2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
14 #include "afs/param.h"
16 #include <afs/param.h>
23 #include "afs/sysincludes.h"
24 #include "afsincludes.h"
30 #include <net/net_globals.h>
31 #endif /* AFS_OSF_ENV */
32 #ifdef AFS_LINUX20_ENV
35 #include "netinet/in.h"
36 #include "afs/afs_args.h"
37 #include "afs/afs_osi.h"
38 #ifdef RX_KERNEL_TRACE
39 #include "rx_kcommon.h"
41 #if (defined(AFS_AUX_ENV) || defined(AFS_AIX_ENV))
45 #undef RXDEBUG /* turn off debugging */
47 #if defined(AFS_SGI_ENV)
48 #include "sys/debug.h"
57 #endif /* AFS_OSF_ENV */
59 #include "afs/sysincludes.h"
60 #include "afsincludes.h"
63 #include "rx_kmutex.h"
64 #include "rx_kernel.h"
68 #include "rx_globals.h"
70 #define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
71 #define AFSOP_STOP_AFS 211 /* Stop AFS process */
72 #define AFSOP_STOP_BKG 212 /* Stop BKG process */
74 extern afs_int32 afs_termState;
76 #include "sys/lockl.h"
77 #include "sys/lock_def.h"
78 #endif /* AFS_AIX41_ENV */
79 # include "rxgen_consts.h"
81 # include <sys/types.h>
86 # include <afs/afsutil.h>
87 # include <WINNT\afsreg.h>
89 # include <sys/socket.h>
90 # include <sys/file.h>
92 # include <sys/stat.h>
93 # include <netinet/in.h>
94 # include <sys/time.h>
104 # include "rx_user.h"
105 # include "rx_clock.h"
106 # include "rx_queue.h"
107 # include "rx_globals.h"
108 # include "rx_trace.h"
109 # include <afs/rxgen_consts.h>
112 int (*registerProgram) () = 0;
113 int (*swapNameProgram) () = 0;
115 /* Local static routines */
116 static void rxi_DestroyConnectionNoLock(register struct rx_connection *conn);
117 #ifdef RX_ENABLE_LOCKS
118 static void rxi_SetAcksInTransmitQueue(register struct rx_call *call);
121 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
123 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
124 afs_int32 rxi_start_in_error;
126 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
129 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
130 * currently allocated within rx. This number is used to allocate the
131 * memory required to return the statistics when queried.
134 static unsigned int rxi_rpc_peer_stat_cnt;
137 * rxi_rpc_process_stat_cnt counts the total number of local process stat
138 * structures currently allocated within rx. The number is used to allocate
139 * the memory required to return the statistics when queried.
142 static unsigned int rxi_rpc_process_stat_cnt;
144 #if !defined(offsetof)
145 #include <stddef.h> /* for definition of offsetof() */
148 #ifdef AFS_PTHREAD_ENV
152 * Use procedural initialization of mutexes/condition variables
156 extern pthread_mutex_t rx_stats_mutex;
157 extern pthread_mutex_t des_init_mutex;
158 extern pthread_mutex_t des_random_mutex;
159 extern pthread_mutex_t rx_clock_mutex;
160 extern pthread_mutex_t rxi_connCacheMutex;
161 extern pthread_mutex_t rx_event_mutex;
162 extern pthread_mutex_t osi_malloc_mutex;
163 extern pthread_mutex_t event_handler_mutex;
164 extern pthread_mutex_t listener_mutex;
165 extern pthread_mutex_t rx_if_init_mutex;
166 extern pthread_mutex_t rx_if_mutex;
167 extern pthread_mutex_t rxkad_client_uid_mutex;
168 extern pthread_mutex_t rxkad_random_mutex;
170 extern pthread_cond_t rx_event_handler_cond;
171 extern pthread_cond_t rx_listener_cond;
173 static pthread_mutex_t epoch_mutex;
174 static pthread_mutex_t rx_init_mutex;
175 static pthread_mutex_t rx_debug_mutex;
178 rxi_InitPthread(void)
180 assert(pthread_mutex_init(&rx_clock_mutex, (const pthread_mutexattr_t *)0)
182 assert(pthread_mutex_init(&rx_stats_mutex, (const pthread_mutexattr_t *)0)
184 assert(pthread_mutex_init
185 (&rxi_connCacheMutex, (const pthread_mutexattr_t *)0) == 0);
186 assert(pthread_mutex_init(&rx_init_mutex, (const pthread_mutexattr_t *)0)
188 assert(pthread_mutex_init(&epoch_mutex, (const pthread_mutexattr_t *)0) ==
190 assert(pthread_mutex_init(&rx_event_mutex, (const pthread_mutexattr_t *)0)
192 assert(pthread_mutex_init(&des_init_mutex, (const pthread_mutexattr_t *)0)
194 assert(pthread_mutex_init
195 (&des_random_mutex, (const pthread_mutexattr_t *)0) == 0);
196 assert(pthread_mutex_init
197 (&osi_malloc_mutex, (const pthread_mutexattr_t *)0) == 0);
198 assert(pthread_mutex_init
199 (&event_handler_mutex, (const pthread_mutexattr_t *)0) == 0);
200 assert(pthread_mutex_init(&listener_mutex, (const pthread_mutexattr_t *)0)
202 assert(pthread_mutex_init
203 (&rx_if_init_mutex, (const pthread_mutexattr_t *)0) == 0);
204 assert(pthread_mutex_init(&rx_if_mutex, (const pthread_mutexattr_t *)0) ==
206 assert(pthread_mutex_init
207 (&rxkad_client_uid_mutex, (const pthread_mutexattr_t *)0) == 0);
208 assert(pthread_mutex_init
209 (&rxkad_random_mutex, (const pthread_mutexattr_t *)0) == 0);
210 assert(pthread_mutex_init(&rx_debug_mutex, (const pthread_mutexattr_t *)0)
213 assert(pthread_cond_init
214 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
215 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
217 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
218 assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
220 rxkad_global_stats_init();
223 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
224 #define INIT_PTHREAD_LOCKS \
225 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
227 * The rx_stats_mutex mutex protects the following global variables:
232 * rxi_lowConnRefCount
233 * rxi_lowPeerRefCount
242 #define INIT_PTHREAD_LOCKS
246 /* Variables for handling the minProcs implementation. availProcs gives the
247 * number of threads available in the pool at this moment (not counting dudes
248 * executing right now). totalMin gives the total number of procs required
249 * for handling all minProcs requests. minDeficit is a dynamic variable
250 * tracking the # of procs required to satisfy all of the remaining minProcs
252 * For fine grain locking to work, the quota check and the reservation of
253 * a server thread has to come while rxi_availProcs and rxi_minDeficit
254 * are locked. To this end, the code has been modified under #ifdef
255 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
256 * same time. A new function, ReturnToServerPool() returns the allocation.
258 * A call can be on several queue's (but only one at a time). When
259 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
260 * that no one else is touching the queue. To this end, we store the address
261 * of the queue lock in the call structure (under the call lock) when we
262 * put the call on a queue, and we clear the call_queue_lock when the
263 * call is removed from a queue (once the call lock has been obtained).
264 * This allows rxi_ResetCall to safely synchronize with others wishing
265 * to manipulate the queue.
268 #ifdef RX_ENABLE_LOCKS
269 static afs_kmutex_t rx_rpc_stats;
270 void rxi_StartUnlocked();
273 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
274 ** pretty good that the next packet coming in is from the same connection
275 ** as the last packet, since we're send multiple packets in a transmit window.
277 struct rx_connection *rxLastConn = 0;
279 #ifdef RX_ENABLE_LOCKS
280 /* The locking hierarchy for rx fine grain locking is composed of these
283 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
284 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
285 * call->lock - locks call data fields.
286 * These are independent of each other:
287 * rx_freeCallQueue_lock
292 * serverQueueEntry->lock
294 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
295 * peer->lock - locks peer data fields.
296 * conn_data_lock - that more than one thread is not updating a conn data
297 * field at the same time.
305 * Do we need a lock to protect the peer field in the conn structure?
306 * conn->peer was previously a constant for all intents and so has no
307 * lock protecting this field. The multihomed client delta introduced
308 * a RX code change : change the peer field in the connection structure
309 * to that remote inetrface from which the last packet for this
310 * connection was sent out. This may become an issue if further changes
313 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
314 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
316 /* rxdb_fileID is used to identify the lock location, along with line#. */
317 static int rxdb_fileID = RXDB_FILE_RX;
318 #endif /* RX_LOCKS_DB */
319 #else /* RX_ENABLE_LOCKS */
320 #define SET_CALL_QUEUE_LOCK(C, L)
321 #define CLEAR_CALL_QUEUE_LOCK(C)
322 #endif /* RX_ENABLE_LOCKS */
323 struct rx_serverQueueEntry *rx_waitForPacket = 0;
324 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
326 /* ------------Exported Interfaces------------- */
328 /* This function allows rxkad to set the epoch to a suitably random number
329 * which rx_NewConnection will use in the future. The principle purpose is to
330 * get rxnull connections to use the same epoch as the rxkad connections do, at
331 * least once the first rxkad connection is established. This is important now
332 * that the host/port addresses aren't used in FindConnection: the uniqueness
333 * of epoch/cid matters and the start time won't do. */
335 #ifdef AFS_PTHREAD_ENV
337 * This mutex protects the following global variables:
341 #define LOCK_EPOCH assert(pthread_mutex_lock(&epoch_mutex)==0)
342 #define UNLOCK_EPOCH assert(pthread_mutex_unlock(&epoch_mutex)==0)
346 #endif /* AFS_PTHREAD_ENV */
349 rx_SetEpoch(afs_uint32 epoch)
356 /* Initialize rx. A port number may be mentioned, in which case this
357 * becomes the default port number for any service installed later.
358 * If 0 is provided for the port number, a random port will be chosen
359 * by the kernel. Whether this will ever overlap anything in
360 * /etc/services is anybody's guess... Returns 0 on success, -1 on
362 static int rxinit_status = 1;
363 #ifdef AFS_PTHREAD_ENV
365 * This mutex protects the following global variables:
369 #define LOCK_RX_INIT assert(pthread_mutex_lock(&rx_init_mutex)==0)
370 #define UNLOCK_RX_INIT assert(pthread_mutex_unlock(&rx_init_mutex)==0)
373 #define UNLOCK_RX_INIT
377 rx_InitHost(u_int host, u_int port)
384 char *htable, *ptable;
387 #if defined(AFS_DJGPP_ENV) && !defined(DEBUG)
388 __djgpp_set_quiet_socket(1);
395 if (rxinit_status == 0) {
396 tmp_status = rxinit_status;
398 return tmp_status; /* Already started; return previous error code. */
404 if (afs_winsockInit() < 0)
410 * Initialize anything necessary to provide a non-premptive threading
413 rxi_InitializeThreadSupport();
416 /* Allocate and initialize a socket for client and perhaps server
419 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
420 if (rx_socket == OSI_NULLSOCKET) {
424 #ifdef RX_ENABLE_LOCKS
427 #endif /* RX_LOCKS_DB */
428 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
429 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
430 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
431 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
432 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
434 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
436 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
438 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
440 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
442 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
444 #if defined(KERNEL) && defined(AFS_HPUX110_ENV)
446 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
447 #endif /* KERNEL && AFS_HPUX110_ENV */
448 #endif /* RX_ENABLE_LOCKS */
451 rx_connDeadTime = 12;
452 rx_tranquil = 0; /* reset flag */
453 memset((char *)&rx_stats, 0, sizeof(struct rx_stats));
455 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
456 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
457 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
458 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
459 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
460 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
462 /* Malloc up a bunch of packets & buffers */
464 queue_Init(&rx_freePacketQueue);
465 rxi_NeedMorePackets = FALSE;
466 #ifdef RX_ENABLE_TSFPQ
467 rx_nPackets = 0; /* in TSFPQ version, rx_nPackets is managed by rxi_MorePackets* */
468 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
469 #else /* RX_ENABLE_TSFPQ */
470 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
471 rxi_MorePackets(rx_nPackets);
472 #endif /* RX_ENABLE_TSFPQ */
479 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
480 tv.tv_sec = clock_now.sec;
481 tv.tv_usec = clock_now.usec;
482 srand((unsigned int)tv.tv_usec);
489 #if defined(KERNEL) && !defined(UKERNEL)
490 /* Really, this should never happen in a real kernel */
493 struct sockaddr_in addr;
494 int addrlen = sizeof(addr);
495 if (getsockname((int)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
499 rx_port = addr.sin_port;
502 rx_stats.minRtt.sec = 9999999;
504 rx_SetEpoch(tv.tv_sec | 0x80000000);
506 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
507 * will provide a randomer value. */
509 MUTEX_ENTER(&rx_stats_mutex);
510 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
511 MUTEX_EXIT(&rx_stats_mutex);
512 /* *Slightly* random start time for the cid. This is just to help
513 * out with the hashing function at the peer */
514 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
515 rx_connHashTable = (struct rx_connection **)htable;
516 rx_peerHashTable = (struct rx_peer **)ptable;
518 rx_lastAckDelay.sec = 0;
519 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
520 rx_hardAckDelay.sec = 0;
521 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
522 rx_softAckDelay.sec = 0;
523 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
525 rxevent_Init(20, rxi_ReScheduleEvents);
527 /* Initialize various global queues */
528 queue_Init(&rx_idleServerQueue);
529 queue_Init(&rx_incomingCallQueue);
530 queue_Init(&rx_freeCallQueue);
532 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
533 /* Initialize our list of usable IP addresses. */
537 /* Start listener process (exact function is dependent on the
538 * implementation environment--kernel or user space) */
542 tmp_status = rxinit_status = 0;
550 return rx_InitHost(htonl(INADDR_ANY), port);
553 /* called with unincremented nRequestsRunning to see if it is OK to start
554 * a new thread in this service. Could be "no" for two reasons: over the
555 * max quota, or would prevent others from reaching their min quota.
557 #ifdef RX_ENABLE_LOCKS
558 /* This verion of QuotaOK reserves quota if it's ok while the
559 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
562 QuotaOK(register struct rx_service *aservice)
564 /* check if over max quota */
565 if (aservice->nRequestsRunning >= aservice->maxProcs) {
569 /* under min quota, we're OK */
570 /* otherwise, can use only if there are enough to allow everyone
571 * to go to their min quota after this guy starts.
573 MUTEX_ENTER(&rx_stats_mutex);
574 if ((aservice->nRequestsRunning < aservice->minProcs)
575 || (rxi_availProcs > rxi_minDeficit)) {
576 aservice->nRequestsRunning++;
577 /* just started call in minProcs pool, need fewer to maintain
579 if (aservice->nRequestsRunning <= aservice->minProcs)
582 MUTEX_EXIT(&rx_stats_mutex);
585 MUTEX_EXIT(&rx_stats_mutex);
591 ReturnToServerPool(register struct rx_service *aservice)
593 aservice->nRequestsRunning--;
594 MUTEX_ENTER(&rx_stats_mutex);
595 if (aservice->nRequestsRunning < aservice->minProcs)
598 MUTEX_EXIT(&rx_stats_mutex);
601 #else /* RX_ENABLE_LOCKS */
603 QuotaOK(register struct rx_service *aservice)
606 /* under min quota, we're OK */
607 if (aservice->nRequestsRunning < aservice->minProcs)
610 /* check if over max quota */
611 if (aservice->nRequestsRunning >= aservice->maxProcs)
614 /* otherwise, can use only if there are enough to allow everyone
615 * to go to their min quota after this guy starts.
617 if (rxi_availProcs > rxi_minDeficit)
621 #endif /* RX_ENABLE_LOCKS */
624 /* Called by rx_StartServer to start up lwp's to service calls.
625 NExistingProcs gives the number of procs already existing, and which
626 therefore needn't be created. */
628 rxi_StartServerProcs(int nExistingProcs)
630 register struct rx_service *service;
635 /* For each service, reserve N processes, where N is the "minimum"
636 * number of processes that MUST be able to execute a request in parallel,
637 * at any time, for that process. Also compute the maximum difference
638 * between any service's maximum number of processes that can run
639 * (i.e. the maximum number that ever will be run, and a guarantee
640 * that this number will run if other services aren't running), and its
641 * minimum number. The result is the extra number of processes that
642 * we need in order to provide the latter guarantee */
643 for (i = 0; i < RX_MAX_SERVICES; i++) {
645 service = rx_services[i];
646 if (service == (struct rx_service *)0)
648 nProcs += service->minProcs;
649 diff = service->maxProcs - service->minProcs;
653 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
654 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
655 for (i = 0; i < nProcs; i++) {
656 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
662 /* This routine is only required on Windows */
664 rx_StartClientThread(void)
666 #ifdef AFS_PTHREAD_ENV
668 pid = (int) pthread_self();
669 #endif /* AFS_PTHREAD_ENV */
671 #endif /* AFS_NT40_ENV */
673 /* This routine must be called if any services are exported. If the
674 * donateMe flag is set, the calling process is donated to the server
677 rx_StartServer(int donateMe)
679 register struct rx_service *service;
685 /* Start server processes, if necessary (exact function is dependent
686 * on the implementation environment--kernel or user space). DonateMe
687 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
688 * case, one less new proc will be created rx_StartServerProcs.
690 rxi_StartServerProcs(donateMe);
692 /* count up the # of threads in minProcs, and add set the min deficit to
693 * be that value, too.
695 for (i = 0; i < RX_MAX_SERVICES; i++) {
696 service = rx_services[i];
697 if (service == (struct rx_service *)0)
699 MUTEX_ENTER(&rx_stats_mutex);
700 rxi_totalMin += service->minProcs;
701 /* below works even if a thread is running, since minDeficit would
702 * still have been decremented and later re-incremented.
704 rxi_minDeficit += service->minProcs;
705 MUTEX_EXIT(&rx_stats_mutex);
708 /* Turn on reaping of idle server connections */
709 rxi_ReapConnections();
718 #ifdef AFS_PTHREAD_ENV
720 pid = (pid_t) pthread_self();
721 #else /* AFS_PTHREAD_ENV */
723 LWP_CurrentProcess(&pid);
724 #endif /* AFS_PTHREAD_ENV */
726 sprintf(name, "srv_%d", ++nProcs);
728 (*registerProgram) (pid, name);
730 #endif /* AFS_NT40_ENV */
731 rx_ServerProc(); /* Never returns */
733 #ifdef RX_ENABLE_TSFPQ
734 /* no use leaving packets around in this thread's local queue if
735 * it isn't getting donated to the server thread pool.
737 rxi_FlushLocalPacketsTSFPQ();
738 #endif /* RX_ENABLE_TSFPQ */
742 /* Create a new client connection to the specified service, using the
743 * specified security object to implement the security model for this
745 struct rx_connection *
746 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
747 register struct rx_securityClass *securityObject,
748 int serviceSecurityIndex)
752 register struct rx_connection *conn;
757 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", ntohl(shost), ntohs(sport), sservice, securityObject, serviceSecurityIndex));
759 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
760 * the case of kmem_alloc? */
761 conn = rxi_AllocConnection();
762 #ifdef RX_ENABLE_LOCKS
763 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
764 MUTEX_INIT(&conn->conn_data_lock, "conn call lock", MUTEX_DEFAULT, 0);
765 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
768 MUTEX_ENTER(&rx_connHashTable_lock);
769 cid = (rx_nextCid += RX_MAXCALLS);
770 conn->type = RX_CLIENT_CONNECTION;
772 conn->epoch = rx_epoch;
773 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
774 conn->serviceId = sservice;
775 conn->securityObject = securityObject;
776 /* This doesn't work in all compilers with void (they're buggy), so fake it
778 conn->securityData = (VOID *) 0;
779 conn->securityIndex = serviceSecurityIndex;
780 rx_SetConnDeadTime(conn, rx_connDeadTime);
781 conn->ackRate = RX_FAST_ACK_RATE;
783 conn->specific = NULL;
784 conn->challengeEvent = NULL;
785 conn->delayedAbortEvent = NULL;
786 conn->abortCount = 0;
789 RXS_NewConnection(securityObject, conn);
791 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
793 conn->refCount++; /* no lock required since only this thread knows... */
794 conn->next = rx_connHashTable[hashindex];
795 rx_connHashTable[hashindex] = conn;
796 MUTEX_ENTER(&rx_stats_mutex);
797 rx_stats.nClientConns++;
798 MUTEX_EXIT(&rx_stats_mutex);
800 MUTEX_EXIT(&rx_connHashTable_lock);
806 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
808 /* The idea is to set the dead time to a value that allows several
809 * keepalives to be dropped without timing out the connection. */
810 conn->secondsUntilDead = MAX(seconds, 6);
811 conn->secondsUntilPing = conn->secondsUntilDead / 6;
814 int rxi_lowPeerRefCount = 0;
815 int rxi_lowConnRefCount = 0;
818 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
819 * NOTE: must not be called with rx_connHashTable_lock held.
822 rxi_CleanupConnection(struct rx_connection *conn)
824 /* Notify the service exporter, if requested, that this connection
825 * is being destroyed */
826 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
827 (*conn->service->destroyConnProc) (conn);
829 /* Notify the security module that this connection is being destroyed */
830 RXS_DestroyConnection(conn->securityObject, conn);
832 /* If this is the last connection using the rx_peer struct, set its
833 * idle time to now. rxi_ReapConnections will reap it if it's still
834 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
836 MUTEX_ENTER(&rx_peerHashTable_lock);
837 if (conn->peer->refCount < 2) {
838 conn->peer->idleWhen = clock_Sec();
839 if (conn->peer->refCount < 1) {
840 conn->peer->refCount = 1;
841 MUTEX_ENTER(&rx_stats_mutex);
842 rxi_lowPeerRefCount++;
843 MUTEX_EXIT(&rx_stats_mutex);
846 conn->peer->refCount--;
847 MUTEX_EXIT(&rx_peerHashTable_lock);
849 MUTEX_ENTER(&rx_stats_mutex);
850 if (conn->type == RX_SERVER_CONNECTION)
851 rx_stats.nServerConns--;
853 rx_stats.nClientConns--;
854 MUTEX_EXIT(&rx_stats_mutex);
857 if (conn->specific) {
859 for (i = 0; i < conn->nSpecific; i++) {
860 if (conn->specific[i] && rxi_keyCreate_destructor[i])
861 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
862 conn->specific[i] = NULL;
864 free(conn->specific);
866 conn->specific = NULL;
870 MUTEX_DESTROY(&conn->conn_call_lock);
871 MUTEX_DESTROY(&conn->conn_data_lock);
872 CV_DESTROY(&conn->conn_call_cv);
874 rxi_FreeConnection(conn);
877 /* Destroy the specified connection */
879 rxi_DestroyConnection(register struct rx_connection *conn)
881 MUTEX_ENTER(&rx_connHashTable_lock);
882 rxi_DestroyConnectionNoLock(conn);
883 /* conn should be at the head of the cleanup list */
884 if (conn == rx_connCleanup_list) {
885 rx_connCleanup_list = rx_connCleanup_list->next;
886 MUTEX_EXIT(&rx_connHashTable_lock);
887 rxi_CleanupConnection(conn);
889 #ifdef RX_ENABLE_LOCKS
891 MUTEX_EXIT(&rx_connHashTable_lock);
893 #endif /* RX_ENABLE_LOCKS */
897 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
899 register struct rx_connection **conn_ptr;
900 register int havecalls = 0;
901 struct rx_packet *packet;
908 MUTEX_ENTER(&conn->conn_data_lock);
909 if (conn->refCount > 0)
912 MUTEX_ENTER(&rx_stats_mutex);
913 rxi_lowConnRefCount++;
914 MUTEX_EXIT(&rx_stats_mutex);
917 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
918 /* Busy; wait till the last guy before proceeding */
919 MUTEX_EXIT(&conn->conn_data_lock);
924 /* If the client previously called rx_NewCall, but it is still
925 * waiting, treat this as a running call, and wait to destroy the
926 * connection later when the call completes. */
927 if ((conn->type == RX_CLIENT_CONNECTION)
928 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
929 conn->flags |= RX_CONN_DESTROY_ME;
930 MUTEX_EXIT(&conn->conn_data_lock);
934 MUTEX_EXIT(&conn->conn_data_lock);
936 /* Check for extant references to this connection */
937 for (i = 0; i < RX_MAXCALLS; i++) {
938 register struct rx_call *call = conn->call[i];
941 if (conn->type == RX_CLIENT_CONNECTION) {
942 MUTEX_ENTER(&call->lock);
943 if (call->delayedAckEvent) {
944 /* Push the final acknowledgment out now--there
945 * won't be a subsequent call to acknowledge the
946 * last reply packets */
947 rxevent_Cancel(call->delayedAckEvent, call,
948 RX_CALL_REFCOUNT_DELAY);
949 if (call->state == RX_STATE_PRECALL
950 || call->state == RX_STATE_ACTIVE) {
951 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
953 rxi_AckAll(NULL, call, 0);
956 MUTEX_EXIT(&call->lock);
960 #ifdef RX_ENABLE_LOCKS
962 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
963 MUTEX_EXIT(&conn->conn_data_lock);
965 /* Someone is accessing a packet right now. */
969 #endif /* RX_ENABLE_LOCKS */
972 /* Don't destroy the connection if there are any call
973 * structures still in use */
974 MUTEX_ENTER(&conn->conn_data_lock);
975 conn->flags |= RX_CONN_DESTROY_ME;
976 MUTEX_EXIT(&conn->conn_data_lock);
981 if (conn->delayedAbortEvent) {
982 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
983 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
985 MUTEX_ENTER(&conn->conn_data_lock);
986 rxi_SendConnectionAbort(conn, packet, 0, 1);
987 MUTEX_EXIT(&conn->conn_data_lock);
988 rxi_FreePacket(packet);
992 /* Remove from connection hash table before proceeding */
994 &rx_connHashTable[CONN_HASH
995 (peer->host, peer->port, conn->cid, conn->epoch,
997 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
998 if (*conn_ptr == conn) {
999 *conn_ptr = conn->next;
1003 /* if the conn that we are destroying was the last connection, then we
1004 * clear rxLastConn as well */
1005 if (rxLastConn == conn)
1008 /* Make sure the connection is completely reset before deleting it. */
1009 /* get rid of pending events that could zap us later */
1010 if (conn->challengeEvent)
1011 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1012 if (conn->checkReachEvent)
1013 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1015 /* Add the connection to the list of destroyed connections that
1016 * need to be cleaned up. This is necessary to avoid deadlocks
1017 * in the routines we call to inform others that this connection is
1018 * being destroyed. */
1019 conn->next = rx_connCleanup_list;
1020 rx_connCleanup_list = conn;
1023 /* Externally available version */
1025 rx_DestroyConnection(register struct rx_connection *conn)
1030 rxi_DestroyConnection(conn);
1035 rx_GetConnection(register struct rx_connection *conn)
1040 MUTEX_ENTER(&conn->conn_data_lock);
1042 MUTEX_EXIT(&conn->conn_data_lock);
1046 /* Wait for the transmit queue to no longer be busy.
1047 * requires the call->lock to be held */
1048 static void rxi_WaitforTQBusy(struct rx_call *call) {
1049 while (call->flags & RX_CALL_TQ_BUSY) {
1050 call->flags |= RX_CALL_TQ_WAIT;
1052 #ifdef RX_ENABLE_LOCKS
1053 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1054 CV_WAIT(&call->cv_tq, &call->lock);
1055 #else /* RX_ENABLE_LOCKS */
1056 osi_rxSleep(&call->tq);
1057 #endif /* RX_ENABLE_LOCKS */
1059 if (call->tqWaiters == 0) {
1060 call->flags &= ~RX_CALL_TQ_WAIT;
1064 /* Start a new rx remote procedure call, on the specified connection.
1065 * If wait is set to 1, wait for a free call channel; otherwise return
1066 * 0. Maxtime gives the maximum number of seconds this call may take,
1067 * after rx_NewCall returns. After this time interval, a call to any
1068 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1069 * For fine grain locking, we hold the conn_call_lock in order to
1070 * to ensure that we don't get signalle after we found a call in an active
1071 * state and before we go to sleep.
1074 rx_NewCall(register struct rx_connection *conn)
1077 register struct rx_call *call;
1078 struct clock queueTime;
1082 dpf(("rx_NewCall(conn %x)\n", conn));
1085 clock_GetTime(&queueTime);
1086 MUTEX_ENTER(&conn->conn_call_lock);
1089 * Check if there are others waiting for a new call.
1090 * If so, let them go first to avoid starving them.
1091 * This is a fairly simple scheme, and might not be
1092 * a complete solution for large numbers of waiters.
1094 * makeCallWaiters keeps track of the number of
1095 * threads waiting to make calls and the
1096 * RX_CONN_MAKECALL_WAITING flag bit is used to
1097 * indicate that there are indeed calls waiting.
1098 * The flag is set when the waiter is incremented.
1099 * It is only cleared in rx_EndCall when
1100 * makeCallWaiters is 0. This prevents us from
1101 * accidently destroying the connection while it
1102 * is potentially about to be used.
1104 MUTEX_ENTER(&conn->conn_data_lock);
1105 if (conn->makeCallWaiters) {
1106 conn->flags |= RX_CONN_MAKECALL_WAITING;
1107 conn->makeCallWaiters++;
1108 MUTEX_EXIT(&conn->conn_data_lock);
1110 #ifdef RX_ENABLE_LOCKS
1111 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1115 MUTEX_ENTER(&conn->conn_data_lock);
1116 conn->makeCallWaiters--;
1118 MUTEX_EXIT(&conn->conn_data_lock);
1121 for (i = 0; i < RX_MAXCALLS; i++) {
1122 call = conn->call[i];
1124 MUTEX_ENTER(&call->lock);
1125 if (call->state == RX_STATE_DALLY) {
1126 rxi_ResetCall(call, 0);
1127 (*call->callNumber)++;
1130 MUTEX_EXIT(&call->lock);
1132 call = rxi_NewCall(conn, i);
1136 if (i < RX_MAXCALLS) {
1139 MUTEX_ENTER(&conn->conn_data_lock);
1140 conn->flags |= RX_CONN_MAKECALL_WAITING;
1141 conn->makeCallWaiters++;
1142 MUTEX_EXIT(&conn->conn_data_lock);
1144 #ifdef RX_ENABLE_LOCKS
1145 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1149 MUTEX_ENTER(&conn->conn_data_lock);
1150 conn->makeCallWaiters--;
1151 MUTEX_EXIT(&conn->conn_data_lock);
1154 * Wake up anyone else who might be giving us a chance to
1155 * run (see code above that avoids resource starvation).
1157 #ifdef RX_ENABLE_LOCKS
1158 CV_BROADCAST(&conn->conn_call_cv);
1163 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1165 /* Client is initially in send mode */
1166 call->state = RX_STATE_ACTIVE;
1167 call->error = conn->error;
1169 call->mode = RX_MODE_ERROR;
1171 call->mode = RX_MODE_SENDING;
1173 /* remember start time for call in case we have hard dead time limit */
1174 call->queueTime = queueTime;
1175 clock_GetTime(&call->startTime);
1176 hzero(call->bytesSent);
1177 hzero(call->bytesRcvd);
1179 /* Turn on busy protocol. */
1180 rxi_KeepAliveOn(call);
1182 MUTEX_EXIT(&call->lock);
1183 MUTEX_EXIT(&conn->conn_call_lock);
1186 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1187 /* Now, if TQ wasn't cleared earlier, do it now. */
1188 MUTEX_ENTER(&call->lock);
1189 rxi_WaitforTQBusy(call);
1190 if (call->flags & RX_CALL_TQ_CLEARME) {
1191 rxi_ClearTransmitQueue(call, 0);
1192 queue_Init(&call->tq);
1194 MUTEX_EXIT(&call->lock);
1195 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1197 dpf(("rx_NewCall(call %x)\n", call));
1202 rxi_HasActiveCalls(register struct rx_connection *aconn)
1205 register struct rx_call *tcall;
1209 for (i = 0; i < RX_MAXCALLS; i++) {
1210 if ((tcall = aconn->call[i])) {
1211 if ((tcall->state == RX_STATE_ACTIVE)
1212 || (tcall->state == RX_STATE_PRECALL)) {
1223 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1224 register afs_int32 * aint32s)
1227 register struct rx_call *tcall;
1231 for (i = 0; i < RX_MAXCALLS; i++) {
1232 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1233 aint32s[i] = aconn->callNumber[i] + 1;
1235 aint32s[i] = aconn->callNumber[i];
1242 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1243 register afs_int32 * aint32s)
1246 register struct rx_call *tcall;
1250 for (i = 0; i < RX_MAXCALLS; i++) {
1251 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1252 aconn->callNumber[i] = aint32s[i] - 1;
1254 aconn->callNumber[i] = aint32s[i];
1260 /* Advertise a new service. A service is named locally by a UDP port
1261 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1264 char *serviceName; Name for identification purposes (e.g. the
1265 service name might be used for probing for
1268 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1269 char *serviceName, struct rx_securityClass **securityObjects,
1270 int nSecurityObjects,
1271 afs_int32(*serviceProc) (struct rx_call * acall))
1273 osi_socket socket = OSI_NULLSOCKET;
1274 register struct rx_service *tservice;
1280 if (serviceId == 0) {
1282 "rx_NewService: service id for service %s is not non-zero.\n",
1289 "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",
1297 tservice = rxi_AllocService();
1299 for (i = 0; i < RX_MAX_SERVICES; i++) {
1300 register struct rx_service *service = rx_services[i];
1302 if (port == service->servicePort && host == service->serviceHost) {
1303 if (service->serviceId == serviceId) {
1304 /* The identical service has already been
1305 * installed; if the caller was intending to
1306 * change the security classes used by this
1307 * service, he/she loses. */
1309 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1310 serviceName, serviceId, service->serviceName);
1312 rxi_FreeService(tservice);
1315 /* Different service, same port: re-use the socket
1316 * which is bound to the same port */
1317 socket = service->socket;
1320 if (socket == OSI_NULLSOCKET) {
1321 /* If we don't already have a socket (from another
1322 * service on same port) get a new one */
1323 socket = rxi_GetHostUDPSocket(host, port);
1324 if (socket == OSI_NULLSOCKET) {
1326 rxi_FreeService(tservice);
1331 service->socket = socket;
1332 service->serviceHost = host;
1333 service->servicePort = port;
1334 service->serviceId = serviceId;
1335 service->serviceName = serviceName;
1336 service->nSecurityObjects = nSecurityObjects;
1337 service->securityObjects = securityObjects;
1338 service->minProcs = 0;
1339 service->maxProcs = 1;
1340 service->idleDeadTime = 60;
1341 service->connDeadTime = rx_connDeadTime;
1342 service->executeRequestProc = serviceProc;
1343 service->checkReach = 0;
1344 rx_services[i] = service; /* not visible until now */
1350 rxi_FreeService(tservice);
1351 (osi_Msg "rx_NewService: cannot support > %d services\n",
1357 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1358 struct rx_securityClass **securityObjects, int nSecurityObjects,
1359 afs_int32(*serviceProc) (struct rx_call * acall))
1361 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1364 /* Generic request processing loop. This routine should be called
1365 * by the implementation dependent rx_ServerProc. If socketp is
1366 * non-null, it will be set to the file descriptor that this thread
1367 * is now listening on. If socketp is null, this routine will never
1370 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1372 register struct rx_call *call;
1373 register afs_int32 code;
1374 register struct rx_service *tservice = NULL;
1381 call = rx_GetCall(threadID, tservice, socketp);
1382 if (socketp && *socketp != OSI_NULLSOCKET) {
1383 /* We are now a listener thread */
1388 /* if server is restarting( typically smooth shutdown) then do not
1389 * allow any new calls.
1392 if (rx_tranquil && (call != NULL)) {
1396 MUTEX_ENTER(&call->lock);
1398 rxi_CallError(call, RX_RESTARTING);
1399 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1401 MUTEX_EXIT(&call->lock);
1405 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1406 #ifdef RX_ENABLE_LOCKS
1408 #endif /* RX_ENABLE_LOCKS */
1409 afs_termState = AFSOP_STOP_AFS;
1410 afs_osi_Wakeup(&afs_termState);
1411 #ifdef RX_ENABLE_LOCKS
1413 #endif /* RX_ENABLE_LOCKS */
1418 tservice = call->conn->service;
1420 if (tservice->beforeProc)
1421 (*tservice->beforeProc) (call);
1423 code = call->conn->service->executeRequestProc(call);
1425 if (tservice->afterProc)
1426 (*tservice->afterProc) (call, code);
1428 rx_EndCall(call, code);
1429 MUTEX_ENTER(&rx_stats_mutex);
1431 MUTEX_EXIT(&rx_stats_mutex);
1437 rx_WakeupServerProcs(void)
1439 struct rx_serverQueueEntry *np, *tqp;
1443 MUTEX_ENTER(&rx_serverPool_lock);
1445 #ifdef RX_ENABLE_LOCKS
1446 if (rx_waitForPacket)
1447 CV_BROADCAST(&rx_waitForPacket->cv);
1448 #else /* RX_ENABLE_LOCKS */
1449 if (rx_waitForPacket)
1450 osi_rxWakeup(rx_waitForPacket);
1451 #endif /* RX_ENABLE_LOCKS */
1452 MUTEX_ENTER(&freeSQEList_lock);
1453 for (np = rx_FreeSQEList; np; np = tqp) {
1454 tqp = *(struct rx_serverQueueEntry **)np;
1455 #ifdef RX_ENABLE_LOCKS
1456 CV_BROADCAST(&np->cv);
1457 #else /* RX_ENABLE_LOCKS */
1459 #endif /* RX_ENABLE_LOCKS */
1461 MUTEX_EXIT(&freeSQEList_lock);
1462 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1463 #ifdef RX_ENABLE_LOCKS
1464 CV_BROADCAST(&np->cv);
1465 #else /* RX_ENABLE_LOCKS */
1467 #endif /* RX_ENABLE_LOCKS */
1469 MUTEX_EXIT(&rx_serverPool_lock);
1474 * One thing that seems to happen is that all the server threads get
1475 * tied up on some empty or slow call, and then a whole bunch of calls
1476 * arrive at once, using up the packet pool, so now there are more
1477 * empty calls. The most critical resources here are server threads
1478 * and the free packet pool. The "doreclaim" code seems to help in
1479 * general. I think that eventually we arrive in this state: there
1480 * are lots of pending calls which do have all their packets present,
1481 * so they won't be reclaimed, are multi-packet calls, so they won't
1482 * be scheduled until later, and thus are tying up most of the free
1483 * packet pool for a very long time.
1485 * 1. schedule multi-packet calls if all the packets are present.
1486 * Probably CPU-bound operation, useful to return packets to pool.
1487 * Do what if there is a full window, but the last packet isn't here?
1488 * 3. preserve one thread which *only* runs "best" calls, otherwise
1489 * it sleeps and waits for that type of call.
1490 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1491 * the current dataquota business is badly broken. The quota isn't adjusted
1492 * to reflect how many packets are presently queued for a running call.
1493 * So, when we schedule a queued call with a full window of packets queued
1494 * up for it, that *should* free up a window full of packets for other 2d-class
1495 * calls to be able to use from the packet pool. But it doesn't.
1497 * NB. Most of the time, this code doesn't run -- since idle server threads
1498 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1499 * as a new call arrives.
1501 /* Sleep until a call arrives. Returns a pointer to the call, ready
1502 * for an rx_Read. */
1503 #ifdef RX_ENABLE_LOCKS
1505 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1507 struct rx_serverQueueEntry *sq;
1508 register struct rx_call *call = (struct rx_call *)0;
1509 struct rx_service *service = NULL;
1512 MUTEX_ENTER(&freeSQEList_lock);
1514 if ((sq = rx_FreeSQEList)) {
1515 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1516 MUTEX_EXIT(&freeSQEList_lock);
1517 } else { /* otherwise allocate a new one and return that */
1518 MUTEX_EXIT(&freeSQEList_lock);
1519 sq = (struct rx_serverQueueEntry *)
1520 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1521 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1522 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1525 MUTEX_ENTER(&rx_serverPool_lock);
1526 if (cur_service != NULL) {
1527 ReturnToServerPool(cur_service);
1530 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1531 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1533 /* Scan for eligible incoming calls. A call is not eligible
1534 * if the maximum number of calls for its service type are
1535 * already executing */
1536 /* One thread will process calls FCFS (to prevent starvation),
1537 * while the other threads may run ahead looking for calls which
1538 * have all their input data available immediately. This helps
1539 * keep threads from blocking, waiting for data from the client. */
1540 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1541 service = tcall->conn->service;
1542 if (!QuotaOK(service)) {
1545 if (tno == rxi_fcfs_thread_num
1546 || !tcall->queue_item_header.next) {
1547 /* If we're the fcfs thread , then we'll just use
1548 * this call. If we haven't been able to find an optimal
1549 * choice, and we're at the end of the list, then use a
1550 * 2d choice if one has been identified. Otherwise... */
1551 call = (choice2 ? choice2 : tcall);
1552 service = call->conn->service;
1553 } else if (!queue_IsEmpty(&tcall->rq)) {
1554 struct rx_packet *rp;
1555 rp = queue_First(&tcall->rq, rx_packet);
1556 if (rp->header.seq == 1) {
1558 || (rp->header.flags & RX_LAST_PACKET)) {
1560 } else if (rxi_2dchoice && !choice2
1561 && !(tcall->flags & RX_CALL_CLEARED)
1562 && (tcall->rprev > rxi_HardAckRate)) {
1571 ReturnToServerPool(service);
1578 MUTEX_EXIT(&rx_serverPool_lock);
1579 MUTEX_ENTER(&call->lock);
1581 if (call->flags & RX_CALL_WAIT_PROC) {
1582 call->flags &= ~RX_CALL_WAIT_PROC;
1583 MUTEX_ENTER(&rx_stats_mutex);
1585 MUTEX_EXIT(&rx_stats_mutex);
1588 if (call->state != RX_STATE_PRECALL || call->error) {
1589 MUTEX_EXIT(&call->lock);
1590 MUTEX_ENTER(&rx_serverPool_lock);
1591 ReturnToServerPool(service);
1596 if (queue_IsEmpty(&call->rq)
1597 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1598 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1600 CLEAR_CALL_QUEUE_LOCK(call);
1603 /* If there are no eligible incoming calls, add this process
1604 * to the idle server queue, to wait for one */
1608 *socketp = OSI_NULLSOCKET;
1610 sq->socketp = socketp;
1611 queue_Append(&rx_idleServerQueue, sq);
1612 #ifndef AFS_AIX41_ENV
1613 rx_waitForPacket = sq;
1615 rx_waitingForPacket = sq;
1616 #endif /* AFS_AIX41_ENV */
1618 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1620 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1621 MUTEX_EXIT(&rx_serverPool_lock);
1622 return (struct rx_call *)0;
1625 } while (!(call = sq->newcall)
1626 && !(socketp && *socketp != OSI_NULLSOCKET));
1627 MUTEX_EXIT(&rx_serverPool_lock);
1629 MUTEX_ENTER(&call->lock);
1635 MUTEX_ENTER(&freeSQEList_lock);
1636 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1637 rx_FreeSQEList = sq;
1638 MUTEX_EXIT(&freeSQEList_lock);
1641 clock_GetTime(&call->startTime);
1642 call->state = RX_STATE_ACTIVE;
1643 call->mode = RX_MODE_RECEIVING;
1644 #ifdef RX_KERNEL_TRACE
1645 if (ICL_SETACTIVE(afs_iclSetp)) {
1646 int glockOwner = ISAFS_GLOCK();
1649 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1650 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1657 rxi_calltrace(RX_CALL_START, call);
1658 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1659 call->conn->service->servicePort, call->conn->service->serviceId,
1662 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1663 MUTEX_EXIT(&call->lock);
1665 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1670 #else /* RX_ENABLE_LOCKS */
1672 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1674 struct rx_serverQueueEntry *sq;
1675 register struct rx_call *call = (struct rx_call *)0, *choice2;
1676 struct rx_service *service = NULL;
1680 MUTEX_ENTER(&freeSQEList_lock);
1682 if ((sq = rx_FreeSQEList)) {
1683 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1684 MUTEX_EXIT(&freeSQEList_lock);
1685 } else { /* otherwise allocate a new one and return that */
1686 MUTEX_EXIT(&freeSQEList_lock);
1687 sq = (struct rx_serverQueueEntry *)
1688 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1689 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1690 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1692 MUTEX_ENTER(&sq->lock);
1694 if (cur_service != NULL) {
1695 cur_service->nRequestsRunning--;
1696 if (cur_service->nRequestsRunning < cur_service->minProcs)
1700 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1701 register struct rx_call *tcall, *ncall;
1702 /* Scan for eligible incoming calls. A call is not eligible
1703 * if the maximum number of calls for its service type are
1704 * already executing */
1705 /* One thread will process calls FCFS (to prevent starvation),
1706 * while the other threads may run ahead looking for calls which
1707 * have all their input data available immediately. This helps
1708 * keep threads from blocking, waiting for data from the client. */
1709 choice2 = (struct rx_call *)0;
1710 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1711 service = tcall->conn->service;
1712 if (QuotaOK(service)) {
1713 if (tno == rxi_fcfs_thread_num
1714 || !tcall->queue_item_header.next) {
1715 /* If we're the fcfs thread, then we'll just use
1716 * this call. If we haven't been able to find an optimal
1717 * choice, and we're at the end of the list, then use a
1718 * 2d choice if one has been identified. Otherwise... */
1719 call = (choice2 ? choice2 : tcall);
1720 service = call->conn->service;
1721 } else if (!queue_IsEmpty(&tcall->rq)) {
1722 struct rx_packet *rp;
1723 rp = queue_First(&tcall->rq, rx_packet);
1724 if (rp->header.seq == 1
1726 || (rp->header.flags & RX_LAST_PACKET))) {
1728 } else if (rxi_2dchoice && !choice2
1729 && !(tcall->flags & RX_CALL_CLEARED)
1730 && (tcall->rprev > rxi_HardAckRate)) {
1743 /* we can't schedule a call if there's no data!!! */
1744 /* send an ack if there's no data, if we're missing the
1745 * first packet, or we're missing something between first
1746 * and last -- there's a "hole" in the incoming data. */
1747 if (queue_IsEmpty(&call->rq)
1748 || queue_First(&call->rq, rx_packet)->header.seq != 1
1749 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1750 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1752 call->flags &= (~RX_CALL_WAIT_PROC);
1753 service->nRequestsRunning++;
1754 /* just started call in minProcs pool, need fewer to maintain
1756 if (service->nRequestsRunning <= service->minProcs)
1760 /* MUTEX_EXIT(&call->lock); */
1762 /* If there are no eligible incoming calls, add this process
1763 * to the idle server queue, to wait for one */
1766 *socketp = OSI_NULLSOCKET;
1768 sq->socketp = socketp;
1769 queue_Append(&rx_idleServerQueue, sq);
1773 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1775 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1776 return (struct rx_call *)0;
1779 } while (!(call = sq->newcall)
1780 && !(socketp && *socketp != OSI_NULLSOCKET));
1782 MUTEX_EXIT(&sq->lock);
1784 MUTEX_ENTER(&freeSQEList_lock);
1785 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1786 rx_FreeSQEList = sq;
1787 MUTEX_EXIT(&freeSQEList_lock);
1790 clock_GetTime(&call->startTime);
1791 call->state = RX_STATE_ACTIVE;
1792 call->mode = RX_MODE_RECEIVING;
1793 #ifdef RX_KERNEL_TRACE
1794 if (ICL_SETACTIVE(afs_iclSetp)) {
1795 int glockOwner = ISAFS_GLOCK();
1798 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1799 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1806 rxi_calltrace(RX_CALL_START, call);
1807 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1808 call->conn->service->servicePort, call->conn->service->serviceId,
1811 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1818 #endif /* RX_ENABLE_LOCKS */
1822 /* Establish a procedure to be called when a packet arrives for a
1823 * call. This routine will be called at most once after each call,
1824 * and will also be called if there is an error condition on the or
1825 * the call is complete. Used by multi rx to build a selection
1826 * function which determines which of several calls is likely to be a
1827 * good one to read from.
1828 * NOTE: the way this is currently implemented it is probably only a
1829 * good idea to (1) use it immediately after a newcall (clients only)
1830 * and (2) only use it once. Other uses currently void your warranty
1833 rx_SetArrivalProc(register struct rx_call *call,
1834 register void (*proc) (register struct rx_call * call,
1836 register int index),
1837 register VOID * handle, register int arg)
1839 call->arrivalProc = proc;
1840 call->arrivalProcHandle = handle;
1841 call->arrivalProcArg = arg;
1844 /* Call is finished (possibly prematurely). Return rc to the peer, if
1845 * appropriate, and return the final error code from the conversation
1849 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1851 register struct rx_connection *conn = call->conn;
1852 register struct rx_service *service;
1858 dpf(("rx_EndCall(call %x rc %d error %d abortCode %d)\n", call, rc, call->error, call->abortCode));
1861 MUTEX_ENTER(&call->lock);
1863 if (rc == 0 && call->error == 0) {
1864 call->abortCode = 0;
1865 call->abortCount = 0;
1868 call->arrivalProc = (void (*)())0;
1869 if (rc && call->error == 0) {
1870 rxi_CallError(call, rc);
1871 /* Send an abort message to the peer if this error code has
1872 * only just been set. If it was set previously, assume the
1873 * peer has already been sent the error code or will request it
1875 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1877 if (conn->type == RX_SERVER_CONNECTION) {
1878 /* Make sure reply or at least dummy reply is sent */
1879 if (call->mode == RX_MODE_RECEIVING) {
1880 rxi_WriteProc(call, 0, 0);
1882 if (call->mode == RX_MODE_SENDING) {
1883 rxi_FlushWrite(call);
1885 service = conn->service;
1886 rxi_calltrace(RX_CALL_END, call);
1887 /* Call goes to hold state until reply packets are acknowledged */
1888 if (call->tfirst + call->nSoftAcked < call->tnext) {
1889 call->state = RX_STATE_HOLD;
1891 call->state = RX_STATE_DALLY;
1892 rxi_ClearTransmitQueue(call, 0);
1893 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1894 rxevent_Cancel(call->keepAliveEvent, call,
1895 RX_CALL_REFCOUNT_ALIVE);
1897 } else { /* Client connection */
1899 /* Make sure server receives input packets, in the case where
1900 * no reply arguments are expected */
1901 if ((call->mode == RX_MODE_SENDING)
1902 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1903 (void)rxi_ReadProc(call, &dummy, 1);
1906 /* If we had an outstanding delayed ack, be nice to the server
1907 * and force-send it now.
1909 if (call->delayedAckEvent) {
1910 rxevent_Cancel(call->delayedAckEvent, call,
1911 RX_CALL_REFCOUNT_DELAY);
1912 call->delayedAckEvent = NULL;
1913 rxi_SendDelayedAck(NULL, call, NULL);
1916 /* We need to release the call lock since it's lower than the
1917 * conn_call_lock and we don't want to hold the conn_call_lock
1918 * over the rx_ReadProc call. The conn_call_lock needs to be held
1919 * here for the case where rx_NewCall is perusing the calls on
1920 * the connection structure. We don't want to signal until
1921 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1922 * have checked this call, found it active and by the time it
1923 * goes to sleep, will have missed the signal.
1925 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
1926 * there are threads waiting to use the conn object.
1928 MUTEX_EXIT(&call->lock);
1929 MUTEX_ENTER(&conn->conn_call_lock);
1930 MUTEX_ENTER(&call->lock);
1931 MUTEX_ENTER(&conn->conn_data_lock);
1932 conn->flags |= RX_CONN_BUSY;
1933 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1934 if (conn->makeCallWaiters == 0)
1935 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1936 MUTEX_EXIT(&conn->conn_data_lock);
1937 #ifdef RX_ENABLE_LOCKS
1938 CV_BROADCAST(&conn->conn_call_cv);
1943 #ifdef RX_ENABLE_LOCKS
1945 MUTEX_EXIT(&conn->conn_data_lock);
1947 #endif /* RX_ENABLE_LOCKS */
1948 call->state = RX_STATE_DALLY;
1950 error = call->error;
1952 /* currentPacket, nLeft, and NFree must be zeroed here, because
1953 * ResetCall cannot: ResetCall may be called at splnet(), in the
1954 * kernel version, and may interrupt the macros rx_Read or
1955 * rx_Write, which run at normal priority for efficiency. */
1956 if (call->currentPacket) {
1957 queue_Prepend(&call->iovq, call->currentPacket);
1958 call->currentPacket = (struct rx_packet *)0;
1961 call->nLeft = call->nFree = call->curlen = 0;
1963 /* Free any packets from the last call to ReadvProc/WritevProc */
1964 rxi_FreePackets(0, &call->iovq);
1966 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1967 MUTEX_EXIT(&call->lock);
1968 if (conn->type == RX_CLIENT_CONNECTION) {
1969 MUTEX_EXIT(&conn->conn_call_lock);
1970 conn->flags &= ~RX_CONN_BUSY;
1974 * Map errors to the local host's errno.h format.
1976 error = ntoh_syserr_conv(error);
1980 #if !defined(KERNEL)
1982 /* Call this routine when shutting down a server or client (especially
1983 * clients). This will allow Rx to gracefully garbage collect server
1984 * connections, and reduce the number of retries that a server might
1985 * make to a dead client.
1986 * This is not quite right, since some calls may still be ongoing and
1987 * we can't lock them to destroy them. */
1991 register struct rx_connection **conn_ptr, **conn_end;
1995 if (rxinit_status == 1) {
1997 return; /* Already shutdown. */
1999 rxi_DeleteCachedConnections();
2000 if (rx_connHashTable) {
2001 MUTEX_ENTER(&rx_connHashTable_lock);
2002 for (conn_ptr = &rx_connHashTable[0], conn_end =
2003 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2005 struct rx_connection *conn, *next;
2006 for (conn = *conn_ptr; conn; conn = next) {
2008 if (conn->type == RX_CLIENT_CONNECTION) {
2009 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2011 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2012 #ifdef RX_ENABLE_LOCKS
2013 rxi_DestroyConnectionNoLock(conn);
2014 #else /* RX_ENABLE_LOCKS */
2015 rxi_DestroyConnection(conn);
2016 #endif /* RX_ENABLE_LOCKS */
2020 #ifdef RX_ENABLE_LOCKS
2021 while (rx_connCleanup_list) {
2022 struct rx_connection *conn;
2023 conn = rx_connCleanup_list;
2024 rx_connCleanup_list = rx_connCleanup_list->next;
2025 MUTEX_EXIT(&rx_connHashTable_lock);
2026 rxi_CleanupConnection(conn);
2027 MUTEX_ENTER(&rx_connHashTable_lock);
2029 MUTEX_EXIT(&rx_connHashTable_lock);
2030 #endif /* RX_ENABLE_LOCKS */
2035 afs_winsockCleanup();
2043 /* if we wakeup packet waiter too often, can get in loop with two
2044 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2046 rxi_PacketsUnWait(void)
2048 if (!rx_waitingForPackets) {
2052 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2053 return; /* still over quota */
2056 rx_waitingForPackets = 0;
2057 #ifdef RX_ENABLE_LOCKS
2058 CV_BROADCAST(&rx_waitingForPackets_cv);
2060 osi_rxWakeup(&rx_waitingForPackets);
2066 /* ------------------Internal interfaces------------------------- */
2068 /* Return this process's service structure for the
2069 * specified socket and service */
2071 rxi_FindService(register osi_socket socket, register u_short serviceId)
2073 register struct rx_service **sp;
2074 for (sp = &rx_services[0]; *sp; sp++) {
2075 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2081 /* Allocate a call structure, for the indicated channel of the
2082 * supplied connection. The mode and state of the call must be set by
2083 * the caller. Returns the call with mutex locked. */
2085 rxi_NewCall(register struct rx_connection *conn, register int channel)
2087 register struct rx_call *call;
2088 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2089 register struct rx_call *cp; /* Call pointer temp */
2090 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2091 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2093 dpf(("rxi_NewCall(conn %x, channel %d)\n", conn, channel));
2095 /* Grab an existing call structure, or allocate a new one.
2096 * Existing call structures are assumed to have been left reset by
2098 MUTEX_ENTER(&rx_freeCallQueue_lock);
2100 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2102 * EXCEPT that the TQ might not yet be cleared out.
2103 * Skip over those with in-use TQs.
2106 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2107 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2113 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2114 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2115 call = queue_First(&rx_freeCallQueue, rx_call);
2116 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2118 MUTEX_ENTER(&rx_stats_mutex);
2119 rx_stats.nFreeCallStructs--;
2120 MUTEX_EXIT(&rx_stats_mutex);
2121 MUTEX_EXIT(&rx_freeCallQueue_lock);
2122 MUTEX_ENTER(&call->lock);
2123 CLEAR_CALL_QUEUE_LOCK(call);
2124 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2125 /* Now, if TQ wasn't cleared earlier, do it now. */
2126 if (call->flags & RX_CALL_TQ_CLEARME) {
2127 rxi_ClearTransmitQueue(call, 0);
2128 queue_Init(&call->tq);
2130 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2131 /* Bind the call to its connection structure */
2133 rxi_ResetCall(call, 1);
2135 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2137 MUTEX_EXIT(&rx_freeCallQueue_lock);
2138 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2139 MUTEX_ENTER(&call->lock);
2140 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2141 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2142 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2144 MUTEX_ENTER(&rx_stats_mutex);
2145 rx_stats.nCallStructs++;
2146 MUTEX_EXIT(&rx_stats_mutex);
2147 /* Initialize once-only items */
2148 queue_Init(&call->tq);
2149 queue_Init(&call->rq);
2150 queue_Init(&call->iovq);
2151 /* Bind the call to its connection structure (prereq for reset) */
2153 rxi_ResetCall(call, 1);
2155 call->channel = channel;
2156 call->callNumber = &conn->callNumber[channel];
2157 /* Note that the next expected call number is retained (in
2158 * conn->callNumber[i]), even if we reallocate the call structure
2160 conn->call[channel] = call;
2161 /* if the channel's never been used (== 0), we should start at 1, otherwise
2162 * the call number is valid from the last time this channel was used */
2163 if (*call->callNumber == 0)
2164 *call->callNumber = 1;
2169 /* A call has been inactive long enough that so we can throw away
2170 * state, including the call structure, which is placed on the call
2172 * Call is locked upon entry.
2173 * haveCTLock set if called from rxi_ReapConnections
2175 #ifdef RX_ENABLE_LOCKS
2177 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2178 #else /* RX_ENABLE_LOCKS */
2180 rxi_FreeCall(register struct rx_call *call)
2181 #endif /* RX_ENABLE_LOCKS */
2183 register int channel = call->channel;
2184 register struct rx_connection *conn = call->conn;
2187 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2188 (*call->callNumber)++;
2189 rxi_ResetCall(call, 0);
2190 call->conn->call[channel] = (struct rx_call *)0;
2192 MUTEX_ENTER(&rx_freeCallQueue_lock);
2193 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2194 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2195 /* A call may be free even though its transmit queue is still in use.
2196 * Since we search the call list from head to tail, put busy calls at
2197 * the head of the list, and idle calls at the tail.
2199 if (call->flags & RX_CALL_TQ_BUSY)
2200 queue_Prepend(&rx_freeCallQueue, call);
2202 queue_Append(&rx_freeCallQueue, call);
2203 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2204 queue_Append(&rx_freeCallQueue, call);
2205 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2206 MUTEX_ENTER(&rx_stats_mutex);
2207 rx_stats.nFreeCallStructs++;
2208 MUTEX_EXIT(&rx_stats_mutex);
2210 MUTEX_EXIT(&rx_freeCallQueue_lock);
2212 /* Destroy the connection if it was previously slated for
2213 * destruction, i.e. the Rx client code previously called
2214 * rx_DestroyConnection (client connections), or
2215 * rxi_ReapConnections called the same routine (server
2216 * connections). Only do this, however, if there are no
2217 * outstanding calls. Note that for fine grain locking, there appears
2218 * to be a deadlock in that rxi_FreeCall has a call locked and
2219 * DestroyConnectionNoLock locks each call in the conn. But note a
2220 * few lines up where we have removed this call from the conn.
2221 * If someone else destroys a connection, they either have no
2222 * call lock held or are going through this section of code.
2224 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2225 MUTEX_ENTER(&conn->conn_data_lock);
2227 MUTEX_EXIT(&conn->conn_data_lock);
2228 #ifdef RX_ENABLE_LOCKS
2230 rxi_DestroyConnectionNoLock(conn);
2232 rxi_DestroyConnection(conn);
2233 #else /* RX_ENABLE_LOCKS */
2234 rxi_DestroyConnection(conn);
2235 #endif /* RX_ENABLE_LOCKS */
2239 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2241 rxi_Alloc(register size_t size)
2245 MUTEX_ENTER(&rx_stats_mutex);
2247 rxi_Allocsize += (afs_int32)size;
2248 MUTEX_EXIT(&rx_stats_mutex);
2250 p = (char *)osi_Alloc(size);
2253 osi_Panic("rxi_Alloc error");
2259 rxi_Free(void *addr, register size_t size)
2261 MUTEX_ENTER(&rx_stats_mutex);
2263 rxi_Allocsize -= (afs_int32)size;
2264 MUTEX_EXIT(&rx_stats_mutex);
2266 osi_Free(addr, size);
2269 /* Find the peer process represented by the supplied (host,port)
2270 * combination. If there is no appropriate active peer structure, a
2271 * new one will be allocated and initialized
2272 * The origPeer, if set, is a pointer to a peer structure on which the
2273 * refcount will be be decremented. This is used to replace the peer
2274 * structure hanging off a connection structure */
2276 rxi_FindPeer(register afs_uint32 host, register u_short port,
2277 struct rx_peer *origPeer, int create)
2279 register struct rx_peer *pp;
2281 hashIndex = PEER_HASH(host, port);
2282 MUTEX_ENTER(&rx_peerHashTable_lock);
2283 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2284 if ((pp->host == host) && (pp->port == port))
2289 pp = rxi_AllocPeer(); /* This bzero's *pp */
2290 pp->host = host; /* set here or in InitPeerParams is zero */
2292 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2293 queue_Init(&pp->congestionQueue);
2294 queue_Init(&pp->rpcStats);
2295 pp->next = rx_peerHashTable[hashIndex];
2296 rx_peerHashTable[hashIndex] = pp;
2297 rxi_InitPeerParams(pp);
2298 MUTEX_ENTER(&rx_stats_mutex);
2299 rx_stats.nPeerStructs++;
2300 MUTEX_EXIT(&rx_stats_mutex);
2307 origPeer->refCount--;
2308 MUTEX_EXIT(&rx_peerHashTable_lock);
2313 /* Find the connection at (host, port) started at epoch, and with the
2314 * given connection id. Creates the server connection if necessary.
2315 * The type specifies whether a client connection or a server
2316 * connection is desired. In both cases, (host, port) specify the
2317 * peer's (host, pair) pair. Client connections are not made
2318 * automatically by this routine. The parameter socket gives the
2319 * socket descriptor on which the packet was received. This is used,
2320 * in the case of server connections, to check that *new* connections
2321 * come via a valid (port, serviceId). Finally, the securityIndex
2322 * parameter must match the existing index for the connection. If a
2323 * server connection is created, it will be created using the supplied
2324 * index, if the index is valid for this service */
2325 struct rx_connection *
2326 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2327 register u_short port, u_short serviceId, afs_uint32 cid,
2328 afs_uint32 epoch, int type, u_int securityIndex)
2330 int hashindex, flag;
2331 register struct rx_connection *conn;
2332 hashindex = CONN_HASH(host, port, cid, epoch, type);
2333 MUTEX_ENTER(&rx_connHashTable_lock);
2334 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2335 rx_connHashTable[hashindex],
2338 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2339 && (epoch == conn->epoch)) {
2340 register struct rx_peer *pp = conn->peer;
2341 if (securityIndex != conn->securityIndex) {
2342 /* this isn't supposed to happen, but someone could forge a packet
2343 * like this, and there seems to be some CM bug that makes this
2344 * happen from time to time -- in which case, the fileserver
2346 MUTEX_EXIT(&rx_connHashTable_lock);
2347 return (struct rx_connection *)0;
2349 if (pp->host == host && pp->port == port)
2351 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2353 /* So what happens when it's a callback connection? */
2354 if ( /*type == RX_CLIENT_CONNECTION && */
2355 (conn->epoch & 0x80000000))
2359 /* the connection rxLastConn that was used the last time is not the
2360 ** one we are looking for now. Hence, start searching in the hash */
2362 conn = rx_connHashTable[hashindex];
2367 struct rx_service *service;
2368 if (type == RX_CLIENT_CONNECTION) {
2369 MUTEX_EXIT(&rx_connHashTable_lock);
2370 return (struct rx_connection *)0;
2372 service = rxi_FindService(socket, serviceId);
2373 if (!service || (securityIndex >= service->nSecurityObjects)
2374 || (service->securityObjects[securityIndex] == 0)) {
2375 MUTEX_EXIT(&rx_connHashTable_lock);
2376 return (struct rx_connection *)0;
2378 conn = rxi_AllocConnection(); /* This bzero's the connection */
2379 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2380 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2381 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2382 conn->next = rx_connHashTable[hashindex];
2383 rx_connHashTable[hashindex] = conn;
2384 conn->peer = rxi_FindPeer(host, port, 0, 1);
2385 conn->type = RX_SERVER_CONNECTION;
2386 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2387 conn->epoch = epoch;
2388 conn->cid = cid & RX_CIDMASK;
2389 /* conn->serial = conn->lastSerial = 0; */
2390 /* conn->timeout = 0; */
2391 conn->ackRate = RX_FAST_ACK_RATE;
2392 conn->service = service;
2393 conn->serviceId = serviceId;
2394 conn->securityIndex = securityIndex;
2395 conn->securityObject = service->securityObjects[securityIndex];
2396 conn->nSpecific = 0;
2397 conn->specific = NULL;
2398 rx_SetConnDeadTime(conn, service->connDeadTime);
2399 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2400 /* Notify security object of the new connection */
2401 RXS_NewConnection(conn->securityObject, conn);
2402 /* XXXX Connection timeout? */
2403 if (service->newConnProc)
2404 (*service->newConnProc) (conn);
2405 MUTEX_ENTER(&rx_stats_mutex);
2406 rx_stats.nServerConns++;
2407 MUTEX_EXIT(&rx_stats_mutex);
2410 MUTEX_ENTER(&conn->conn_data_lock);
2412 MUTEX_EXIT(&conn->conn_data_lock);
2414 rxLastConn = conn; /* store this connection as the last conn used */
2415 MUTEX_EXIT(&rx_connHashTable_lock);
2419 /* There are two packet tracing routines available for testing and monitoring
2420 * Rx. One is called just after every packet is received and the other is
2421 * called just before every packet is sent. Received packets, have had their
2422 * headers decoded, and packets to be sent have not yet had their headers
2423 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2424 * containing the network address. Both can be modified. The return value, if
2425 * non-zero, indicates that the packet should be dropped. */
2427 int (*rx_justReceived) () = 0;
2428 int (*rx_almostSent) () = 0;
2430 /* A packet has been received off the interface. Np is the packet, socket is
2431 * the socket number it was received from (useful in determining which service
2432 * this packet corresponds to), and (host, port) reflect the host,port of the
2433 * sender. This call returns the packet to the caller if it is finished with
2434 * it, rather than de-allocating it, just as a small performance hack */
2437 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2438 afs_uint32 host, u_short port, int *tnop,
2439 struct rx_call **newcallp)
2441 register struct rx_call *call;
2442 register struct rx_connection *conn;
2444 afs_uint32 currentCallNumber;
2450 struct rx_packet *tnp;
2453 /* We don't print out the packet until now because (1) the time may not be
2454 * accurate enough until now in the lwp implementation (rx_Listener only gets
2455 * the time after the packet is read) and (2) from a protocol point of view,
2456 * this is the first time the packet has been seen */
2457 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2458 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2459 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2460 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2461 np->header.epoch, np->header.cid, np->header.callNumber,
2462 np->header.seq, np->header.flags, np));
2465 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2466 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2469 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2470 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2473 /* If an input tracer function is defined, call it with the packet and
2474 * network address. Note this function may modify its arguments. */
2475 if (rx_justReceived) {
2476 struct sockaddr_in addr;
2478 addr.sin_family = AF_INET;
2479 addr.sin_port = port;
2480 addr.sin_addr.s_addr = host;
2481 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2482 addr.sin_len = sizeof(addr);
2483 #endif /* AFS_OSF_ENV */
2484 drop = (*rx_justReceived) (np, &addr);
2485 /* drop packet if return value is non-zero */
2488 port = addr.sin_port; /* in case fcn changed addr */
2489 host = addr.sin_addr.s_addr;
2493 /* If packet was not sent by the client, then *we* must be the client */
2494 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2495 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2497 /* Find the connection (or fabricate one, if we're the server & if
2498 * necessary) associated with this packet */
2500 rxi_FindConnection(socket, host, port, np->header.serviceId,
2501 np->header.cid, np->header.epoch, type,
2502 np->header.securityIndex);
2505 /* If no connection found or fabricated, just ignore the packet.
2506 * (An argument could be made for sending an abort packet for
2511 MUTEX_ENTER(&conn->conn_data_lock);
2512 if (conn->maxSerial < np->header.serial)
2513 conn->maxSerial = np->header.serial;
2514 MUTEX_EXIT(&conn->conn_data_lock);
2516 /* If the connection is in an error state, send an abort packet and ignore
2517 * the incoming packet */
2519 /* Don't respond to an abort packet--we don't want loops! */
2520 MUTEX_ENTER(&conn->conn_data_lock);
2521 if (np->header.type != RX_PACKET_TYPE_ABORT)
2522 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2524 MUTEX_EXIT(&conn->conn_data_lock);
2528 /* Check for connection-only requests (i.e. not call specific). */
2529 if (np->header.callNumber == 0) {
2530 switch (np->header.type) {
2531 case RX_PACKET_TYPE_ABORT: {
2532 /* What if the supplied error is zero? */
2533 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2534 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2535 rxi_ConnectionError(conn, errcode);
2536 MUTEX_ENTER(&conn->conn_data_lock);
2538 MUTEX_EXIT(&conn->conn_data_lock);
2541 case RX_PACKET_TYPE_CHALLENGE:
2542 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2543 MUTEX_ENTER(&conn->conn_data_lock);
2545 MUTEX_EXIT(&conn->conn_data_lock);
2547 case RX_PACKET_TYPE_RESPONSE:
2548 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2549 MUTEX_ENTER(&conn->conn_data_lock);
2551 MUTEX_EXIT(&conn->conn_data_lock);
2553 case RX_PACKET_TYPE_PARAMS:
2554 case RX_PACKET_TYPE_PARAMS + 1:
2555 case RX_PACKET_TYPE_PARAMS + 2:
2556 /* ignore these packet types for now */
2557 MUTEX_ENTER(&conn->conn_data_lock);
2559 MUTEX_EXIT(&conn->conn_data_lock);
2564 /* Should not reach here, unless the peer is broken: send an
2566 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2567 MUTEX_ENTER(&conn->conn_data_lock);
2568 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2570 MUTEX_EXIT(&conn->conn_data_lock);
2575 channel = np->header.cid & RX_CHANNELMASK;
2576 call = conn->call[channel];
2577 #ifdef RX_ENABLE_LOCKS
2579 MUTEX_ENTER(&call->lock);
2580 /* Test to see if call struct is still attached to conn. */
2581 if (call != conn->call[channel]) {
2583 MUTEX_EXIT(&call->lock);
2584 if (type == RX_SERVER_CONNECTION) {
2585 call = conn->call[channel];
2586 /* If we started with no call attached and there is one now,
2587 * another thread is also running this routine and has gotten
2588 * the connection channel. We should drop this packet in the tests
2589 * below. If there was a call on this connection and it's now
2590 * gone, then we'll be making a new call below.
2591 * If there was previously a call and it's now different then
2592 * the old call was freed and another thread running this routine
2593 * has created a call on this channel. One of these two threads
2594 * has a packet for the old call and the code below handles those
2598 MUTEX_ENTER(&call->lock);
2600 /* This packet can't be for this call. If the new call address is
2601 * 0 then no call is running on this channel. If there is a call
2602 * then, since this is a client connection we're getting data for
2603 * it must be for the previous call.
2605 MUTEX_ENTER(&rx_stats_mutex);
2606 rx_stats.spuriousPacketsRead++;
2607 MUTEX_EXIT(&rx_stats_mutex);
2608 MUTEX_ENTER(&conn->conn_data_lock);
2610 MUTEX_EXIT(&conn->conn_data_lock);
2615 currentCallNumber = conn->callNumber[channel];
2617 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2618 if (np->header.callNumber < currentCallNumber) {
2619 MUTEX_ENTER(&rx_stats_mutex);
2620 rx_stats.spuriousPacketsRead++;
2621 MUTEX_EXIT(&rx_stats_mutex);
2622 #ifdef RX_ENABLE_LOCKS
2624 MUTEX_EXIT(&call->lock);
2626 MUTEX_ENTER(&conn->conn_data_lock);
2628 MUTEX_EXIT(&conn->conn_data_lock);
2632 MUTEX_ENTER(&conn->conn_call_lock);
2633 call = rxi_NewCall(conn, channel);
2634 MUTEX_EXIT(&conn->conn_call_lock);
2635 *call->callNumber = np->header.callNumber;
2636 if (np->header.callNumber == 0)
2637 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));
2639 call->state = RX_STATE_PRECALL;
2640 clock_GetTime(&call->queueTime);
2641 hzero(call->bytesSent);
2642 hzero(call->bytesRcvd);
2644 * If the number of queued calls exceeds the overload
2645 * threshold then abort this call.
2647 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2648 struct rx_packet *tp;
2650 rxi_CallError(call, rx_BusyError);
2651 tp = rxi_SendCallAbort(call, np, 1, 0);
2652 MUTEX_EXIT(&call->lock);
2653 MUTEX_ENTER(&conn->conn_data_lock);
2655 MUTEX_EXIT(&conn->conn_data_lock);
2656 MUTEX_ENTER(&rx_stats_mutex);
2658 MUTEX_EXIT(&rx_stats_mutex);
2661 rxi_KeepAliveOn(call);
2662 } else if (np->header.callNumber != currentCallNumber) {
2663 /* Wait until the transmit queue is idle before deciding
2664 * whether to reset the current call. Chances are that the
2665 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2668 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2669 while ((call->state == RX_STATE_ACTIVE)
2670 && (call->flags & RX_CALL_TQ_BUSY)) {
2671 call->flags |= RX_CALL_TQ_WAIT;
2673 #ifdef RX_ENABLE_LOCKS
2674 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2675 CV_WAIT(&call->cv_tq, &call->lock);
2676 #else /* RX_ENABLE_LOCKS */
2677 osi_rxSleep(&call->tq);
2678 #endif /* RX_ENABLE_LOCKS */
2680 if (call->tqWaiters == 0)
2681 call->flags &= ~RX_CALL_TQ_WAIT;
2683 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2684 /* If the new call cannot be taken right now send a busy and set
2685 * the error condition in this call, so that it terminates as
2686 * quickly as possible */
2687 if (call->state == RX_STATE_ACTIVE) {
2688 struct rx_packet *tp;
2690 rxi_CallError(call, RX_CALL_DEAD);
2691 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2693 MUTEX_EXIT(&call->lock);
2694 MUTEX_ENTER(&conn->conn_data_lock);
2696 MUTEX_EXIT(&conn->conn_data_lock);
2699 rxi_ResetCall(call, 0);
2700 *call->callNumber = np->header.callNumber;
2701 if (np->header.callNumber == 0)
2702 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));
2704 call->state = RX_STATE_PRECALL;
2705 clock_GetTime(&call->queueTime);
2706 hzero(call->bytesSent);
2707 hzero(call->bytesRcvd);
2709 * If the number of queued calls exceeds the overload
2710 * threshold then abort this call.
2712 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2713 struct rx_packet *tp;
2715 rxi_CallError(call, rx_BusyError);
2716 tp = rxi_SendCallAbort(call, np, 1, 0);
2717 MUTEX_EXIT(&call->lock);
2718 MUTEX_ENTER(&conn->conn_data_lock);
2720 MUTEX_EXIT(&conn->conn_data_lock);
2721 MUTEX_ENTER(&rx_stats_mutex);
2723 MUTEX_EXIT(&rx_stats_mutex);
2726 rxi_KeepAliveOn(call);
2728 /* Continuing call; do nothing here. */
2730 } else { /* we're the client */
2731 /* Ignore all incoming acknowledgements for calls in DALLY state */
2732 if (call && (call->state == RX_STATE_DALLY)
2733 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2734 MUTEX_ENTER(&rx_stats_mutex);
2735 rx_stats.ignorePacketDally++;
2736 MUTEX_EXIT(&rx_stats_mutex);
2737 #ifdef RX_ENABLE_LOCKS
2739 MUTEX_EXIT(&call->lock);
2742 MUTEX_ENTER(&conn->conn_data_lock);
2744 MUTEX_EXIT(&conn->conn_data_lock);
2748 /* Ignore anything that's not relevant to the current call. If there
2749 * isn't a current call, then no packet is relevant. */
2750 if (!call || (np->header.callNumber != currentCallNumber)) {
2751 MUTEX_ENTER(&rx_stats_mutex);
2752 rx_stats.spuriousPacketsRead++;
2753 MUTEX_EXIT(&rx_stats_mutex);
2754 #ifdef RX_ENABLE_LOCKS
2756 MUTEX_EXIT(&call->lock);
2759 MUTEX_ENTER(&conn->conn_data_lock);
2761 MUTEX_EXIT(&conn->conn_data_lock);
2764 /* If the service security object index stamped in the packet does not
2765 * match the connection's security index, ignore the packet */
2766 if (np->header.securityIndex != conn->securityIndex) {
2767 #ifdef RX_ENABLE_LOCKS
2768 MUTEX_EXIT(&call->lock);
2770 MUTEX_ENTER(&conn->conn_data_lock);
2772 MUTEX_EXIT(&conn->conn_data_lock);
2776 /* If we're receiving the response, then all transmit packets are
2777 * implicitly acknowledged. Get rid of them. */
2778 if (np->header.type == RX_PACKET_TYPE_DATA) {
2779 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2780 /* XXX Hack. Because we must release the global rx lock when
2781 * sending packets (osi_NetSend) we drop all acks while we're
2782 * traversing the tq in rxi_Start sending packets out because
2783 * packets may move to the freePacketQueue as result of being here!
2784 * So we drop these packets until we're safely out of the
2785 * traversing. Really ugly!
2786 * For fine grain RX locking, we set the acked field in the
2787 * packets and let rxi_Start remove them from the transmit queue.
2789 if (call->flags & RX_CALL_TQ_BUSY) {
2790 #ifdef RX_ENABLE_LOCKS
2791 rxi_SetAcksInTransmitQueue(call);
2794 return np; /* xmitting; drop packet */
2797 rxi_ClearTransmitQueue(call, 0);
2799 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2800 rxi_ClearTransmitQueue(call, 0);
2801 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2803 if (np->header.type == RX_PACKET_TYPE_ACK) {
2804 /* now check to see if this is an ack packet acknowledging that the
2805 * server actually *lost* some hard-acked data. If this happens we
2806 * ignore this packet, as it may indicate that the server restarted in
2807 * the middle of a call. It is also possible that this is an old ack
2808 * packet. We don't abort the connection in this case, because this
2809 * *might* just be an old ack packet. The right way to detect a server
2810 * restart in the midst of a call is to notice that the server epoch
2812 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2813 * XXX unacknowledged. I think that this is off-by-one, but
2814 * XXX I don't dare change it just yet, since it will
2815 * XXX interact badly with the server-restart detection
2816 * XXX code in receiveackpacket. */
2817 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2818 MUTEX_ENTER(&rx_stats_mutex);
2819 rx_stats.spuriousPacketsRead++;
2820 MUTEX_EXIT(&rx_stats_mutex);
2821 MUTEX_EXIT(&call->lock);
2822 MUTEX_ENTER(&conn->conn_data_lock);
2824 MUTEX_EXIT(&conn->conn_data_lock);
2828 } /* else not a data packet */
2831 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2832 /* Set remote user defined status from packet */
2833 call->remoteStatus = np->header.userStatus;
2835 /* Note the gap between the expected next packet and the actual
2836 * packet that arrived, when the new packet has a smaller serial number
2837 * than expected. Rioses frequently reorder packets all by themselves,
2838 * so this will be quite important with very large window sizes.
2839 * Skew is checked against 0 here to avoid any dependence on the type of
2840 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2842 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2843 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2844 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2846 MUTEX_ENTER(&conn->conn_data_lock);
2847 skew = conn->lastSerial - np->header.serial;
2848 conn->lastSerial = np->header.serial;
2849 MUTEX_EXIT(&conn->conn_data_lock);
2851 register struct rx_peer *peer;
2853 if (skew > peer->inPacketSkew) {
2854 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2856 peer->inPacketSkew = skew;
2860 /* Now do packet type-specific processing */
2861 switch (np->header.type) {
2862 case RX_PACKET_TYPE_DATA:
2863 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2866 case RX_PACKET_TYPE_ACK:
2867 /* Respond immediately to ack packets requesting acknowledgement
2869 if (np->header.flags & RX_REQUEST_ACK) {
2871 (void)rxi_SendCallAbort(call, 0, 1, 0);
2873 (void)rxi_SendAck(call, 0, np->header.serial,
2874 RX_ACK_PING_RESPONSE, 1);
2876 np = rxi_ReceiveAckPacket(call, np, 1);
2878 case RX_PACKET_TYPE_ABORT: {
2879 /* An abort packet: reset the call, passing the error up to the user. */
2880 /* What if error is zero? */
2881 /* What if the error is -1? the application will treat it as a timeout. */
2882 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
2883 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
2884 rxi_CallError(call, errdata);
2885 MUTEX_EXIT(&call->lock);
2886 MUTEX_ENTER(&conn->conn_data_lock);
2888 MUTEX_EXIT(&conn->conn_data_lock);
2889 return np; /* xmitting; drop packet */
2891 case RX_PACKET_TYPE_BUSY:
2894 case RX_PACKET_TYPE_ACKALL:
2895 /* All packets acknowledged, so we can drop all packets previously
2896 * readied for sending */
2897 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2898 /* XXX Hack. We because we can't release the global rx lock when
2899 * sending packets (osi_NetSend) we drop all ack pkts while we're
2900 * traversing the tq in rxi_Start sending packets out because
2901 * packets may move to the freePacketQueue as result of being
2902 * here! So we drop these packets until we're safely out of the
2903 * traversing. Really ugly!
2904 * For fine grain RX locking, we set the acked field in the packets
2905 * and let rxi_Start remove the packets from the transmit queue.
2907 if (call->flags & RX_CALL_TQ_BUSY) {
2908 #ifdef RX_ENABLE_LOCKS
2909 rxi_SetAcksInTransmitQueue(call);
2911 #else /* RX_ENABLE_LOCKS */
2912 MUTEX_EXIT(&call->lock);
2913 MUTEX_ENTER(&conn->conn_data_lock);
2915 MUTEX_EXIT(&conn->conn_data_lock);
2916 return np; /* xmitting; drop packet */
2917 #endif /* RX_ENABLE_LOCKS */
2919 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2920 rxi_ClearTransmitQueue(call, 0);
2923 /* Should not reach here, unless the peer is broken: send an abort
2925 rxi_CallError(call, RX_PROTOCOL_ERROR);
2926 np = rxi_SendCallAbort(call, np, 1, 0);
2929 /* Note when this last legitimate packet was received, for keep-alive
2930 * processing. Note, we delay getting the time until now in the hope that
2931 * the packet will be delivered to the user before any get time is required
2932 * (if not, then the time won't actually be re-evaluated here). */
2933 call->lastReceiveTime = clock_Sec();
2934 MUTEX_EXIT(&call->lock);
2935 MUTEX_ENTER(&conn->conn_data_lock);
2937 MUTEX_EXIT(&conn->conn_data_lock);
2941 /* return true if this is an "interesting" connection from the point of view
2942 of someone trying to debug the system */
2944 rxi_IsConnInteresting(struct rx_connection *aconn)
2947 register struct rx_call *tcall;
2949 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2951 for (i = 0; i < RX_MAXCALLS; i++) {
2952 tcall = aconn->call[i];
2954 if ((tcall->state == RX_STATE_PRECALL)
2955 || (tcall->state == RX_STATE_ACTIVE))
2957 if ((tcall->mode == RX_MODE_SENDING)
2958 || (tcall->mode == RX_MODE_RECEIVING))
2966 /* if this is one of the last few packets AND it wouldn't be used by the
2967 receiving call to immediately satisfy a read request, then drop it on
2968 the floor, since accepting it might prevent a lock-holding thread from
2969 making progress in its reading. If a call has been cleared while in
2970 the precall state then ignore all subsequent packets until the call
2971 is assigned to a thread. */
2974 TooLow(struct rx_packet *ap, struct rx_call *acall)
2977 MUTEX_ENTER(&rx_stats_mutex);
2978 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
2979 && (acall->state == RX_STATE_PRECALL))
2980 || ((rx_nFreePackets < rxi_dataQuota + 2)
2981 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
2982 && (acall->flags & RX_CALL_READER_WAIT)))) {
2985 MUTEX_EXIT(&rx_stats_mutex);
2991 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
2992 struct rx_call *acall)
2994 struct rx_call *call = acall;
2998 MUTEX_ENTER(&conn->conn_data_lock);
2999 conn->checkReachEvent = NULL;
3000 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3003 MUTEX_EXIT(&conn->conn_data_lock);
3007 MUTEX_ENTER(&conn->conn_call_lock);
3008 MUTEX_ENTER(&conn->conn_data_lock);
3009 for (i = 0; i < RX_MAXCALLS; i++) {
3010 struct rx_call *tc = conn->call[i];
3011 if (tc && tc->state == RX_STATE_PRECALL) {
3017 /* Indicate that rxi_CheckReachEvent is no longer running by
3018 * clearing the flag. Must be atomic under conn_data_lock to
3019 * avoid a new call slipping by: rxi_CheckConnReach holds
3020 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3022 conn->flags &= ~RX_CONN_ATTACHWAIT;
3023 MUTEX_EXIT(&conn->conn_data_lock);
3024 MUTEX_EXIT(&conn->conn_call_lock);
3029 MUTEX_ENTER(&call->lock);
3030 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3032 MUTEX_EXIT(&call->lock);
3034 clock_GetTime(&when);
3035 when.sec += RX_CHECKREACH_TIMEOUT;
3036 MUTEX_ENTER(&conn->conn_data_lock);
3037 if (!conn->checkReachEvent) {
3039 conn->checkReachEvent =
3040 rxevent_Post(&when, rxi_CheckReachEvent, conn, NULL);
3042 MUTEX_EXIT(&conn->conn_data_lock);
3048 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3050 struct rx_service *service = conn->service;
3051 struct rx_peer *peer = conn->peer;
3052 afs_uint32 now, lastReach;
3054 if (service->checkReach == 0)
3058 MUTEX_ENTER(&peer->peer_lock);
3059 lastReach = peer->lastReachTime;
3060 MUTEX_EXIT(&peer->peer_lock);
3061 if (now - lastReach < RX_CHECKREACH_TTL)
3064 MUTEX_ENTER(&conn->conn_data_lock);
3065 if (conn->flags & RX_CONN_ATTACHWAIT) {
3066 MUTEX_EXIT(&conn->conn_data_lock);
3069 conn->flags |= RX_CONN_ATTACHWAIT;
3070 MUTEX_EXIT(&conn->conn_data_lock);
3071 if (!conn->checkReachEvent)
3072 rxi_CheckReachEvent(NULL, conn, call);
3077 /* try to attach call, if authentication is complete */
3079 TryAttach(register struct rx_call *acall, register osi_socket socket,
3080 register int *tnop, register struct rx_call **newcallp,
3083 struct rx_connection *conn = acall->conn;
3085 if (conn->type == RX_SERVER_CONNECTION
3086 && acall->state == RX_STATE_PRECALL) {
3087 /* Don't attach until we have any req'd. authentication. */
3088 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3089 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3090 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3091 /* Note: this does not necessarily succeed; there
3092 * may not any proc available
3095 rxi_ChallengeOn(acall->conn);
3100 /* A data packet has been received off the interface. This packet is
3101 * appropriate to the call (the call is in the right state, etc.). This
3102 * routine can return a packet to the caller, for re-use */
3105 rxi_ReceiveDataPacket(register struct rx_call *call,
3106 register struct rx_packet *np, int istack,
3107 osi_socket socket, afs_uint32 host, u_short port,
3108 int *tnop, struct rx_call **newcallp)
3110 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3114 afs_uint32 seq, serial, flags;
3116 struct rx_packet *tnp;
3118 MUTEX_ENTER(&rx_stats_mutex);
3119 rx_stats.dataPacketsRead++;
3120 MUTEX_EXIT(&rx_stats_mutex);
3123 /* If there are no packet buffers, drop this new packet, unless we can find
3124 * packet buffers from inactive calls */
3126 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3127 MUTEX_ENTER(&rx_freePktQ_lock);
3128 rxi_NeedMorePackets = TRUE;
3129 MUTEX_EXIT(&rx_freePktQ_lock);
3130 MUTEX_ENTER(&rx_stats_mutex);
3131 rx_stats.noPacketBuffersOnRead++;
3132 MUTEX_EXIT(&rx_stats_mutex);
3133 call->rprev = np->header.serial;
3134 rxi_calltrace(RX_TRACE_DROP, call);
3135 dpf(("packet %x dropped on receipt - quota problems", np));
3137 rxi_ClearReceiveQueue(call);
3138 clock_GetTime(&when);
3139 clock_Add(&when, &rx_softAckDelay);
3140 if (!call->delayedAckEvent
3141 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3142 rxevent_Cancel(call->delayedAckEvent, call,
3143 RX_CALL_REFCOUNT_DELAY);
3144 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3145 call->delayedAckEvent =
3146 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3148 /* we've damaged this call already, might as well do it in. */
3154 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3155 * packet is one of several packets transmitted as a single
3156 * datagram. Do not send any soft or hard acks until all packets
3157 * in a jumbogram have been processed. Send negative acks right away.
3159 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3160 /* tnp is non-null when there are more packets in the
3161 * current jumbo gram */
3168 seq = np->header.seq;
3169 serial = np->header.serial;
3170 flags = np->header.flags;
3172 /* If the call is in an error state, send an abort message */
3174 return rxi_SendCallAbort(call, np, istack, 0);
3176 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3177 * AFS 3.5 jumbogram. */
3178 if (flags & RX_JUMBO_PACKET) {
3179 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3184 if (np->header.spare != 0) {
3185 MUTEX_ENTER(&call->conn->conn_data_lock);
3186 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3187 MUTEX_EXIT(&call->conn->conn_data_lock);
3190 /* The usual case is that this is the expected next packet */
3191 if (seq == call->rnext) {
3193 /* Check to make sure it is not a duplicate of one already queued */
3194 if (queue_IsNotEmpty(&call->rq)
3195 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3196 MUTEX_ENTER(&rx_stats_mutex);
3197 rx_stats.dupPacketsRead++;
3198 MUTEX_EXIT(&rx_stats_mutex);
3199 dpf(("packet %x dropped on receipt - duplicate", np));
3200 rxevent_Cancel(call->delayedAckEvent, call,
3201 RX_CALL_REFCOUNT_DELAY);
3202 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3208 /* It's the next packet. Stick it on the receive queue
3209 * for this call. Set newPackets to make sure we wake
3210 * the reader once all packets have been processed */
3211 queue_Prepend(&call->rq, np);
3213 np = NULL; /* We can't use this anymore */
3216 /* If an ack is requested then set a flag to make sure we
3217 * send an acknowledgement for this packet */
3218 if (flags & RX_REQUEST_ACK) {
3219 ackNeeded = RX_ACK_REQUESTED;
3222 /* Keep track of whether we have received the last packet */
3223 if (flags & RX_LAST_PACKET) {
3224 call->flags |= RX_CALL_HAVE_LAST;
3228 /* Check whether we have all of the packets for this call */
3229 if (call->flags & RX_CALL_HAVE_LAST) {
3230 afs_uint32 tseq; /* temporary sequence number */
3231 struct rx_packet *tp; /* Temporary packet pointer */
3232 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3234 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3235 if (tseq != tp->header.seq)
3237 if (tp->header.flags & RX_LAST_PACKET) {
3238 call->flags |= RX_CALL_RECEIVE_DONE;
3245 /* Provide asynchronous notification for those who want it
3246 * (e.g. multi rx) */
3247 if (call->arrivalProc) {
3248 (*call->arrivalProc) (call, call->arrivalProcHandle,
3249 call->arrivalProcArg);
3250 call->arrivalProc = (void (*)())0;
3253 /* Update last packet received */
3256 /* If there is no server process serving this call, grab
3257 * one, if available. We only need to do this once. If a
3258 * server thread is available, this thread becomes a server
3259 * thread and the server thread becomes a listener thread. */
3261 TryAttach(call, socket, tnop, newcallp, 0);
3264 /* This is not the expected next packet. */
3266 /* Determine whether this is a new or old packet, and if it's
3267 * a new one, whether it fits into the current receive window.
3268 * Also figure out whether the packet was delivered in sequence.
3269 * We use the prev variable to determine whether the new packet
3270 * is the successor of its immediate predecessor in the
3271 * receive queue, and the missing flag to determine whether
3272 * any of this packets predecessors are missing. */
3274 afs_uint32 prev; /* "Previous packet" sequence number */
3275 struct rx_packet *tp; /* Temporary packet pointer */
3276 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3277 int missing; /* Are any predecessors missing? */
3279 /* If the new packet's sequence number has been sent to the
3280 * application already, then this is a duplicate */
3281 if (seq < call->rnext) {
3282 MUTEX_ENTER(&rx_stats_mutex);
3283 rx_stats.dupPacketsRead++;
3284 MUTEX_EXIT(&rx_stats_mutex);
3285 rxevent_Cancel(call->delayedAckEvent, call,
3286 RX_CALL_REFCOUNT_DELAY);
3287 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3293 /* If the sequence number is greater than what can be
3294 * accomodated by the current window, then send a negative
3295 * acknowledge and drop the packet */
3296 if ((call->rnext + call->rwind) <= seq) {
3297 rxevent_Cancel(call->delayedAckEvent, call,
3298 RX_CALL_REFCOUNT_DELAY);
3299 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3306 /* Look for the packet in the queue of old received packets */
3307 for (prev = call->rnext - 1, missing =
3308 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3309 /*Check for duplicate packet */
3310 if (seq == tp->header.seq) {
3311 MUTEX_ENTER(&rx_stats_mutex);
3312 rx_stats.dupPacketsRead++;
3313 MUTEX_EXIT(&rx_stats_mutex);
3314 rxevent_Cancel(call->delayedAckEvent, call,
3315 RX_CALL_REFCOUNT_DELAY);
3316 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3322 /* If we find a higher sequence packet, break out and
3323 * insert the new packet here. */
3324 if (seq < tp->header.seq)
3326 /* Check for missing packet */
3327 if (tp->header.seq != prev + 1) {
3331 prev = tp->header.seq;
3334 /* Keep track of whether we have received the last packet. */
3335 if (flags & RX_LAST_PACKET) {
3336 call->flags |= RX_CALL_HAVE_LAST;
3339 /* It's within the window: add it to the the receive queue.
3340 * tp is left by the previous loop either pointing at the
3341 * packet before which to insert the new packet, or at the
3342 * queue head if the queue is empty or the packet should be
3344 queue_InsertBefore(tp, np);
3348 /* Check whether we have all of the packets for this call */
3349 if ((call->flags & RX_CALL_HAVE_LAST)
3350 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3351 afs_uint32 tseq; /* temporary sequence number */
3354 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3355 if (tseq != tp->header.seq)
3357 if (tp->header.flags & RX_LAST_PACKET) {
3358 call->flags |= RX_CALL_RECEIVE_DONE;
3365 /* We need to send an ack of the packet is out of sequence,
3366 * or if an ack was requested by the peer. */
3367 if (seq != prev + 1 || missing) {
3368 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3369 } else if (flags & RX_REQUEST_ACK) {
3370 ackNeeded = RX_ACK_REQUESTED;
3373 /* Acknowledge the last packet for each call */
3374 if (flags & RX_LAST_PACKET) {
3385 * If the receiver is waiting for an iovec, fill the iovec
3386 * using the data from the receive queue */
3387 if (call->flags & RX_CALL_IOVEC_WAIT) {
3388 didHardAck = rxi_FillReadVec(call, serial);
3389 /* the call may have been aborted */
3398 /* Wakeup the reader if any */
3399 if ((call->flags & RX_CALL_READER_WAIT)
3400 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3401 || (call->iovNext >= call->iovMax)
3402 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3403 call->flags &= ~RX_CALL_READER_WAIT;
3404 #ifdef RX_ENABLE_LOCKS
3405 CV_BROADCAST(&call->cv_rq);
3407 osi_rxWakeup(&call->rq);
3413 * Send an ack when requested by the peer, or once every
3414 * rxi_SoftAckRate packets until the last packet has been
3415 * received. Always send a soft ack for the last packet in
3416 * the server's reply. */
3418 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3419 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3420 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3421 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3422 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3423 } else if (call->nSoftAcks) {
3424 clock_GetTime(&when);
3425 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3426 clock_Add(&when, &rx_lastAckDelay);
3428 clock_Add(&when, &rx_softAckDelay);
3430 if (!call->delayedAckEvent
3431 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3432 rxevent_Cancel(call->delayedAckEvent, call,
3433 RX_CALL_REFCOUNT_DELAY);
3434 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3435 call->delayedAckEvent =
3436 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3438 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3439 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3446 static void rxi_ComputeRate();
3450 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3452 struct rx_peer *peer = conn->peer;
3454 MUTEX_ENTER(&peer->peer_lock);
3455 peer->lastReachTime = clock_Sec();
3456 MUTEX_EXIT(&peer->peer_lock);
3458 MUTEX_ENTER(&conn->conn_data_lock);
3459 if (conn->flags & RX_CONN_ATTACHWAIT) {
3462 conn->flags &= ~RX_CONN_ATTACHWAIT;
3463 MUTEX_EXIT(&conn->conn_data_lock);
3465 for (i = 0; i < RX_MAXCALLS; i++) {
3466 struct rx_call *call = conn->call[i];
3469 MUTEX_ENTER(&call->lock);
3470 /* tnop can be null if newcallp is null */
3471 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3473 MUTEX_EXIT(&call->lock);
3477 MUTEX_EXIT(&conn->conn_data_lock);
3481 rx_ack_reason(int reason)
3484 case RX_ACK_REQUESTED:
3486 case RX_ACK_DUPLICATE:
3488 case RX_ACK_OUT_OF_SEQUENCE:
3490 case RX_ACK_EXCEEDS_WINDOW:
3492 case RX_ACK_NOSPACE:
3496 case RX_ACK_PING_RESPONSE:
3508 /* rxi_ComputePeerNetStats
3510 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3511 * estimates (like RTT and throughput) based on ack packets. Caller
3512 * must ensure that the packet in question is the right one (i.e.
3513 * serial number matches).
3516 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3517 struct rx_ackPacket *ap, struct rx_packet *np)
3519 struct rx_peer *peer = call->conn->peer;
3521 /* Use RTT if not delayed by client. */
3522 if (ap->reason != RX_ACK_DELAY)
3523 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3525 rxi_ComputeRate(peer, call, p, np, ap->reason);
3529 /* The real smarts of the whole thing. */
3531 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3534 struct rx_ackPacket *ap;
3536 register struct rx_packet *tp;
3537 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3538 register struct rx_connection *conn = call->conn;
3539 struct rx_peer *peer = conn->peer;
3542 /* because there are CM's that are bogus, sending weird values for this. */
3543 afs_uint32 skew = 0;
3548 int newAckCount = 0;
3549 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3550 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3552 MUTEX_ENTER(&rx_stats_mutex);
3553 rx_stats.ackPacketsRead++;
3554 MUTEX_EXIT(&rx_stats_mutex);
3555 ap = (struct rx_ackPacket *)rx_DataOf(np);
3556 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3558 return np; /* truncated ack packet */
3560 /* depends on ack packet struct */
3561 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3562 first = ntohl(ap->firstPacket);
3563 serial = ntohl(ap->serial);
3564 /* temporarily disabled -- needs to degrade over time
3565 * skew = ntohs(ap->maxSkew); */
3567 /* Ignore ack packets received out of order */
3568 if (first < call->tfirst) {
3572 if (np->header.flags & RX_SLOW_START_OK) {
3573 call->flags |= RX_CALL_SLOW_START_OK;
3576 if (ap->reason == RX_ACK_PING_RESPONSE)
3577 rxi_UpdatePeerReach(conn, call);
3581 if (rxdebug_active) {
3585 len = _snprintf(msg, sizeof(msg),
3586 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3587 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3588 ntohl(ap->serial), ntohl(ap->previousPacket),
3589 (unsigned int)np->header.seq, (unsigned int)skew,
3590 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3594 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3595 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3599 OutputDebugString(msg);
3601 #else /* AFS_NT40_ENV */
3604 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3605 ap->reason, ntohl(ap->previousPacket),
3606 (unsigned int)np->header.seq, (unsigned int)serial,
3607 (unsigned int)skew, ntohl(ap->firstPacket));
3610 for (offset = 0; offset < nAcks; offset++)
3611 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3616 #endif /* AFS_NT40_ENV */
3619 /* Update the outgoing packet skew value to the latest value of
3620 * the peer's incoming packet skew value. The ack packet, of
3621 * course, could arrive out of order, but that won't affect things
3623 MUTEX_ENTER(&peer->peer_lock);
3624 peer->outPacketSkew = skew;
3626 /* Check for packets that no longer need to be transmitted, and
3627 * discard them. This only applies to packets positively
3628 * acknowledged as having been sent to the peer's upper level.
3629 * All other packets must be retained. So only packets with
3630 * sequence numbers < ap->firstPacket are candidates. */
3631 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3632 if (tp->header.seq >= first)
3634 call->tfirst = tp->header.seq + 1;
3636 && (tp->header.serial == serial || tp->firstSerial == serial))
3637 rxi_ComputePeerNetStats(call, tp, ap, np);
3638 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3641 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3642 /* XXX Hack. Because we have to release the global rx lock when sending
3643 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3644 * in rxi_Start sending packets out because packets may move to the
3645 * freePacketQueue as result of being here! So we drop these packets until
3646 * we're safely out of the traversing. Really ugly!
3647 * To make it even uglier, if we're using fine grain locking, we can
3648 * set the ack bits in the packets and have rxi_Start remove the packets
3649 * when it's done transmitting.
3651 if (call->flags & RX_CALL_TQ_BUSY) {
3652 #ifdef RX_ENABLE_LOCKS
3653 tp->flags |= RX_PKTFLAG_ACKED;
3654 call->flags |= RX_CALL_TQ_SOME_ACKED;
3655 #else /* RX_ENABLE_LOCKS */
3657 #endif /* RX_ENABLE_LOCKS */
3659 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3662 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3667 /* Give rate detector a chance to respond to ping requests */
3668 if (ap->reason == RX_ACK_PING_RESPONSE) {
3669 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3673 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3675 /* Now go through explicit acks/nacks and record the results in
3676 * the waiting packets. These are packets that can't be released
3677 * yet, even with a positive acknowledge. This positive
3678 * acknowledge only means the packet has been received by the
3679 * peer, not that it will be retained long enough to be sent to
3680 * the peer's upper level. In addition, reset the transmit timers
3681 * of any missing packets (those packets that must be missing
3682 * because this packet was out of sequence) */
3684 call->nSoftAcked = 0;
3685 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3686 /* Update round trip time if the ack was stimulated on receipt
3688 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3689 #ifdef RX_ENABLE_LOCKS
3690 if (tp->header.seq >= first)
3691 #endif /* RX_ENABLE_LOCKS */
3692 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3694 && (tp->header.serial == serial || tp->firstSerial == serial))
3695 rxi_ComputePeerNetStats(call, tp, ap, np);
3697 /* Set the acknowledge flag per packet based on the
3698 * information in the ack packet. An acknowlegded packet can
3699 * be downgraded when the server has discarded a packet it
3700 * soacked previously, or when an ack packet is received
3701 * out of sequence. */
3702 if (tp->header.seq < first) {
3703 /* Implicit ack information */
3704 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3707 tp->flags |= RX_PKTFLAG_ACKED;
3708 } else if (tp->header.seq < first + nAcks) {
3709 /* Explicit ack information: set it in the packet appropriately */
3710 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3711 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3713 tp->flags |= RX_PKTFLAG_ACKED;
3720 } else /* RX_ACK_TYPE_NACK */ {
3721 tp->flags &= ~RX_PKTFLAG_ACKED;
3725 tp->flags &= ~RX_PKTFLAG_ACKED;
3729 /* If packet isn't yet acked, and it has been transmitted at least
3730 * once, reset retransmit time using latest timeout
3731 * ie, this should readjust the retransmit timer for all outstanding
3732 * packets... So we don't just retransmit when we should know better*/
3734 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3735 tp->retryTime = tp->timeSent;
3736 clock_Add(&tp->retryTime, &peer->timeout);
3737 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3738 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3742 /* If the window has been extended by this acknowledge packet,
3743 * then wakeup a sender waiting in alloc for window space, or try
3744 * sending packets now, if he's been sitting on packets due to
3745 * lack of window space */
3746 if (call->tnext < (call->tfirst + call->twind)) {
3747 #ifdef RX_ENABLE_LOCKS
3748 CV_SIGNAL(&call->cv_twind);
3750 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3751 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3752 osi_rxWakeup(&call->twind);
3755 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3756 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3760 /* if the ack packet has a receivelen field hanging off it,
3761 * update our state */
3762 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3765 /* If the ack packet has a "recommended" size that is less than
3766 * what I am using now, reduce my size to match */
3767 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3768 (int)sizeof(afs_int32), &tSize);
3769 tSize = (afs_uint32) ntohl(tSize);
3770 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3772 /* Get the maximum packet size to send to this peer */
3773 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
3775 tSize = (afs_uint32) ntohl(tSize);
3776 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3777 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3779 /* sanity check - peer might have restarted with different params.
3780 * If peer says "send less", dammit, send less... Peer should never
3781 * be unable to accept packets of the size that prior AFS versions would
3782 * send without asking. */
3783 if (peer->maxMTU != tSize) {
3784 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
3786 peer->maxMTU = tSize;
3787 peer->MTU = MIN(tSize, peer->MTU);
3788 call->MTU = MIN(call->MTU, tSize);
3791 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3794 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3795 (int)sizeof(afs_int32), &tSize);
3796 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3797 if (tSize < call->twind) { /* smaller than our send */
3798 call->twind = tSize; /* window, we must send less... */
3799 call->ssthresh = MIN(call->twind, call->ssthresh);
3802 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3803 * network MTU confused with the loopback MTU. Calculate the
3804 * maximum MTU here for use in the slow start code below.
3806 maxMTU = peer->maxMTU;
3807 /* Did peer restart with older RX version? */
3808 if (peer->maxDgramPackets > 1) {
3809 peer->maxDgramPackets = 1;
3811 } else if (np->length >=
3812 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3815 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3816 sizeof(afs_int32), &tSize);
3817 tSize = (afs_uint32) ntohl(tSize);
3819 * As of AFS 3.5 we set the send window to match the receive window.
3821 if (tSize < call->twind) {
3822 call->twind = tSize;
3823 call->ssthresh = MIN(call->twind, call->ssthresh);
3824 } else if (tSize > call->twind) {
3825 call->twind = tSize;
3829 * As of AFS 3.5, a jumbogram is more than one fixed size
3830 * packet transmitted in a single UDP datagram. If the remote
3831 * MTU is smaller than our local MTU then never send a datagram
3832 * larger than the natural MTU.
3835 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3836 sizeof(afs_int32), &tSize);
3837 maxDgramPackets = (afs_uint32) ntohl(tSize);
3838 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3840 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
3841 maxDgramPackets = MIN(maxDgramPackets, tSize);
3842 if (maxDgramPackets > 1) {
3843 peer->maxDgramPackets = maxDgramPackets;
3844 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3846 peer->maxDgramPackets = 1;
3847 call->MTU = peer->natMTU;
3849 } else if (peer->maxDgramPackets > 1) {
3850 /* Restarted with lower version of RX */
3851 peer->maxDgramPackets = 1;
3853 } else if (peer->maxDgramPackets > 1
3854 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3855 /* Restarted with lower version of RX */
3856 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3857 peer->natMTU = OLD_MAX_PACKET_SIZE;
3858 peer->MTU = OLD_MAX_PACKET_SIZE;
3859 peer->maxDgramPackets = 1;
3860 peer->nDgramPackets = 1;
3862 call->MTU = OLD_MAX_PACKET_SIZE;
3867 * Calculate how many datagrams were successfully received after
3868 * the first missing packet and adjust the negative ack counter
3873 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3874 if (call->nNacks < nNacked) {
3875 call->nNacks = nNacked;
3884 if (call->flags & RX_CALL_FAST_RECOVER) {
3886 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3888 call->flags &= ~RX_CALL_FAST_RECOVER;
3889 call->cwind = call->nextCwind;
3890 call->nextCwind = 0;
3893 call->nCwindAcks = 0;
3894 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3895 /* Three negative acks in a row trigger congestion recovery */
3896 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3897 MUTEX_EXIT(&peer->peer_lock);
3898 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3899 /* someone else is waiting to start recovery */
3902 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3903 rxi_WaitforTQBusy(call);
3904 MUTEX_ENTER(&peer->peer_lock);
3905 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3906 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3907 call->flags |= RX_CALL_FAST_RECOVER;
3908 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3910 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3911 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3912 call->nextCwind = call->ssthresh;
3915 peer->MTU = call->MTU;
3916 peer->cwind = call->nextCwind;
3917 peer->nDgramPackets = call->nDgramPackets;
3919 call->congestSeq = peer->congestSeq;
3920 /* Reset the resend times on the packets that were nacked
3921 * so we will retransmit as soon as the window permits*/
3922 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3924 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3925 clock_Zero(&tp->retryTime);
3927 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3932 /* If cwind is smaller than ssthresh, then increase
3933 * the window one packet for each ack we receive (exponential
3935 * If cwind is greater than or equal to ssthresh then increase
3936 * the congestion window by one packet for each cwind acks we
3937 * receive (linear growth). */
3938 if (call->cwind < call->ssthresh) {
3940 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3941 call->nCwindAcks = 0;
3943 call->nCwindAcks += newAckCount;
3944 if (call->nCwindAcks >= call->cwind) {
3945 call->nCwindAcks = 0;
3946 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3950 * If we have received several acknowledgements in a row then
3951 * it is time to increase the size of our datagrams
3953 if ((int)call->nAcks > rx_nDgramThreshold) {
3954 if (peer->maxDgramPackets > 1) {
3955 if (call->nDgramPackets < peer->maxDgramPackets) {
3956 call->nDgramPackets++;
3958 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3959 } else if (call->MTU < peer->maxMTU) {
3960 call->MTU += peer->natMTU;
3961 call->MTU = MIN(call->MTU, peer->maxMTU);
3967 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3969 /* Servers need to hold the call until all response packets have
3970 * been acknowledged. Soft acks are good enough since clients
3971 * are not allowed to clear their receive queues. */
3972 if (call->state == RX_STATE_HOLD
3973 && call->tfirst + call->nSoftAcked >= call->tnext) {
3974 call->state = RX_STATE_DALLY;
3975 rxi_ClearTransmitQueue(call, 0);
3976 } else if (!queue_IsEmpty(&call->tq)) {
3977 rxi_Start(0, call, 0, istack);
3982 /* Received a response to a challenge packet */
3984 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
3985 register struct rx_packet *np, int istack)
3989 /* Ignore the packet if we're the client */
3990 if (conn->type == RX_CLIENT_CONNECTION)
3993 /* If already authenticated, ignore the packet (it's probably a retry) */
3994 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
3997 /* Otherwise, have the security object evaluate the response packet */
3998 error = RXS_CheckResponse(conn->securityObject, conn, np);
4000 /* If the response is invalid, reset the connection, sending
4001 * an abort to the peer */
4005 rxi_ConnectionError(conn, error);
4006 MUTEX_ENTER(&conn->conn_data_lock);
4007 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4008 MUTEX_EXIT(&conn->conn_data_lock);
4011 /* If the response is valid, any calls waiting to attach
4012 * servers can now do so */
4015 for (i = 0; i < RX_MAXCALLS; i++) {
4016 struct rx_call *call = conn->call[i];
4018 MUTEX_ENTER(&call->lock);
4019 if (call->state == RX_STATE_PRECALL)
4020 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4021 /* tnop can be null if newcallp is null */
4022 MUTEX_EXIT(&call->lock);
4026 /* Update the peer reachability information, just in case
4027 * some calls went into attach-wait while we were waiting
4028 * for authentication..
4030 rxi_UpdatePeerReach(conn, NULL);
4035 /* A client has received an authentication challenge: the security
4036 * object is asked to cough up a respectable response packet to send
4037 * back to the server. The server is responsible for retrying the
4038 * challenge if it fails to get a response. */
4041 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
4042 register struct rx_packet *np, int istack)
4046 /* Ignore the challenge if we're the server */
4047 if (conn->type == RX_SERVER_CONNECTION)
4050 /* Ignore the challenge if the connection is otherwise idle; someone's
4051 * trying to use us as an oracle. */
4052 if (!rxi_HasActiveCalls(conn))
4055 /* Send the security object the challenge packet. It is expected to fill
4056 * in the response. */
4057 error = RXS_GetResponse(conn->securityObject, conn, np);
4059 /* If the security object is unable to return a valid response, reset the
4060 * connection and send an abort to the peer. Otherwise send the response
4061 * packet to the peer connection. */
4063 rxi_ConnectionError(conn, error);
4064 MUTEX_ENTER(&conn->conn_data_lock);
4065 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4066 MUTEX_EXIT(&conn->conn_data_lock);
4068 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4069 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4075 /* Find an available server process to service the current request in
4076 * the given call structure. If one isn't available, queue up this
4077 * call so it eventually gets one */
4079 rxi_AttachServerProc(register struct rx_call *call,
4080 register osi_socket socket, register int *tnop,
4081 register struct rx_call **newcallp)
4083 register struct rx_serverQueueEntry *sq;
4084 register struct rx_service *service = call->conn->service;
4085 register int haveQuota = 0;
4087 /* May already be attached */
4088 if (call->state == RX_STATE_ACTIVE)
4091 MUTEX_ENTER(&rx_serverPool_lock);
4093 haveQuota = QuotaOK(service);
4094 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4095 /* If there are no processes available to service this call,
4096 * put the call on the incoming call queue (unless it's
4097 * already on the queue).
4099 #ifdef RX_ENABLE_LOCKS
4101 ReturnToServerPool(service);
4102 #endif /* RX_ENABLE_LOCKS */
4104 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4105 call->flags |= RX_CALL_WAIT_PROC;
4106 MUTEX_ENTER(&rx_stats_mutex);
4109 MUTEX_EXIT(&rx_stats_mutex);
4110 rxi_calltrace(RX_CALL_ARRIVAL, call);
4111 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4112 queue_Append(&rx_incomingCallQueue, call);
4115 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4117 /* If hot threads are enabled, and both newcallp and sq->socketp
4118 * are non-null, then this thread will process the call, and the
4119 * idle server thread will start listening on this threads socket.
4122 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4125 *sq->socketp = socket;
4126 clock_GetTime(&call->startTime);
4127 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4131 if (call->flags & RX_CALL_WAIT_PROC) {
4132 /* Conservative: I don't think this should happen */
4133 call->flags &= ~RX_CALL_WAIT_PROC;
4134 if (queue_IsOnQueue(call)) {
4136 MUTEX_ENTER(&rx_stats_mutex);
4138 MUTEX_EXIT(&rx_stats_mutex);
4141 call->state = RX_STATE_ACTIVE;
4142 call->mode = RX_MODE_RECEIVING;
4143 #ifdef RX_KERNEL_TRACE
4145 int glockOwner = ISAFS_GLOCK();
4148 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4149 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4155 if (call->flags & RX_CALL_CLEARED) {
4156 /* send an ack now to start the packet flow up again */
4157 call->flags &= ~RX_CALL_CLEARED;
4158 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4160 #ifdef RX_ENABLE_LOCKS
4163 service->nRequestsRunning++;
4164 if (service->nRequestsRunning <= service->minProcs)
4170 MUTEX_EXIT(&rx_serverPool_lock);
4173 /* Delay the sending of an acknowledge event for a short while, while
4174 * a new call is being prepared (in the case of a client) or a reply
4175 * is being prepared (in the case of a server). Rather than sending
4176 * an ack packet, an ACKALL packet is sent. */
4178 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4180 #ifdef RX_ENABLE_LOCKS
4182 MUTEX_ENTER(&call->lock);
4183 call->delayedAckEvent = NULL;
4184 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4186 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4187 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4189 MUTEX_EXIT(&call->lock);
4190 #else /* RX_ENABLE_LOCKS */
4192 call->delayedAckEvent = NULL;
4193 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4194 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4195 #endif /* RX_ENABLE_LOCKS */
4199 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4202 #ifdef RX_ENABLE_LOCKS
4204 MUTEX_ENTER(&call->lock);
4205 if (event == call->delayedAckEvent)
4206 call->delayedAckEvent = NULL;
4207 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4209 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4211 MUTEX_EXIT(&call->lock);
4212 #else /* RX_ENABLE_LOCKS */
4214 call->delayedAckEvent = NULL;
4215 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4216 #endif /* RX_ENABLE_LOCKS */
4220 #ifdef RX_ENABLE_LOCKS
4221 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4222 * clearing them out.
4225 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4227 register struct rx_packet *p, *tp;
4230 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4231 p->flags |= RX_PKTFLAG_ACKED;
4235 call->flags |= RX_CALL_TQ_CLEARME;
4236 call->flags |= RX_CALL_TQ_SOME_ACKED;
4239 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4240 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4241 call->tfirst = call->tnext;
4242 call->nSoftAcked = 0;
4244 if (call->flags & RX_CALL_FAST_RECOVER) {
4245 call->flags &= ~RX_CALL_FAST_RECOVER;
4246 call->cwind = call->nextCwind;
4247 call->nextCwind = 0;
4250 CV_SIGNAL(&call->cv_twind);
4252 #endif /* RX_ENABLE_LOCKS */
4254 /* Clear out the transmit queue for the current call (all packets have
4255 * been received by peer) */
4257 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4259 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4260 register struct rx_packet *p, *tp;
4262 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4264 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4265 p->flags |= RX_PKTFLAG_ACKED;
4269 call->flags |= RX_CALL_TQ_CLEARME;
4270 call->flags |= RX_CALL_TQ_SOME_ACKED;
4273 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4274 rxi_FreePackets(0, &call->tq);
4275 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4276 call->flags &= ~RX_CALL_TQ_CLEARME;
4278 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4280 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4281 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4282 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4283 call->nSoftAcked = 0;
4285 if (call->flags & RX_CALL_FAST_RECOVER) {
4286 call->flags &= ~RX_CALL_FAST_RECOVER;
4287 call->cwind = call->nextCwind;
4289 #ifdef RX_ENABLE_LOCKS
4290 CV_SIGNAL(&call->cv_twind);
4292 osi_rxWakeup(&call->twind);
4297 rxi_ClearReceiveQueue(register struct rx_call *call)
4299 if (queue_IsNotEmpty(&call->rq)) {
4300 rx_packetReclaims += rxi_FreePackets(0, &call->rq);
4301 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4303 if (call->state == RX_STATE_PRECALL) {
4304 call->flags |= RX_CALL_CLEARED;
4308 /* Send an abort packet for the specified call */
4310 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4311 int istack, int force)
4319 /* Clients should never delay abort messages */
4320 if (rx_IsClientConn(call->conn))
4323 if (call->abortCode != call->error) {
4324 call->abortCode = call->error;
4325 call->abortCount = 0;
4328 if (force || rxi_callAbortThreshhold == 0
4329 || call->abortCount < rxi_callAbortThreshhold) {
4330 if (call->delayedAbortEvent) {
4331 rxevent_Cancel(call->delayedAbortEvent, call,
4332 RX_CALL_REFCOUNT_ABORT);
4334 error = htonl(call->error);
4337 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4338 (char *)&error, sizeof(error), istack);
4339 } else if (!call->delayedAbortEvent) {
4340 clock_GetTime(&when);
4341 clock_Addmsec(&when, rxi_callAbortDelay);
4342 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4343 call->delayedAbortEvent =
4344 rxevent_Post(&when, rxi_SendDelayedCallAbort, call, 0);
4349 /* Send an abort packet for the specified connection. Packet is an
4350 * optional pointer to a packet that can be used to send the abort.
4351 * Once the number of abort messages reaches the threshhold, an
4352 * event is scheduled to send the abort. Setting the force flag
4353 * overrides sending delayed abort messages.
4355 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4356 * to send the abort packet.
4359 rxi_SendConnectionAbort(register struct rx_connection *conn,
4360 struct rx_packet *packet, int istack, int force)
4368 /* Clients should never delay abort messages */
4369 if (rx_IsClientConn(conn))
4372 if (force || rxi_connAbortThreshhold == 0
4373 || conn->abortCount < rxi_connAbortThreshhold) {
4374 if (conn->delayedAbortEvent) {
4375 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4377 error = htonl(conn->error);
4379 MUTEX_EXIT(&conn->conn_data_lock);
4381 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4382 RX_PACKET_TYPE_ABORT, (char *)&error,
4383 sizeof(error), istack);
4384 MUTEX_ENTER(&conn->conn_data_lock);
4385 } else if (!conn->delayedAbortEvent) {
4386 clock_GetTime(&when);
4387 clock_Addmsec(&when, rxi_connAbortDelay);
4388 conn->delayedAbortEvent =
4389 rxevent_Post(&when, rxi_SendDelayedConnAbort, conn, 0);
4394 /* Associate an error all of the calls owned by a connection. Called
4395 * with error non-zero. This is only for really fatal things, like
4396 * bad authentication responses. The connection itself is set in
4397 * error at this point, so that future packets received will be
4400 rxi_ConnectionError(register struct rx_connection *conn,
4401 register afs_int32 error)
4406 dpf(("rxi_ConnectionError conn %x error %d", conn, error));
4408 MUTEX_ENTER(&conn->conn_data_lock);
4409 if (conn->challengeEvent)
4410 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4411 if (conn->checkReachEvent) {
4412 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4413 conn->checkReachEvent = 0;
4414 conn->flags &= ~RX_CONN_ATTACHWAIT;
4417 MUTEX_EXIT(&conn->conn_data_lock);
4418 for (i = 0; i < RX_MAXCALLS; i++) {
4419 struct rx_call *call = conn->call[i];
4421 MUTEX_ENTER(&call->lock);
4422 rxi_CallError(call, error);
4423 MUTEX_EXIT(&call->lock);
4426 conn->error = error;
4427 MUTEX_ENTER(&rx_stats_mutex);
4428 rx_stats.fatalErrors++;
4429 MUTEX_EXIT(&rx_stats_mutex);
4434 rxi_CallError(register struct rx_call *call, afs_int32 error)
4436 dpf(("rxi_CallError call %x error %d call->error %d", call, error, call->error));
4438 error = call->error;
4440 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4441 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4442 rxi_ResetCall(call, 0);
4445 rxi_ResetCall(call, 0);
4447 call->error = error;
4448 call->mode = RX_MODE_ERROR;
4451 /* Reset various fields in a call structure, and wakeup waiting
4452 * processes. Some fields aren't changed: state & mode are not
4453 * touched (these must be set by the caller), and bufptr, nLeft, and
4454 * nFree are not reset, since these fields are manipulated by
4455 * unprotected macros, and may only be reset by non-interrupting code.
4458 /* this code requires that call->conn be set properly as a pre-condition. */
4459 #endif /* ADAPT_WINDOW */
4462 rxi_ResetCall(register struct rx_call *call, register int newcall)
4465 register struct rx_peer *peer;
4466 struct rx_packet *packet;
4468 dpf(("rxi_ResetCall(call %x, newcall %d)\n", call, newcall));
4470 /* Notify anyone who is waiting for asynchronous packet arrival */
4471 if (call->arrivalProc) {
4472 (*call->arrivalProc) (call, call->arrivalProcHandle,
4473 call->arrivalProcArg);
4474 call->arrivalProc = (void (*)())0;
4477 if (call->delayedAbortEvent) {
4478 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4479 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4481 rxi_SendCallAbort(call, packet, 0, 1);
4482 rxi_FreePacket(packet);
4487 * Update the peer with the congestion information in this call
4488 * so other calls on this connection can pick up where this call
4489 * left off. If the congestion sequence numbers don't match then
4490 * another call experienced a retransmission.
4492 peer = call->conn->peer;
4493 MUTEX_ENTER(&peer->peer_lock);
4495 if (call->congestSeq == peer->congestSeq) {
4496 peer->cwind = MAX(peer->cwind, call->cwind);
4497 peer->MTU = MAX(peer->MTU, call->MTU);
4498 peer->nDgramPackets =
4499 MAX(peer->nDgramPackets, call->nDgramPackets);
4502 call->abortCode = 0;
4503 call->abortCount = 0;
4505 if (peer->maxDgramPackets > 1) {
4506 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4508 call->MTU = peer->MTU;
4510 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4511 call->ssthresh = rx_maxSendWindow;
4512 call->nDgramPackets = peer->nDgramPackets;
4513 call->congestSeq = peer->congestSeq;
4514 MUTEX_EXIT(&peer->peer_lock);
4516 flags = call->flags;
4517 rxi_ClearReceiveQueue(call);
4518 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4519 if (flags & RX_CALL_TQ_BUSY) {
4520 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4521 call->flags |= (flags & RX_CALL_TQ_WAIT);
4523 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4525 rxi_ClearTransmitQueue(call, 0);
4526 queue_Init(&call->tq);
4527 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4528 dpf(("rcall %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4531 while (call->tqWaiters) {
4532 #ifdef RX_ENABLE_LOCKS
4533 CV_BROADCAST(&call->cv_tq);
4534 #else /* RX_ENABLE_LOCKS */
4535 osi_rxWakeup(&call->tq);
4536 #endif /* RX_ENABLE_LOCKS */
4540 queue_Init(&call->rq);
4542 call->rwind = rx_initReceiveWindow;
4543 call->twind = rx_initSendWindow;
4544 call->nSoftAcked = 0;
4545 call->nextCwind = 0;
4548 call->nCwindAcks = 0;
4549 call->nSoftAcks = 0;
4550 call->nHardAcks = 0;
4552 call->tfirst = call->rnext = call->tnext = 1;
4554 call->lastAcked = 0;
4555 call->localStatus = call->remoteStatus = 0;
4557 if (flags & RX_CALL_READER_WAIT) {
4558 #ifdef RX_ENABLE_LOCKS
4559 CV_BROADCAST(&call->cv_rq);
4561 osi_rxWakeup(&call->rq);
4564 if (flags & RX_CALL_WAIT_PACKETS) {
4565 MUTEX_ENTER(&rx_freePktQ_lock);
4566 rxi_PacketsUnWait(); /* XXX */
4567 MUTEX_EXIT(&rx_freePktQ_lock);
4569 #ifdef RX_ENABLE_LOCKS
4570 CV_SIGNAL(&call->cv_twind);
4572 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4573 osi_rxWakeup(&call->twind);
4576 #ifdef RX_ENABLE_LOCKS
4577 /* The following ensures that we don't mess with any queue while some
4578 * other thread might also be doing so. The call_queue_lock field is
4579 * is only modified under the call lock. If the call is in the process
4580 * of being removed from a queue, the call is not locked until the
4581 * the queue lock is dropped and only then is the call_queue_lock field
4582 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4583 * Note that any other routine which removes a call from a queue has to
4584 * obtain the queue lock before examing the queue and removing the call.
4586 if (call->call_queue_lock) {
4587 MUTEX_ENTER(call->call_queue_lock);
4588 if (queue_IsOnQueue(call)) {
4590 if (flags & RX_CALL_WAIT_PROC) {
4591 MUTEX_ENTER(&rx_stats_mutex);
4593 MUTEX_EXIT(&rx_stats_mutex);
4596 MUTEX_EXIT(call->call_queue_lock);
4597 CLEAR_CALL_QUEUE_LOCK(call);
4599 #else /* RX_ENABLE_LOCKS */
4600 if (queue_IsOnQueue(call)) {
4602 if (flags & RX_CALL_WAIT_PROC)
4605 #endif /* RX_ENABLE_LOCKS */
4607 rxi_KeepAliveOff(call);
4608 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4611 /* Send an acknowledge for the indicated packet (seq,serial) of the
4612 * indicated call, for the indicated reason (reason). This
4613 * acknowledge will specifically acknowledge receiving the packet, and
4614 * will also specify which other packets for this call have been
4615 * received. This routine returns the packet that was used to the
4616 * caller. The caller is responsible for freeing it or re-using it.
4617 * This acknowledgement also returns the highest sequence number
4618 * actually read out by the higher level to the sender; the sender
4619 * promises to keep around packets that have not been read by the
4620 * higher level yet (unless, of course, the sender decides to abort
4621 * the call altogether). Any of p, seq, serial, pflags, or reason may
4622 * be set to zero without ill effect. That is, if they are zero, they
4623 * will not convey any information.
4624 * NOW there is a trailer field, after the ack where it will safely be
4625 * ignored by mundanes, which indicates the maximum size packet this
4626 * host can swallow. */
4628 register struct rx_packet *optionalPacket; use to send ack (or null)
4629 int seq; Sequence number of the packet we are acking
4630 int serial; Serial number of the packet
4631 int pflags; Flags field from packet header
4632 int reason; Reason an acknowledge was prompted
4636 rxi_SendAck(register struct rx_call *call,
4637 register struct rx_packet *optionalPacket, int serial, int reason,
4640 struct rx_ackPacket *ap;
4641 register struct rx_packet *rqp;
4642 register struct rx_packet *nxp; /* For queue_Scan */
4643 register struct rx_packet *p;
4646 #ifdef RX_ENABLE_TSFPQ
4647 struct rx_ts_info_t * rx_ts_info;
4651 * Open the receive window once a thread starts reading packets
4653 if (call->rnext > 1) {
4654 call->rwind = rx_maxReceiveWindow;
4657 call->nHardAcks = 0;
4658 call->nSoftAcks = 0;
4659 if (call->rnext > call->lastAcked)
4660 call->lastAcked = call->rnext;
4664 rx_computelen(p, p->length); /* reset length, you never know */
4665 } /* where that's been... */
4666 #ifdef RX_ENABLE_TSFPQ
4668 RX_TS_INFO_GET(rx_ts_info);
4669 if ((p = rx_ts_info->local_special_packet)) {
4670 rx_computelen(p, p->length);
4671 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4672 rx_ts_info->local_special_packet = p;
4673 } else { /* We won't send the ack, but don't panic. */
4674 return optionalPacket;
4678 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4679 /* We won't send the ack, but don't panic. */
4680 return optionalPacket;
4685 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4688 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4689 #ifndef RX_ENABLE_TSFPQ
4690 if (!optionalPacket)
4693 return optionalPacket;
4695 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4696 if (rx_Contiguous(p) < templ) {
4697 #ifndef RX_ENABLE_TSFPQ
4698 if (!optionalPacket)
4701 return optionalPacket;
4706 /* MTUXXX failing to send an ack is very serious. We should */
4707 /* try as hard as possible to send even a partial ack; it's */
4708 /* better than nothing. */
4709 ap = (struct rx_ackPacket *)rx_DataOf(p);
4710 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4711 ap->reason = reason;
4713 /* The skew computation used to be bogus, I think it's better now. */
4714 /* We should start paying attention to skew. XXX */
4715 ap->serial = htonl(serial);
4716 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4718 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4719 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4721 /* No fear of running out of ack packet here because there can only be at most
4722 * one window full of unacknowledged packets. The window size must be constrained
4723 * to be less than the maximum ack size, of course. Also, an ack should always
4724 * fit into a single packet -- it should not ever be fragmented. */
4725 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4726 if (!rqp || !call->rq.next
4727 || (rqp->header.seq > (call->rnext + call->rwind))) {
4728 #ifndef RX_ENABLE_TSFPQ
4729 if (!optionalPacket)
4732 rxi_CallError(call, RX_CALL_DEAD);
4733 return optionalPacket;
4736 while (rqp->header.seq > call->rnext + offset)
4737 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4738 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4740 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4741 #ifndef RX_ENABLE_TSFPQ
4742 if (!optionalPacket)
4745 rxi_CallError(call, RX_CALL_DEAD);
4746 return optionalPacket;
4751 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4753 /* these are new for AFS 3.3 */
4754 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4755 templ = htonl(templ);
4756 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4757 templ = htonl(call->conn->peer->ifMTU);
4758 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4759 sizeof(afs_int32), &templ);
4761 /* new for AFS 3.4 */
4762 templ = htonl(call->rwind);
4763 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4764 sizeof(afs_int32), &templ);
4766 /* new for AFS 3.5 */
4767 templ = htonl(call->conn->peer->ifDgramPackets);
4768 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4769 sizeof(afs_int32), &templ);
4771 p->header.serviceId = call->conn->serviceId;
4772 p->header.cid = (call->conn->cid | call->channel);
4773 p->header.callNumber = *call->callNumber;
4775 p->header.securityIndex = call->conn->securityIndex;
4776 p->header.epoch = call->conn->epoch;
4777 p->header.type = RX_PACKET_TYPE_ACK;
4778 p->header.flags = RX_SLOW_START_OK;
4779 if (reason == RX_ACK_PING) {
4780 p->header.flags |= RX_REQUEST_ACK;
4782 clock_GetTime(&call->pingRequestTime);
4785 if (call->conn->type == RX_CLIENT_CONNECTION)
4786 p->header.flags |= RX_CLIENT_INITIATED;
4790 if (rxdebug_active) {
4794 len = _snprintf(msg, sizeof(msg),
4795 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4796 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4797 ntohl(ap->serial), ntohl(ap->previousPacket),
4798 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
4799 ap->nAcks, ntohs(ap->bufferSpace) );
4803 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
4804 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4808 OutputDebugString(msg);
4810 #else /* AFS_NT40_ENV */
4812 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
4813 ap->reason, ntohl(ap->previousPacket),
4814 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4816 for (offset = 0; offset < ap->nAcks; offset++)
4817 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4822 #endif /* AFS_NT40_ENV */
4825 register int i, nbytes = p->length;
4827 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4828 if (nbytes <= p->wirevec[i].iov_len) {
4829 register int savelen, saven;
4831 savelen = p->wirevec[i].iov_len;
4833 p->wirevec[i].iov_len = nbytes;
4835 rxi_Send(call, p, istack);
4836 p->wirevec[i].iov_len = savelen;
4840 nbytes -= p->wirevec[i].iov_len;
4843 MUTEX_ENTER(&rx_stats_mutex);
4844 rx_stats.ackPacketsSent++;
4845 MUTEX_EXIT(&rx_stats_mutex);
4846 #ifndef RX_ENABLE_TSFPQ
4847 if (!optionalPacket)
4850 return optionalPacket; /* Return packet for re-use by caller */
4853 /* Send all of the packets in the list in single datagram */
4855 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4856 int istack, int moreFlag, struct clock *now,
4857 struct clock *retryTime, int resending)
4862 struct rx_connection *conn = call->conn;
4863 struct rx_peer *peer = conn->peer;
4865 MUTEX_ENTER(&peer->peer_lock);
4868 peer->reSends += len;
4869 MUTEX_ENTER(&rx_stats_mutex);
4870 rx_stats.dataPacketsSent += len;
4871 MUTEX_EXIT(&rx_stats_mutex);
4872 MUTEX_EXIT(&peer->peer_lock);
4874 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4878 /* Set the packet flags and schedule the resend events */
4879 /* Only request an ack for the last packet in the list */
4880 for (i = 0; i < len; i++) {
4881 list[i]->retryTime = *retryTime;
4882 if (list[i]->header.serial) {
4883 /* Exponentially backoff retry times */
4884 if (list[i]->backoff < MAXBACKOFF) {
4885 /* so it can't stay == 0 */
4886 list[i]->backoff = (list[i]->backoff << 1) + 1;
4889 clock_Addmsec(&(list[i]->retryTime),
4890 ((afs_uint32) list[i]->backoff) << 8);
4893 /* Wait a little extra for the ack on the last packet */
4894 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4895 clock_Addmsec(&(list[i]->retryTime), 400);
4898 /* Record the time sent */
4899 list[i]->timeSent = *now;
4901 /* Ask for an ack on retransmitted packets, on every other packet
4902 * if the peer doesn't support slow start. Ask for an ack on every
4903 * packet until the congestion window reaches the ack rate. */
4904 if (list[i]->header.serial) {
4906 MUTEX_ENTER(&rx_stats_mutex);
4907 rx_stats.dataPacketsReSent++;
4908 MUTEX_EXIT(&rx_stats_mutex);
4910 /* improved RTO calculation- not Karn */
4911 list[i]->firstSent = *now;
4912 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4913 || (!(call->flags & RX_CALL_SLOW_START_OK)
4914 && (list[i]->header.seq & 1)))) {
4919 MUTEX_ENTER(&peer->peer_lock);
4923 MUTEX_ENTER(&rx_stats_mutex);
4924 rx_stats.dataPacketsSent++;
4925 MUTEX_EXIT(&rx_stats_mutex);
4926 MUTEX_EXIT(&peer->peer_lock);
4928 /* Tag this packet as not being the last in this group,
4929 * for the receiver's benefit */
4930 if (i < len - 1 || moreFlag) {
4931 list[i]->header.flags |= RX_MORE_PACKETS;
4934 /* Install the new retransmit time for the packet, and
4935 * record the time sent */
4936 list[i]->timeSent = *now;
4940 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4943 /* Since we're about to send a data packet to the peer, it's
4944 * safe to nuke any scheduled end-of-packets ack */
4945 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4947 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4948 MUTEX_EXIT(&call->lock);
4950 rxi_SendPacketList(call, conn, list, len, istack);
4952 rxi_SendPacket(call, conn, list[0], istack);
4954 MUTEX_ENTER(&call->lock);
4955 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4957 /* Update last send time for this call (for keep-alive
4958 * processing), and for the connection (so that we can discover
4959 * idle connections) */
4960 conn->lastSendTime = call->lastSendTime = clock_Sec();
4963 /* When sending packets we need to follow these rules:
4964 * 1. Never send more than maxDgramPackets in a jumbogram.
4965 * 2. Never send a packet with more than two iovecs in a jumbogram.
4966 * 3. Never send a retransmitted packet in a jumbogram.
4967 * 4. Never send more than cwind/4 packets in a jumbogram
4968 * We always keep the last list we should have sent so we
4969 * can set the RX_MORE_PACKETS flags correctly.
4972 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4973 int istack, struct clock *now, struct clock *retryTime,
4976 int i, cnt, lastCnt = 0;
4977 struct rx_packet **listP, **lastP = 0;
4978 struct rx_peer *peer = call->conn->peer;
4979 int morePackets = 0;
4981 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4982 /* Does the current packet force us to flush the current list? */
4984 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4985 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4987 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
4989 /* If the call enters an error state stop sending, or if
4990 * we entered congestion recovery mode, stop sending */
4991 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4999 /* Add the current packet to the list if it hasn't been acked.
5000 * Otherwise adjust the list pointer to skip the current packet. */
5001 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5003 /* Do we need to flush the list? */
5004 if (cnt >= (int)peer->maxDgramPackets
5005 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5006 || list[i]->header.serial
5007 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5009 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5010 retryTime, resending);
5011 /* If the call enters an error state stop sending, or if
5012 * we entered congestion recovery mode, stop sending */
5014 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5019 listP = &list[i + 1];
5024 osi_Panic("rxi_SendList error");
5026 listP = &list[i + 1];
5030 /* Send the whole list when the call is in receive mode, when
5031 * the call is in eof mode, when we are in fast recovery mode,
5032 * and when we have the last packet */
5033 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5034 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5035 || (call->flags & RX_CALL_FAST_RECOVER)) {
5036 /* Check for the case where the current list contains
5037 * an acked packet. Since we always send retransmissions
5038 * in a separate packet, we only need to check the first
5039 * packet in the list */
5040 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5044 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5045 retryTime, resending);
5046 /* If the call enters an error state stop sending, or if
5047 * we entered congestion recovery mode, stop sending */
5048 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5052 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5055 } else if (lastCnt > 0) {
5056 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5061 #ifdef RX_ENABLE_LOCKS
5062 /* Call rxi_Start, below, but with the call lock held. */
5064 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
5065 void *arg1, int istack)
5067 MUTEX_ENTER(&call->lock);
5068 rxi_Start(event, call, arg1, istack);
5069 MUTEX_EXIT(&call->lock);
5071 #endif /* RX_ENABLE_LOCKS */
5073 /* This routine is called when new packets are readied for
5074 * transmission and when retransmission may be necessary, or when the
5075 * transmission window or burst count are favourable. This should be
5076 * better optimized for new packets, the usual case, now that we've
5077 * got rid of queues of send packets. XXXXXXXXXXX */
5079 rxi_Start(struct rxevent *event, register struct rx_call *call,
5080 void *arg1, int istack)
5082 struct rx_packet *p;
5083 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
5084 struct rx_peer *peer = call->conn->peer;
5085 struct clock now, retryTime;
5089 struct rx_packet **xmitList;
5092 /* If rxi_Start is being called as a result of a resend event,
5093 * then make sure that the event pointer is removed from the call
5094 * structure, since there is no longer a per-call retransmission
5096 if (event && event == call->resendEvent) {
5097 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5098 call->resendEvent = NULL;
5100 if (queue_IsEmpty(&call->tq)) {
5104 /* Timeouts trigger congestion recovery */
5105 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5106 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5107 /* someone else is waiting to start recovery */
5110 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5111 rxi_WaitforTQBusy(call);
5112 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5113 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5114 call->flags |= RX_CALL_FAST_RECOVER;
5115 if (peer->maxDgramPackets > 1) {
5116 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5118 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5120 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5121 call->nDgramPackets = 1;
5123 call->nextCwind = 1;
5126 MUTEX_ENTER(&peer->peer_lock);
5127 peer->MTU = call->MTU;
5128 peer->cwind = call->cwind;
5129 peer->nDgramPackets = 1;
5131 call->congestSeq = peer->congestSeq;
5132 MUTEX_EXIT(&peer->peer_lock);
5133 /* Clear retry times on packets. Otherwise, it's possible for
5134 * some packets in the queue to force resends at rates faster
5135 * than recovery rates.
5137 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5138 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5139 clock_Zero(&p->retryTime);
5144 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5145 MUTEX_ENTER(&rx_stats_mutex);
5146 rx_tq_debug.rxi_start_in_error++;
5147 MUTEX_EXIT(&rx_stats_mutex);
5152 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5153 /* Get clock to compute the re-transmit time for any packets
5154 * in this burst. Note, if we back off, it's reasonable to
5155 * back off all of the packets in the same manner, even if
5156 * some of them have been retransmitted more times than more
5157 * recent additions */
5158 clock_GetTime(&now);
5159 retryTime = now; /* initialize before use */
5160 MUTEX_ENTER(&peer->peer_lock);
5161 clock_Add(&retryTime, &peer->timeout);
5162 MUTEX_EXIT(&peer->peer_lock);
5164 /* Send (or resend) any packets that need it, subject to
5165 * window restrictions and congestion burst control
5166 * restrictions. Ask for an ack on the last packet sent in
5167 * this burst. For now, we're relying upon the window being
5168 * considerably bigger than the largest number of packets that
5169 * are typically sent at once by one initial call to
5170 * rxi_Start. This is probably bogus (perhaps we should ask
5171 * for an ack when we're half way through the current
5172 * window?). Also, for non file transfer applications, this
5173 * may end up asking for an ack for every packet. Bogus. XXXX
5176 * But check whether we're here recursively, and let the other guy
5179 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5180 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5181 call->flags |= RX_CALL_TQ_BUSY;
5183 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5185 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5186 call->flags &= ~RX_CALL_NEED_START;
5187 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5189 maxXmitPackets = MIN(call->twind, call->cwind);
5190 xmitList = (struct rx_packet **)
5191 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5192 if (xmitList == NULL)
5193 osi_Panic("rxi_Start, failed to allocate xmit list");
5194 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5195 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5196 /* We shouldn't be sending packets if a thread is waiting
5197 * to initiate congestion recovery */
5201 && (call->flags & RX_CALL_FAST_RECOVER)) {
5202 /* Only send one packet during fast recovery */
5205 if ((p->flags & RX_PKTFLAG_FREE)
5206 || (!queue_IsEnd(&call->tq, nxp)
5207 && (nxp->flags & RX_PKTFLAG_FREE))
5208 || (p == (struct rx_packet *)&rx_freePacketQueue)
5209 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5210 osi_Panic("rxi_Start: xmit queue clobbered");
5212 if (p->flags & RX_PKTFLAG_ACKED) {
5213 MUTEX_ENTER(&rx_stats_mutex);
5214 rx_stats.ignoreAckedPacket++;
5215 MUTEX_EXIT(&rx_stats_mutex);
5216 continue; /* Ignore this packet if it has been acknowledged */
5219 /* Turn off all flags except these ones, which are the same
5220 * on each transmission */
5221 p->header.flags &= RX_PRESET_FLAGS;
5223 if (p->header.seq >=
5224 call->tfirst + MIN((int)call->twind,
5225 (int)(call->nSoftAcked +
5227 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5228 /* Note: if we're waiting for more window space, we can
5229 * still send retransmits; hence we don't return here, but
5230 * break out to schedule a retransmit event */
5231 dpf(("call %d waiting for window",
5232 *(call->callNumber)));
5236 /* Transmit the packet if it needs to be sent. */
5237 if (!clock_Lt(&now, &p->retryTime)) {
5238 if (nXmitPackets == maxXmitPackets) {
5239 rxi_SendXmitList(call, xmitList, nXmitPackets,
5240 istack, &now, &retryTime,
5242 osi_Free(xmitList, maxXmitPackets *
5243 sizeof(struct rx_packet *));
5246 xmitList[nXmitPackets++] = p;
5250 /* xmitList now hold pointers to all of the packets that are
5251 * ready to send. Now we loop to send the packets */
5252 if (nXmitPackets > 0) {
5253 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5254 &now, &retryTime, resending);
5257 maxXmitPackets * sizeof(struct rx_packet *));
5259 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5261 * TQ references no longer protected by this flag; they must remain
5262 * protected by the global lock.
5264 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5265 call->flags &= ~RX_CALL_TQ_BUSY;
5266 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5267 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5268 #ifdef RX_ENABLE_LOCKS
5269 osirx_AssertMine(&call->lock, "rxi_Start start");
5270 CV_BROADCAST(&call->cv_tq);
5271 #else /* RX_ENABLE_LOCKS */
5272 osi_rxWakeup(&call->tq);
5273 #endif /* RX_ENABLE_LOCKS */
5278 /* We went into the error state while sending packets. Now is
5279 * the time to reset the call. This will also inform the using
5280 * process that the call is in an error state.
5282 MUTEX_ENTER(&rx_stats_mutex);
5283 rx_tq_debug.rxi_start_aborted++;
5284 MUTEX_EXIT(&rx_stats_mutex);
5285 call->flags &= ~RX_CALL_TQ_BUSY;
5286 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5287 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5288 #ifdef RX_ENABLE_LOCKS
5289 osirx_AssertMine(&call->lock, "rxi_Start middle");
5290 CV_BROADCAST(&call->cv_tq);
5291 #else /* RX_ENABLE_LOCKS */
5292 osi_rxWakeup(&call->tq);
5293 #endif /* RX_ENABLE_LOCKS */
5295 rxi_CallError(call, call->error);
5298 #ifdef RX_ENABLE_LOCKS
5299 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5300 register int missing;
5301 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5302 /* Some packets have received acks. If they all have, we can clear
5303 * the transmit queue.
5306 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5307 if (p->header.seq < call->tfirst
5308 && (p->flags & RX_PKTFLAG_ACKED)) {
5315 call->flags |= RX_CALL_TQ_CLEARME;
5317 #endif /* RX_ENABLE_LOCKS */
5318 /* Don't bother doing retransmits if the TQ is cleared. */
5319 if (call->flags & RX_CALL_TQ_CLEARME) {
5320 rxi_ClearTransmitQueue(call, 1);
5322 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5325 /* Always post a resend event, if there is anything in the
5326 * queue, and resend is possible. There should be at least
5327 * one unacknowledged packet in the queue ... otherwise none
5328 * of these packets should be on the queue in the first place.
5330 if (call->resendEvent) {
5331 /* Cancel the existing event and post a new one */
5332 rxevent_Cancel(call->resendEvent, call,
5333 RX_CALL_REFCOUNT_RESEND);
5336 /* The retry time is the retry time on the first unacknowledged
5337 * packet inside the current window */
5339 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5340 /* Don't set timers for packets outside the window */
5341 if (p->header.seq >= call->tfirst + call->twind) {
5345 if (!(p->flags & RX_PKTFLAG_ACKED)
5346 && !clock_IsZero(&p->retryTime)) {
5348 retryTime = p->retryTime;
5353 /* Post a new event to re-run rxi_Start when retries may be needed */
5354 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5355 #ifdef RX_ENABLE_LOCKS
5356 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5358 rxevent_Post2(&retryTime, rxi_StartUnlocked,
5359 (void *)call, 0, istack);
5360 #else /* RX_ENABLE_LOCKS */
5362 rxevent_Post2(&retryTime, rxi_Start, (void *)call,
5364 #endif /* RX_ENABLE_LOCKS */
5367 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5368 } while (call->flags & RX_CALL_NEED_START);
5370 * TQ references no longer protected by this flag; they must remain
5371 * protected by the global lock.
5373 call->flags &= ~RX_CALL_TQ_BUSY;
5374 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5375 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5376 #ifdef RX_ENABLE_LOCKS
5377 osirx_AssertMine(&call->lock, "rxi_Start end");
5378 CV_BROADCAST(&call->cv_tq);
5379 #else /* RX_ENABLE_LOCKS */
5380 osi_rxWakeup(&call->tq);
5381 #endif /* RX_ENABLE_LOCKS */
5384 call->flags |= RX_CALL_NEED_START;
5386 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5388 if (call->resendEvent) {
5389 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5394 /* Also adjusts the keep alive parameters for the call, to reflect
5395 * that we have just sent a packet (so keep alives aren't sent
5398 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5401 register struct rx_connection *conn = call->conn;
5403 /* Stamp each packet with the user supplied status */
5404 p->header.userStatus = call->localStatus;
5406 /* Allow the security object controlling this call's security to
5407 * make any last-minute changes to the packet */
5408 RXS_SendPacket(conn->securityObject, call, p);
5410 /* Since we're about to send SOME sort of packet to the peer, it's
5411 * safe to nuke any scheduled end-of-packets ack */
5412 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5414 /* Actually send the packet, filling in more connection-specific fields */
5415 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5416 MUTEX_EXIT(&call->lock);
5417 rxi_SendPacket(call, conn, p, istack);
5418 MUTEX_ENTER(&call->lock);
5419 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5421 /* Update last send time for this call (for keep-alive
5422 * processing), and for the connection (so that we can discover
5423 * idle connections) */
5424 conn->lastSendTime = call->lastSendTime = clock_Sec();
5428 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5429 * that things are fine. Also called periodically to guarantee that nothing
5430 * falls through the cracks (e.g. (error + dally) connections have keepalive
5431 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5433 * haveCTLock Set if calling from rxi_ReapConnections
5435 #ifdef RX_ENABLE_LOCKS
5437 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5438 #else /* RX_ENABLE_LOCKS */
5440 rxi_CheckCall(register struct rx_call *call)
5441 #endif /* RX_ENABLE_LOCKS */
5443 register struct rx_connection *conn = call->conn;
5445 afs_uint32 deadTime;
5447 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5448 if (call->flags & RX_CALL_TQ_BUSY) {
5449 /* Call is active and will be reset by rxi_Start if it's
5450 * in an error state.
5455 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5457 (((afs_uint32) conn->secondsUntilDead << 10) +
5458 ((afs_uint32) conn->peer->rtt >> 3) +
5459 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5461 /* These are computed to the second (+- 1 second). But that's
5462 * good enough for these values, which should be a significant
5463 * number of seconds. */
5464 if (now > (call->lastReceiveTime + deadTime)) {
5465 if (call->state == RX_STATE_ACTIVE) {
5466 rxi_CallError(call, RX_CALL_DEAD);
5469 #ifdef RX_ENABLE_LOCKS
5470 /* Cancel pending events */
5471 rxevent_Cancel(call->delayedAckEvent, call,
5472 RX_CALL_REFCOUNT_DELAY);
5473 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5474 rxevent_Cancel(call->keepAliveEvent, call,
5475 RX_CALL_REFCOUNT_ALIVE);
5476 if (call->refCount == 0) {
5477 rxi_FreeCall(call, haveCTLock);
5481 #else /* RX_ENABLE_LOCKS */
5484 #endif /* RX_ENABLE_LOCKS */
5486 /* Non-active calls are destroyed if they are not responding
5487 * to pings; active calls are simply flagged in error, so the
5488 * attached process can die reasonably gracefully. */
5490 /* see if we have a non-activity timeout */
5491 if (call->startWait && conn->idleDeadTime
5492 && ((call->startWait + conn->idleDeadTime) < now)) {
5493 if (call->state == RX_STATE_ACTIVE) {
5494 rxi_CallError(call, RX_CALL_TIMEOUT);
5498 /* see if we have a hard timeout */
5499 if (conn->hardDeadTime
5500 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5501 if (call->state == RX_STATE_ACTIVE)
5502 rxi_CallError(call, RX_CALL_TIMEOUT);
5509 /* When a call is in progress, this routine is called occasionally to
5510 * make sure that some traffic has arrived (or been sent to) the peer.
5511 * If nothing has arrived in a reasonable amount of time, the call is
5512 * declared dead; if nothing has been sent for a while, we send a
5513 * keep-alive packet (if we're actually trying to keep the call alive)
5516 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5519 struct rx_connection *conn;
5522 MUTEX_ENTER(&call->lock);
5523 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5524 if (event == call->keepAliveEvent)
5525 call->keepAliveEvent = NULL;
5528 #ifdef RX_ENABLE_LOCKS
5529 if (rxi_CheckCall(call, 0)) {
5530 MUTEX_EXIT(&call->lock);
5533 #else /* RX_ENABLE_LOCKS */
5534 if (rxi_CheckCall(call))
5536 #endif /* RX_ENABLE_LOCKS */
5538 /* Don't try to keep alive dallying calls */
5539 if (call->state == RX_STATE_DALLY) {
5540 MUTEX_EXIT(&call->lock);
5545 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5546 /* Don't try to send keepalives if there is unacknowledged data */
5547 /* the rexmit code should be good enough, this little hack
5548 * doesn't quite work XXX */
5549 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5551 rxi_ScheduleKeepAliveEvent(call);
5552 MUTEX_EXIT(&call->lock);
5557 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5559 if (!call->keepAliveEvent) {
5561 clock_GetTime(&when);
5562 when.sec += call->conn->secondsUntilPing;
5563 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5564 call->keepAliveEvent =
5565 rxevent_Post(&when, rxi_KeepAliveEvent, call, 0);
5569 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5571 rxi_KeepAliveOn(register struct rx_call *call)
5573 /* Pretend last packet received was received now--i.e. if another
5574 * packet isn't received within the keep alive time, then the call
5575 * will die; Initialize last send time to the current time--even
5576 * if a packet hasn't been sent yet. This will guarantee that a
5577 * keep-alive is sent within the ping time */
5578 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5579 rxi_ScheduleKeepAliveEvent(call);
5582 /* This routine is called to send connection abort messages
5583 * that have been delayed to throttle looping clients. */
5585 rxi_SendDelayedConnAbort(struct rxevent *event,
5586 register struct rx_connection *conn, char *dummy)
5589 struct rx_packet *packet;
5591 MUTEX_ENTER(&conn->conn_data_lock);
5592 conn->delayedAbortEvent = NULL;
5593 error = htonl(conn->error);
5595 MUTEX_EXIT(&conn->conn_data_lock);
5596 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5599 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5600 RX_PACKET_TYPE_ABORT, (char *)&error,
5602 rxi_FreePacket(packet);
5606 /* This routine is called to send call abort messages
5607 * that have been delayed to throttle looping clients. */
5609 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5613 struct rx_packet *packet;
5615 MUTEX_ENTER(&call->lock);
5616 call->delayedAbortEvent = NULL;
5617 error = htonl(call->error);
5619 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5622 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5623 (char *)&error, sizeof(error), 0);
5624 rxi_FreePacket(packet);
5626 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5627 MUTEX_EXIT(&call->lock);
5630 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5631 * seconds) to ask the client to authenticate itself. The routine
5632 * issues a challenge to the client, which is obtained from the
5633 * security object associated with the connection */
5635 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5636 void *arg1, int tries)
5638 conn->challengeEvent = NULL;
5639 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5640 register struct rx_packet *packet;
5644 /* We've failed to authenticate for too long.
5645 * Reset any calls waiting for authentication;
5646 * they are all in RX_STATE_PRECALL.
5650 MUTEX_ENTER(&conn->conn_call_lock);
5651 for (i = 0; i < RX_MAXCALLS; i++) {
5652 struct rx_call *call = conn->call[i];
5654 MUTEX_ENTER(&call->lock);
5655 if (call->state == RX_STATE_PRECALL) {
5656 rxi_CallError(call, RX_CALL_DEAD);
5657 rxi_SendCallAbort(call, NULL, 0, 0);
5659 MUTEX_EXIT(&call->lock);
5662 MUTEX_EXIT(&conn->conn_call_lock);
5666 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5668 /* If there's no packet available, do this later. */
5669 RXS_GetChallenge(conn->securityObject, conn, packet);
5670 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5671 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5672 rxi_FreePacket(packet);
5674 clock_GetTime(&when);
5675 when.sec += RX_CHALLENGE_TIMEOUT;
5676 conn->challengeEvent =
5677 rxevent_Post2(&when, rxi_ChallengeEvent, conn, 0,
5682 /* Call this routine to start requesting the client to authenticate
5683 * itself. This will continue until authentication is established,
5684 * the call times out, or an invalid response is returned. The
5685 * security object associated with the connection is asked to create
5686 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5687 * defined earlier. */
5689 rxi_ChallengeOn(register struct rx_connection *conn)
5691 if (!conn->challengeEvent) {
5692 RXS_CreateChallenge(conn->securityObject, conn);
5693 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5698 /* Compute round trip time of the packet provided, in *rttp.
5701 /* rxi_ComputeRoundTripTime is called with peer locked. */
5702 /* sentp and/or peer may be null */
5704 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5705 register struct clock *sentp,
5706 register struct rx_peer *peer)
5708 struct clock thisRtt, *rttp = &thisRtt;
5710 register int rtt_timeout;
5712 clock_GetTime(rttp);
5714 if (clock_Lt(rttp, sentp)) {
5716 return; /* somebody set the clock back, don't count this time. */
5718 clock_Sub(rttp, sentp);
5719 MUTEX_ENTER(&rx_stats_mutex);
5720 if (clock_Lt(rttp, &rx_stats.minRtt))
5721 rx_stats.minRtt = *rttp;
5722 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5723 if (rttp->sec > 60) {
5724 MUTEX_EXIT(&rx_stats_mutex);
5725 return; /* somebody set the clock ahead */
5727 rx_stats.maxRtt = *rttp;
5729 clock_Add(&rx_stats.totalRtt, rttp);
5730 rx_stats.nRttSamples++;
5731 MUTEX_EXIT(&rx_stats_mutex);
5733 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5735 /* Apply VanJacobson round-trip estimations */
5740 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5741 * srtt is stored as fixed point with 3 bits after the binary
5742 * point (i.e., scaled by 8). The following magic is
5743 * equivalent to the smoothing algorithm in rfc793 with an
5744 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5745 * srtt*8 = srtt*8 + rtt - srtt
5746 * srtt = srtt + rtt/8 - srtt/8
5749 delta = MSEC(rttp) - (peer->rtt >> 3);
5753 * We accumulate a smoothed rtt variance (actually, a smoothed
5754 * mean difference), then set the retransmit timer to smoothed
5755 * rtt + 4 times the smoothed variance (was 2x in van's original
5756 * paper, but 4x works better for me, and apparently for him as
5758 * rttvar is stored as
5759 * fixed point with 2 bits after the binary point (scaled by
5760 * 4). The following is equivalent to rfc793 smoothing with
5761 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5762 * replaces rfc793's wired-in beta.
5763 * dev*4 = dev*4 + (|actual - expected| - dev)
5769 delta -= (peer->rtt_dev >> 2);
5770 peer->rtt_dev += delta;
5772 /* I don't have a stored RTT so I start with this value. Since I'm
5773 * probably just starting a call, and will be pushing more data down
5774 * this, I expect congestion to increase rapidly. So I fudge a
5775 * little, and I set deviance to half the rtt. In practice,
5776 * deviance tends to approach something a little less than
5777 * half the smoothed rtt. */
5778 peer->rtt = (MSEC(rttp) << 3) + 8;
5779 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5781 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5782 * the other of these connections is usually in a user process, and can
5783 * be switched and/or swapped out. So on fast, reliable networks, the
5784 * timeout would otherwise be too short.
5786 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5787 clock_Zero(&(peer->timeout));
5788 clock_Addmsec(&(peer->timeout), rtt_timeout);
5790 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)));
5794 /* Find all server connections that have not been active for a long time, and
5797 rxi_ReapConnections(void)
5800 clock_GetTime(&now);
5802 /* Find server connection structures that haven't been used for
5803 * greater than rx_idleConnectionTime */
5805 struct rx_connection **conn_ptr, **conn_end;
5806 int i, havecalls = 0;
5807 MUTEX_ENTER(&rx_connHashTable_lock);
5808 for (conn_ptr = &rx_connHashTable[0], conn_end =
5809 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5811 struct rx_connection *conn, *next;
5812 struct rx_call *call;
5816 for (conn = *conn_ptr; conn; conn = next) {
5817 /* XXX -- Shouldn't the connection be locked? */
5820 for (i = 0; i < RX_MAXCALLS; i++) {
5821 call = conn->call[i];
5824 MUTEX_ENTER(&call->lock);
5825 #ifdef RX_ENABLE_LOCKS
5826 result = rxi_CheckCall(call, 1);
5827 #else /* RX_ENABLE_LOCKS */
5828 result = rxi_CheckCall(call);
5829 #endif /* RX_ENABLE_LOCKS */
5830 MUTEX_EXIT(&call->lock);
5832 /* If CheckCall freed the call, it might
5833 * have destroyed the connection as well,
5834 * which screws up the linked lists.
5840 if (conn->type == RX_SERVER_CONNECTION) {
5841 /* This only actually destroys the connection if
5842 * there are no outstanding calls */
5843 MUTEX_ENTER(&conn->conn_data_lock);
5844 if (!havecalls && !conn->refCount
5845 && ((conn->lastSendTime + rx_idleConnectionTime) <
5847 conn->refCount++; /* it will be decr in rx_DestroyConn */
5848 MUTEX_EXIT(&conn->conn_data_lock);
5849 #ifdef RX_ENABLE_LOCKS
5850 rxi_DestroyConnectionNoLock(conn);
5851 #else /* RX_ENABLE_LOCKS */
5852 rxi_DestroyConnection(conn);
5853 #endif /* RX_ENABLE_LOCKS */
5855 #ifdef RX_ENABLE_LOCKS
5857 MUTEX_EXIT(&conn->conn_data_lock);
5859 #endif /* RX_ENABLE_LOCKS */
5863 #ifdef RX_ENABLE_LOCKS
5864 while (rx_connCleanup_list) {
5865 struct rx_connection *conn;
5866 conn = rx_connCleanup_list;
5867 rx_connCleanup_list = rx_connCleanup_list->next;
5868 MUTEX_EXIT(&rx_connHashTable_lock);
5869 rxi_CleanupConnection(conn);
5870 MUTEX_ENTER(&rx_connHashTable_lock);
5872 MUTEX_EXIT(&rx_connHashTable_lock);
5873 #endif /* RX_ENABLE_LOCKS */
5876 /* Find any peer structures that haven't been used (haven't had an
5877 * associated connection) for greater than rx_idlePeerTime */
5879 struct rx_peer **peer_ptr, **peer_end;
5881 MUTEX_ENTER(&rx_rpc_stats);
5882 MUTEX_ENTER(&rx_peerHashTable_lock);
5883 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5884 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5886 struct rx_peer *peer, *next, *prev;
5887 for (prev = peer = *peer_ptr; peer; peer = next) {
5889 code = MUTEX_TRYENTER(&peer->peer_lock);
5890 if ((code) && (peer->refCount == 0)
5891 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5892 rx_interface_stat_p rpc_stat, nrpc_stat;
5894 MUTEX_EXIT(&peer->peer_lock);
5895 MUTEX_DESTROY(&peer->peer_lock);
5897 (&peer->rpcStats, rpc_stat, nrpc_stat,
5898 rx_interface_stat)) {
5899 unsigned int num_funcs;
5902 queue_Remove(&rpc_stat->queue_header);
5903 queue_Remove(&rpc_stat->all_peers);
5904 num_funcs = rpc_stat->stats[0].func_total;
5906 sizeof(rx_interface_stat_t) +
5907 rpc_stat->stats[0].func_total *
5908 sizeof(rx_function_entry_v1_t);
5910 rxi_Free(rpc_stat, space);
5911 rxi_rpc_peer_stat_cnt -= num_funcs;
5914 MUTEX_ENTER(&rx_stats_mutex);
5915 rx_stats.nPeerStructs--;
5916 MUTEX_EXIT(&rx_stats_mutex);
5917 if (peer == *peer_ptr) {
5924 MUTEX_EXIT(&peer->peer_lock);
5930 MUTEX_EXIT(&rx_peerHashTable_lock);
5931 MUTEX_EXIT(&rx_rpc_stats);
5934 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5935 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5936 * GC, just below. Really, we shouldn't have to keep moving packets from
5937 * one place to another, but instead ought to always know if we can
5938 * afford to hold onto a packet in its particular use. */
5939 MUTEX_ENTER(&rx_freePktQ_lock);
5940 if (rx_waitingForPackets) {
5941 rx_waitingForPackets = 0;
5942 #ifdef RX_ENABLE_LOCKS
5943 CV_BROADCAST(&rx_waitingForPackets_cv);
5945 osi_rxWakeup(&rx_waitingForPackets);
5948 MUTEX_EXIT(&rx_freePktQ_lock);
5950 now.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5951 rxevent_Post(&now, rxi_ReapConnections, 0, 0);
5955 /* rxs_Release - This isn't strictly necessary but, since the macro name from
5956 * rx.h is sort of strange this is better. This is called with a security
5957 * object before it is discarded. Each connection using a security object has
5958 * its own refcount to the object so it won't actually be freed until the last
5959 * connection is destroyed.
5961 * This is the only rxs module call. A hold could also be written but no one
5965 rxs_Release(struct rx_securityClass *aobj)
5967 return RXS_Close(aobj);
5971 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
5972 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
5973 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
5974 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
5976 /* Adjust our estimate of the transmission rate to this peer, given
5977 * that the packet p was just acked. We can adjust peer->timeout and
5978 * call->twind. Pragmatically, this is called
5979 * only with packets of maximal length.
5980 * Called with peer and call locked.
5984 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
5985 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
5987 afs_int32 xferSize, xferMs;
5988 register afs_int32 minTime;
5991 /* Count down packets */
5992 if (peer->rateFlag > 0)
5994 /* Do nothing until we're enabled */
5995 if (peer->rateFlag != 0)
6000 /* Count only when the ack seems legitimate */
6001 switch (ackReason) {
6002 case RX_ACK_REQUESTED:
6004 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6008 case RX_ACK_PING_RESPONSE:
6009 if (p) /* want the response to ping-request, not data send */
6011 clock_GetTime(&newTO);
6012 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6013 clock_Sub(&newTO, &call->pingRequestTime);
6014 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6018 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6025 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));
6027 /* Track only packets that are big enough. */
6028 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6032 /* absorb RTT data (in milliseconds) for these big packets */
6033 if (peer->smRtt == 0) {
6034 peer->smRtt = xferMs;
6036 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6041 if (peer->countDown) {
6045 peer->countDown = 10; /* recalculate only every so often */
6047 /* In practice, we can measure only the RTT for full packets,
6048 * because of the way Rx acks the data that it receives. (If it's
6049 * smaller than a full packet, it often gets implicitly acked
6050 * either by the call response (from a server) or by the next call
6051 * (from a client), and either case confuses transmission times
6052 * with processing times.) Therefore, replace the above
6053 * more-sophisticated processing with a simpler version, where the
6054 * smoothed RTT is kept for full-size packets, and the time to
6055 * transmit a windowful of full-size packets is simply RTT *
6056 * windowSize. Again, we take two steps:
6057 - ensure the timeout is large enough for a single packet's RTT;
6058 - ensure that the window is small enough to fit in the desired timeout.*/
6060 /* First, the timeout check. */
6061 minTime = peer->smRtt;
6062 /* Get a reasonable estimate for a timeout period */
6064 newTO.sec = minTime / 1000;
6065 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6067 /* Increase the timeout period so that we can always do at least
6068 * one packet exchange */
6069 if (clock_Gt(&newTO, &peer->timeout)) {
6071 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));
6073 peer->timeout = newTO;
6076 /* Now, get an estimate for the transmit window size. */
6077 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6078 /* Now, convert to the number of full packets that could fit in a
6079 * reasonable fraction of that interval */
6080 minTime /= (peer->smRtt << 1);
6081 xferSize = minTime; /* (make a copy) */
6083 /* Now clamp the size to reasonable bounds. */
6086 else if (minTime > rx_Window)
6087 minTime = rx_Window;
6088 /* if (minTime != peer->maxWindow) {
6089 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6090 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6091 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6093 peer->maxWindow = minTime;
6094 elide... call->twind = minTime;
6098 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6099 * Discern this by calculating the timeout necessary for rx_Window
6101 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6102 /* calculate estimate for transmission interval in milliseconds */
6103 minTime = rx_Window * peer->smRtt;
6104 if (minTime < 1000) {
6105 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
6106 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6107 peer->timeout.usec, peer->smRtt, peer->packetSize));
6109 newTO.sec = 0; /* cut back on timeout by half a second */
6110 newTO.usec = 500000;
6111 clock_Sub(&peer->timeout, &newTO);
6116 } /* end of rxi_ComputeRate */
6117 #endif /* ADAPT_WINDOW */
6125 #define TRACE_OPTION_DEBUGLOG 4
6133 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6134 0, KEY_QUERY_VALUE, &parmKey);
6135 if (code != ERROR_SUCCESS)
6138 dummyLen = sizeof(TraceOption);
6139 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6140 (BYTE *) &TraceOption, &dummyLen);
6141 if (code == ERROR_SUCCESS) {
6142 rxdebug_active = (TraceOption & TRACE_OPTION_DEBUGLOG) ? 1 : 0;
6144 RegCloseKey (parmKey);
6145 #endif /* AFS_NT40_ENV */
6150 rx_DebugOnOff(int on)
6152 rxdebug_active = on;
6154 #endif /* AFS_NT40_ENV */
6157 /* Don't call this debugging routine directly; use dpf */
6159 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
6160 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
6168 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6171 len = _snprintf(msg, sizeof(msg)-2,
6172 tformat, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10,
6173 a11, a12, a13, a14, a15);
6175 if (msg[len-1] != '\n') {
6179 OutputDebugString(msg);
6184 clock_GetTime(&now);
6185 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
6186 (unsigned int)now.usec / 1000);
6187 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
6194 * This function is used to process the rx_stats structure that is local
6195 * to a process as well as an rx_stats structure received from a remote
6196 * process (via rxdebug). Therefore, it needs to do minimal version
6200 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
6201 afs_int32 freePackets, char version)
6205 if (size != sizeof(struct rx_stats)) {
6207 "Unexpected size of stats structure: was %d, expected %d\n",
6208 size, sizeof(struct rx_stats));
6211 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6214 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6215 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6216 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6217 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6218 s->specialPktAllocFailures);
6220 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6221 s->receivePktAllocFailures, s->sendPktAllocFailures,
6222 s->specialPktAllocFailures);
6226 " greedy %d, " "bogusReads %d (last from host %x), "
6227 "noPackets %d, " "noBuffers %d, " "selects %d, "
6228 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6229 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6230 s->selects, s->sendSelects);
6232 fprintf(file, " packets read: ");
6233 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6234 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6236 fprintf(file, "\n");
6239 " other read counters: data %d, " "ack %d, " "dup %d "
6240 "spurious %d " "dally %d\n", s->dataPacketsRead,
6241 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6242 s->ignorePacketDally);
6244 fprintf(file, " packets sent: ");
6245 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6246 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6248 fprintf(file, "\n");
6251 " other send counters: ack %d, " "data %d (not resends), "
6252 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6253 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6254 s->dataPacketsPushed, s->ignoreAckedPacket);
6257 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6258 s->netSendFailures, (int)s->fatalErrors);
6260 if (s->nRttSamples) {
6261 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6262 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6264 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6265 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6269 " %d server connections, " "%d client connections, "
6270 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6271 s->nServerConns, s->nClientConns, s->nPeerStructs,
6272 s->nCallStructs, s->nFreeCallStructs);
6274 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6275 fprintf(file, " %d clock updates\n", clock_nUpdates);
6280 /* for backward compatibility */
6282 rx_PrintStats(FILE * file)
6284 MUTEX_ENTER(&rx_stats_mutex);
6285 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6287 MUTEX_EXIT(&rx_stats_mutex);
6291 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6293 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6294 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6295 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6298 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6299 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6300 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6303 " Packet size %d, " "max in packet skew %d, "
6304 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6305 (int)peer->outPacketSkew);
6308 #ifdef AFS_PTHREAD_ENV
6310 * This mutex protects the following static variables:
6314 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6315 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6317 #define LOCK_RX_DEBUG
6318 #define UNLOCK_RX_DEBUG
6319 #endif /* AFS_PTHREAD_ENV */
6322 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6323 u_char type, void *inputData, size_t inputLength,
6324 void *outputData, size_t outputLength)
6326 static afs_int32 counter = 100;
6327 time_t waitTime, waitCount, startTime, endTime;
6328 struct rx_header theader;
6330 register afs_int32 code;
6331 struct timeval tv_now, tv_wake, tv_delta;
6332 struct sockaddr_in taddr, faddr;
6337 startTime = time(0);
6343 tp = &tbuffer[sizeof(struct rx_header)];
6344 taddr.sin_family = AF_INET;
6345 taddr.sin_port = remotePort;
6346 taddr.sin_addr.s_addr = remoteAddr;
6347 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6348 taddr.sin_len = sizeof(struct sockaddr_in);
6351 memset(&theader, 0, sizeof(theader));
6352 theader.epoch = htonl(999);
6354 theader.callNumber = htonl(counter);
6357 theader.type = type;
6358 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6359 theader.serviceId = 0;
6361 memcpy(tbuffer, &theader, sizeof(theader));
6362 memcpy(tp, inputData, inputLength);
6364 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6365 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6367 /* see if there's a packet available */
6368 gettimeofday(&tv_wake,0);
6369 tv_wake.tv_sec += waitTime;
6372 FD_SET(socket, &imask);
6373 tv_delta.tv_sec = tv_wake.tv_sec;
6374 tv_delta.tv_usec = tv_wake.tv_usec;
6375 gettimeofday(&tv_now, 0);
6377 if (tv_delta.tv_usec < tv_now.tv_usec) {
6379 tv_delta.tv_usec += 1000000;
6382 tv_delta.tv_usec -= tv_now.tv_usec;
6384 if (tv_delta.tv_sec < tv_now.tv_sec) {
6388 tv_delta.tv_sec -= tv_now.tv_sec;
6390 code = select(socket + 1, &imask, 0, 0, &tv_delta);
6391 if (code == 1 && FD_ISSET(socket, &imask)) {
6392 /* now receive a packet */
6393 faddrLen = sizeof(struct sockaddr_in);
6395 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6396 (struct sockaddr *)&faddr, &faddrLen);
6399 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6400 if (counter == ntohl(theader.callNumber))
6408 /* see if we've timed out */
6416 code -= sizeof(struct rx_header);
6417 if (code > outputLength)
6418 code = outputLength;
6419 memcpy(outputData, tp, code);
6424 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6425 afs_uint16 remotePort, struct rx_debugStats * stat,
6426 afs_uint32 * supportedValues)
6428 struct rx_debugIn in;
6431 *supportedValues = 0;
6432 in.type = htonl(RX_DEBUGI_GETSTATS);
6435 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6436 &in, sizeof(in), stat, sizeof(*stat));
6439 * If the call was successful, fixup the version and indicate
6440 * what contents of the stat structure are valid.
6441 * Also do net to host conversion of fields here.
6445 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6446 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6448 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6449 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6451 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6452 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6454 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6455 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6457 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6458 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6460 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6461 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6463 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6464 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6466 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6467 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6470 stat->nFreePackets = ntohl(stat->nFreePackets);
6471 stat->packetReclaims = ntohl(stat->packetReclaims);
6472 stat->callsExecuted = ntohl(stat->callsExecuted);
6473 stat->nWaiting = ntohl(stat->nWaiting);
6474 stat->idleThreads = ntohl(stat->idleThreads);
6481 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6482 afs_uint16 remotePort, struct rx_stats * stat,
6483 afs_uint32 * supportedValues)
6485 struct rx_debugIn in;
6486 afs_int32 *lp = (afs_int32 *) stat;
6491 * supportedValues is currently unused, but added to allow future
6492 * versioning of this function.
6495 *supportedValues = 0;
6496 in.type = htonl(RX_DEBUGI_RXSTATS);
6498 memset(stat, 0, sizeof(*stat));
6500 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6501 &in, sizeof(in), stat, sizeof(*stat));
6506 * Do net to host conversion here
6509 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6518 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6519 afs_uint16 remotePort, size_t version_length,
6523 return MakeDebugCall(socket, remoteAddr, remotePort,
6524 RX_PACKET_TYPE_VERSION, a, 1, version,
6529 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6530 afs_uint16 remotePort, afs_int32 * nextConnection,
6531 int allConnections, afs_uint32 debugSupportedValues,
6532 struct rx_debugConn * conn,
6533 afs_uint32 * supportedValues)
6535 struct rx_debugIn in;
6540 * supportedValues is currently unused, but added to allow future
6541 * versioning of this function.
6544 *supportedValues = 0;
6545 if (allConnections) {
6546 in.type = htonl(RX_DEBUGI_GETALLCONN);
6548 in.type = htonl(RX_DEBUGI_GETCONN);
6550 in.index = htonl(*nextConnection);
6551 memset(conn, 0, sizeof(*conn));
6553 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6554 &in, sizeof(in), conn, sizeof(*conn));
6557 *nextConnection += 1;
6560 * Convert old connection format to new structure.
6563 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6564 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6565 #define MOVEvL(a) (conn->a = vL->a)
6567 /* any old or unrecognized version... */
6568 for (i = 0; i < RX_MAXCALLS; i++) {
6569 MOVEvL(callState[i]);
6570 MOVEvL(callMode[i]);
6571 MOVEvL(callFlags[i]);
6572 MOVEvL(callOther[i]);
6574 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6575 MOVEvL(secStats.type);
6576 MOVEvL(secStats.level);
6577 MOVEvL(secStats.flags);
6578 MOVEvL(secStats.expires);
6579 MOVEvL(secStats.packetsReceived);
6580 MOVEvL(secStats.packetsSent);
6581 MOVEvL(secStats.bytesReceived);
6582 MOVEvL(secStats.bytesSent);
6587 * Do net to host conversion here
6589 * I don't convert host or port since we are most likely
6590 * going to want these in NBO.
6592 conn->cid = ntohl(conn->cid);
6593 conn->serial = ntohl(conn->serial);
6594 for (i = 0; i < RX_MAXCALLS; i++) {
6595 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6597 conn->error = ntohl(conn->error);
6598 conn->secStats.flags = ntohl(conn->secStats.flags);
6599 conn->secStats.expires = ntohl(conn->secStats.expires);
6600 conn->secStats.packetsReceived =
6601 ntohl(conn->secStats.packetsReceived);
6602 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6603 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6604 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6605 conn->epoch = ntohl(conn->epoch);
6606 conn->natMTU = ntohl(conn->natMTU);
6613 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6614 afs_uint16 remotePort, afs_int32 * nextPeer,
6615 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6616 afs_uint32 * supportedValues)
6618 struct rx_debugIn in;
6622 * supportedValues is currently unused, but added to allow future
6623 * versioning of this function.
6626 *supportedValues = 0;
6627 in.type = htonl(RX_DEBUGI_GETPEER);
6628 in.index = htonl(*nextPeer);
6629 memset(peer, 0, sizeof(*peer));
6631 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6632 &in, sizeof(in), peer, sizeof(*peer));
6638 * Do net to host conversion here
6640 * I don't convert host or port since we are most likely
6641 * going to want these in NBO.
6643 peer->ifMTU = ntohs(peer->ifMTU);
6644 peer->idleWhen = ntohl(peer->idleWhen);
6645 peer->refCount = ntohs(peer->refCount);
6646 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6647 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6648 peer->rtt = ntohl(peer->rtt);
6649 peer->rtt_dev = ntohl(peer->rtt_dev);
6650 peer->timeout.sec = ntohl(peer->timeout.sec);
6651 peer->timeout.usec = ntohl(peer->timeout.usec);
6652 peer->nSent = ntohl(peer->nSent);
6653 peer->reSends = ntohl(peer->reSends);
6654 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6655 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6656 peer->rateFlag = ntohl(peer->rateFlag);
6657 peer->natMTU = ntohs(peer->natMTU);
6658 peer->maxMTU = ntohs(peer->maxMTU);
6659 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6660 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6661 peer->MTU = ntohs(peer->MTU);
6662 peer->cwind = ntohs(peer->cwind);
6663 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6664 peer->congestSeq = ntohs(peer->congestSeq);
6665 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6666 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6667 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6668 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6673 #endif /* RXDEBUG */
6678 struct rx_serverQueueEntry *np;
6681 register struct rx_call *call;
6682 register struct rx_serverQueueEntry *sq;
6686 if (rxinit_status == 1) {
6688 return; /* Already shutdown. */
6692 #ifndef AFS_PTHREAD_ENV
6693 FD_ZERO(&rx_selectMask);
6694 #endif /* AFS_PTHREAD_ENV */
6695 rxi_dataQuota = RX_MAX_QUOTA;
6696 #ifndef AFS_PTHREAD_ENV
6698 #endif /* AFS_PTHREAD_ENV */
6701 #ifndef AFS_PTHREAD_ENV
6702 #ifndef AFS_USE_GETTIMEOFDAY
6704 #endif /* AFS_USE_GETTIMEOFDAY */
6705 #endif /* AFS_PTHREAD_ENV */
6707 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6708 call = queue_First(&rx_freeCallQueue, rx_call);
6710 rxi_Free(call, sizeof(struct rx_call));
6713 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6714 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6720 struct rx_peer **peer_ptr, **peer_end;
6721 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6722 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6724 struct rx_peer *peer, *next;
6725 for (peer = *peer_ptr; peer; peer = next) {
6726 rx_interface_stat_p rpc_stat, nrpc_stat;
6729 (&peer->rpcStats, rpc_stat, nrpc_stat,
6730 rx_interface_stat)) {
6731 unsigned int num_funcs;
6734 queue_Remove(&rpc_stat->queue_header);
6735 queue_Remove(&rpc_stat->all_peers);
6736 num_funcs = rpc_stat->stats[0].func_total;
6738 sizeof(rx_interface_stat_t) +
6739 rpc_stat->stats[0].func_total *
6740 sizeof(rx_function_entry_v1_t);
6742 rxi_Free(rpc_stat, space);
6743 MUTEX_ENTER(&rx_rpc_stats);
6744 rxi_rpc_peer_stat_cnt -= num_funcs;
6745 MUTEX_EXIT(&rx_rpc_stats);
6749 MUTEX_ENTER(&rx_stats_mutex);
6750 rx_stats.nPeerStructs--;
6751 MUTEX_EXIT(&rx_stats_mutex);
6755 for (i = 0; i < RX_MAX_SERVICES; i++) {
6757 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6759 for (i = 0; i < rx_hashTableSize; i++) {
6760 register struct rx_connection *tc, *ntc;
6761 MUTEX_ENTER(&rx_connHashTable_lock);
6762 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6764 for (j = 0; j < RX_MAXCALLS; j++) {
6766 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6769 rxi_Free(tc, sizeof(*tc));
6771 MUTEX_EXIT(&rx_connHashTable_lock);
6774 MUTEX_ENTER(&freeSQEList_lock);
6776 while ((np = rx_FreeSQEList)) {
6777 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6778 MUTEX_DESTROY(&np->lock);
6779 rxi_Free(np, sizeof(*np));
6782 MUTEX_EXIT(&freeSQEList_lock);
6783 MUTEX_DESTROY(&freeSQEList_lock);
6784 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6785 MUTEX_DESTROY(&rx_connHashTable_lock);
6786 MUTEX_DESTROY(&rx_peerHashTable_lock);
6787 MUTEX_DESTROY(&rx_serverPool_lock);
6789 osi_Free(rx_connHashTable,
6790 rx_hashTableSize * sizeof(struct rx_connection *));
6791 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6793 UNPIN(rx_connHashTable,
6794 rx_hashTableSize * sizeof(struct rx_connection *));
6795 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6797 rxi_FreeAllPackets();
6799 MUTEX_ENTER(&rx_stats_mutex);
6800 rxi_dataQuota = RX_MAX_QUOTA;
6801 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6802 MUTEX_EXIT(&rx_stats_mutex);
6808 #ifdef RX_ENABLE_LOCKS
6810 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6812 if (!MUTEX_ISMINE(lockaddr))
6813 osi_Panic("Lock not held: %s", msg);
6815 #endif /* RX_ENABLE_LOCKS */
6820 * Routines to implement connection specific data.
6824 rx_KeyCreate(rx_destructor_t rtn)
6827 MUTEX_ENTER(&rxi_keyCreate_lock);
6828 key = rxi_keyCreate_counter++;
6829 rxi_keyCreate_destructor = (rx_destructor_t *)
6830 realloc((void *)rxi_keyCreate_destructor,
6831 (key + 1) * sizeof(rx_destructor_t));
6832 rxi_keyCreate_destructor[key] = rtn;
6833 MUTEX_EXIT(&rxi_keyCreate_lock);
6838 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6841 MUTEX_ENTER(&conn->conn_data_lock);
6842 if (!conn->specific) {
6843 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6844 for (i = 0; i < key; i++)
6845 conn->specific[i] = NULL;
6846 conn->nSpecific = key + 1;
6847 conn->specific[key] = ptr;
6848 } else if (key >= conn->nSpecific) {
6849 conn->specific = (void **)
6850 realloc(conn->specific, (key + 1) * sizeof(void *));
6851 for (i = conn->nSpecific; i < key; i++)
6852 conn->specific[i] = NULL;
6853 conn->nSpecific = key + 1;
6854 conn->specific[key] = ptr;
6856 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6857 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6858 conn->specific[key] = ptr;
6860 MUTEX_EXIT(&conn->conn_data_lock);
6864 rx_GetSpecific(struct rx_connection *conn, int key)
6867 MUTEX_ENTER(&conn->conn_data_lock);
6868 if (key >= conn->nSpecific)
6871 ptr = conn->specific[key];
6872 MUTEX_EXIT(&conn->conn_data_lock);
6876 #endif /* !KERNEL */
6879 * processStats is a queue used to store the statistics for the local
6880 * process. Its contents are similar to the contents of the rpcStats
6881 * queue on a rx_peer structure, but the actual data stored within
6882 * this queue contains totals across the lifetime of the process (assuming
6883 * the stats have not been reset) - unlike the per peer structures
6884 * which can come and go based upon the peer lifetime.
6887 static struct rx_queue processStats = { &processStats, &processStats };
6890 * peerStats is a queue used to store the statistics for all peer structs.
6891 * Its contents are the union of all the peer rpcStats queues.
6894 static struct rx_queue peerStats = { &peerStats, &peerStats };
6897 * rxi_monitor_processStats is used to turn process wide stat collection
6901 static int rxi_monitor_processStats = 0;
6904 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6907 static int rxi_monitor_peerStats = 0;
6910 * rxi_AddRpcStat - given all of the information for a particular rpc
6911 * call, create (if needed) and update the stat totals for the rpc.
6915 * IN stats - the queue of stats that will be updated with the new value
6917 * IN rxInterface - a unique number that identifies the rpc interface
6919 * IN currentFunc - the index of the function being invoked
6921 * IN totalFunc - the total number of functions in this interface
6923 * IN queueTime - the amount of time this function waited for a thread
6925 * IN execTime - the amount of time this function invocation took to execute
6927 * IN bytesSent - the number bytes sent by this invocation
6929 * IN bytesRcvd - the number bytes received by this invocation
6931 * IN isServer - if true, this invocation was made to a server
6933 * IN remoteHost - the ip address of the remote host
6935 * IN remotePort - the port of the remote host
6937 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6939 * INOUT counter - if a new stats structure is allocated, the counter will
6940 * be updated with the new number of allocated stat structures
6948 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6949 afs_uint32 currentFunc, afs_uint32 totalFunc,
6950 struct clock *queueTime, struct clock *execTime,
6951 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6952 afs_uint32 remoteHost, afs_uint32 remotePort,
6953 int addToPeerList, unsigned int *counter)
6956 rx_interface_stat_p rpc_stat, nrpc_stat;
6959 * See if there's already a structure for this interface
6962 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6963 if ((rpc_stat->stats[0].interfaceId == rxInterface)
6964 && (rpc_stat->stats[0].remote_is_server == isServer))
6969 * Didn't find a match so allocate a new structure and add it to the
6973 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
6974 || (rpc_stat->stats[0].interfaceId != rxInterface)
6975 || (rpc_stat->stats[0].remote_is_server != isServer)) {
6980 sizeof(rx_interface_stat_t) +
6981 totalFunc * sizeof(rx_function_entry_v1_t);
6983 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
6984 if (rpc_stat == NULL) {
6988 *counter += totalFunc;
6989 for (i = 0; i < totalFunc; i++) {
6990 rpc_stat->stats[i].remote_peer = remoteHost;
6991 rpc_stat->stats[i].remote_port = remotePort;
6992 rpc_stat->stats[i].remote_is_server = isServer;
6993 rpc_stat->stats[i].interfaceId = rxInterface;
6994 rpc_stat->stats[i].func_total = totalFunc;
6995 rpc_stat->stats[i].func_index = i;
6996 hzero(rpc_stat->stats[i].invocations);
6997 hzero(rpc_stat->stats[i].bytes_sent);
6998 hzero(rpc_stat->stats[i].bytes_rcvd);
6999 rpc_stat->stats[i].queue_time_sum.sec = 0;
7000 rpc_stat->stats[i].queue_time_sum.usec = 0;
7001 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7002 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7003 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7004 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7005 rpc_stat->stats[i].queue_time_max.sec = 0;
7006 rpc_stat->stats[i].queue_time_max.usec = 0;
7007 rpc_stat->stats[i].execution_time_sum.sec = 0;
7008 rpc_stat->stats[i].execution_time_sum.usec = 0;
7009 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7010 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7011 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7012 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7013 rpc_stat->stats[i].execution_time_max.sec = 0;
7014 rpc_stat->stats[i].execution_time_max.usec = 0;
7016 queue_Prepend(stats, rpc_stat);
7017 if (addToPeerList) {
7018 queue_Prepend(&peerStats, &rpc_stat->all_peers);
7023 * Increment the stats for this function
7026 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
7027 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
7028 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
7029 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
7030 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7031 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7032 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7034 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7035 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7037 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7038 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7040 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7041 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7043 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7044 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7052 * rx_IncrementTimeAndCount - increment the times and count for a particular
7057 * IN peer - the peer who invoked the rpc
7059 * IN rxInterface - a unique number that identifies the rpc interface
7061 * IN currentFunc - the index of the function being invoked
7063 * IN totalFunc - the total number of functions in this interface
7065 * IN queueTime - the amount of time this function waited for a thread
7067 * IN execTime - the amount of time this function invocation took to execute
7069 * IN bytesSent - the number bytes sent by this invocation
7071 * IN bytesRcvd - the number bytes received by this invocation
7073 * IN isServer - if true, this invocation was made to a server
7081 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7082 afs_uint32 currentFunc, afs_uint32 totalFunc,
7083 struct clock *queueTime, struct clock *execTime,
7084 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7088 MUTEX_ENTER(&rx_rpc_stats);
7089 MUTEX_ENTER(&peer->peer_lock);
7091 if (rxi_monitor_peerStats) {
7092 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7093 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7094 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7097 if (rxi_monitor_processStats) {
7098 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7099 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7100 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7103 MUTEX_EXIT(&peer->peer_lock);
7104 MUTEX_EXIT(&rx_rpc_stats);
7109 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7113 * IN callerVersion - the rpc stat version of the caller.
7115 * IN count - the number of entries to marshall.
7117 * IN stats - pointer to stats to be marshalled.
7119 * OUT ptr - Where to store the marshalled data.
7126 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7127 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7133 * We only support the first version
7135 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7136 *(ptr++) = stats->remote_peer;
7137 *(ptr++) = stats->remote_port;
7138 *(ptr++) = stats->remote_is_server;
7139 *(ptr++) = stats->interfaceId;
7140 *(ptr++) = stats->func_total;
7141 *(ptr++) = stats->func_index;
7142 *(ptr++) = hgethi(stats->invocations);
7143 *(ptr++) = hgetlo(stats->invocations);
7144 *(ptr++) = hgethi(stats->bytes_sent);
7145 *(ptr++) = hgetlo(stats->bytes_sent);
7146 *(ptr++) = hgethi(stats->bytes_rcvd);
7147 *(ptr++) = hgetlo(stats->bytes_rcvd);
7148 *(ptr++) = stats->queue_time_sum.sec;
7149 *(ptr++) = stats->queue_time_sum.usec;
7150 *(ptr++) = stats->queue_time_sum_sqr.sec;
7151 *(ptr++) = stats->queue_time_sum_sqr.usec;
7152 *(ptr++) = stats->queue_time_min.sec;
7153 *(ptr++) = stats->queue_time_min.usec;
7154 *(ptr++) = stats->queue_time_max.sec;
7155 *(ptr++) = stats->queue_time_max.usec;
7156 *(ptr++) = stats->execution_time_sum.sec;
7157 *(ptr++) = stats->execution_time_sum.usec;
7158 *(ptr++) = stats->execution_time_sum_sqr.sec;
7159 *(ptr++) = stats->execution_time_sum_sqr.usec;
7160 *(ptr++) = stats->execution_time_min.sec;
7161 *(ptr++) = stats->execution_time_min.usec;
7162 *(ptr++) = stats->execution_time_max.sec;
7163 *(ptr++) = stats->execution_time_max.usec;
7169 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7174 * IN callerVersion - the rpc stat version of the caller
7176 * OUT myVersion - the rpc stat version of this function
7178 * OUT clock_sec - local time seconds
7180 * OUT clock_usec - local time microseconds
7182 * OUT allocSize - the number of bytes allocated to contain stats
7184 * OUT statCount - the number stats retrieved from this process.
7186 * OUT stats - the actual stats retrieved from this process.
7190 * Returns void. If successful, stats will != NULL.
7194 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7195 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7196 size_t * allocSize, afs_uint32 * statCount,
7197 afs_uint32 ** stats)
7207 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7210 * Check to see if stats are enabled
7213 MUTEX_ENTER(&rx_rpc_stats);
7214 if (!rxi_monitor_processStats) {
7215 MUTEX_EXIT(&rx_rpc_stats);
7219 clock_GetTime(&now);
7220 *clock_sec = now.sec;
7221 *clock_usec = now.usec;
7224 * Allocate the space based upon the caller version
7226 * If the client is at an older version than we are,
7227 * we return the statistic data in the older data format, but
7228 * we still return our version number so the client knows we
7229 * are maintaining more data than it can retrieve.
7232 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7233 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7234 *statCount = rxi_rpc_process_stat_cnt;
7237 * This can't happen yet, but in the future version changes
7238 * can be handled by adding additional code here
7242 if (space > (size_t) 0) {
7244 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7247 rx_interface_stat_p rpc_stat, nrpc_stat;
7251 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7253 * Copy the data based upon the caller version
7255 rx_MarshallProcessRPCStats(callerVersion,
7256 rpc_stat->stats[0].func_total,
7257 rpc_stat->stats, &ptr);
7263 MUTEX_EXIT(&rx_rpc_stats);
7268 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7272 * IN callerVersion - the rpc stat version of the caller
7274 * OUT myVersion - the rpc stat version of this function
7276 * OUT clock_sec - local time seconds
7278 * OUT clock_usec - local time microseconds
7280 * OUT allocSize - the number of bytes allocated to contain stats
7282 * OUT statCount - the number of stats retrieved from the individual
7285 * OUT stats - the actual stats retrieved from the individual peer structures.
7289 * Returns void. If successful, stats will != NULL.
7293 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7294 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7295 size_t * allocSize, afs_uint32 * statCount,
7296 afs_uint32 ** stats)
7306 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7309 * Check to see if stats are enabled
7312 MUTEX_ENTER(&rx_rpc_stats);
7313 if (!rxi_monitor_peerStats) {
7314 MUTEX_EXIT(&rx_rpc_stats);
7318 clock_GetTime(&now);
7319 *clock_sec = now.sec;
7320 *clock_usec = now.usec;
7323 * Allocate the space based upon the caller version
7325 * If the client is at an older version than we are,
7326 * we return the statistic data in the older data format, but
7327 * we still return our version number so the client knows we
7328 * are maintaining more data than it can retrieve.
7331 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7332 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7333 *statCount = rxi_rpc_peer_stat_cnt;
7336 * This can't happen yet, but in the future version changes
7337 * can be handled by adding additional code here
7341 if (space > (size_t) 0) {
7343 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7346 rx_interface_stat_p rpc_stat, nrpc_stat;
7350 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7352 * We have to fix the offset of rpc_stat since we are
7353 * keeping this structure on two rx_queues. The rx_queue
7354 * package assumes that the rx_queue member is the first
7355 * member of the structure. That is, rx_queue assumes that
7356 * any one item is only on one queue at a time. We are
7357 * breaking that assumption and so we have to do a little
7358 * math to fix our pointers.
7361 fix_offset = (char *)rpc_stat;
7362 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7363 rpc_stat = (rx_interface_stat_p) fix_offset;
7366 * Copy the data based upon the caller version
7368 rx_MarshallProcessRPCStats(callerVersion,
7369 rpc_stat->stats[0].func_total,
7370 rpc_stat->stats, &ptr);
7376 MUTEX_EXIT(&rx_rpc_stats);
7381 * rx_FreeRPCStats - free memory allocated by
7382 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7386 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7387 * rx_RetrievePeerRPCStats
7389 * IN allocSize - the number of bytes in stats.
7397 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7399 rxi_Free(stats, allocSize);
7403 * rx_queryProcessRPCStats - see if process rpc stat collection is
7404 * currently enabled.
7410 * Returns 0 if stats are not enabled != 0 otherwise
7414 rx_queryProcessRPCStats(void)
7417 MUTEX_ENTER(&rx_rpc_stats);
7418 rc = rxi_monitor_processStats;
7419 MUTEX_EXIT(&rx_rpc_stats);
7424 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7430 * Returns 0 if stats are not enabled != 0 otherwise
7434 rx_queryPeerRPCStats(void)
7437 MUTEX_ENTER(&rx_rpc_stats);
7438 rc = rxi_monitor_peerStats;
7439 MUTEX_EXIT(&rx_rpc_stats);
7444 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7454 rx_enableProcessRPCStats(void)
7456 MUTEX_ENTER(&rx_rpc_stats);
7457 rx_enable_stats = 1;
7458 rxi_monitor_processStats = 1;
7459 MUTEX_EXIT(&rx_rpc_stats);
7463 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7473 rx_enablePeerRPCStats(void)
7475 MUTEX_ENTER(&rx_rpc_stats);
7476 rx_enable_stats = 1;
7477 rxi_monitor_peerStats = 1;
7478 MUTEX_EXIT(&rx_rpc_stats);
7482 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7492 rx_disableProcessRPCStats(void)
7494 rx_interface_stat_p rpc_stat, nrpc_stat;
7497 MUTEX_ENTER(&rx_rpc_stats);
7500 * Turn off process statistics and if peer stats is also off, turn
7504 rxi_monitor_processStats = 0;
7505 if (rxi_monitor_peerStats == 0) {
7506 rx_enable_stats = 0;
7509 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7510 unsigned int num_funcs = 0;
7513 queue_Remove(rpc_stat);
7514 num_funcs = rpc_stat->stats[0].func_total;
7516 sizeof(rx_interface_stat_t) +
7517 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7519 rxi_Free(rpc_stat, space);
7520 rxi_rpc_process_stat_cnt -= num_funcs;
7522 MUTEX_EXIT(&rx_rpc_stats);
7526 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7536 rx_disablePeerRPCStats(void)
7538 struct rx_peer **peer_ptr, **peer_end;
7541 MUTEX_ENTER(&rx_rpc_stats);
7544 * Turn off peer statistics and if process stats is also off, turn
7548 rxi_monitor_peerStats = 0;
7549 if (rxi_monitor_processStats == 0) {
7550 rx_enable_stats = 0;
7553 MUTEX_ENTER(&rx_peerHashTable_lock);
7554 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7555 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7557 struct rx_peer *peer, *next, *prev;
7558 for (prev = peer = *peer_ptr; peer; peer = next) {
7560 code = MUTEX_TRYENTER(&peer->peer_lock);
7562 rx_interface_stat_p rpc_stat, nrpc_stat;
7565 (&peer->rpcStats, rpc_stat, nrpc_stat,
7566 rx_interface_stat)) {
7567 unsigned int num_funcs = 0;
7570 queue_Remove(&rpc_stat->queue_header);
7571 queue_Remove(&rpc_stat->all_peers);
7572 num_funcs = rpc_stat->stats[0].func_total;
7574 sizeof(rx_interface_stat_t) +
7575 rpc_stat->stats[0].func_total *
7576 sizeof(rx_function_entry_v1_t);
7578 rxi_Free(rpc_stat, space);
7579 rxi_rpc_peer_stat_cnt -= num_funcs;
7581 MUTEX_EXIT(&peer->peer_lock);
7582 if (prev == *peer_ptr) {
7592 MUTEX_EXIT(&rx_peerHashTable_lock);
7593 MUTEX_EXIT(&rx_rpc_stats);
7597 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7602 * IN clearFlag - flag indicating which stats to clear
7610 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7612 rx_interface_stat_p rpc_stat, nrpc_stat;
7614 MUTEX_ENTER(&rx_rpc_stats);
7616 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7617 unsigned int num_funcs = 0, i;
7618 num_funcs = rpc_stat->stats[0].func_total;
7619 for (i = 0; i < num_funcs; i++) {
7620 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7621 hzero(rpc_stat->stats[i].invocations);
7623 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7624 hzero(rpc_stat->stats[i].bytes_sent);
7626 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7627 hzero(rpc_stat->stats[i].bytes_rcvd);
7629 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7630 rpc_stat->stats[i].queue_time_sum.sec = 0;
7631 rpc_stat->stats[i].queue_time_sum.usec = 0;
7633 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7634 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7635 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7637 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7638 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7639 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7641 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7642 rpc_stat->stats[i].queue_time_max.sec = 0;
7643 rpc_stat->stats[i].queue_time_max.usec = 0;
7645 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7646 rpc_stat->stats[i].execution_time_sum.sec = 0;
7647 rpc_stat->stats[i].execution_time_sum.usec = 0;
7649 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7650 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7651 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7653 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7654 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7655 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7657 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7658 rpc_stat->stats[i].execution_time_max.sec = 0;
7659 rpc_stat->stats[i].execution_time_max.usec = 0;
7664 MUTEX_EXIT(&rx_rpc_stats);
7668 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7673 * IN clearFlag - flag indicating which stats to clear
7681 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7683 rx_interface_stat_p rpc_stat, nrpc_stat;
7685 MUTEX_ENTER(&rx_rpc_stats);
7687 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7688 unsigned int num_funcs = 0, i;
7691 * We have to fix the offset of rpc_stat since we are
7692 * keeping this structure on two rx_queues. The rx_queue
7693 * package assumes that the rx_queue member is the first
7694 * member of the structure. That is, rx_queue assumes that
7695 * any one item is only on one queue at a time. We are
7696 * breaking that assumption and so we have to do a little
7697 * math to fix our pointers.
7700 fix_offset = (char *)rpc_stat;
7701 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7702 rpc_stat = (rx_interface_stat_p) fix_offset;
7704 num_funcs = rpc_stat->stats[0].func_total;
7705 for (i = 0; i < num_funcs; i++) {
7706 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7707 hzero(rpc_stat->stats[i].invocations);
7709 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7710 hzero(rpc_stat->stats[i].bytes_sent);
7712 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7713 hzero(rpc_stat->stats[i].bytes_rcvd);
7715 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7716 rpc_stat->stats[i].queue_time_sum.sec = 0;
7717 rpc_stat->stats[i].queue_time_sum.usec = 0;
7719 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7720 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7721 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7723 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7724 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7725 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7727 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7728 rpc_stat->stats[i].queue_time_max.sec = 0;
7729 rpc_stat->stats[i].queue_time_max.usec = 0;
7731 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7732 rpc_stat->stats[i].execution_time_sum.sec = 0;
7733 rpc_stat->stats[i].execution_time_sum.usec = 0;
7735 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7736 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7737 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7739 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7740 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7741 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7743 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7744 rpc_stat->stats[i].execution_time_max.sec = 0;
7745 rpc_stat->stats[i].execution_time_max.usec = 0;
7750 MUTEX_EXIT(&rx_rpc_stats);
7754 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7755 * is authorized to enable/disable/clear RX statistics.
7757 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7760 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7762 rxi_rxstat_userok = proc;
7766 rx_RxStatUserOk(struct rx_call *call)
7768 if (!rxi_rxstat_userok)
7770 return rxi_rxstat_userok(call);
7775 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
7776 * function in the MSVC runtime DLL (msvcrt.dll).
7778 * Note: the system serializes calls to this function.
7781 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
7782 DWORD reason, /* reason function is being called */
7783 LPVOID reserved) /* reserved for future use */
7786 case DLL_PROCESS_ATTACH:
7787 /* library is being attached to a process */
7791 case DLL_PROCESS_DETACH: