465 #ifdef RX_ENABLE_LOCKS
5466 /* Cancel pending events */
5467 rxevent_Cancel(call->delayedAckEvent, call,
5468 RX_CALL_REFCOUNT_DELAY);
5469 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5470 rxevent_Cancel(call->keepAliveEvent, call,
5471 RX_CALL_REFCOUNT_ALIVE);
5472 if (call->refCount == 0) {
5473 rxi_FreeCall(call, haveCTLock);
5477 #else /* RX_ENABLE_LOCKS */
5480 #endif /* RX_ENABLE_LOCKS */
5482 /* Non-active calls are destroyed if they are not responding
5483 * to pings; active calls are simply flagged in error, so the
5484 * attached process can die reasonably gracefully. */
5486 /* see if we have a non-activity timeout */
5487 if (call->startWait && conn->idleDeadTime
5488 && ((call->startWait + conn->idleDeadTime) < now)) {
5489 if (call->state == RX_STATE_ACTIVE) {
5490 rxi_CallError(call, RX_CALL_TIMEOUT);
5494 /* see if we have a hard timeout */
5495 if (conn->hardDeadTime
5496 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5497 if (call->state == RX_STATE_ACTIVE)
5498 rxi_CallError(call, RX_CALL_TIMEOUT);
5505 /* When a call is in progress, this routine is called occasionally to
5506 * make sure that some traffic has arrived (or been sent to) the peer.
5507 * If nothing has arrived in a reasonable amount of time, the call is
5508 * declared dead; if nothing has been sent for a while, we send a
5509 * keep-alive packet (if we're actually trying to keep the call alive)
5512 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5515 struct rx_connection *conn;
5518 MUTEX_ENTER(&call->lock);
5519 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5520 if (event == call->keepAliveEvent)
5521 call->keepAliveEvent = NULL;
5524 #ifdef RX_ENABLE_LOCKS
5525 if (rxi_CheckCall(call, 0)) {
5526 MUTEX_EXIT(&call->lock);
5529 #else /* RX_ENABLE_LOCKS */
5530 if (rxi_CheckCall(call))
5532 #endif /* RX_ENABLE_LOCKS */
5534 /* Don't try to keep alive dallying calls */
5535 if (call->state == RX_STATE_DALLY) {
5536 MUTEX_EXIT(&call->lock);
5541 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5542 /* Don't try to send keepalives if there is unacknowledged data */
5543 /* the rexmit code should be good enough, this little hack
5544 * doesn't quite work XXX */
5545 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5547 rxi_ScheduleKeepAliveEvent(call);
5548 MUTEX_EXIT(&call->lock);
5553 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5555 if (!call->keepAliveEvent) {
5557 clock_GetTime(&when);
5558 when.sec += call->conn->secondsUntilPing;
5559 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5560 call->keepAliveEvent =
5561 rxevent_Post(&when, rxi_KeepAliveEvent, call, 0);
5565 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5567 rxi_KeepAliveOn(register struct rx_call *call)
5569 /* Pretend last packet received was received now--i.e. if another
5570 * packet isn't received within the keep alive time, then the call
5571 * will die; Initialize last send time to the current time--even
5572 * if a packet hasn't been sent yet. This will guarantee that a
5573 * keep-alive is sent within the ping time */
5574 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5575 rxi_ScheduleKeepAliveEvent(call);
5578 /* This routine is called to send connection abort messages
5579 * that have been delayed to throttle looping clients. */
5581 rxi_SendDelayedConnAbort(struct rxevent *event,
5582 register struct rx_connection *conn, char *dummy)
5585 struct rx_packet *packet;
5587 MUTEX_ENTER(&conn->conn_data_lock);
5588 conn->delayedAbortEvent = NULL;
5589 error = htonl(conn->error);
5591 MUTEX_EXIT(&conn->conn_data_lock);
5592 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5595 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5596 RX_PACKET_TYPE_ABORT, (char *)&error,
5598 rxi_FreePacket(packet);
5602 /* This routine is called to send call abort messages
5603 * that have been delayed to throttle looping clients. */
5605 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5609 struct rx_packet *packet;
5611 MUTEX_ENTER(&call->lock);
5612 call->delayedAbortEvent = NULL;
5613 error = htonl(call->error);
5615 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5618 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5619 (char *)&error, sizeof(error), 0);
5620 rxi_FreePacket(packet);
5622 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5623 MUTEX_EXIT(&call->lock);
5626 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5627 * seconds) to ask the client to authenticate itself. The routine
5628 * issues a challenge to the client, which is obtained from the
5629 * security object associated with the connection */
5631 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5632 void *arg1, int tries)
5634 conn->challengeEvent = NULL;
5635 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5636 register struct rx_packet *packet;
5640 /* We've failed to authenticate for too long.
5641 * Reset any calls waiting for authentication;
5642 * they are all in RX_STATE_PRECALL.
5646 MUTEX_ENTER(&conn->conn_call_lock);
5647 for (i = 0; i < RX_MAXCALLS; i++) {
5648 struct rx_call *call = conn->call[i];
5650 MUTEX_ENTER(&call->lock);
5651 if (call->state == RX_STATE_PRECALL) {
5652 rxi_CallError(call, RX_CALL_DEAD);
5653 rxi_SendCallAbort(call, NULL, 0, 0);
5655 MUTEX_EXIT(&call->lock);
5658 MUTEX_EXIT(&conn->conn_call_lock);
5662 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5664 /* If there's no packet available, do this later. */
5665 RXS_GetChallenge(conn->securityObject, conn, packet);
5666 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5667 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5668 rxi_FreePacket(packet);
5670 clock_GetTime(&when);
5671 when.sec += RX_CHALLENGE_TIMEOUT;
5672 conn->challengeEvent =
5673 rxevent_Post2(&when, rxi_ChallengeEvent, conn, 0,
5678 /* Call this routine to start requesting the client to authenticate
5679 * itself. This will continue until authentication is established,
5680 * the call times out, or an invalid response is returned. The
5681 * security object associated with the connection is asked to create
5682 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5683 * defined earlier. */
5685 rxi_ChallengeOn(register struct rx_connection *conn)
5687 if (!conn->challengeEvent) {
5688 RXS_CreateChallenge(conn->securityObject, conn);
5689 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5694 /* Compute round trip time of the packet provided, in *rttp.
5697 /* rxi_ComputeRoundTripTime is called with peer locked. */
5698 /* sentp and/or peer may be null */
5700 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5701 register struct clock *sentp,
5702 register struct rx_peer *peer)
5704 struct clock thisRtt, *rttp = &thisRtt;
5706 register int rtt_timeout;
5708 clock_GetTime(rttp);
5710 if (clock_Lt(rttp, sentp)) {
5712 return; /* somebody set the clock back, don't count this time. */
5714 clock_Sub(rttp, sentp);
5715 MUTEX_ENTER(&rx_stats_mutex);
5716 if (clock_Lt(rttp, &rx_stats.minRtt))
5717 rx_stats.minRtt = *rttp;
5718 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5719 if (rttp->sec > 60) {
5720 MUTEX_EXIT(&rx_stats_mutex);
5721 return; /* somebody set the clock ahead */
5723 rx_stats.maxRtt = *rttp;
5725 clock_Add(&rx_stats.totalRtt, rttp);
5726 rx_stats.nRttSamples++;
5727 MUTEX_EXIT(&rx_stats_mutex);
5729 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5731 /* Apply VanJacobson round-trip estimations */
5736 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5737 * srtt is stored as fixed point with 3 bits after the binary
5738 * point (i.e., scaled by 8). The following magic is
5739 * equivalent to the smoothing algorithm in rfc793 with an
5740 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5741 * srtt*8 = srtt*8 + rtt - srtt
5742 * srtt = srtt + rtt/8 - srtt/8
5745 delta = MSEC(rttp) - (peer->rtt >> 3);
5749 * We accumulate a smoothed rtt variance (actually, a smoothed
5750 * mean difference), then set the retransmit timer to smoothed
5751 * rtt + 4 times the smoothed variance (was 2x in van's original
5752 * paper, but 4x works better for me, and apparently for him as
5754 * rttvar is stored as
5755 * fixed point with 2 bits after the binary point (scaled by
5756 * 4). The following is equivalent to rfc793 smoothing with
5757 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5758 * replaces rfc793's wired-in beta.
5759 * dev*4 = dev*4 + (|actual - expected| - dev)
5765 delta -= (peer->rtt_dev >> 2);
5766 peer->rtt_dev += delta;
5768 /* I don't have a stored RTT so I start with this value. Since I'm
5769 * probably just starting a call, and will be pushing more data down
5770 * this, I expect congestion to increase rapidly. So I fudge a
5771 * little, and I set deviance to half the rtt. In practice,
5772 * deviance tends to approach something a little less than
5773 * half the smoothed rtt. */
5774 peer->rtt = (MSEC(rttp) << 3) + 8;
5775 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5777 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5778 * the other of these connections is usually in a user process, and can
5779 * be switched and/or swapped out. So on fast, reliable networks, the
5780 * timeout would otherwise be too short.
5782 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5783 clock_Zero(&(peer->timeout));
5784 clock_Addmsec(&(peer->timeout), rtt_timeout);
5786 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)));
5790 /* Find all server connections that have not been active for a long time, and
5793 rxi_ReapConnections(void)
5796 clock_GetTime(&now);
5798 /* Find server connection structures that haven't been used for
5799 * greater than rx_idleConnectionTime */
5801 struct rx_connection **conn_ptr, **conn_end;
5802 int i, havecalls = 0;
5803 MUTEX_ENTER(&rx_connHashTable_lock);
5804 for (conn_ptr = &rx_connHashTable[0], conn_end =
5805 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5807 struct rx_connection *conn, *next;
5808 struct rx_call *call;
5812 for (conn = *conn_ptr; conn; conn = next) {
5813 /* XXX -- Shouldn't the connection be locked? */
5816 for (i = 0; i < RX_MAXCALLS; i++) {
5817 call = conn->call[i];
5820 MUTEX_ENTER(&call->lock);
5821 #ifdef RX_ENABLE_LOCKS
5822 result = rxi_CheckCall(call, 1);
5823 #else /* RX_ENABLE_LOCKS */
5824 result = rxi_CheckCall(call);
5825 #endif /* RX_ENABLE_LOCKS */
5826 MUTEX_EXIT(&call->lock);
5828 /* If CheckCall freed the call, it might
5829 * have destroyed the connection as well,
5830 * which screws up the linked lists.
5836 if (conn->type == RX_SERVER_CONNECTION) {
5837 /* This only actually destroys the connection if
5838 * there are no outstanding calls */
5839 MUTEX_ENTER(&conn->conn_data_lock);
5840 if (!havecalls && !conn->refCount
5841 && ((conn->lastSendTime + rx_idleConnectionTime) <
5843 conn->refCount++; /* it will be decr in rx_DestroyConn */
5844 MUTEX_EXIT(&conn->conn_data_lock);
5845 #ifdef RX_ENABLE_LOCKS
5846 rxi_DestroyConnectionNoLock(conn);
5847 #else /* RX_ENABLE_LOCKS */
5848 rxi_DestroyConnection(conn);
5849 #endif /* RX_ENABLE_LOCKS */
5851 #ifdef RX_ENABLE_LOCKS
5853 MUTEX_EXIT(&conn->conn_data_lock);
5855 #endif /* RX_ENABLE_LOCKS */
5859 #ifdef RX_ENABLE_LOCKS
5860 while (rx_connCleanup_list) {
5861 struct rx_connection *conn;
5862 conn = rx_connCleanup_list;
5863 rx_connCleanup_list = rx_connCleanup_list->next;
5864 MUTEX_EXIT(&rx_connHashTable_lock);
5865 rxi_CleanupConnection(conn);
5866 MUTEX_ENTER(&rx_connHashTable_lock);
5868 MUTEX_EXIT(&rx_connHashTable_lock);
5869 #endif /* RX_ENABLE_LOCKS */
5872 /* Find any peer structures that haven't been used (haven't had an
5873 * associated connection) for greater than rx_idlePeerTime */
5875 struct rx_peer **peer_ptr, **peer_end;
5877 MUTEX_ENTER(&rx_rpc_stats);
5878 MUTEX_ENTER(&rx_peerHashTable_lock);
5879 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5880 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5882 struct rx_peer *peer, *next, *prev;
5883 for (prev = peer = *peer_ptr; peer; peer = next) {
5885 code = MUTEX_TRYENTER(&peer->peer_lock);
5886 if ((code) && (peer->refCount == 0)
5887 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5888 rx_interface_stat_p rpc_stat, nrpc_stat;
5890 MUTEX_EXIT(&peer->peer_lock);
5891 MUTEX_DESTROY(&peer->peer_lock);
5893 (&peer->rpcStats, rpc_stat, nrpc_stat,
5894 rx_interface_stat)) {
5895 unsigned int num_funcs;
5898 queue_Remove(&rpc_stat->queue_header);
5899 queue_Remove(&rpc_stat->all_peers);
5900 num_funcs = rpc_stat->stats[0].func_total;
5902 sizeof(rx_interface_stat_t) +
5903 rpc_stat->stats[0].func_total *
5904 sizeof(rx_function_entry_v1_t);
5906 rxi_Free(rpc_stat, space);
5907 rxi_rpc_peer_stat_cnt -= num_funcs;
5910 MUTEX_ENTER(&rx_stats_mutex);
5911 rx_stats.nPeerStructs--;
5912 MUTEX_EXIT(&rx_stats_mutex);
5913 if (peer == *peer_ptr) {
5920 MUTEX_EXIT(&peer->peer_lock);
5926 MUTEX_EXIT(&rx_peerHashTable_lock);
5927 MUTEX_EXIT(&rx_rpc_stats);
5930 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5931 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5932 * GC, just below. Really, we shouldn't have to keep moving packets from
5933 * one place to another, but instead ought to always know if we can
5934 * afford to hold onto a packet in its particular use. */
5935 MUTEX_ENTER(&rx_freePktQ_lock);
5936 if (rx_waitingForPackets) {
5937 rx_waitingForPackets = 0;
5938 #ifdef RX_ENABLE_LOCKS
5939 CV_BROADCAST(&rx_waitingForPackets_cv);
5941 osi_rxWakeup(&rx_waitingForPackets);
5944 MUTEX_EXIT(&rx_freePktQ_lock);
5946 now.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5947 rxevent_Post(&now, rxi_ReapConnections, 0, 0);
5951 /* rxs_Release - This isn't strictly necessary but, since the macro name from
5952 * rx.h is sort of strange this is better. This is called with a security
5953 * object before it is discarded. Each connection using a security object has
5954 * its own refcount to the object so it won't actually be freed until the last
5955 * connection is destroyed.
5957 * This is the only rxs module call. A hold could also be written but no one
5961 rxs_Release(struct rx_securityClass *aobj)
5963 return RXS_Close(aobj);
5967 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
5968 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
5969 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
5970 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
5972 /* Adjust our estimate of the transmission rate to this peer, given
5973 * that the packet p was just acked. We can adjust peer->timeout and
5974 * call->twind. Pragmatically, this is called
5975 * only with packets of maximal length.
5976 * Called with peer and call locked.
5980 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
5981 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
5983 afs_int32 xferSize, xferMs;
5984 register afs_int32 minTime;
5987 /* Count down packets */
5988 if (peer->rateFlag > 0)
5990 /* Do nothing until we're enabled */
5991 if (peer->rateFlag != 0)
5996 /* Count only when the ack seems legitimate */
5997 switch (ackReason) {
5998 case RX_ACK_REQUESTED:
6000 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6004 case RX_ACK_PING_RESPONSE:
6005 if (p) /* want the response to ping-request, not data send */
6007 clock_GetTime(&newTO);
6008 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6009 clock_Sub(&newTO, &call->pingRequestTime);
6010 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6014 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6021 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));
6023 /* Track only packets that are big enough. */
6024 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6028 /* absorb RTT data (in milliseconds) for these big packets */
6029 if (peer->smRtt == 0) {
6030 peer->smRtt = xferMs;
6032 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6037 if (peer->countDown) {
6041 peer->countDown = 10; /* recalculate only every so often */
6043 /* In practice, we can measure only the RTT for full packets,
6044 * because of the way Rx acks the data that it receives. (If it's
6045 * smaller than a full packet, it often gets implicitly acked
6046 * either by the call response (from a server) or by the next call
6047 * (from a client), and either case confuses transmission times
6048 * with processing times.) Therefore, replace the above
6049 * more-sophisticated processing with a simpler version, where the
6050 * smoothed RTT is kept for full-size packets, and the time to
6051 * transmit a windowful of full-size packets is simply RTT *
6052 * windowSize. Again, we take two steps:
6053 - ensure the timeout is large enough for a single packet's RTT;
6054 - ensure that the window is small enough to fit in the desired timeout.*/
6056 /* First, the timeout check. */
6057 minTime = peer->smRtt;
6058 /* Get a reasonable estimate for a timeout period */
6060 newTO.sec = minTime / 1000;
6061 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6063 /* Increase the timeout period so that we can always do at least
6064 * one packet exchange */
6065 if (clock_Gt(&newTO, &peer->timeout)) {
6067 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));
6069 peer->timeout = newTO;
6072 /* Now, get an estimate for the transmit window size. */
6073 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6074 /* Now, convert to the number of full packets that could fit in a
6075 * reasonable fraction of that interval */
6076 minTime /= (peer->smRtt << 1);
6077 xferSize = minTime; /* (make a copy) */
6079 /* Now clamp the size to reasonable bounds. */
6082 else if (minTime > rx_Window)
6083 minTime = rx_Window;
6084 /* if (minTime != peer->maxWindow) {
6085 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6086 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6087 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6089 peer->maxWindow = minTime;
6090 elide... call->twind = minTime;
6094 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6095 * Discern this by calculating the timeout necessary for rx_Window
6097 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6098 /* calculate estimate for transmission interval in milliseconds */
6099 minTime = rx_Window * peer->smRtt;
6100 if (minTime < 1000) {
6101 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
6102 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6103 peer->timeout.usec, peer->smRtt, peer->packetSize));
6105 newTO.sec = 0; /* cut back on timeout by half a second */
6106 newTO.usec = 500000;
6107 clock_Sub(&peer->timeout, &newTO);
6112 } /* end of rxi_ComputeRate */
6113 #endif /* ADAPT_WINDOW */
6121 #define TRACE_OPTION_DEBUGLOG 4
6129 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6130 0, KEY_QUERY_VALUE, &parmKey);
6131 if (code != ERROR_SUCCESS)
6134 dummyLen = sizeof(TraceOption);
6135 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6136 (BYTE *) &TraceOption, &dummyLen);
6137 if (code == ERROR_SUCCESS) {
6138 rxdebug_active = (TraceOption & TRACE_OPTION_DEBUGLOG) ? 1 : 0;
6140 RegCloseKey (parmKey);
6141 #endif /* AFS_NT40_ENV */
6146 rx_DebugOnOff(int on)
6148 rxdebug_active = on;
6150 #endif /* AFS_NT40_ENV */
6153 /* Don't call this debugging routine directly; use dpf */
6155 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
6156 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
6164 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6167 len = _snprintf(msg, sizeof(msg)-2,
6168 tformat, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10,
6169 a11, a12, a13, a14, a15);
6171 if (msg[len-1] != '\n') {
6175 OutputDebugString(msg);
6180 clock_GetTime(&now);
6181 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
6182 (unsigned int)now.usec / 1000);
6183 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
6190 * This function is used to process the rx_stats structure that is local
6191 * to a process as well as an rx_stats structure received from a remote
6192 * process (via rxdebug). Therefore, it needs to do minimal version
6196 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
6197 afs_int32 freePackets, char version)
6201 if (size != sizeof(struct rx_stats)) {
6203 "Unexpected size of stats structure: was %d, expected %d\n",
6204 size, sizeof(struct rx_stats));
6207 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6210 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6211 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6212 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6213 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6214 s->specialPktAllocFailures);
6216 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6217 s->receivePktAllocFailures, s->sendPktAllocFailures,
6218 s->specialPktAllocFailures);
6222 " greedy %d, " "bogusReads %d (last from host %x), "
6223 "noPackets %d, " "noBuffers %d, " "selects %d, "
6224 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6225 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6226 s->selects, s->sendSelects);
6228 fprintf(file, " packets read: ");
6229 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6230 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6232 fprintf(file, "\n");
6235 " other read counters: data %d, " "ack %d, " "dup %d "
6236 "spurious %d " "dally %d\n", s->dataPacketsRead,
6237 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6238 s->ignorePacketDally);
6240 fprintf(file, " packets sent: ");
6241 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6242 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6244 fprintf(file, "\n");
6247 " other send counters: ack %d, " "data %d (not resends), "
6248 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6249 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6250 s->dataPacketsPushed, s->ignoreAckedPacket);
6253 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6254 s->netSendFailures, (int)s->fatalErrors);
6256 if (s->nRttSamples) {
6257 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6258 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6260 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6261 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6265 " %d server connections, " "%d client connections, "
6266 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6267 s->nServerConns, s->nClientConns, s->nPeerStructs,
6268 s->nCallStructs, s->nFreeCallStructs);
6270 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6271 fprintf(file, " %d clock updates\n", clock_nUpdates);
6276 /* for backward compatibility */
6278 rx_PrintStats(FILE * file)
6280 MUTEX_ENTER(&rx_stats_mutex);
6281 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6283 MUTEX_EXIT(&rx_stats_mutex);
6287 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6289 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6290 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6291 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6294 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6295 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6296 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6299 " Packet size %d, " "max in packet skew %d, "
6300 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6301 (int)peer->outPacketSkew);
6304 #ifdef AFS_PTHREAD_ENV
6306 * This mutex protects the following static variables:
6310 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6311 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6313 #define LOCK_RX_DEBUG
6314 #define UNLOCK_RX_DEBUG
6315 #endif /* AFS_PTHREAD_ENV */
6318 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6319 u_char type, void *inputData, size_t inputLength,
6320 void *outputData, size_t outputLength)
6322 static afs_int32 counter = 100;
6323 time_t waitTime, waitCount, startTime, endTime;
6324 struct rx_header theader;
6326 register afs_int32 code;
6327 struct timeval tv_now, tv_wake, tv_delta;
6328 struct sockaddr_in taddr, faddr;
6333 startTime = time(0);
6339 tp = &tbuffer[sizeof(struct rx_header)];
6340 taddr.sin_family = AF_INET;
6341 taddr.sin_port = remotePort;
6342 taddr.sin_addr.s_addr = remoteAddr;
6343 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6344 taddr.sin_len = sizeof(struct sockaddr_in);
6347 memset(&theader, 0, sizeof(theader));
6348 theader.epoch = htonl(999);
6350 theader.callNumber = htonl(counter);
6353 theader.type = type;
6354 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6355 theader.serviceId = 0;
6357 memcpy(tbuffer, &theader, sizeof(theader));
6358 memcpy(tp, inputData, inputLength);
6360 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6361 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6363 /* see if there's a packet available */
6364 gettimeofday(&tv_wake,0);
6365 tv_wake.tv_sec += waitTime;
6368 FD_SET(socket, &imask);
6369 tv_delta.tv_sec = tv_wake.tv_sec;
6370 tv_delta.tv_usec = tv_wake.tv_usec;
6371 gettimeofday(&tv_now, 0);
6373 if (tv_delta.tv_usec < tv_now.tv_usec) {
6375 tv_delta.tv_usec += 1000000;
6378 tv_delta.tv_usec -= tv_now.tv_usec;
6380 if (tv_delta.tv_sec < tv_now.tv_sec) {
6384 tv_delta.tv_sec -= tv_now.tv_sec;
6386 code = select(socket + 1, &imask, 0, 0, &tv_delta);
6387 if (code == 1 && FD_ISSET(socket, &imask)) {
6388 /* now receive a packet */
6389 faddrLen = sizeof(struct sockaddr_in);
6391 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6392 (struct sockaddr *)&faddr, &faddrLen);
6395 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6396 if (counter == ntohl(theader.callNumber))
6404 /* see if we've timed out */
6412 code -= sizeof(struct rx_header);
6413 if (code > outputLength)
6414 code = outputLength;
6415 memcpy(outputData, tp, code);
6420 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6421 afs_uint16 remotePort, struct rx_debugStats * stat,
6422 afs_uint32 * supportedValues)
6424 struct rx_debugIn in;
6427 *supportedValues = 0;
6428 in.type = htonl(RX_DEBUGI_GETSTATS);
6431 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6432 &in, sizeof(in), stat, sizeof(*stat));
6435 * If the call was successful, fixup the version and indicate
6436 * what contents of the stat structure are valid.
6437 * Also do net to host conversion of fields here.
6441 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6442 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6444 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6445 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6447 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6448 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6450 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6451 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6453 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6454 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6456 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6457 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6459 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6460 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6462 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6463 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6466 stat->nFreePackets = ntohl(stat->nFreePackets);
6467 stat->packetReclaims = ntohl(stat->packetReclaims);
6468 stat->callsExecuted = ntohl(stat->callsExecuted);
6469 stat->nWaiting = ntohl(stat->nWaiting);
6470 stat->idleThreads = ntohl(stat->idleThreads);
6477 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6478 afs_uint16 remotePort, struct rx_stats * stat,
6479 afs_uint32 * supportedValues)
6481 struct rx_debugIn in;
6482 afs_int32 *lp = (afs_int32 *) stat;
6487 * supportedValues is currently unused, but added to allow future
6488 * versioning of this function.
6491 *supportedValues = 0;
6492 in.type = htonl(RX_DEBUGI_RXSTATS);
6494 memset(stat, 0, sizeof(*stat));
6496 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6497 &in, sizeof(in), stat, sizeof(*stat));
6502 * Do net to host conversion here
6505 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6514 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6515 afs_uint16 remotePort, size_t version_length,
6519 return MakeDebugCall(socket, remoteAddr, remotePort,
6520 RX_PACKET_TYPE_VERSION, a, 1, version,
6525 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6526 afs_uint16 remotePort, afs_int32 * nextConnection,
6527 int allConnections, afs_uint32 debugSupportedValues,
6528 struct rx_debugConn * conn,
6529 afs_uint32 * supportedValues)
6531 struct rx_debugIn in;
6536 * supportedValues is currently unused, but added to allow future
6537 * versioning of this function.
6540 *supportedValues = 0;
6541 if (allConnections) {
6542 in.type = htonl(RX_DEBUGI_GETALLCONN);
6544 in.type = htonl(RX_DEBUGI_GETCONN);
6546 in.index = htonl(*nextConnection);
6547 memset(conn, 0, sizeof(*conn));
6549 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6550 &in, sizeof(in), conn, sizeof(*conn));
6553 *nextConnection += 1;
6556 * Convert old connection format to new structure.
6559 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6560 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6561 #define MOVEvL(a) (conn->a = vL->a)
6563 /* any old or unrecognized version... */
6564 for (i = 0; i < RX_MAXCALLS; i++) {
6565 MOVEvL(callState[i]);
6566 MOVEvL(callMode[i]);
6567 MOVEvL(callFlags[i]);
6568 MOVEvL(callOther[i]);
6570 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6571 MOVEvL(secStats.type);
6572 MOVEvL(secStats.level);
6573 MOVEvL(secStats.flags);
6574 MOVEvL(secStats.expires);
6575 MOVEvL(secStats.packetsReceived);
6576 MOVEvL(secStats.packetsSent);
6577 MOVEvL(secStats.bytesReceived);
6578 MOVEvL(secStats.bytesSent);
6583 * Do net to host conversion here
6585 * I don't convert host or port since we are most likely
6586 * going to want these in NBO.
6588 conn->cid = ntohl(conn->cid);
6589 conn->serial = ntohl(conn->serial);
6590 for (i = 0; i < RX_MAXCALLS; i++) {
6591 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6593 conn->error = ntohl(conn->error);
6594 conn->secStats.flags = ntohl(conn->secStats.flags);
6595 conn->secStats.expires = ntohl(conn->secStats.expires);
6596 conn->secStats.packetsReceived =
6597 ntohl(conn->secStats.packetsReceived);
6598 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6599 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6600 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6601 conn->epoch = ntohl(conn->epoch);
6602 conn->natMTU = ntohl(conn->natMTU);
6609 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6610 afs_uint16 remotePort, afs_int32 * nextPeer,
6611 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6612 afs_uint32 * supportedValues)
6614 struct rx_debugIn in;
6618 * supportedValues is currently unused, but added to allow future
6619 * versioning of this function.
6622 *supportedValues = 0;
6623 in.type = htonl(RX_DEBUGI_GETPEER);
6624 in.index = htonl(*nextPeer);
6625 memset(peer, 0, sizeof(*peer));
6627 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6628 &in, sizeof(in), peer, sizeof(*peer));
6634 * Do net to host conversion here
6636 * I don't convert host or port since we are most likely
6637 * going to want these in NBO.
6639 peer->ifMTU = ntohs(peer->ifMTU);
6640 peer->idleWhen = ntohl(peer->idleWhen);
6641 peer->refCount = ntohs(peer->refCount);
6642 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6643 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6644 peer->rtt = ntohl(peer->rtt);
6645 peer->rtt_dev = ntohl(peer->rtt_dev);
6646 peer->timeout.sec = ntohl(peer->timeout.sec);
6647 peer->timeout.usec = ntohl(peer->timeout.usec);
6648 peer->nSent = ntohl(peer->nSent);
6649 peer->reSends = ntohl(peer->reSends);
6650 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6651 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6652 peer->rateFlag = ntohl(peer->rateFlag);
6653 peer->natMTU = ntohs(peer->natMTU);
6654 peer->maxMTU = ntohs(peer->maxMTU);
6655 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6656 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6657 peer->MTU = ntohs(peer->MTU);
6658 peer->cwind = ntohs(peer->cwind);
6659 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6660 peer->congestSeq = ntohs(peer->congestSeq);
6661 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6662 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6663 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6664 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6669 #endif /* RXDEBUG */
6674 struct rx_serverQueueEntry *np;
6677 register struct rx_call *call;
6678 register struct rx_serverQueueEntry *sq;
6682 if (rxinit_status == 1) {
6684 return; /* Already shutdown. */
6688 #ifndef AFS_PTHREAD_ENV
6689 FD_ZERO(&rx_selectMask);
6690 #endif /* AFS_PTHREAD_ENV */
6691 rxi_dataQuota = RX_MAX_QUOTA;
6692 #ifndef AFS_PTHREAD_ENV
6694 #endif /* AFS_PTHREAD_ENV */
6697 #ifndef AFS_PTHREAD_ENV
6698 #ifndef AFS_USE_GETTIMEOFDAY
6700 #endif /* AFS_USE_GETTIMEOFDAY */
6701 #endif /* AFS_PTHREAD_ENV */
6703 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6704 call = queue_First(&rx_freeCallQueue, rx_call);
6706 rxi_Free(call, sizeof(struct rx_call));
6709 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6710 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6716 struct rx_peer **peer_ptr, **peer_end;
6717 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6718 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6720 struct rx_peer *peer, *next;
6721 for (peer = *peer_ptr; peer; peer = next) {
6722 rx_interface_stat_p rpc_stat, nrpc_stat;
6725 (&peer->rpcStats, rpc_stat, nrpc_stat,
6726 rx_interface_stat)) {
6727 unsigned int num_funcs;
6730 queue_Remove(&rpc_stat->queue_header);
6731 queue_Remove(&rpc_stat->all_peers);
6732 num_funcs = rpc_stat->stats[0].func_total;
6734 sizeof(rx_interface_stat_t) +
6735 rpc_stat->stats[0].func_total *
6736 sizeof(rx_function_entry_v1_t);
6738 rxi_Free(rpc_stat, space);
6739 MUTEX_ENTER(&rx_rpc_stats);
6740 rxi_rpc_peer_stat_cnt -= num_funcs;
6741 MUTEX_EXIT(&rx_rpc_stats);
6745 MUTEX_ENTER(&rx_stats_mutex);
6746 rx_stats.nPeerStructs--;
6747 MUTEX_EXIT(&rx_stats_mutex);
6751 for (i = 0; i < RX_MAX_SERVICES; i++) {
6753 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6755 for (i = 0; i < rx_hashTableSize; i++) {
6756 register struct rx_connection *tc, *ntc;
6757 MUTEX_ENTER(&rx_connHashTable_lock);
6758 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6760 for (j = 0; j < RX_MAXCALLS; j++) {
6762 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6765 rxi_Free(tc, sizeof(*tc));
6767 MUTEX_EXIT(&rx_connHashTable_lock);
6770 MUTEX_ENTER(&freeSQEList_lock);
6772 while ((np = rx_FreeSQEList)) {
6773 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6774 MUTEX_DESTROY(&np->lock);
6775 rxi_Free(np, sizeof(*np));
6778 MUTEX_EXIT(&freeSQEList_lock);
6779 MUTEX_DESTROY(&freeSQEList_lock);
6780 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6781 MUTEX_DESTROY(&rx_connHashTable_lock);
6782 MUTEX_DESTROY(&rx_peerHashTable_lock);
6783 MUTEX_DESTROY(&rx_serverPool_lock);
6785 osi_Free(rx_connHashTable,
6786 rx_hashTableSize * sizeof(struct rx_connection *));
6787 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6789 UNPIN(rx_connHashTable,
6790 rx_hashTableSize * sizeof(struct rx_connection *));
6791 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6793 rxi_FreeAllPackets();
6795 MUTEX_ENTER(&rx_stats_mutex);
6796 rxi_dataQuota = RX_MAX_QUOTA;
6797 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6798 MUTEX_EXIT(&rx_stats_mutex);
6804 #ifdef RX_ENABLE_LOCKS
6806 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6808 if (!MUTEX_ISMINE(lockaddr))
6809 osi_Panic("Lock not held: %s", msg);
6811 #endif /* RX_ENABLE_LOCKS */
6816 * Routines to implement connection specific data.
6820 rx_KeyCreate(rx_destructor_t rtn)
6823 MUTEX_ENTER(&rxi_keyCreate_lock);
6824 key = rxi_keyCreate_counter++;
6825 rxi_keyCreate_destructor = (rx_destructor_t *)
6826 realloc((void *)rxi_keyCreate_destructor,
6827 (key + 1) * sizeof(rx_destructor_t));
6828 rxi_keyCreate_destructor[key] = rtn;
6829 MUTEX_EXIT(&rxi_keyCreate_lock);
6834 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6837 MUTEX_ENTER(&conn->conn_data_lock);
6838 if (!conn->specific) {
6839 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6840 for (i = 0; i < key; i++)
6841 conn->specific[i] = NULL;
6842 conn->nSpecific = key + 1;
6843 conn->specific[key] = ptr;
6844 } else if (key >= conn->nSpecific) {
6845 conn->specific = (void **)
6846 realloc(conn->specific, (key + 1) * sizeof(void *));
6847 for (i = conn->nSpecific; i < key; i++)
6848 conn->specific[i] = NULL;
6849 conn->nSpecific = key + 1;
6850 conn->specific[key] = ptr;
6852 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6853 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6854 conn->specific[key] = ptr;
6856 MUTEX_EXIT(&conn->conn_data_lock);
6860 rx_GetSpecific(struct rx_connection *conn, int key)
6863 MUTEX_ENTER(&conn->conn_data_lock);
6864 if (key >= conn->nSpecific)
6867 ptr = conn->specific[key];
6868 MUTEX_EXIT(&conn->conn_data_lock);
6872 #endif /* !KERNEL */
6875 * processStats is a queue used to store the statistics for the local
6876 * process. Its contents are similar to the contents of the rpcStats
6877 * queue on a rx_peer structure, but the actual data stored within
6878 * this queue contains totals across the lifetime of the process (assuming
6879 * the stats have not been reset) - unlike the per peer structures
6880 * which can come and go based upon the peer lifetime.
6883 static struct rx_queue processStats = { &processStats, &processStats };
6886 * peerStats is a queue used to store the statistics for all peer structs.
6887 * Its contents are the union of all the peer rpcStats queues.
6890 static struct rx_queue peerStats = { &peerStats, &peerStats };
6893 * rxi_monitor_processStats is used to turn process wide stat collection
6897 static int rxi_monitor_processStats = 0;
6900 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6903 static int rxi_monitor_peerStats = 0;
6906 * rxi_AddRpcStat - given all of the information for a particular rpc
6907 * call, create (if needed) and update the stat totals for the rpc.
6911 * IN stats - the queue of stats that will be updated with the new value
6913 * IN rxInterface - a unique number that identifies the rpc interface
6915 * IN currentFunc - the index of the function being invoked
6917 * IN totalFunc - the total number of functions in this interface
6919 * IN queueTime - the amount of time this function waited for a thread
6921 * IN execTime - the amount of time this function invocation took to execute
6923 * IN bytesSent - the number bytes sent by this invocation
6925 * IN bytesRcvd - the number bytes received by this invocation
6927 * IN isServer - if true, this invocation was made to a server
6929 * IN remoteHost - the ip address of the remote host
6931 * IN remotePort - the port of the remote host
6933 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6935 * INOUT counter - if a new stats structure is allocated, the counter will
6936 * be updated with the new number of allocated stat structures
6944 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6945 afs_uint32 currentFunc, afs_uint32 totalFunc,
6946 struct clock *queueTime, struct clock *execTime,
6947 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6948 afs_uint32 remoteHost, afs_uint32 remotePort,
6949 int addToPeerList, unsigned int *counter)
6952 rx_interface_stat_p rpc_stat, nrpc_stat;
6955 * See if there's already a structure for this interface
6958 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6959 if ((rpc_stat->stats[0].interfaceId == rxInterface)
6960 && (rpc_stat->stats[0].remote_is_server == isServer))
6965 * Didn't find a match so allocate a new structure and add it to the
6969 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
6970 || (rpc_stat->stats[0].interfaceId != rxInterface)
6971 || (rpc_stat->stats[0].remote_is_server != isServer)) {
6976 sizeof(rx_interface_stat_t) +
6977 totalFunc * sizeof(rx_function_entry_v1_t);
6979 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
6980 if (rpc_stat == NULL) {
6984 *counter += totalFunc;
6985 for (i = 0; i < totalFunc; i++) {
6986 rpc_stat->stats[i].remote_peer = remoteHost;
6987 rpc_stat->stats[i].remote_port = remotePort;
6988 rpc_stat->stats[i].remote_is_server = isServer;
6989 rpc_stat->stats[i].interfaceId = rxInterface;
6990 rpc_stat->stats[i].func_total = totalFunc;
6991 rpc_stat->stats[i].func_index = i;
6992 hzero(rpc_stat->stats[i].invocations);
6993 hzero(rpc_stat->stats[i].bytes_sent);
6994 hzero(rpc_stat->stats[i].bytes_rcvd);
6995 rpc_stat->stats[i].queue_time_sum.sec = 0;
6996 rpc_stat->stats[i].queue_time_sum.usec = 0;
6997 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
6998 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
6999 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7000 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7001 rpc_stat->stats[i].queue_time_max.sec = 0;
7002 rpc_stat->stats[i].queue_time_max.usec = 0;
7003 rpc_stat->stats[i].execution_time_sum.sec = 0;
7004 rpc_stat->stats[i].execution_time_sum.usec = 0;
7005 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7006 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7007 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7008 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7009 rpc_stat->stats[i].execution_time_max.sec = 0;
7010 rpc_stat->stats[i].execution_time_max.usec = 0;
7012 queue_Prepend(stats, rpc_stat);
7013 if (addToPeerList) {
7014 queue_Prepend(&peerStats, &rpc_stat->all_peers);
7019 * Increment the stats for this function
7022 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
7023 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
7024 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
7025 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
7026 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7027 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7028 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7030 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7031 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7033 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7034 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7036 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7037 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7039 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7040 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7048 * rx_IncrementTimeAndCount - increment the times and count for a particular
7053 * IN peer - the peer who invoked the rpc
7055 * IN rxInterface - a unique number that identifies the rpc interface
7057 * IN currentFunc - the index of the function being invoked
7059 * IN totalFunc - the total number of functions in this interface
7061 * IN queueTime - the amount of time this function waited for a thread
7063 * IN execTime - the amount of time this function invocation took to execute
7065 * IN bytesSent - the number bytes sent by this invocation
7067 * IN bytesRcvd - the number bytes received by this invocation
7069 * IN isServer - if true, this invocation was made to a server
7077 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7078 afs_uint32 currentFunc, afs_uint32 totalFunc,
7079 struct clock *queueTime, struct clock *execTime,
7080 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7084 MUTEX_ENTER(&rx_rpc_stats);
7085 MUTEX_ENTER(&peer->peer_lock);
7087 if (rxi_monitor_peerStats) {
7088 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7089 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7090 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7093 if (rxi_monitor_processStats) {
7094 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7095 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7096 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7099 MUTEX_EXIT(&peer->peer_lock);
7100 MUTEX_EXIT(&rx_rpc_stats);
7105 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7109 * IN callerVersion - the rpc stat version of the caller.
7111 * IN count - the number of entries to marshall.
7113 * IN stats - pointer to stats to be marshalled.
7115 * OUT ptr - Where to store the marshalled data.
7122 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7123 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7129 * We only support the first version
7131 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7132 *(ptr++) = stats->remote_peer;
7133 *(ptr++) = stats->remote_port;
7134 *(ptr++) = stats->remote_is_server;
7135 *(ptr++) = stats->interfaceId;
7136 *(ptr++) = stats->func_total;
7137 *(ptr++) = stats->func_index;
7138 *(ptr++) = hgethi(stats->invocations);
7139 *(ptr++) = hgetlo(stats->invocations);
7140 *(ptr++) = hgethi(stats->bytes_sent);
7141 *(ptr++) = hgetlo(stats->bytes_sent);
7142 *(ptr++) = hgethi(stats->bytes_rcvd);
7143 *(ptr++) = hgetlo(stats->bytes_rcvd);
7144 *(ptr++) = stats->queue_time_sum.sec;
7145 *(ptr++) = stats->queue_time_sum.usec;
7146 *(ptr++) = stats->queue_time_sum_sqr.sec;
7147 *(ptr++) = stats->queue_time_sum_sqr.usec;
7148 *(ptr++) = stats->queue_time_min.sec;
7149 *(ptr++) = stats->queue_time_min.usec;
7150 *(ptr++) = stats->queue_time_max.sec;
7151 *(ptr++) = stats->queue_time_max.usec;
7152 *(ptr++) = stats->execution_time_sum.sec;
7153 *(ptr++) = stats->execution_time_sum.usec;
7154 *(ptr++) = stats->execution_time_sum_sqr.sec;
7155 *(ptr++) = stats->execution_time_sum_sqr.usec;
7156 *(ptr++) = stats->execution_time_min.sec;
7157 *(ptr++) = stats->execution_time_min.usec;
7158 *(ptr++) = stats->execution_time_max.sec;
7159 *(ptr++) = stats->execution_time_max.usec;
7165 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7170 * IN callerVersion - the rpc stat version of the caller
7172 * OUT myVersion - the rpc stat version of this function
7174 * OUT clock_sec - local time seconds
7176 * OUT clock_usec - local time microseconds
7178 * OUT allocSize - the number of bytes allocated to contain stats
7180 * OUT statCount - the number stats retrieved from this process.
7182 * OUT stats - the actual stats retrieved from this process.
7186 * Returns void. If successful, stats will != NULL.
7190 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7191 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7192 size_t * allocSize, afs_uint32 * statCount,
7193 afs_uint32 ** stats)
7203 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7206 * Check to see if stats are enabled
7209 MUTEX_ENTER(&rx_rpc_stats);
7210 if (!rxi_monitor_processStats) {
7211 MUTEX_EXIT(&rx_rpc_stats);
7215 clock_GetTime(&now);
7216 *clock_sec = now.sec;
7217 *clock_usec = now.usec;
7220 * Allocate the space based upon the caller version
7222 * If the client is at an older version than we are,
7223 * we return the statistic data in the older data format, but
7224 * we still return our version number so the client knows we
7225 * are maintaining more data than it can retrieve.
7228 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7229 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7230 *statCount = rxi_rpc_process_stat_cnt;
7233 * This can't happen yet, but in the future version changes
7234 * can be handled by adding additional code here
7238 if (space > (size_t) 0) {
7240 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7243 rx_interface_stat_p rpc_stat, nrpc_stat;
7247 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7249 * Copy the data based upon the caller version
7251 rx_MarshallProcessRPCStats(callerVersion,
7252 rpc_stat->stats[0].func_total,
7253 rpc_stat->stats, &ptr);
7259 MUTEX_EXIT(&rx_rpc_stats);
7264 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7268 * IN callerVersion - the rpc stat version of the caller
7270 * OUT myVersion - the rpc stat version of this function
7272 * OUT clock_sec - local time seconds
7274 * OUT clock_usec - local time microseconds
7276 * OUT allocSize - the number of bytes allocated to contain stats
7278 * OUT statCount - the number of stats retrieved from the individual
7281 * OUT stats - the actual stats retrieved from the individual peer structures.
7285 * Returns void. If successful, stats will != NULL.
7289 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7290 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7291 size_t * allocSize, afs_uint32 * statCount,
7292 afs_uint32 ** stats)
7302 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7305 * Check to see if stats are enabled
7308 MUTEX_ENTER(&rx_rpc_stats);
7309 if (!rxi_monitor_peerStats) {
7310 MUTEX_EXIT(&rx_rpc_stats);
7314 clock_GetTime(&now);
7315 *clock_sec = now.sec;
7316 *clock_usec = now.usec;
7319 * Allocate the space based upon the caller version
7321 * If the client is at an older version than we are,
7322 * we return the statistic data in the older data format, but
7323 * we still return our version number so the client knows we
7324 * are maintaining more data than it can retrieve.
7327 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7328 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7329 *statCount = rxi_rpc_peer_stat_cnt;
7332 * This can't happen yet, but in the future version changes
7333 * can be handled by adding additional code here
7337 if (space > (size_t) 0) {
7339 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7342 rx_interface_stat_p rpc_stat, nrpc_stat;
7346 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7348 * We have to fix the offset of rpc_stat since we are
7349 * keeping this structure on two rx_queues. The rx_queue
7350 * package assumes that the rx_queue member is the first
7351 * member of the structure. That is, rx_queue assumes that
7352 * any one item is only on one queue at a time. We are
7353 * breaking that assumption and so we have to do a little
7354 * math to fix our pointers.
7357 fix_offset = (char *)rpc_stat;
7358 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7359 rpc_stat = (rx_interface_stat_p) fix_offset;
7362 * Copy the data based upon the caller version
7364 rx_MarshallProcessRPCStats(callerVersion,
7365 rpc_stat->stats[0].func_total,
7366 rpc_stat->stats, &ptr);
7372 MUTEX_EXIT(&rx_rpc_stats);
7377 * rx_FreeRPCStats - free memory allocated by
7378 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7382 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7383 * rx_RetrievePeerRPCStats
7385 * IN allocSize - the number of bytes in stats.
7393 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7395 rxi_Free(stats, allocSize);
7399 * rx_queryProcessRPCStats - see if process rpc stat collection is
7400 * currently enabled.
7406 * Returns 0 if stats are not enabled != 0 otherwise
7410 rx_queryProcessRPCStats(void)
7413 MUTEX_ENTER(&rx_rpc_stats);
7414 rc = rxi_monitor_processStats;
7415 MUTEX_EXIT(&rx_rpc_stats);
7420 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7426 * Returns 0 if stats are not enabled != 0 otherwise
7430 rx_queryPeerRPCStats(void)
7433 MUTEX_ENTER(&rx_rpc_stats);
7434 rc = rxi_monitor_peerStats;
7435 MUTEX_EXIT(&rx_rpc_stats);
7440 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7450 rx_enableProcessRPCStats(void)
7452 MUTEX_ENTER(&rx_rpc_stats);
7453 rx_enable_stats = 1;
7454 rxi_monitor_processStats = 1;
7455 MUTEX_EXIT(&rx_rpc_stats);
7459 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7469 rx_enablePeerRPCStats(void)
7471 MUTEX_ENTER(&rx_rpc_stats);
7472 rx_enable_stats = 1;
7473 rxi_monitor_peerStats = 1;
7474 MUTEX_EXIT(&rx_rpc_stats);
7478 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7488 rx_disableProcessRPCStats(void)
7490 rx_interface_stat_p rpc_stat, nrpc_stat;
7493 MUTEX_ENTER(&rx_rpc_stats);
7496 * Turn off process statistics and if peer stats is also off, turn
7500 rxi_monitor_processStats = 0;
7501 if (rxi_monitor_peerStats == 0) {
7502 rx_enable_stats = 0;
7505 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7506 unsigned int num_funcs = 0;
7509 queue_Remove(rpc_stat);
7510 num_funcs = rpc_stat->stats[0].func_total;
7512 sizeof(rx_interface_stat_t) +
7513 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7515 rxi_Free(rpc_stat, space);
7516 rxi_rpc_process_stat_cnt -= num_funcs;
7518 MUTEX_EXIT(&rx_rpc_stats);
7522 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7532 rx_disablePeerRPCStats(void)
7534 struct rx_peer **peer_ptr, **peer_end;
7537 MUTEX_ENTER(&rx_rpc_stats);
7540 * Turn off peer statistics and if process stats is also off, turn
7544 rxi_monitor_peerStats = 0;
7545 if (rxi_monitor_processStats == 0) {
7546 rx_enable_stats = 0;
7549 MUTEX_ENTER(&rx_peerHashTable_lock);
7550 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7551 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7553 struct rx_peer *peer, *next, *prev;
7554 for (prev = peer = *peer_ptr; peer; peer = next) {
7556 code = MUTEX_TRYENTER(&peer->peer_lock);
7558 rx_interface_stat_p rpc_stat, nrpc_stat;
7561 (&peer->rpcStats, rpc_stat, nrpc_stat,
7562 rx_interface_stat)) {
7563 unsigned int num_funcs = 0;
7566 queue_Remove(&rpc_stat->queue_header);
7567 queue_Remove(&rpc_stat->all_peers);
7568 num_funcs = rpc_stat->stats[0].func_total;
7570 sizeof(rx_interface_stat_t) +
7571 rpc_stat->stats[0].func_total *
7572 sizeof(rx_function_entry_v1_t);
7574 rxi_Free(rpc_stat, space);
7575 rxi_rpc_peer_stat_cnt -= num_funcs;
7577 MUTEX_EXIT(&peer->peer_lock);
7578 if (prev == *peer_ptr) {
7588 MUTEX_EXIT(&rx_peerHashTable_lock);
7589 MUTEX_EXIT(&rx_rpc_stats);
7593 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7598 * IN clearFlag - flag indicating which stats to clear
7606 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7608 rx_interface_stat_p rpc_stat, nrpc_stat;
7610 MUTEX_ENTER(&rx_rpc_stats);
7612 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7613 unsigned int num_funcs = 0, i;
7614 num_funcs = rpc_stat->stats[0].func_total;
7615 for (i = 0; i < num_funcs; i++) {
7616 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7617 hzero(rpc_stat->stats[i].invocations);
7619 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7620 hzero(rpc_stat->stats[i].bytes_sent);
7622 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7623 hzero(rpc_stat->stats[i].bytes_rcvd);
7625 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7626 rpc_stat->stats[i].queue_time_sum.sec = 0;
7627 rpc_stat->stats[i].queue_time_sum.usec = 0;
7629 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7630 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7631 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7633 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7634 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7635 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7637 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7638 rpc_stat->stats[i].queue_time_max.sec = 0;
7639 rpc_stat->stats[i].queue_time_max.usec = 0;
7641 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7642 rpc_stat->stats[i].execution_time_sum.sec = 0;
7643 rpc_stat->stats[i].execution_time_sum.usec = 0;
7645 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7646 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7647 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7649 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7650 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7651 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7653 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7654 rpc_stat->stats[i].execution_time_max.sec = 0;
7655 rpc_stat->stats[i].execution_time_max.usec = 0;
7660 MUTEX_EXIT(&rx_rpc_stats);
7664 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7669 * IN clearFlag - flag indicating which stats to clear
7677 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7679 rx_interface_stat_p rpc_stat, nrpc_stat;
7681 MUTEX_ENTER(&rx_rpc_stats);
7683 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7684 unsigned int num_funcs = 0, i;
7687 * We have to fix the offset of rpc_stat since we are
7688 * keeping this structure on two rx_queues. The rx_queue
7689 * package assumes that the rx_queue member is the first
7690 * member of the structure. That is, rx_queue assumes that
7691 * any one item is only on one queue at a time. We are
7692 * breaking that assumption and so we have to do a little
7693 * math to fix our pointers.
7696 fix_offset = (char *)rpc_stat;
7697 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7698 rpc_stat = (rx_interface_stat_p) fix_offset;
7700 num_funcs = rpc_stat->stats[0].func_total;
7701 for (i = 0; i < num_funcs; i++) {
7702 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7703 hzero(rpc_stat->stats[i].invocations);
7705 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7706 hzero(rpc_stat->stats[i].bytes_sent);
7708 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7709 hzero(rpc_stat->stats[i].bytes_rcvd);
7711 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7712 rpc_stat->stats[i].queue_time_sum.sec = 0;
7713 rpc_stat->stats[i].queue_time_sum.usec = 0;
7715 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7716 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7717 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7719 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7720 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7721 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7723 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7724 rpc_stat->stats[i].queue_time_max.sec = 0;
7725 rpc_stat->stats[i].queue_time_max.usec = 0;
7727 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7728 rpc_stat->stats[i].execution_time_sum.sec = 0;
7729 rpc_stat->stats[i].execution_time_sum.usec = 0;
7731 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7732 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7733 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7735 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7736 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7737 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7739 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7740 rpc_stat->stats[i].execution_time_max.sec = 0;
7741 rpc_stat->stats[i].execution_time_max.usec = 0;
7746 MUTEX_EXIT(&rx_rpc_stats);
7750 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7751 * is authorized to enable/disable/clear RX statistics.
7753 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7756 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7758 rxi_rxstat_userok = proc;
7762 rx_RxStatUserOk(struct rx_call *call)
7764 if (!rxi_rxstat_userok)
7766 return rxi_rxstat_userok(call);
7771 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
7772 * function in the MSVC runtime DLL (msvcrt.dll).
7774 * Note: the system serializes calls to this function.
7777 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
7778 DWORD reason, /* reason function is being called */
7779 LPVOID reserved) /* reserved for future use */
7782 case DLL_PROCESS_ATTACH:
7783 /* library is being attached to a process */
7787 case DLL_PROCESS_DETACH: