2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
14 #include "afs/param.h"
16 #include <afs/param.h>
23 #include "afs/sysincludes.h"
24 #include "afsincludes.h"
30 #include <net/net_globals.h>
31 #endif /* AFS_OSF_ENV */
32 #ifdef AFS_LINUX20_ENV
35 #include "netinet/in.h"
36 #include "afs/afs_args.h"
37 #include "afs/afs_osi.h"
38 #ifdef RX_KERNEL_TRACE
39 #include "rx_kcommon.h"
41 #if (defined(AFS_AUX_ENV) || defined(AFS_AIX_ENV))
45 #undef RXDEBUG /* turn off debugging */
47 #if defined(AFS_SGI_ENV)
48 #include "sys/debug.h"
57 #endif /* AFS_OSF_ENV */
59 #include "afs/sysincludes.h"
60 #include "afsincludes.h"
63 #include "rx_kmutex.h"
64 #include "rx_kernel.h"
68 #include "rx_globals.h"
70 #define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
71 #define AFSOP_STOP_AFS 211 /* Stop AFS process */
72 #define AFSOP_STOP_BKG 212 /* Stop BKG process */
74 extern afs_int32 afs_termState;
76 #include "sys/lockl.h"
77 #include "sys/lock_def.h"
78 #endif /* AFS_AIX41_ENV */
79 # include "rxgen_consts.h"
81 # include <sys/types.h>
86 # include <afs/afsutil.h>
87 # include <WINNT\afsreg.h>
89 # include <sys/socket.h>
90 # include <sys/file.h>
92 # include <sys/stat.h>
93 # include <netinet/in.h>
94 # include <sys/time.h>
104 # include "rx_user.h"
105 # include "rx_clock.h"
106 # include "rx_queue.h"
107 # include "rx_globals.h"
108 # include "rx_trace.h"
109 # include <afs/rxgen_consts.h>
112 int (*registerProgram) () = 0;
113 int (*swapNameProgram) () = 0;
115 /* Local static routines */
116 static void rxi_DestroyConnectionNoLock(register struct rx_connection *conn);
117 #ifdef RX_ENABLE_LOCKS
118 static void rxi_SetAcksInTransmitQueue(register struct rx_call *call);
121 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
123 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
124 afs_int32 rxi_start_in_error;
126 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
129 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
130 * currently allocated within rx. This number is used to allocate the
131 * memory required to return the statistics when queried.
134 static unsigned int rxi_rpc_peer_stat_cnt;
137 * rxi_rpc_process_stat_cnt counts the total number of local process stat
138 * structures currently allocated within rx. The number is used to allocate
139 * the memory required to return the statistics when queried.
142 static unsigned int rxi_rpc_process_stat_cnt;
144 #if !defined(offsetof)
145 #include <stddef.h> /* for definition of offsetof() */
148 #ifdef AFS_PTHREAD_ENV
152 * Use procedural initialization of mutexes/condition variables
156 extern pthread_mutex_t rx_stats_mutex;
157 extern pthread_mutex_t des_init_mutex;
158 extern pthread_mutex_t des_random_mutex;
159 extern pthread_mutex_t rx_clock_mutex;
160 extern pthread_mutex_t rxi_connCacheMutex;
161 extern pthread_mutex_t rx_event_mutex;
162 extern pthread_mutex_t osi_malloc_mutex;
163 extern pthread_mutex_t event_handler_mutex;
164 extern pthread_mutex_t listener_mutex;
165 extern pthread_mutex_t rx_if_init_mutex;
166 extern pthread_mutex_t rx_if_mutex;
167 extern pthread_mutex_t rxkad_client_uid_mutex;
168 extern pthread_mutex_t rxkad_random_mutex;
170 extern pthread_cond_t rx_event_handler_cond;
171 extern pthread_cond_t rx_listener_cond;
173 static pthread_mutex_t epoch_mutex;
174 static pthread_mutex_t rx_init_mutex;
175 static pthread_mutex_t rx_debug_mutex;
178 rxi_InitPthread(void)
180 assert(pthread_mutex_init(&rx_clock_mutex, (const pthread_mutexattr_t *)0)
182 assert(pthread_mutex_init(&rx_stats_mutex, (const pthread_mutexattr_t *)0)
184 assert(pthread_mutex_init
185 (&rxi_connCacheMutex, (const pthread_mutexattr_t *)0) == 0);
186 assert(pthread_mutex_init(&rx_init_mutex, (const pthread_mutexattr_t *)0)
188 assert(pthread_mutex_init(&epoch_mutex, (const pthread_mutexattr_t *)0) ==
190 assert(pthread_mutex_init(&rx_event_mutex, (const pthread_mutexattr_t *)0)
192 assert(pthread_mutex_init(&des_init_mutex, (const pthread_mutexattr_t *)0)
194 assert(pthread_mutex_init
195 (&des_random_mutex, (const pthread_mutexattr_t *)0) == 0);
196 assert(pthread_mutex_init
197 (&osi_malloc_mutex, (const pthread_mutexattr_t *)0) == 0);
198 assert(pthread_mutex_init
199 (&event_handler_mutex, (const pthread_mutexattr_t *)0) == 0);
200 assert(pthread_mutex_init(&listener_mutex, (const pthread_mutexattr_t *)0)
202 assert(pthread_mutex_init
203 (&rx_if_init_mutex, (const pthread_mutexattr_t *)0) == 0);
204 assert(pthread_mutex_init(&rx_if_mutex, (const pthread_mutexattr_t *)0) ==
206 assert(pthread_mutex_init
207 (&rxkad_client_uid_mutex, (const pthread_mutexattr_t *)0) == 0);
208 assert(pthread_mutex_init
209 (&rxkad_random_mutex, (const pthread_mutexattr_t *)0) == 0);
210 assert(pthread_mutex_init(&rx_debug_mutex, (const pthread_mutexattr_t *)0)
213 assert(pthread_cond_init
214 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
215 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
217 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
218 assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
220 rxkad_global_stats_init();
223 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
224 #define INIT_PTHREAD_LOCKS \
225 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
227 * The rx_stats_mutex mutex protects the following global variables:
232 * rxi_lowConnRefCount
233 * rxi_lowPeerRefCount
242 #define INIT_PTHREAD_LOCKS
246 /* Variables for handling the minProcs implementation. availProcs gives the
247 * number of threads available in the pool at this moment (not counting dudes
248 * executing right now). totalMin gives the total number of procs required
249 * for handling all minProcs requests. minDeficit is a dynamic variable
250 * tracking the # of procs required to satisfy all of the remaining minProcs
252 * For fine grain locking to work, the quota check and the reservation of
253 * a server thread has to come while rxi_availProcs and rxi_minDeficit
254 * are locked. To this end, the code has been modified under #ifdef
255 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
256 * same time. A new function, ReturnToServerPool() returns the allocation.
258 * A call can be on several queue's (but only one at a time). When
259 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
260 * that no one else is touching the queue. To this end, we store the address
261 * of the queue lock in the call structure (under the call lock) when we
262 * put the call on a queue, and we clear the call_queue_lock when the
263 * call is removed from a queue (once the call lock has been obtained).
264 * This allows rxi_ResetCall to safely synchronize with others wishing
265 * to manipulate the queue.
268 #ifdef RX_ENABLE_LOCKS
269 static afs_kmutex_t rx_rpc_stats;
270 void rxi_StartUnlocked();
273 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
274 ** pretty good that the next packet coming in is from the same connection
275 ** as the last packet, since we're send multiple packets in a transmit window.
277 struct rx_connection *rxLastConn = 0;
279 #ifdef RX_ENABLE_LOCKS
280 /* The locking hierarchy for rx fine grain locking is composed of these
283 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
284 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
285 * call->lock - locks call data fields.
286 * These are independent of each other:
287 * rx_freeCallQueue_lock
292 * serverQueueEntry->lock
294 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
295 * peer->lock - locks peer data fields.
296 * conn_data_lock - that more than one thread is not updating a conn data
297 * field at the same time.
305 * Do we need a lock to protect the peer field in the conn structure?
306 * conn->peer was previously a constant for all intents and so has no
307 * lock protecting this field. The multihomed client delta introduced
308 * a RX code change : change the peer field in the connection structure
309 * to that remote inetrface from which the last packet for this
310 * connection was sent out. This may become an issue if further changes
313 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
314 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
316 /* rxdb_fileID is used to identify the lock location, along with line#. */
317 static int rxdb_fileID = RXDB_FILE_RX;
318 #endif /* RX_LOCKS_DB */
319 #else /* RX_ENABLE_LOCKS */
320 #define SET_CALL_QUEUE_LOCK(C, L)
321 #define CLEAR_CALL_QUEUE_LOCK(C)
322 #endif /* RX_ENABLE_LOCKS */
323 struct rx_serverQueueEntry *rx_waitForPacket = 0;
324 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
326 /* ------------Exported Interfaces------------- */
328 /* This function allows rxkad to set the epoch to a suitably random number
329 * which rx_NewConnection will use in the future. The principle purpose is to
330 * get rxnull connections to use the same epoch as the rxkad connections do, at
331 * least once the first rxkad connection is established. This is important now
332 * that the host/port addresses aren't used in FindConnection: the uniqueness
333 * of epoch/cid matters and the start time won't do. */
335 #ifdef AFS_PTHREAD_ENV
337 * This mutex protects the following global variables:
341 #define LOCK_EPOCH assert(pthread_mutex_lock(&epoch_mutex)==0)
342 #define UNLOCK_EPOCH assert(pthread_mutex_unlock(&epoch_mutex)==0)
346 #endif /* AFS_PTHREAD_ENV */
349 rx_SetEpoch(afs_uint32 epoch)
356 /* Initialize rx. A port number may be mentioned, in which case this
357 * becomes the default port number for any service installed later.
358 * If 0 is provided for the port number, a random port will be chosen
359 * by the kernel. Whether this will ever overlap anything in
360 * /etc/services is anybody's guess... Returns 0 on success, -1 on
362 static int rxinit_status = 1;
363 #ifdef AFS_PTHREAD_ENV
365 * This mutex protects the following global variables:
369 #define LOCK_RX_INIT assert(pthread_mutex_lock(&rx_init_mutex)==0)
370 #define UNLOCK_RX_INIT assert(pthread_mutex_unlock(&rx_init_mutex)==0)
373 #define UNLOCK_RX_INIT
377 rx_InitHost(u_int host, u_int port)
384 char *htable, *ptable;
387 #if defined(AFS_DJGPP_ENV) && !defined(DEBUG)
388 __djgpp_set_quiet_socket(1);
395 if (rxinit_status == 0) {
396 tmp_status = rxinit_status;
398 return tmp_status; /* Already started; return previous error code. */
404 if (afs_winsockInit() < 0)
410 * Initialize anything necessary to provide a non-premptive threading
413 rxi_InitializeThreadSupport();
416 /* Allocate and initialize a socket for client and perhaps server
419 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
420 if (rx_socket == OSI_NULLSOCKET) {
424 #ifdef RX_ENABLE_LOCKS
427 #endif /* RX_LOCKS_DB */
428 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
429 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
430 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
431 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
432 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
434 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
436 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
438 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
440 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
442 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
444 #if defined(KERNEL) && defined(AFS_HPUX110_ENV)
446 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
447 #endif /* KERNEL && AFS_HPUX110_ENV */
448 #endif /* RX_ENABLE_LOCKS */
451 rx_connDeadTime = 12;
452 rx_tranquil = 0; /* reset flag */
453 memset((char *)&rx_stats, 0, sizeof(struct rx_stats));
455 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
456 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
457 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
458 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
459 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
460 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
462 /* Malloc up a bunch of packets & buffers */
464 queue_Init(&rx_freePacketQueue);
465 rxi_NeedMorePackets = FALSE;
466 #ifdef RX_ENABLE_TSFPQ
467 rx_nPackets = 0; /* in TSFPQ version, rx_nPackets is managed by rxi_MorePackets* */
468 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
469 #else /* RX_ENABLE_TSFPQ */
470 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
471 rxi_MorePackets(rx_nPackets);
472 #endif /* RX_ENABLE_TSFPQ */
479 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
480 tv.tv_sec = clock_now.sec;
481 tv.tv_usec = clock_now.usec;
482 srand((unsigned int)tv.tv_usec);
489 #if defined(KERNEL) && !defined(UKERNEL)
490 /* Really, this should never happen in a real kernel */
493 struct sockaddr_in addr;
494 int addrlen = sizeof(addr);
495 if (getsockname((int)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
499 rx_port = addr.sin_port;
502 rx_stats.minRtt.sec = 9999999;
504 rx_SetEpoch(tv.tv_sec | 0x80000000);
506 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
507 * will provide a randomer value. */
509 MUTEX_ENTER(&rx_stats_mutex);
510 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
511 MUTEX_EXIT(&rx_stats_mutex);
512 /* *Slightly* random start time for the cid. This is just to help
513 * out with the hashing function at the peer */
514 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
515 rx_connHashTable = (struct rx_connection **)htable;
516 rx_peerHashTable = (struct rx_peer **)ptable;
518 rx_lastAckDelay.sec = 0;
519 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
520 rx_hardAckDelay.sec = 0;
521 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
522 rx_softAckDelay.sec = 0;
523 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
525 rxevent_Init(20, rxi_ReScheduleEvents);
527 /* Initialize various global queues */
528 queue_Init(&rx_idleServerQueue);
529 queue_Init(&rx_incomingCallQueue);
530 queue_Init(&rx_freeCallQueue);
532 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
533 /* Initialize our list of usable IP addresses. */
537 /* Start listener process (exact function is dependent on the
538 * implementation environment--kernel or user space) */
542 tmp_status = rxinit_status = 0;
550 return rx_InitHost(htonl(INADDR_ANY), port);
553 /* called with unincremented nRequestsRunning to see if it is OK to start
554 * a new thread in this service. Could be "no" for two reasons: over the
555 * max quota, or would prevent others from reaching their min quota.
557 #ifdef RX_ENABLE_LOCKS
558 /* This verion of QuotaOK reserves quota if it's ok while the
559 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
562 QuotaOK(register struct rx_service *aservice)
564 /* check if over max quota */
565 if (aservice->nRequestsRunning >= aservice->maxProcs) {
569 /* under min quota, we're OK */
570 /* otherwise, can use only if there are enough to allow everyone
571 * to go to their min quota after this guy starts.
573 MUTEX_ENTER(&rx_stats_mutex);
574 if ((aservice->nRequestsRunning < aservice->minProcs)
575 || (rxi_availProcs > rxi_minDeficit)) {
576 aservice->nRequestsRunning++;
577 /* just started call in minProcs pool, need fewer to maintain
579 if (aservice->nRequestsRunning <= aservice->minProcs)
582 MUTEX_EXIT(&rx_stats_mutex);
585 MUTEX_EXIT(&rx_stats_mutex);
591 ReturnToServerPool(register struct rx_service *aservice)
593 aservice->nRequestsRunning--;
594 MUTEX_ENTER(&rx_stats_mutex);
595 if (aservice->nRequestsRunning < aservice->minProcs)
598 MUTEX_EXIT(&rx_stats_mutex);
601 #else /* RX_ENABLE_LOCKS */
603 QuotaOK(register struct rx_service *aservice)
606 /* under min quota, we're OK */
607 if (aservice->nRequestsRunning < aservice->minProcs)
610 /* check if over max quota */
611 if (aservice->nRequestsRunning >= aservice->maxProcs)
614 /* otherwise, can use only if there are enough to allow everyone
615 * to go to their min quota after this guy starts.
617 if (rxi_availProcs > rxi_minDeficit)
621 #endif /* RX_ENABLE_LOCKS */
624 /* Called by rx_StartServer to start up lwp's to service calls.
625 NExistingProcs gives the number of procs already existing, and which
626 therefore needn't be created. */
628 rxi_StartServerProcs(int nExistingProcs)
630 register struct rx_service *service;
635 /* For each service, reserve N processes, where N is the "minimum"
636 * number of processes that MUST be able to execute a request in parallel,
637 * at any time, for that process. Also compute the maximum difference
638 * between any service's maximum number of processes that can run
639 * (i.e. the maximum number that ever will be run, and a guarantee
640 * that this number will run if other services aren't running), and its
641 * minimum number. The result is the extra number of processes that
642 * we need in order to provide the latter guarantee */
643 for (i = 0; i < RX_MAX_SERVICES; i++) {
645 service = rx_services[i];
646 if (service == (struct rx_service *)0)
648 nProcs += service->minProcs;
649 diff = service->maxProcs - service->minProcs;
653 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
654 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
655 for (i = 0; i < nProcs; i++) {
656 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
662 /* This routine is only required on Windows */
664 rx_StartClientThread(void)
666 #ifdef AFS_PTHREAD_ENV
668 pid = (int) pthread_self();
669 #endif /* AFS_PTHREAD_ENV */
671 #endif /* AFS_NT40_ENV */
673 /* This routine must be called if any services are exported. If the
674 * donateMe flag is set, the calling process is donated to the server
677 rx_StartServer(int donateMe)
679 register struct rx_service *service;
685 /* Start server processes, if necessary (exact function is dependent
686 * on the implementation environment--kernel or user space). DonateMe
687 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
688 * case, one less new proc will be created rx_StartServerProcs.
690 rxi_StartServerProcs(donateMe);
692 /* count up the # of threads in minProcs, and add set the min deficit to
693 * be that value, too.
695 for (i = 0; i < RX_MAX_SERVICES; i++) {
696 service = rx_services[i];
697 if (service == (struct rx_service *)0)
699 MUTEX_ENTER(&rx_stats_mutex);
700 rxi_totalMin += service->minProcs;
701 /* below works even if a thread is running, since minDeficit would
702 * still have been decremented and later re-incremented.
704 rxi_minDeficit += service->minProcs;
705 MUTEX_EXIT(&rx_stats_mutex);
708 /* Turn on reaping of idle server connections */
709 rxi_ReapConnections();
718 #ifdef AFS_PTHREAD_ENV
720 pid = (pid_t) pthread_self();
721 #else /* AFS_PTHREAD_ENV */
723 LWP_CurrentProcess(&pid);
724 #endif /* AFS_PTHREAD_ENV */
726 sprintf(name, "srv_%d", ++nProcs);
728 (*registerProgram) (pid, name);
730 #endif /* AFS_NT40_ENV */
731 rx_ServerProc(); /* Never returns */
733 #ifdef RX_ENABLE_TSFPQ
734 /* no use leaving packets around in this thread's local queue if
735 * it isn't getting donated to the server thread pool.
737 rxi_FlushLocalPacketsTSFPQ();
738 #endif /* RX_ENABLE_TSFPQ */
742 /* Create a new client connection to the specified service, using the
743 * specified security object to implement the security model for this
745 struct rx_connection *
746 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
747 register struct rx_securityClass *securityObject,
748 int serviceSecurityIndex)
752 register struct rx_connection *conn;
757 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", ntohl(shost), ntohs(sport), sservice, securityObject, serviceSecurityIndex));
759 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
760 * the case of kmem_alloc? */
761 conn = rxi_AllocConnection();
762 #ifdef RX_ENABLE_LOCKS
763 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
764 MUTEX_INIT(&conn->conn_data_lock, "conn call lock", MUTEX_DEFAULT, 0);
765 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
768 MUTEX_ENTER(&rx_connHashTable_lock);
769 cid = (rx_nextCid += RX_MAXCALLS);
770 conn->type = RX_CLIENT_CONNECTION;
772 conn->epoch = rx_epoch;
773 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
774 conn->serviceId = sservice;
775 conn->securityObject = securityObject;
776 /* This doesn't work in all compilers with void (they're buggy), so fake it
778 conn->securityData = (VOID *) 0;
779 conn->securityIndex = serviceSecurityIndex;
780 rx_SetConnDeadTime(conn, rx_connDeadTime);
781 conn->ackRate = RX_FAST_ACK_RATE;
783 conn->specific = NULL;
784 conn->challengeEvent = NULL;
785 conn->delayedAbortEvent = NULL;
786 conn->abortCount = 0;
789 RXS_NewConnection(securityObject, conn);
791 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
793 conn->refCount++; /* no lock required since only this thread knows... */
794 conn->next = rx_connHashTable[hashindex];
795 rx_connHashTable[hashindex] = conn;
796 MUTEX_ENTER(&rx_stats_mutex);
797 rx_stats.nClientConns++;
798 MUTEX_EXIT(&rx_stats_mutex);
800 MUTEX_EXIT(&rx_connHashTable_lock);
806 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
808 /* The idea is to set the dead time to a value that allows several
809 * keepalives to be dropped without timing out the connection. */
810 conn->secondsUntilDead = MAX(seconds, 6);
811 conn->secondsUntilPing = conn->secondsUntilDead / 6;
814 int rxi_lowPeerRefCount = 0;
815 int rxi_lowConnRefCount = 0;
818 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
819 * NOTE: must not be called with rx_connHashTable_lock held.
822 rxi_CleanupConnection(struct rx_connection *conn)
824 /* Notify the service exporter, if requested, that this connection
825 * is being destroyed */
826 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
827 (*conn->service->destroyConnProc) (conn);
829 /* Notify the security module that this connection is being destroyed */
830 RXS_DestroyConnection(conn->securityObject, conn);
832 /* If this is the last connection using the rx_peer struct, set its
833 * idle time to now. rxi_ReapConnections will reap it if it's still
834 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
836 MUTEX_ENTER(&rx_peerHashTable_lock);
837 if (conn->peer->refCount < 2) {
838 conn->peer->idleWhen = clock_Sec();
839 if (conn->peer->refCount < 1) {
840 conn->peer->refCount = 1;
841 MUTEX_ENTER(&rx_stats_mutex);
842 rxi_lowPeerRefCount++;
843 MUTEX_EXIT(&rx_stats_mutex);
846 conn->peer->refCount--;
847 MUTEX_EXIT(&rx_peerHashTable_lock);
849 MUTEX_ENTER(&rx_stats_mutex);
850 if (conn->type == RX_SERVER_CONNECTION)
851 rx_stats.nServerConns--;
853 rx_stats.nClientConns--;
854 MUTEX_EXIT(&rx_stats_mutex);
857 if (conn->specific) {
859 for (i = 0; i < conn->nSpecific; i++) {
860 if (conn->specific[i] && rxi_keyCreate_destructor[i])
861 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
862 conn->specific[i] = NULL;
864 free(conn->specific);
866 conn->specific = NULL;
870 MUTEX_DESTROY(&conn->conn_call_lock);
871 MUTEX_DESTROY(&conn->conn_data_lock);
872 CV_DESTROY(&conn->conn_call_cv);
874 rxi_FreeConnection(conn);
877 /* Destroy the specified connection */
879 rxi_DestroyConnection(register struct rx_connection *conn)
881 MUTEX_ENTER(&rx_connHashTable_lock);
882 rxi_DestroyConnectionNoLock(conn);
883 /* conn should be at the head of the cleanup list */
884 if (conn == rx_connCleanup_list) {
885 rx_connCleanup_list = rx_connCleanup_list->next;
886 MUTEX_EXIT(&rx_connHashTable_lock);
887 rxi_CleanupConnection(conn);
889 #ifdef RX_ENABLE_LOCKS
891 MUTEX_EXIT(&rx_connHashTable_lock);
893 #endif /* RX_ENABLE_LOCKS */
897 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
899 register struct rx_connection **conn_ptr;
900 register int havecalls = 0;
901 struct rx_packet *packet;
908 MUTEX_ENTER(&conn->conn_data_lock);
909 if (conn->refCount > 0)
912 MUTEX_ENTER(&rx_stats_mutex);
913 rxi_lowConnRefCount++;
914 MUTEX_EXIT(&rx_stats_mutex);
917 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
918 /* Busy; wait till the last guy before proceeding */
919 MUTEX_EXIT(&conn->conn_data_lock);
924 /* If the client previously called rx_NewCall, but it is still
925 * waiting, treat this as a running call, and wait to destroy the
926 * connection later when the call completes. */
927 if ((conn->type == RX_CLIENT_CONNECTION)
928 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
929 conn->flags |= RX_CONN_DESTROY_ME;
930 MUTEX_EXIT(&conn->conn_data_lock);
934 MUTEX_EXIT(&conn->conn_data_lock);
936 /* Check for extant references to this connection */
937 for (i = 0; i < RX_MAXCALLS; i++) {
938 register struct rx_call *call = conn->call[i];
941 if (conn->type == RX_CLIENT_CONNECTION) {
942 MUTEX_ENTER(&call->lock);
943 if (call->delayedAckEvent) {
944 /* Push the final acknowledgment out now--there
945 * won't be a subsequent call to acknowledge the
946 * last reply packets */
947 rxevent_Cancel(call->delayedAckEvent, call,
948 RX_CALL_REFCOUNT_DELAY);
949 if (call->state == RX_STATE_PRECALL
950 || call->state == RX_STATE_ACTIVE) {
951 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
953 rxi_AckAll(NULL, call, 0);
956 MUTEX_EXIT(&call->lock);
960 #ifdef RX_ENABLE_LOCKS
962 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
963 MUTEX_EXIT(&conn->conn_data_lock);
965 /* Someone is accessing a packet right now. */
969 #endif /* RX_ENABLE_LOCKS */
972 /* Don't destroy the connection if there are any call
973 * structures still in use */
974 MUTEX_ENTER(&conn->conn_data_lock);
975 conn->flags |= RX_CONN_DESTROY_ME;
976 MUTEX_EXIT(&conn->conn_data_lock);
981 if (conn->delayedAbortEvent) {
982 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
983 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
985 MUTEX_ENTER(&conn->conn_data_lock);
986 rxi_SendConnectionAbort(conn, packet, 0, 1);
987 MUTEX_EXIT(&conn->conn_data_lock);
988 rxi_FreePacket(packet);
992 /* Remove from connection hash table before proceeding */
994 &rx_connHashTable[CONN_HASH
995 (peer->host, peer->port, conn->cid, conn->epoch,
997 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
998 if (*conn_ptr == conn) {
999 *conn_ptr = conn->next;
1003 /* if the conn that we are destroying was the last connection, then we
1004 * clear rxLastConn as well */
1005 if (rxLastConn == conn)
1008 /* Make sure the connection is completely reset before deleting it. */
1009 /* get rid of pending events that could zap us later */
1010 if (conn->challengeEvent)
1011 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1012 if (conn->checkReachEvent)
1013 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1015 /* Add the connection to the list of destroyed connections that
1016 * need to be cleaned up. This is necessary to avoid deadlocks
1017 * in the routines we call to inform others that this connection is
1018 * being destroyed. */
1019 conn->next = rx_connCleanup_list;
1020 rx_connCleanup_list = conn;
1023 /* Externally available version */
1025 rx_DestroyConnection(register struct rx_connection *conn)
1030 rxi_DestroyConnection(conn);
1035 rx_GetConnection(register struct rx_connection *conn)
1040 MUTEX_ENTER(&conn->conn_data_lock);
1042 MUTEX_EXIT(&conn->conn_data_lock);
1046 /* Start a new rx remote procedure call, on the specified connection.
1047 * If wait is set to 1, wait for a free call channel; otherwise return
1048 * 0. Maxtime gives the maximum number of seconds this call may take,
1049 * after rx_MakeCall returns. After this time interval, a call to any
1050 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1051 * For fine grain locking, we hold the conn_call_lock in order to
1052 * to ensure that we don't get signalle after we found a call in an active
1053 * state and before we go to sleep.
1056 rx_NewCall(register struct rx_connection *conn)
1059 register struct rx_call *call;
1060 struct clock queueTime;
1064 dpf(("rx_MakeCall(conn %x)\n", conn));
1067 clock_GetTime(&queueTime);
1068 MUTEX_ENTER(&conn->conn_call_lock);
1071 * Check if there are others waiting for a new call.
1072 * If so, let them go first to avoid starving them.
1073 * This is a fairly simple scheme, and might not be
1074 * a complete solution for large numbers of waiters.
1076 * makeCallWaiters keeps track of the number of
1077 * threads waiting to make calls and the
1078 * RX_CONN_MAKECALL_WAITING flag bit is used to
1079 * indicate that there are indeed calls waiting.
1080 * The flag is set when the waiter is incremented.
1081 * It is only cleared in rx_EndCall when
1082 * makeCallWaiters is 0. This prevents us from
1083 * accidently destroying the connection while it
1084 * is potentially about to be used.
1086 MUTEX_ENTER(&conn->conn_data_lock);
1087 if (conn->makeCallWaiters) {
1088 conn->flags |= RX_CONN_MAKECALL_WAITING;
1089 conn->makeCallWaiters++;
1090 MUTEX_EXIT(&conn->conn_data_lock);
1092 #ifdef RX_ENABLE_LOCKS
1093 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1097 MUTEX_ENTER(&conn->conn_data_lock);
1098 conn->makeCallWaiters--;
1100 MUTEX_EXIT(&conn->conn_data_lock);
1103 for (i = 0; i < RX_MAXCALLS; i++) {
1104 call = conn->call[i];
1106 MUTEX_ENTER(&call->lock);
1107 if (call->state == RX_STATE_DALLY) {
1108 rxi_ResetCall(call, 0);
1109 (*call->callNumber)++;
1112 MUTEX_EXIT(&call->lock);
1114 call = rxi_NewCall(conn, i);
1118 if (i < RX_MAXCALLS) {
1121 MUTEX_ENTER(&conn->conn_data_lock);
1122 conn->flags |= RX_CONN_MAKECALL_WAITING;
1123 conn->makeCallWaiters++;
1124 MUTEX_EXIT(&conn->conn_data_lock);
1126 #ifdef RX_ENABLE_LOCKS
1127 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1131 MUTEX_ENTER(&conn->conn_data_lock);
1132 conn->makeCallWaiters--;
1133 MUTEX_EXIT(&conn->conn_data_lock);
1136 * Wake up anyone else who might be giving us a chance to
1137 * run (see code above that avoids resource starvation).
1139 #ifdef RX_ENABLE_LOCKS
1140 CV_BROADCAST(&conn->conn_call_cv);
1145 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1147 /* Client is initially in send mode */
1148 call->state = RX_STATE_ACTIVE;
1149 call->error = conn->error;
1151 call->mode = RX_MODE_ERROR;
1153 call->mode = RX_MODE_SENDING;
1155 /* remember start time for call in case we have hard dead time limit */
1156 call->queueTime = queueTime;
1157 clock_GetTime(&call->startTime);
1158 hzero(call->bytesSent);
1159 hzero(call->bytesRcvd);
1161 /* Turn on busy protocol. */
1162 rxi_KeepAliveOn(call);
1164 MUTEX_EXIT(&call->lock);
1165 MUTEX_EXIT(&conn->conn_call_lock);
1168 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1169 /* Now, if TQ wasn't cleared earlier, do it now. */
1170 MUTEX_ENTER(&call->lock);
1171 while (call->flags & RX_CALL_TQ_BUSY) {
1172 call->flags |= RX_CALL_TQ_WAIT;
1174 #ifdef RX_ENABLE_LOCKS
1175 osirx_AssertMine(&call->lock, "rxi_Start lock4");
1176 CV_WAIT(&call->cv_tq, &call->lock);
1177 #else /* RX_ENABLE_LOCKS */
1178 osi_rxSleep(&call->tq);
1179 #endif /* RX_ENABLE_LOCKS */
1181 if (call->tqWaiters == 0) {
1182 call->flags &= ~RX_CALL_TQ_WAIT;
1185 if (call->flags & RX_CALL_TQ_CLEARME) {
1186 rxi_ClearTransmitQueue(call, 0);
1187 queue_Init(&call->tq);
1189 MUTEX_EXIT(&call->lock);
1190 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1196 rxi_HasActiveCalls(register struct rx_connection *aconn)
1199 register struct rx_call *tcall;
1203 for (i = 0; i < RX_MAXCALLS; i++) {
1204 if ((tcall = aconn->call[i])) {
1205 if ((tcall->state == RX_STATE_ACTIVE)
1206 || (tcall->state == RX_STATE_PRECALL)) {
1217 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1218 register afs_int32 * aint32s)
1221 register struct rx_call *tcall;
1225 for (i = 0; i < RX_MAXCALLS; i++) {
1226 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1227 aint32s[i] = aconn->callNumber[i] + 1;
1229 aint32s[i] = aconn->callNumber[i];
1236 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1237 register afs_int32 * aint32s)
1240 register struct rx_call *tcall;
1244 for (i = 0; i < RX_MAXCALLS; i++) {
1245 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1246 aconn->callNumber[i] = aint32s[i] - 1;
1248 aconn->callNumber[i] = aint32s[i];
1254 /* Advertise a new service. A service is named locally by a UDP port
1255 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1258 char *serviceName; Name for identification purposes (e.g. the
1259 service name might be used for probing for
1262 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1263 struct rx_securityClass **securityObjects, int nSecurityObjects,
1264 afs_int32(*serviceProc) (struct rx_call * acall))
1266 osi_socket socket = OSI_NULLSOCKET;
1267 register struct rx_service *tservice;
1273 if (serviceId == 0) {
1275 "rx_NewService: service id for service %s is not non-zero.\n",
1282 "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",
1290 tservice = rxi_AllocService();
1292 for (i = 0; i < RX_MAX_SERVICES; i++) {
1293 register struct rx_service *service = rx_services[i];
1295 if (port == service->servicePort) {
1296 if (service->serviceId == serviceId) {
1297 /* The identical service has already been
1298 * installed; if the caller was intending to
1299 * change the security classes used by this
1300 * service, he/she loses. */
1302 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1303 serviceName, serviceId, service->serviceName);
1305 rxi_FreeService(tservice);
1308 /* Different service, same port: re-use the socket
1309 * which is bound to the same port */
1310 socket = service->socket;
1313 if (socket == OSI_NULLSOCKET) {
1314 /* If we don't already have a socket (from another
1315 * service on same port) get a new one */
1316 socket = rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
1317 if (socket == OSI_NULLSOCKET) {
1319 rxi_FreeService(tservice);
1324 service->socket = socket;
1325 service->servicePort = port;
1326 service->serviceId = serviceId;
1327 service->serviceName = serviceName;
1328 service->nSecurityObjects = nSecurityObjects;
1329 service->securityObjects = securityObjects;
1330 service->minProcs = 0;
1331 service->maxProcs = 1;
1332 service->idleDeadTime = 60;
1333 service->connDeadTime = rx_connDeadTime;
1334 service->executeRequestProc = serviceProc;
1335 service->checkReach = 0;
1336 rx_services[i] = service; /* not visible until now */
1342 rxi_FreeService(tservice);
1343 (osi_Msg "rx_NewService: cannot support > %d services\n",
1348 /* Generic request processing loop. This routine should be called
1349 * by the implementation dependent rx_ServerProc. If socketp is
1350 * non-null, it will be set to the file descriptor that this thread
1351 * is now listening on. If socketp is null, this routine will never
1354 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1356 register struct rx_call *call;
1357 register afs_int32 code;
1358 register struct rx_service *tservice = NULL;
1365 call = rx_GetCall(threadID, tservice, socketp);
1366 if (socketp && *socketp != OSI_NULLSOCKET) {
1367 /* We are now a listener thread */
1372 /* if server is restarting( typically smooth shutdown) then do not
1373 * allow any new calls.
1376 if (rx_tranquil && (call != NULL)) {
1380 MUTEX_ENTER(&call->lock);
1382 rxi_CallError(call, RX_RESTARTING);
1383 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1385 MUTEX_EXIT(&call->lock);
1389 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1390 #ifdef RX_ENABLE_LOCKS
1392 #endif /* RX_ENABLE_LOCKS */
1393 afs_termState = AFSOP_STOP_AFS;
1394 afs_osi_Wakeup(&afs_termState);
1395 #ifdef RX_ENABLE_LOCKS
1397 #endif /* RX_ENABLE_LOCKS */
1402 tservice = call->conn->service;
1404 if (tservice->beforeProc)
1405 (*tservice->beforeProc) (call);
1407 code = call->conn->service->executeRequestProc(call);
1409 if (tservice->afterProc)
1410 (*tservice->afterProc) (call, code);
1412 rx_EndCall(call, code);
1413 MUTEX_ENTER(&rx_stats_mutex);
1415 MUTEX_EXIT(&rx_stats_mutex);
1421 rx_WakeupServerProcs(void)
1423 struct rx_serverQueueEntry *np, *tqp;
1427 MUTEX_ENTER(&rx_serverPool_lock);
1429 #ifdef RX_ENABLE_LOCKS
1430 if (rx_waitForPacket)
1431 CV_BROADCAST(&rx_waitForPacket->cv);
1432 #else /* RX_ENABLE_LOCKS */
1433 if (rx_waitForPacket)
1434 osi_rxWakeup(rx_waitForPacket);
1435 #endif /* RX_ENABLE_LOCKS */
1436 MUTEX_ENTER(&freeSQEList_lock);
1437 for (np = rx_FreeSQEList; np; np = tqp) {
1438 tqp = *(struct rx_serverQueueEntry **)np;
1439 #ifdef RX_ENABLE_LOCKS
1440 CV_BROADCAST(&np->cv);
1441 #else /* RX_ENABLE_LOCKS */
1443 #endif /* RX_ENABLE_LOCKS */
1445 MUTEX_EXIT(&freeSQEList_lock);
1446 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1447 #ifdef RX_ENABLE_LOCKS
1448 CV_BROADCAST(&np->cv);
1449 #else /* RX_ENABLE_LOCKS */
1451 #endif /* RX_ENABLE_LOCKS */
1453 MUTEX_EXIT(&rx_serverPool_lock);
1458 * One thing that seems to happen is that all the server threads get
1459 * tied up on some empty or slow call, and then a whole bunch of calls
1460 * arrive at once, using up the packet pool, so now there are more
1461 * empty calls. The most critical resources here are server threads
1462 * and the free packet pool. The "doreclaim" code seems to help in
1463 * general. I think that eventually we arrive in this state: there
1464 * are lots of pending calls which do have all their packets present,
1465 * so they won't be reclaimed, are multi-packet calls, so they won't
1466 * be scheduled until later, and thus are tying up most of the free
1467 * packet pool for a very long time.
1469 * 1. schedule multi-packet calls if all the packets are present.
1470 * Probably CPU-bound operation, useful to return packets to pool.
1471 * Do what if there is a full window, but the last packet isn't here?
1472 * 3. preserve one thread which *only* runs "best" calls, otherwise
1473 * it sleeps and waits for that type of call.
1474 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1475 * the current dataquota business is badly broken. The quota isn't adjusted
1476 * to reflect how many packets are presently queued for a running call.
1477 * So, when we schedule a queued call with a full window of packets queued
1478 * up for it, that *should* free up a window full of packets for other 2d-class
1479 * calls to be able to use from the packet pool. But it doesn't.
1481 * NB. Most of the time, this code doesn't run -- since idle server threads
1482 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1483 * as a new call arrives.
1485 /* Sleep until a call arrives. Returns a pointer to the call, ready
1486 * for an rx_Read. */
1487 #ifdef RX_ENABLE_LOCKS
1489 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1491 struct rx_serverQueueEntry *sq;
1492 register struct rx_call *call = (struct rx_call *)0;
1493 struct rx_service *service = NULL;
1496 MUTEX_ENTER(&freeSQEList_lock);
1498 if ((sq = rx_FreeSQEList)) {
1499 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1500 MUTEX_EXIT(&freeSQEList_lock);
1501 } else { /* otherwise allocate a new one and return that */
1502 MUTEX_EXIT(&freeSQEList_lock);
1503 sq = (struct rx_serverQueueEntry *)
1504 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1505 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1506 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1509 MUTEX_ENTER(&rx_serverPool_lock);
1510 if (cur_service != NULL) {
1511 ReturnToServerPool(cur_service);
1514 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1515 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1517 /* Scan for eligible incoming calls. A call is not eligible
1518 * if the maximum number of calls for its service type are
1519 * already executing */
1520 /* One thread will process calls FCFS (to prevent starvation),
1521 * while the other threads may run ahead looking for calls which
1522 * have all their input data available immediately. This helps
1523 * keep threads from blocking, waiting for data from the client. */
1524 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1525 service = tcall->conn->service;
1526 if (!QuotaOK(service)) {
1529 if (tno == rxi_fcfs_thread_num
1530 || !tcall->queue_item_header.next) {
1531 /* If we're the fcfs thread , then we'll just use
1532 * this call. If we haven't been able to find an optimal
1533 * choice, and we're at the end of the list, then use a
1534 * 2d choice if one has been identified. Otherwise... */
1535 call = (choice2 ? choice2 : tcall);
1536 service = call->conn->service;
1537 } else if (!queue_IsEmpty(&tcall->rq)) {
1538 struct rx_packet *rp;
1539 rp = queue_First(&tcall->rq, rx_packet);
1540 if (rp->header.seq == 1) {
1542 || (rp->header.flags & RX_LAST_PACKET)) {
1544 } else if (rxi_2dchoice && !choice2
1545 && !(tcall->flags & RX_CALL_CLEARED)
1546 && (tcall->rprev > rxi_HardAckRate)) {
1555 ReturnToServerPool(service);
1562 MUTEX_EXIT(&rx_serverPool_lock);
1563 MUTEX_ENTER(&call->lock);
1565 if (call->flags & RX_CALL_WAIT_PROC) {
1566 call->flags &= ~RX_CALL_WAIT_PROC;
1567 MUTEX_ENTER(&rx_stats_mutex);
1569 MUTEX_EXIT(&rx_stats_mutex);
1572 if (call->state != RX_STATE_PRECALL || call->error) {
1573 MUTEX_EXIT(&call->lock);
1574 MUTEX_ENTER(&rx_serverPool_lock);
1575 ReturnToServerPool(service);
1580 if (queue_IsEmpty(&call->rq)
1581 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1582 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1584 CLEAR_CALL_QUEUE_LOCK(call);
1587 /* If there are no eligible incoming calls, add this process
1588 * to the idle server queue, to wait for one */
1592 *socketp = OSI_NULLSOCKET;
1594 sq->socketp = socketp;
1595 queue_Append(&rx_idleServerQueue, sq);
1596 #ifndef AFS_AIX41_ENV
1597 rx_waitForPacket = sq;
1599 rx_waitingForPacket = sq;
1600 #endif /* AFS_AIX41_ENV */
1602 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1604 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1605 MUTEX_EXIT(&rx_serverPool_lock);
1606 return (struct rx_call *)0;
1609 } while (!(call = sq->newcall)
1610 && !(socketp && *socketp != OSI_NULLSOCKET));
1611 MUTEX_EXIT(&rx_serverPool_lock);
1613 MUTEX_ENTER(&call->lock);
1619 MUTEX_ENTER(&freeSQEList_lock);
1620 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1621 rx_FreeSQEList = sq;
1622 MUTEX_EXIT(&freeSQEList_lock);
1625 clock_GetTime(&call->startTime);
1626 call->state = RX_STATE_ACTIVE;
1627 call->mode = RX_MODE_RECEIVING;
1628 #ifdef RX_KERNEL_TRACE
1629 if (ICL_SETACTIVE(afs_iclSetp)) {
1630 int glockOwner = ISAFS_GLOCK();
1633 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1634 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1641 rxi_calltrace(RX_CALL_START, call);
1642 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1643 call->conn->service->servicePort, call->conn->service->serviceId,
1646 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1647 MUTEX_EXIT(&call->lock);
1649 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1654 #else /* RX_ENABLE_LOCKS */
1656 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1658 struct rx_serverQueueEntry *sq;
1659 register struct rx_call *call = (struct rx_call *)0, *choice2;
1660 struct rx_service *service = NULL;
1664 MUTEX_ENTER(&freeSQEList_lock);
1666 if ((sq = rx_FreeSQEList)) {
1667 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1668 MUTEX_EXIT(&freeSQEList_lock);
1669 } else { /* otherwise allocate a new one and return that */
1670 MUTEX_EXIT(&freeSQEList_lock);
1671 sq = (struct rx_serverQueueEntry *)
1672 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1673 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1674 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1676 MUTEX_ENTER(&sq->lock);
1678 if (cur_service != NULL) {
1679 cur_service->nRequestsRunning--;
1680 if (cur_service->nRequestsRunning < cur_service->minProcs)
1684 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1685 register struct rx_call *tcall, *ncall;
1686 /* Scan for eligible incoming calls. A call is not eligible
1687 * if the maximum number of calls for its service type are
1688 * already executing */
1689 /* One thread will process calls FCFS (to prevent starvation),
1690 * while the other threads may run ahead looking for calls which
1691 * have all their input data available immediately. This helps
1692 * keep threads from blocking, waiting for data from the client. */
1693 choice2 = (struct rx_call *)0;
1694 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1695 service = tcall->conn->service;
1696 if (QuotaOK(service)) {
1697 if (tno == rxi_fcfs_thread_num
1698 || !tcall->queue_item_header.next) {
1699 /* If we're the fcfs thread, then we'll just use
1700 * this call. If we haven't been able to find an optimal
1701 * choice, and we're at the end of the list, then use a
1702 * 2d choice if one has been identified. Otherwise... */
1703 call = (choice2 ? choice2 : tcall);
1704 service = call->conn->service;
1705 } else if (!queue_IsEmpty(&tcall->rq)) {
1706 struct rx_packet *rp;
1707 rp = queue_First(&tcall->rq, rx_packet);
1708 if (rp->header.seq == 1
1710 || (rp->header.flags & RX_LAST_PACKET))) {
1712 } else if (rxi_2dchoice && !choice2
1713 && !(tcall->flags & RX_CALL_CLEARED)
1714 && (tcall->rprev > rxi_HardAckRate)) {
1727 /* we can't schedule a call if there's no data!!! */
1728 /* send an ack if there's no data, if we're missing the
1729 * first packet, or we're missing something between first
1730 * and last -- there's a "hole" in the incoming data. */
1731 if (queue_IsEmpty(&call->rq)
1732 || queue_First(&call->rq, rx_packet)->header.seq != 1
1733 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1734 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1736 call->flags &= (~RX_CALL_WAIT_PROC);
1737 service->nRequestsRunning++;
1738 /* just started call in minProcs pool, need fewer to maintain
1740 if (service->nRequestsRunning <= service->minProcs)
1744 /* MUTEX_EXIT(&call->lock); */
1746 /* If there are no eligible incoming calls, add this process
1747 * to the idle server queue, to wait for one */
1750 *socketp = OSI_NULLSOCKET;
1752 sq->socketp = socketp;
1753 queue_Append(&rx_idleServerQueue, sq);
1757 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1759 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1760 return (struct rx_call *)0;
1763 } while (!(call = sq->newcall)
1764 && !(socketp && *socketp != OSI_NULLSOCKET));
1766 MUTEX_EXIT(&sq->lock);
1768 MUTEX_ENTER(&freeSQEList_lock);
1769 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1770 rx_FreeSQEList = sq;
1771 MUTEX_EXIT(&freeSQEList_lock);
1774 clock_GetTime(&call->startTime);
1775 call->state = RX_STATE_ACTIVE;
1776 call->mode = RX_MODE_RECEIVING;
1777 #ifdef RX_KERNEL_TRACE
1778 if (ICL_SETACTIVE(afs_iclSetp)) {
1779 int glockOwner = ISAFS_GLOCK();
1782 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1783 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1790 rxi_calltrace(RX_CALL_START, call);
1791 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1792 call->conn->service->servicePort, call->conn->service->serviceId,
1795 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1802 #endif /* RX_ENABLE_LOCKS */
1806 /* Establish a procedure to be called when a packet arrives for a
1807 * call. This routine will be called at most once after each call,
1808 * and will also be called if there is an error condition on the or
1809 * the call is complete. Used by multi rx to build a selection
1810 * function which determines which of several calls is likely to be a
1811 * good one to read from.
1812 * NOTE: the way this is currently implemented it is probably only a
1813 * good idea to (1) use it immediately after a newcall (clients only)
1814 * and (2) only use it once. Other uses currently void your warranty
1817 rx_SetArrivalProc(register struct rx_call *call,
1818 register void (*proc) (register struct rx_call * call,
1820 register int index),
1821 register VOID * handle, register int arg)
1823 call->arrivalProc = proc;
1824 call->arrivalProcHandle = handle;
1825 call->arrivalProcArg = arg;
1828 /* Call is finished (possibly prematurely). Return rc to the peer, if
1829 * appropriate, and return the final error code from the conversation
1833 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1835 register struct rx_connection *conn = call->conn;
1836 register struct rx_service *service;
1842 dpf(("rx_EndCall(call %x rc %d error %d abortCode %d)\n", call, rc, call->error, call->abortCode));
1845 MUTEX_ENTER(&call->lock);
1847 if (rc == 0 && call->error == 0) {
1848 call->abortCode = 0;
1849 call->abortCount = 0;
1852 call->arrivalProc = (void (*)())0;
1853 if (rc && call->error == 0) {
1854 rxi_CallError(call, rc);
1855 /* Send an abort message to the peer if this error code has
1856 * only just been set. If it was set previously, assume the
1857 * peer has already been sent the error code or will request it
1859 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1861 if (conn->type == RX_SERVER_CONNECTION) {
1862 /* Make sure reply or at least dummy reply is sent */
1863 if (call->mode == RX_MODE_RECEIVING) {
1864 rxi_WriteProc(call, 0, 0);
1866 if (call->mode == RX_MODE_SENDING) {
1867 rxi_FlushWrite(call);
1869 service = conn->service;
1870 rxi_calltrace(RX_CALL_END, call);
1871 /* Call goes to hold state until reply packets are acknowledged */
1872 if (call->tfirst + call->nSoftAcked < call->tnext) {
1873 call->state = RX_STATE_HOLD;
1875 call->state = RX_STATE_DALLY;
1876 rxi_ClearTransmitQueue(call, 0);
1877 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1878 rxevent_Cancel(call->keepAliveEvent, call,
1879 RX_CALL_REFCOUNT_ALIVE);
1881 } else { /* Client connection */
1883 /* Make sure server receives input packets, in the case where
1884 * no reply arguments are expected */
1885 if ((call->mode == RX_MODE_SENDING)
1886 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1887 (void)rxi_ReadProc(call, &dummy, 1);
1890 /* If we had an outstanding delayed ack, be nice to the server
1891 * and force-send it now.
1893 if (call->delayedAckEvent) {
1894 rxevent_Cancel(call->delayedAckEvent, call,
1895 RX_CALL_REFCOUNT_DELAY);
1896 call->delayedAckEvent = NULL;
1897 rxi_SendDelayedAck(NULL, call, NULL);
1900 /* We need to release the call lock since it's lower than the
1901 * conn_call_lock and we don't want to hold the conn_call_lock
1902 * over the rx_ReadProc call. The conn_call_lock needs to be held
1903 * here for the case where rx_NewCall is perusing the calls on
1904 * the connection structure. We don't want to signal until
1905 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1906 * have checked this call, found it active and by the time it
1907 * goes to sleep, will have missed the signal.
1909 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
1910 * there are threads waiting to use the conn object.
1912 MUTEX_EXIT(&call->lock);
1913 MUTEX_ENTER(&conn->conn_call_lock);
1914 MUTEX_ENTER(&call->lock);
1915 MUTEX_ENTER(&conn->conn_data_lock);
1916 conn->flags |= RX_CONN_BUSY;
1917 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1918 if (conn->makeCallWaiters == 0)
1919 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1920 MUTEX_EXIT(&conn->conn_data_lock);
1921 #ifdef RX_ENABLE_LOCKS
1922 CV_BROADCAST(&conn->conn_call_cv);
1927 #ifdef RX_ENABLE_LOCKS
1929 MUTEX_EXIT(&conn->conn_data_lock);
1931 #endif /* RX_ENABLE_LOCKS */
1932 call->state = RX_STATE_DALLY;
1934 error = call->error;
1936 /* currentPacket, nLeft, and NFree must be zeroed here, because
1937 * ResetCall cannot: ResetCall may be called at splnet(), in the
1938 * kernel version, and may interrupt the macros rx_Read or
1939 * rx_Write, which run at normal priority for efficiency. */
1940 if (call->currentPacket) {
1941 queue_Prepend(&call->iovq, call->currentPacket);
1942 call->currentPacket = (struct rx_packet *)0;
1945 call->nLeft = call->nFree = call->curlen = 0;
1947 /* Free any packets from the last call to ReadvProc/WritevProc */
1948 rxi_FreePackets(0, &call->iovq);
1950 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1951 MUTEX_EXIT(&call->lock);
1952 if (conn->type == RX_CLIENT_CONNECTION) {
1953 MUTEX_EXIT(&conn->conn_call_lock);
1954 conn->flags &= ~RX_CONN_BUSY;
1958 * Map errors to the local host's errno.h format.
1960 error = ntoh_syserr_conv(error);
1964 #if !defined(KERNEL)
1966 /* Call this routine when shutting down a server or client (especially
1967 * clients). This will allow Rx to gracefully garbage collect server
1968 * connections, and reduce the number of retries that a server might
1969 * make to a dead client.
1970 * This is not quite right, since some calls may still be ongoing and
1971 * we can't lock them to destroy them. */
1975 register struct rx_connection **conn_ptr, **conn_end;
1979 if (rxinit_status == 1) {
1981 return; /* Already shutdown. */
1983 rxi_DeleteCachedConnections();
1984 if (rx_connHashTable) {
1985 MUTEX_ENTER(&rx_connHashTable_lock);
1986 for (conn_ptr = &rx_connHashTable[0], conn_end =
1987 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
1989 struct rx_connection *conn, *next;
1990 for (conn = *conn_ptr; conn; conn = next) {
1992 if (conn->type == RX_CLIENT_CONNECTION) {
1993 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
1995 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
1996 #ifdef RX_ENABLE_LOCKS
1997 rxi_DestroyConnectionNoLock(conn);
1998 #else /* RX_ENABLE_LOCKS */
1999 rxi_DestroyConnection(conn);
2000 #endif /* RX_ENABLE_LOCKS */
2004 #ifdef RX_ENABLE_LOCKS
2005 while (rx_connCleanup_list) {
2006 struct rx_connection *conn;
2007 conn = rx_connCleanup_list;
2008 rx_connCleanup_list = rx_connCleanup_list->next;
2009 MUTEX_EXIT(&rx_connHashTable_lock);
2010 rxi_CleanupConnection(conn);
2011 MUTEX_ENTER(&rx_connHashTable_lock);
2013 MUTEX_EXIT(&rx_connHashTable_lock);
2014 #endif /* RX_ENABLE_LOCKS */
2023 /* if we wakeup packet waiter too often, can get in loop with two
2024 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2026 rxi_PacketsUnWait(void)
2028 if (!rx_waitingForPackets) {
2032 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2033 return; /* still over quota */
2036 rx_waitingForPackets = 0;
2037 #ifdef RX_ENABLE_LOCKS
2038 CV_BROADCAST(&rx_waitingForPackets_cv);
2040 osi_rxWakeup(&rx_waitingForPackets);
2046 /* ------------------Internal interfaces------------------------- */
2048 /* Return this process's service structure for the
2049 * specified socket and service */
2051 rxi_FindService(register osi_socket socket, register u_short serviceId)
2053 register struct rx_service **sp;
2054 for (sp = &rx_services[0]; *sp; sp++) {
2055 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2061 /* Allocate a call structure, for the indicated channel of the
2062 * supplied connection. The mode and state of the call must be set by
2063 * the caller. Returns the call with mutex locked. */
2065 rxi_NewCall(register struct rx_connection *conn, register int channel)
2067 register struct rx_call *call;
2068 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2069 register struct rx_call *cp; /* Call pointer temp */
2070 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2071 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2073 /* Grab an existing call structure, or allocate a new one.
2074 * Existing call structures are assumed to have been left reset by
2076 MUTEX_ENTER(&rx_freeCallQueue_lock);
2078 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2080 * EXCEPT that the TQ might not yet be cleared out.
2081 * Skip over those with in-use TQs.
2084 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2085 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2091 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2092 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2093 call = queue_First(&rx_freeCallQueue, rx_call);
2094 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2096 MUTEX_ENTER(&rx_stats_mutex);
2097 rx_stats.nFreeCallStructs--;
2098 MUTEX_EXIT(&rx_stats_mutex);
2099 MUTEX_EXIT(&rx_freeCallQueue_lock);
2100 MUTEX_ENTER(&call->lock);
2101 CLEAR_CALL_QUEUE_LOCK(call);
2102 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2103 /* Now, if TQ wasn't cleared earlier, do it now. */
2104 if (call->flags & RX_CALL_TQ_CLEARME) {
2105 rxi_ClearTransmitQueue(call, 0);
2106 queue_Init(&call->tq);
2108 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2109 /* Bind the call to its connection structure */
2111 rxi_ResetCall(call, 1);
2113 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2115 MUTEX_EXIT(&rx_freeCallQueue_lock);
2116 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2117 MUTEX_ENTER(&call->lock);
2118 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2119 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2120 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2122 MUTEX_ENTER(&rx_stats_mutex);
2123 rx_stats.nCallStructs++;
2124 MUTEX_EXIT(&rx_stats_mutex);
2125 /* Initialize once-only items */
2126 queue_Init(&call->tq);
2127 queue_Init(&call->rq);
2128 queue_Init(&call->iovq);
2129 /* Bind the call to its connection structure (prereq for reset) */
2131 rxi_ResetCall(call, 1);
2133 call->channel = channel;
2134 call->callNumber = &conn->callNumber[channel];
2135 /* Note that the next expected call number is retained (in
2136 * conn->callNumber[i]), even if we reallocate the call structure
2138 conn->call[channel] = call;
2139 /* if the channel's never been used (== 0), we should start at 1, otherwise
2140 * the call number is valid from the last time this channel was used */
2141 if (*call->callNumber == 0)
2142 *call->callNumber = 1;
2147 /* A call has been inactive long enough that so we can throw away
2148 * state, including the call structure, which is placed on the call
2150 * Call is locked upon entry.
2151 * haveCTLock set if called from rxi_ReapConnections
2153 #ifdef RX_ENABLE_LOCKS
2155 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2156 #else /* RX_ENABLE_LOCKS */
2158 rxi_FreeCall(register struct rx_call *call)
2159 #endif /* RX_ENABLE_LOCKS */
2161 register int channel = call->channel;
2162 register struct rx_connection *conn = call->conn;
2165 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2166 (*call->callNumber)++;
2167 rxi_ResetCall(call, 0);
2168 call->conn->call[channel] = (struct rx_call *)0;
2170 MUTEX_ENTER(&rx_freeCallQueue_lock);
2171 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2172 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2173 /* A call may be free even though its transmit queue is still in use.
2174 * Since we search the call list from head to tail, put busy calls at
2175 * the head of the list, and idle calls at the tail.
2177 if (call->flags & RX_CALL_TQ_BUSY)
2178 queue_Prepend(&rx_freeCallQueue, call);
2180 queue_Append(&rx_freeCallQueue, call);
2181 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2182 queue_Append(&rx_freeCallQueue, call);
2183 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2184 MUTEX_ENTER(&rx_stats_mutex);
2185 rx_stats.nFreeCallStructs++;
2186 MUTEX_EXIT(&rx_stats_mutex);
2188 MUTEX_EXIT(&rx_freeCallQueue_lock);
2190 /* Destroy the connection if it was previously slated for
2191 * destruction, i.e. the Rx client code previously called
2192 * rx_DestroyConnection (client connections), or
2193 * rxi_ReapConnections called the same routine (server
2194 * connections). Only do this, however, if there are no
2195 * outstanding calls. Note that for fine grain locking, there appears
2196 * to be a deadlock in that rxi_FreeCall has a call locked and
2197 * DestroyConnectionNoLock locks each call in the conn. But note a
2198 * few lines up where we have removed this call from the conn.
2199 * If someone else destroys a connection, they either have no
2200 * call lock held or are going through this section of code.
2202 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2203 MUTEX_ENTER(&conn->conn_data_lock);
2205 MUTEX_EXIT(&conn->conn_data_lock);
2206 #ifdef RX_ENABLE_LOCKS
2208 rxi_DestroyConnectionNoLock(conn);
2210 rxi_DestroyConnection(conn);
2211 #else /* RX_ENABLE_LOCKS */
2212 rxi_DestroyConnection(conn);
2213 #endif /* RX_ENABLE_LOCKS */
2217 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2219 rxi_Alloc(register size_t size)
2223 MUTEX_ENTER(&rx_stats_mutex);
2225 rxi_Allocsize += (afs_int32)size;
2226 MUTEX_EXIT(&rx_stats_mutex);
2228 p = (char *)osi_Alloc(size);
2231 osi_Panic("rxi_Alloc error");
2237 rxi_Free(void *addr, register size_t size)
2239 MUTEX_ENTER(&rx_stats_mutex);
2241 rxi_Allocsize -= (afs_int32)size;
2242 MUTEX_EXIT(&rx_stats_mutex);
2244 osi_Free(addr, size);
2247 /* Find the peer process represented by the supplied (host,port)
2248 * combination. If there is no appropriate active peer structure, a
2249 * new one will be allocated and initialized
2250 * The origPeer, if set, is a pointer to a peer structure on which the
2251 * refcount will be be decremented. This is used to replace the peer
2252 * structure hanging off a connection structure */
2254 rxi_FindPeer(register afs_uint32 host, register u_short port,
2255 struct rx_peer *origPeer, int create)
2257 register struct rx_peer *pp;
2259 hashIndex = PEER_HASH(host, port);
2260 MUTEX_ENTER(&rx_peerHashTable_lock);
2261 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2262 if ((pp->host == host) && (pp->port == port))
2267 pp = rxi_AllocPeer(); /* This bzero's *pp */
2268 pp->host = host; /* set here or in InitPeerParams is zero */
2270 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2271 queue_Init(&pp->congestionQueue);
2272 queue_Init(&pp->rpcStats);
2273 pp->next = rx_peerHashTable[hashIndex];
2274 rx_peerHashTable[hashIndex] = pp;
2275 rxi_InitPeerParams(pp);
2276 MUTEX_ENTER(&rx_stats_mutex);
2277 rx_stats.nPeerStructs++;
2278 MUTEX_EXIT(&rx_stats_mutex);
2285 origPeer->refCount--;
2286 MUTEX_EXIT(&rx_peerHashTable_lock);
2291 /* Find the connection at (host, port) started at epoch, and with the
2292 * given connection id. Creates the server connection if necessary.
2293 * The type specifies whether a client connection or a server
2294 * connection is desired. In both cases, (host, port) specify the
2295 * peer's (host, pair) pair. Client connections are not made
2296 * automatically by this routine. The parameter socket gives the
2297 * socket descriptor on which the packet was received. This is used,
2298 * in the case of server connections, to check that *new* connections
2299 * come via a valid (port, serviceId). Finally, the securityIndex
2300 * parameter must match the existing index for the connection. If a
2301 * server connection is created, it will be created using the supplied
2302 * index, if the index is valid for this service */
2303 struct rx_connection *
2304 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2305 register u_short port, u_short serviceId, afs_uint32 cid,
2306 afs_uint32 epoch, int type, u_int securityIndex)
2308 int hashindex, flag;
2309 register struct rx_connection *conn;
2310 hashindex = CONN_HASH(host, port, cid, epoch, type);
2311 MUTEX_ENTER(&rx_connHashTable_lock);
2312 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2313 rx_connHashTable[hashindex],
2316 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2317 && (epoch == conn->epoch)) {
2318 register struct rx_peer *pp = conn->peer;
2319 if (securityIndex != conn->securityIndex) {
2320 /* this isn't supposed to happen, but someone could forge a packet
2321 * like this, and there seems to be some CM bug that makes this
2322 * happen from time to time -- in which case, the fileserver
2324 MUTEX_EXIT(&rx_connHashTable_lock);
2325 return (struct rx_connection *)0;
2327 if (pp->host == host && pp->port == port)
2329 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2331 /* So what happens when it's a callback connection? */
2332 if ( /*type == RX_CLIENT_CONNECTION && */
2333 (conn->epoch & 0x80000000))
2337 /* the connection rxLastConn that was used the last time is not the
2338 ** one we are looking for now. Hence, start searching in the hash */
2340 conn = rx_connHashTable[hashindex];
2345 struct rx_service *service;
2346 if (type == RX_CLIENT_CONNECTION) {
2347 MUTEX_EXIT(&rx_connHashTable_lock);
2348 return (struct rx_connection *)0;
2350 service = rxi_FindService(socket, serviceId);
2351 if (!service || (securityIndex >= service->nSecurityObjects)
2352 || (service->securityObjects[securityIndex] == 0)) {
2353 MUTEX_EXIT(&rx_connHashTable_lock);
2354 return (struct rx_connection *)0;
2356 conn = rxi_AllocConnection(); /* This bzero's the connection */
2357 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2358 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2359 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2360 conn->next = rx_connHashTable[hashindex];
2361 rx_connHashTable[hashindex] = conn;
2362 conn->peer = rxi_FindPeer(host, port, 0, 1);
2363 conn->type = RX_SERVER_CONNECTION;
2364 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2365 conn->epoch = epoch;
2366 conn->cid = cid & RX_CIDMASK;
2367 /* conn->serial = conn->lastSerial = 0; */
2368 /* conn->timeout = 0; */
2369 conn->ackRate = RX_FAST_ACK_RATE;
2370 conn->service = service;
2371 conn->serviceId = serviceId;
2372 conn->securityIndex = securityIndex;
2373 conn->securityObject = service->securityObjects[securityIndex];
2374 conn->nSpecific = 0;
2375 conn->specific = NULL;
2376 rx_SetConnDeadTime(conn, service->connDeadTime);
2377 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2378 /* Notify security object of the new connection */
2379 RXS_NewConnection(conn->securityObject, conn);
2380 /* XXXX Connection timeout? */
2381 if (service->newConnProc)
2382 (*service->newConnProc) (conn);
2383 MUTEX_ENTER(&rx_stats_mutex);
2384 rx_stats.nServerConns++;
2385 MUTEX_EXIT(&rx_stats_mutex);
2388 MUTEX_ENTER(&conn->conn_data_lock);
2390 MUTEX_EXIT(&conn->conn_data_lock);
2392 rxLastConn = conn; /* store this connection as the last conn used */
2393 MUTEX_EXIT(&rx_connHashTable_lock);
2397 /* There are two packet tracing routines available for testing and monitoring
2398 * Rx. One is called just after every packet is received and the other is
2399 * called just before every packet is sent. Received packets, have had their
2400 * headers decoded, and packets to be sent have not yet had their headers
2401 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2402 * containing the network address. Both can be modified. The return value, if
2403 * non-zero, indicates that the packet should be dropped. */
2405 int (*rx_justReceived) () = 0;
2406 int (*rx_almostSent) () = 0;
2408 /* A packet has been received off the interface. Np is the packet, socket is
2409 * the socket number it was received from (useful in determining which service
2410 * this packet corresponds to), and (host, port) reflect the host,port of the
2411 * sender. This call returns the packet to the caller if it is finished with
2412 * it, rather than de-allocating it, just as a small performance hack */
2415 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2416 afs_uint32 host, u_short port, int *tnop,
2417 struct rx_call **newcallp)
2419 register struct rx_call *call;
2420 register struct rx_connection *conn;
2422 afs_uint32 currentCallNumber;
2428 struct rx_packet *tnp;
2431 /* We don't print out the packet until now because (1) the time may not be
2432 * accurate enough until now in the lwp implementation (rx_Listener only gets
2433 * the time after the packet is read) and (2) from a protocol point of view,
2434 * this is the first time the packet has been seen */
2435 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2436 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2437 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2438 np->header.serial, packetType, ntohl(host), htohs(port), np->header.serviceId,
2439 np->header.epoch, np->header.cid, np->header.callNumber,
2440 np->header.seq, np->header.flags, np));
2443 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2444 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2447 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2448 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2451 /* If an input tracer function is defined, call it with the packet and
2452 * network address. Note this function may modify its arguments. */
2453 if (rx_justReceived) {
2454 struct sockaddr_in addr;
2456 addr.sin_family = AF_INET;
2457 addr.sin_port = port;
2458 addr.sin_addr.s_addr = host;
2459 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2460 addr.sin_len = sizeof(addr);
2461 #endif /* AFS_OSF_ENV */
2462 drop = (*rx_justReceived) (np, &addr);
2463 /* drop packet if return value is non-zero */
2466 port = addr.sin_port; /* in case fcn changed addr */
2467 host = addr.sin_addr.s_addr;
2471 /* If packet was not sent by the client, then *we* must be the client */
2472 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2473 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2475 /* Find the connection (or fabricate one, if we're the server & if
2476 * necessary) associated with this packet */
2478 rxi_FindConnection(socket, host, port, np->header.serviceId,
2479 np->header.cid, np->header.epoch, type,
2480 np->header.securityIndex);
2483 /* If no connection found or fabricated, just ignore the packet.
2484 * (An argument could be made for sending an abort packet for
2489 MUTEX_ENTER(&conn->conn_data_lock);
2490 if (conn->maxSerial < np->header.serial)
2491 conn->maxSerial = np->header.serial;
2492 MUTEX_EXIT(&conn->conn_data_lock);
2494 /* If the connection is in an error state, send an abort packet and ignore
2495 * the incoming packet */
2497 /* Don't respond to an abort packet--we don't want loops! */
2498 MUTEX_ENTER(&conn->conn_data_lock);
2499 if (np->header.type != RX_PACKET_TYPE_ABORT)
2500 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2502 MUTEX_EXIT(&conn->conn_data_lock);
2506 /* Check for connection-only requests (i.e. not call specific). */
2507 if (np->header.callNumber == 0) {
2508 switch (np->header.type) {
2509 case RX_PACKET_TYPE_ABORT: {
2510 /* What if the supplied error is zero? */
2511 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2512 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2513 rxi_ConnectionError(conn, errcode);
2514 MUTEX_ENTER(&conn->conn_data_lock);
2516 MUTEX_EXIT(&conn->conn_data_lock);
2519 case RX_PACKET_TYPE_CHALLENGE:
2520 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2521 MUTEX_ENTER(&conn->conn_data_lock);
2523 MUTEX_EXIT(&conn->conn_data_lock);
2525 case RX_PACKET_TYPE_RESPONSE:
2526 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2527 MUTEX_ENTER(&conn->conn_data_lock);
2529 MUTEX_EXIT(&conn->conn_data_lock);
2531 case RX_PACKET_TYPE_PARAMS:
2532 case RX_PACKET_TYPE_PARAMS + 1:
2533 case RX_PACKET_TYPE_PARAMS + 2:
2534 /* ignore these packet types for now */
2535 MUTEX_ENTER(&conn->conn_data_lock);
2537 MUTEX_EXIT(&conn->conn_data_lock);
2542 /* Should not reach here, unless the peer is broken: send an
2544 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2545 MUTEX_ENTER(&conn->conn_data_lock);
2546 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2548 MUTEX_EXIT(&conn->conn_data_lock);
2553 channel = np->header.cid & RX_CHANNELMASK;
2554 call = conn->call[channel];
2555 #ifdef RX_ENABLE_LOCKS
2557 MUTEX_ENTER(&call->lock);
2558 /* Test to see if call struct is still attached to conn. */
2559 if (call != conn->call[channel]) {
2561 MUTEX_EXIT(&call->lock);
2562 if (type == RX_SERVER_CONNECTION) {
2563 call = conn->call[channel];
2564 /* If we started with no call attached and there is one now,
2565 * another thread is also running this routine and has gotten
2566 * the connection channel. We should drop this packet in the tests
2567 * below. If there was a call on this connection and it's now
2568 * gone, then we'll be making a new call below.
2569 * If there was previously a call and it's now different then
2570 * the old call was freed and another thread running this routine
2571 * has created a call on this channel. One of these two threads
2572 * has a packet for the old call and the code below handles those
2576 MUTEX_ENTER(&call->lock);
2578 /* This packet can't be for this call. If the new call address is
2579 * 0 then no call is running on this channel. If there is a call
2580 * then, since this is a client connection we're getting data for
2581 * it must be for the previous call.
2583 MUTEX_ENTER(&rx_stats_mutex);
2584 rx_stats.spuriousPacketsRead++;
2585 MUTEX_EXIT(&rx_stats_mutex);
2586 MUTEX_ENTER(&conn->conn_data_lock);
2588 MUTEX_EXIT(&conn->conn_data_lock);
2593 currentCallNumber = conn->callNumber[channel];
2595 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2596 if (np->header.callNumber < currentCallNumber) {
2597 MUTEX_ENTER(&rx_stats_mutex);
2598 rx_stats.spuriousPacketsRead++;
2599 MUTEX_EXIT(&rx_stats_mutex);
2600 #ifdef RX_ENABLE_LOCKS
2602 MUTEX_EXIT(&call->lock);
2604 MUTEX_ENTER(&conn->conn_data_lock);
2606 MUTEX_EXIT(&conn->conn_data_lock);
2610 MUTEX_ENTER(&conn->conn_call_lock);
2611 call = rxi_NewCall(conn, channel);
2612 MUTEX_EXIT(&conn->conn_call_lock);
2613 *call->callNumber = np->header.callNumber;
2614 if (np->header.callNumber == 0)
2615 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));
2617 call->state = RX_STATE_PRECALL;
2618 clock_GetTime(&call->queueTime);
2619 hzero(call->bytesSent);
2620 hzero(call->bytesRcvd);
2622 * If the number of queued calls exceeds the overload
2623 * threshold then abort this call.
2625 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2626 struct rx_packet *tp;
2628 rxi_CallError(call, rx_BusyError);
2629 tp = rxi_SendCallAbort(call, np, 1, 0);
2630 MUTEX_EXIT(&call->lock);
2631 MUTEX_ENTER(&conn->conn_data_lock);
2633 MUTEX_EXIT(&conn->conn_data_lock);
2634 MUTEX_ENTER(&rx_stats_mutex);
2636 MUTEX_EXIT(&rx_stats_mutex);
2639 rxi_KeepAliveOn(call);
2640 } else if (np->header.callNumber != currentCallNumber) {
2641 /* Wait until the transmit queue is idle before deciding
2642 * whether to reset the current call. Chances are that the
2643 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2646 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2647 while ((call->state == RX_STATE_ACTIVE)
2648 && (call->flags & RX_CALL_TQ_BUSY)) {
2649 call->flags |= RX_CALL_TQ_WAIT;
2651 #ifdef RX_ENABLE_LOCKS
2652 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2653 CV_WAIT(&call->cv_tq, &call->lock);
2654 #else /* RX_ENABLE_LOCKS */
2655 osi_rxSleep(&call->tq);
2656 #endif /* RX_ENABLE_LOCKS */
2658 if (call->tqWaiters == 0)
2659 call->flags &= ~RX_CALL_TQ_WAIT;
2661 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2662 /* If the new call cannot be taken right now send a busy and set
2663 * the error condition in this call, so that it terminates as
2664 * quickly as possible */
2665 if (call->state == RX_STATE_ACTIVE) {
2666 struct rx_packet *tp;
2668 rxi_CallError(call, RX_CALL_DEAD);
2669 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2671 MUTEX_EXIT(&call->lock);
2672 MUTEX_ENTER(&conn->conn_data_lock);
2674 MUTEX_EXIT(&conn->conn_data_lock);
2677 rxi_ResetCall(call, 0);
2678 *call->callNumber = np->header.callNumber;
2679 if (np->header.callNumber == 0)
2680 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));
2682 call->state = RX_STATE_PRECALL;
2683 clock_GetTime(&call->queueTime);
2684 hzero(call->bytesSent);
2685 hzero(call->bytesRcvd);
2687 * If the number of queued calls exceeds the overload
2688 * threshold then abort this call.
2690 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2691 struct rx_packet *tp;
2693 rxi_CallError(call, rx_BusyError);
2694 tp = rxi_SendCallAbort(call, np, 1, 0);
2695 MUTEX_EXIT(&call->lock);
2696 MUTEX_ENTER(&conn->conn_data_lock);
2698 MUTEX_EXIT(&conn->conn_data_lock);
2699 MUTEX_ENTER(&rx_stats_mutex);
2701 MUTEX_EXIT(&rx_stats_mutex);
2704 rxi_KeepAliveOn(call);
2706 /* Continuing call; do nothing here. */
2708 } else { /* we're the client */
2709 /* Ignore all incoming acknowledgements for calls in DALLY state */
2710 if (call && (call->state == RX_STATE_DALLY)
2711 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2712 MUTEX_ENTER(&rx_stats_mutex);
2713 rx_stats.ignorePacketDally++;
2714 MUTEX_EXIT(&rx_stats_mutex);
2715 #ifdef RX_ENABLE_LOCKS
2717 MUTEX_EXIT(&call->lock);
2720 MUTEX_ENTER(&conn->conn_data_lock);
2722 MUTEX_EXIT(&conn->conn_data_lock);
2726 /* Ignore anything that's not relevant to the current call. If there
2727 * isn't a current call, then no packet is relevant. */
2728 if (!call || (np->header.callNumber != currentCallNumber)) {
2729 MUTEX_ENTER(&rx_stats_mutex);
2730 rx_stats.spuriousPacketsRead++;
2731 MUTEX_EXIT(&rx_stats_mutex);
2732 #ifdef RX_ENABLE_LOCKS
2734 MUTEX_EXIT(&call->lock);
2737 MUTEX_ENTER(&conn->conn_data_lock);
2739 MUTEX_EXIT(&conn->conn_data_lock);
2742 /* If the service security object index stamped in the packet does not
2743 * match the connection's security index, ignore the packet */
2744 if (np->header.securityIndex != conn->securityIndex) {
2745 #ifdef RX_ENABLE_LOCKS
2746 MUTEX_EXIT(&call->lock);
2748 MUTEX_ENTER(&conn->conn_data_lock);
2750 MUTEX_EXIT(&conn->conn_data_lock);
2754 /* If we're receiving the response, then all transmit packets are
2755 * implicitly acknowledged. Get rid of them. */
2756 if (np->header.type == RX_PACKET_TYPE_DATA) {
2757 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2758 /* XXX Hack. Because we must release the global rx lock when
2759 * sending packets (osi_NetSend) we drop all acks while we're
2760 * traversing the tq in rxi_Start sending packets out because
2761 * packets may move to the freePacketQueue as result of being here!
2762 * So we drop these packets until we're safely out of the
2763 * traversing. Really ugly!
2764 * For fine grain RX locking, we set the acked field in the
2765 * packets and let rxi_Start remove them from the transmit queue.
2767 if (call->flags & RX_CALL_TQ_BUSY) {
2768 #ifdef RX_ENABLE_LOCKS
2769 rxi_SetAcksInTransmitQueue(call);
2772 return np; /* xmitting; drop packet */
2775 rxi_ClearTransmitQueue(call, 0);
2777 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2778 rxi_ClearTransmitQueue(call, 0);
2779 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2781 if (np->header.type == RX_PACKET_TYPE_ACK) {
2782 /* now check to see if this is an ack packet acknowledging that the
2783 * server actually *lost* some hard-acked data. If this happens we
2784 * ignore this packet, as it may indicate that the server restarted in
2785 * the middle of a call. It is also possible that this is an old ack
2786 * packet. We don't abort the connection in this case, because this
2787 * *might* just be an old ack packet. The right way to detect a server
2788 * restart in the midst of a call is to notice that the server epoch
2790 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2791 * XXX unacknowledged. I think that this is off-by-one, but
2792 * XXX I don't dare change it just yet, since it will
2793 * XXX interact badly with the server-restart detection
2794 * XXX code in receiveackpacket. */
2795 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2796 MUTEX_ENTER(&rx_stats_mutex);
2797 rx_stats.spuriousPacketsRead++;
2798 MUTEX_EXIT(&rx_stats_mutex);
2799 MUTEX_EXIT(&call->lock);
2800 MUTEX_ENTER(&conn->conn_data_lock);
2802 MUTEX_EXIT(&conn->conn_data_lock);
2806 } /* else not a data packet */
2809 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2810 /* Set remote user defined status from packet */
2811 call->remoteStatus = np->header.userStatus;
2813 /* Note the gap between the expected next packet and the actual
2814 * packet that arrived, when the new packet has a smaller serial number
2815 * than expected. Rioses frequently reorder packets all by themselves,
2816 * so this will be quite important with very large window sizes.
2817 * Skew is checked against 0 here to avoid any dependence on the type of
2818 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2820 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2821 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2822 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2824 MUTEX_ENTER(&conn->conn_data_lock);
2825 skew = conn->lastSerial - np->header.serial;
2826 conn->lastSerial = np->header.serial;
2827 MUTEX_EXIT(&conn->conn_data_lock);
2829 register struct rx_peer *peer;
2831 if (skew > peer->inPacketSkew) {
2832 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2834 peer->inPacketSkew = skew;
2838 /* Now do packet type-specific processing */
2839 switch (np->header.type) {
2840 case RX_PACKET_TYPE_DATA:
2841 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2844 case RX_PACKET_TYPE_ACK:
2845 /* Respond immediately to ack packets requesting acknowledgement
2847 if (np->header.flags & RX_REQUEST_ACK) {
2849 (void)rxi_SendCallAbort(call, 0, 1, 0);
2851 (void)rxi_SendAck(call, 0, np->header.serial,
2852 RX_ACK_PING_RESPONSE, 1);
2854 np = rxi_ReceiveAckPacket(call, np, 1);
2856 case RX_PACKET_TYPE_ABORT: {
2857 /* An abort packet: reset the call, passing the error up to the user. */
2858 /* What if error is zero? */
2859 /* What if the error is -1? the application will treat it as a timeout. */
2860 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
2861 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
2862 rxi_CallError(call, errdata);
2863 MUTEX_EXIT(&call->lock);
2864 MUTEX_ENTER(&conn->conn_data_lock);
2866 MUTEX_EXIT(&conn->conn_data_lock);
2867 return np; /* xmitting; drop packet */
2869 case RX_PACKET_TYPE_BUSY:
2872 case RX_PACKET_TYPE_ACKALL:
2873 /* All packets acknowledged, so we can drop all packets previously
2874 * readied for sending */
2875 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2876 /* XXX Hack. We because we can't release the global rx lock when
2877 * sending packets (osi_NetSend) we drop all ack pkts while we're
2878 * traversing the tq in rxi_Start sending packets out because
2879 * packets may move to the freePacketQueue as result of being
2880 * here! So we drop these packets until we're safely out of the
2881 * traversing. Really ugly!
2882 * For fine grain RX locking, we set the acked field in the packets
2883 * and let rxi_Start remove the packets from the transmit queue.
2885 if (call->flags & RX_CALL_TQ_BUSY) {
2886 #ifdef RX_ENABLE_LOCKS
2887 rxi_SetAcksInTransmitQueue(call);
2889 #else /* RX_ENABLE_LOCKS */
2890 MUTEX_EXIT(&call->lock);
2891 MUTEX_ENTER(&conn->conn_data_lock);
2893 MUTEX_EXIT(&conn->conn_data_lock);
2894 return np; /* xmitting; drop packet */
2895 #endif /* RX_ENABLE_LOCKS */
2897 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2898 rxi_ClearTransmitQueue(call, 0);
2901 /* Should not reach here, unless the peer is broken: send an abort
2903 rxi_CallError(call, RX_PROTOCOL_ERROR);
2904 np = rxi_SendCallAbort(call, np, 1, 0);
2907 /* Note when this last legitimate packet was received, for keep-alive
2908 * processing. Note, we delay getting the time until now in the hope that
2909 * the packet will be delivered to the user before any get time is required
2910 * (if not, then the time won't actually be re-evaluated here). */
2911 call->lastReceiveTime = clock_Sec();
2912 MUTEX_EXIT(&call->lock);
2913 MUTEX_ENTER(&conn->conn_data_lock);
2915 MUTEX_EXIT(&conn->conn_data_lock);
2919 /* return true if this is an "interesting" connection from the point of view
2920 of someone trying to debug the system */
2922 rxi_IsConnInteresting(struct rx_connection *aconn)
2925 register struct rx_call *tcall;
2927 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2929 for (i = 0; i < RX_MAXCALLS; i++) {
2930 tcall = aconn->call[i];
2932 if ((tcall->state == RX_STATE_PRECALL)
2933 || (tcall->state == RX_STATE_ACTIVE))
2935 if ((tcall->mode == RX_MODE_SENDING)
2936 || (tcall->mode == RX_MODE_RECEIVING))
2944 /* if this is one of the last few packets AND it wouldn't be used by the
2945 receiving call to immediately satisfy a read request, then drop it on
2946 the floor, since accepting it might prevent a lock-holding thread from
2947 making progress in its reading. If a call has been cleared while in
2948 the precall state then ignore all subsequent packets until the call
2949 is assigned to a thread. */
2952 TooLow(struct rx_packet *ap, struct rx_call *acall)
2955 MUTEX_ENTER(&rx_stats_mutex);
2956 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
2957 && (acall->state == RX_STATE_PRECALL))
2958 || ((rx_nFreePackets < rxi_dataQuota + 2)
2959 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
2960 && (acall->flags & RX_CALL_READER_WAIT)))) {
2963 MUTEX_EXIT(&rx_stats_mutex);
2969 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
2970 struct rx_call *acall)
2972 struct rx_call *call = acall;
2976 MUTEX_ENTER(&conn->conn_data_lock);
2977 conn->checkReachEvent = NULL;
2978 waiting = conn->flags & RX_CONN_ATTACHWAIT;
2981 MUTEX_EXIT(&conn->conn_data_lock);
2985 MUTEX_ENTER(&conn->conn_call_lock);
2986 MUTEX_ENTER(&conn->conn_data_lock);
2987 for (i = 0; i < RX_MAXCALLS; i++) {
2988 struct rx_call *tc = conn->call[i];
2989 if (tc && tc->state == RX_STATE_PRECALL) {
2995 /* Indicate that rxi_CheckReachEvent is no longer running by
2996 * clearing the flag. Must be atomic under conn_data_lock to
2997 * avoid a new call slipping by: rxi_CheckConnReach holds
2998 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3000 conn->flags &= ~RX_CONN_ATTACHWAIT;
3001 MUTEX_EXIT(&conn->conn_data_lock);
3002 MUTEX_EXIT(&conn->conn_call_lock);
3007 MUTEX_ENTER(&call->lock);
3008 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3010 MUTEX_EXIT(&call->lock);
3012 clock_GetTime(&when);
3013 when.sec += RX_CHECKREACH_TIMEOUT;
3014 MUTEX_ENTER(&conn->conn_data_lock);
3015 if (!conn->checkReachEvent) {
3017 conn->checkReachEvent =
3018 rxevent_Post(&when, rxi_CheckReachEvent, conn, NULL);
3020 MUTEX_EXIT(&conn->conn_data_lock);
3026 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3028 struct rx_service *service = conn->service;
3029 struct rx_peer *peer = conn->peer;
3030 afs_uint32 now, lastReach;
3032 if (service->checkReach == 0)
3036 MUTEX_ENTER(&peer->peer_lock);
3037 lastReach = peer->lastReachTime;
3038 MUTEX_EXIT(&peer->peer_lock);
3039 if (now - lastReach < RX_CHECKREACH_TTL)
3042 MUTEX_ENTER(&conn->conn_data_lock);
3043 if (conn->flags & RX_CONN_ATTACHWAIT) {
3044 MUTEX_EXIT(&conn->conn_data_lock);
3047 conn->flags |= RX_CONN_ATTACHWAIT;
3048 MUTEX_EXIT(&conn->conn_data_lock);
3049 if (!conn->checkReachEvent)
3050 rxi_CheckReachEvent(NULL, conn, call);
3055 /* try to attach call, if authentication is complete */
3057 TryAttach(register struct rx_call *acall, register osi_socket socket,
3058 register int *tnop, register struct rx_call **newcallp,
3061 struct rx_connection *conn = acall->conn;
3063 if (conn->type == RX_SERVER_CONNECTION
3064 && acall->state == RX_STATE_PRECALL) {
3065 /* Don't attach until we have any req'd. authentication. */
3066 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3067 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3068 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3069 /* Note: this does not necessarily succeed; there
3070 * may not any proc available
3073 rxi_ChallengeOn(acall->conn);
3078 /* A data packet has been received off the interface. This packet is
3079 * appropriate to the call (the call is in the right state, etc.). This
3080 * routine can return a packet to the caller, for re-use */
3083 rxi_ReceiveDataPacket(register struct rx_call *call,
3084 register struct rx_packet *np, int istack,
3085 osi_socket socket, afs_uint32 host, u_short port,
3086 int *tnop, struct rx_call **newcallp)
3088 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3092 afs_uint32 seq, serial, flags;
3094 struct rx_packet *tnp;
3096 MUTEX_ENTER(&rx_stats_mutex);
3097 rx_stats.dataPacketsRead++;
3098 MUTEX_EXIT(&rx_stats_mutex);
3101 /* If there are no packet buffers, drop this new packet, unless we can find
3102 * packet buffers from inactive calls */
3104 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3105 MUTEX_ENTER(&rx_freePktQ_lock);
3106 rxi_NeedMorePackets = TRUE;
3107 MUTEX_EXIT(&rx_freePktQ_lock);
3108 MUTEX_ENTER(&rx_stats_mutex);
3109 rx_stats.noPacketBuffersOnRead++;
3110 MUTEX_EXIT(&rx_stats_mutex);
3111 call->rprev = np->header.serial;
3112 rxi_calltrace(RX_TRACE_DROP, call);
3113 dpf(("packet %x dropped on receipt - quota problems", np));
3115 rxi_ClearReceiveQueue(call);
3116 clock_GetTime(&when);
3117 clock_Add(&when, &rx_softAckDelay);
3118 if (!call->delayedAckEvent
3119 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3120 rxevent_Cancel(call->delayedAckEvent, call,
3121 RX_CALL_REFCOUNT_DELAY);
3122 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3123 call->delayedAckEvent =
3124 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3126 /* we've damaged this call already, might as well do it in. */
3132 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3133 * packet is one of several packets transmitted as a single
3134 * datagram. Do not send any soft or hard acks until all packets
3135 * in a jumbogram have been processed. Send negative acks right away.
3137 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3138 /* tnp is non-null when there are more packets in the
3139 * current jumbo gram */
3146 seq = np->header.seq;
3147 serial = np->header.serial;
3148 flags = np->header.flags;
3150 /* If the call is in an error state, send an abort message */
3152 return rxi_SendCallAbort(call, np, istack, 0);
3154 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3155 * AFS 3.5 jumbogram. */
3156 if (flags & RX_JUMBO_PACKET) {
3157 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3162 if (np->header.spare != 0) {
3163 MUTEX_ENTER(&call->conn->conn_data_lock);
3164 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3165 MUTEX_EXIT(&call->conn->conn_data_lock);
3168 /* The usual case is that this is the expected next packet */
3169 if (seq == call->rnext) {
3171 /* Check to make sure it is not a duplicate of one already queued */
3172 if (queue_IsNotEmpty(&call->rq)
3173 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3174 MUTEX_ENTER(&rx_stats_mutex);
3175 rx_stats.dupPacketsRead++;
3176 MUTEX_EXIT(&rx_stats_mutex);
3177 dpf(("packet %x dropped on receipt - duplicate", np));
3178 rxevent_Cancel(call->delayedAckEvent, call,
3179 RX_CALL_REFCOUNT_DELAY);
3180 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3186 /* It's the next packet. Stick it on the receive queue
3187 * for this call. Set newPackets to make sure we wake
3188 * the reader once all packets have been processed */
3189 queue_Prepend(&call->rq, np);
3191 np = NULL; /* We can't use this anymore */
3194 /* If an ack is requested then set a flag to make sure we
3195 * send an acknowledgement for this packet */
3196 if (flags & RX_REQUEST_ACK) {
3197 ackNeeded = RX_ACK_REQUESTED;
3200 /* Keep track of whether we have received the last packet */
3201 if (flags & RX_LAST_PACKET) {
3202 call->flags |= RX_CALL_HAVE_LAST;
3206 /* Check whether we have all of the packets for this call */
3207 if (call->flags & RX_CALL_HAVE_LAST) {
3208 afs_uint32 tseq; /* temporary sequence number */
3209 struct rx_packet *tp; /* Temporary packet pointer */
3210 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3212 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3213 if (tseq != tp->header.seq)
3215 if (tp->header.flags & RX_LAST_PACKET) {
3216 call->flags |= RX_CALL_RECEIVE_DONE;
3223 /* Provide asynchronous notification for those who want it
3224 * (e.g. multi rx) */
3225 if (call->arrivalProc) {
3226 (*call->arrivalProc) (call, call->arrivalProcHandle,
3227 call->arrivalProcArg);
3228 call->arrivalProc = (void (*)())0;
3231 /* Update last packet received */
3234 /* If there is no server process serving this call, grab
3235 * one, if available. We only need to do this once. If a
3236 * server thread is available, this thread becomes a server
3237 * thread and the server thread becomes a listener thread. */
3239 TryAttach(call, socket, tnop, newcallp, 0);
3242 /* This is not the expected next packet. */
3244 /* Determine whether this is a new or old packet, and if it's
3245 * a new one, whether it fits into the current receive window.
3246 * Also figure out whether the packet was delivered in sequence.
3247 * We use the prev variable to determine whether the new packet
3248 * is the successor of its immediate predecessor in the
3249 * receive queue, and the missing flag to determine whether
3250 * any of this packets predecessors are missing. */
3252 afs_uint32 prev; /* "Previous packet" sequence number */
3253 struct rx_packet *tp; /* Temporary packet pointer */
3254 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3255 int missing; /* Are any predecessors missing? */
3257 /* If the new packet's sequence number has been sent to the
3258 * application already, then this is a duplicate */
3259 if (seq < call->rnext) {
3260 MUTEX_ENTER(&rx_stats_mutex);
3261 rx_stats.dupPacketsRead++;
3262 MUTEX_EXIT(&rx_stats_mutex);
3263 rxevent_Cancel(call->delayedAckEvent, call,
3264 RX_CALL_REFCOUNT_DELAY);
3265 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3271 /* If the sequence number is greater than what can be
3272 * accomodated by the current window, then send a negative
3273 * acknowledge and drop the packet */
3274 if ((call->rnext + call->rwind) <= seq) {
3275 rxevent_Cancel(call->delayedAckEvent, call,
3276 RX_CALL_REFCOUNT_DELAY);
3277 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3284 /* Look for the packet in the queue of old received packets */
3285 for (prev = call->rnext - 1, missing =
3286 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3287 /*Check for duplicate packet */
3288 if (seq == tp->header.seq) {
3289 MUTEX_ENTER(&rx_stats_mutex);
3290 rx_stats.dupPacketsRead++;
3291 MUTEX_EXIT(&rx_stats_mutex);
3292 rxevent_Cancel(call->delayedAckEvent, call,
3293 RX_CALL_REFCOUNT_DELAY);
3294 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3300 /* If we find a higher sequence packet, break out and
3301 * insert the new packet here. */
3302 if (seq < tp->header.seq)
3304 /* Check for missing packet */
3305 if (tp->header.seq != prev + 1) {
3309 prev = tp->header.seq;
3312 /* Keep track of whether we have received the last packet. */
3313 if (flags & RX_LAST_PACKET) {
3314 call->flags |= RX_CALL_HAVE_LAST;
3317 /* It's within the window: add it to the the receive queue.
3318 * tp is left by the previous loop either pointing at the
3319 * packet before which to insert the new packet, or at the
3320 * queue head if the queue is empty or the packet should be
3322 queue_InsertBefore(tp, np);
3326 /* Check whether we have all of the packets for this call */
3327 if ((call->flags & RX_CALL_HAVE_LAST)
3328 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3329 afs_uint32 tseq; /* temporary sequence number */
3332 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3333 if (tseq != tp->header.seq)
3335 if (tp->header.flags & RX_LAST_PACKET) {
3336 call->flags |= RX_CALL_RECEIVE_DONE;
3343 /* We need to send an ack of the packet is out of sequence,
3344 * or if an ack was requested by the peer. */
3345 if (seq != prev + 1 || missing || (flags & RX_REQUEST_ACK)) {
3346 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3349 /* Acknowledge the last packet for each call */
3350 if (flags & RX_LAST_PACKET) {
3361 * If the receiver is waiting for an iovec, fill the iovec
3362 * using the data from the receive queue */
3363 if (call->flags & RX_CALL_IOVEC_WAIT) {
3364 didHardAck = rxi_FillReadVec(call, serial);
3365 /* the call may have been aborted */
3374 /* Wakeup the reader if any */
3375 if ((call->flags & RX_CALL_READER_WAIT)
3376 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3377 || (call->iovNext >= call->iovMax)
3378 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3379 call->flags &= ~RX_CALL_READER_WAIT;
3380 #ifdef RX_ENABLE_LOCKS
3381 CV_BROADCAST(&call->cv_rq);
3383 osi_rxWakeup(&call->rq);
3389 * Send an ack when requested by the peer, or once every
3390 * rxi_SoftAckRate packets until the last packet has been
3391 * received. Always send a soft ack for the last packet in
3392 * the server's reply. */
3394 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3395 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3396 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3397 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3398 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3399 } else if (call->nSoftAcks) {
3400 clock_GetTime(&when);
3401 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3402 clock_Add(&when, &rx_lastAckDelay);
3404 clock_Add(&when, &rx_softAckDelay);
3406 if (!call->delayedAckEvent
3407 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3408 rxevent_Cancel(call->delayedAckEvent, call,
3409 RX_CALL_REFCOUNT_DELAY);
3410 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3411 call->delayedAckEvent =
3412 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3414 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3415 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3422 static void rxi_ComputeRate();
3426 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3428 struct rx_peer *peer = conn->peer;
3430 MUTEX_ENTER(&peer->peer_lock);
3431 peer->lastReachTime = clock_Sec();
3432 MUTEX_EXIT(&peer->peer_lock);
3434 MUTEX_ENTER(&conn->conn_data_lock);
3435 if (conn->flags & RX_CONN_ATTACHWAIT) {
3438 conn->flags &= ~RX_CONN_ATTACHWAIT;
3439 MUTEX_EXIT(&conn->conn_data_lock);
3441 for (i = 0; i < RX_MAXCALLS; i++) {
3442 struct rx_call *call = conn->call[i];
3445 MUTEX_ENTER(&call->lock);
3446 /* tnop can be null if newcallp is null */
3447 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3449 MUTEX_EXIT(&call->lock);
3453 MUTEX_EXIT(&conn->conn_data_lock);
3457 rx_ack_reason(int reason)
3460 case RX_ACK_REQUESTED:
3462 case RX_ACK_DUPLICATE:
3464 case RX_ACK_OUT_OF_SEQUENCE:
3466 case RX_ACK_EXCEEDS_WINDOW:
3468 case RX_ACK_NOSPACE:
3472 case RX_ACK_PING_RESPONSE:
3484 /* rxi_ComputePeerNetStats
3486 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3487 * estimates (like RTT and throughput) based on ack packets. Caller
3488 * must ensure that the packet in question is the right one (i.e.
3489 * serial number matches).
3492 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3493 struct rx_ackPacket *ap, struct rx_packet *np)
3495 struct rx_peer *peer = call->conn->peer;
3497 /* Use RTT if not delayed by client. */
3498 if (ap->reason != RX_ACK_DELAY)
3499 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3501 rxi_ComputeRate(peer, call, p, np, ap->reason);
3505 /* The real smarts of the whole thing. */
3507 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3510 struct rx_ackPacket *ap;
3512 register struct rx_packet *tp;
3513 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3514 register struct rx_connection *conn = call->conn;
3515 struct rx_peer *peer = conn->peer;
3518 /* because there are CM's that are bogus, sending weird values for this. */
3519 afs_uint32 skew = 0;
3524 int newAckCount = 0;
3525 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3526 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3528 MUTEX_ENTER(&rx_stats_mutex);
3529 rx_stats.ackPacketsRead++;
3530 MUTEX_EXIT(&rx_stats_mutex);
3531 ap = (struct rx_ackPacket *)rx_DataOf(np);
3532 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3534 return np; /* truncated ack packet */
3536 /* depends on ack packet struct */
3537 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3538 first = ntohl(ap->firstPacket);
3539 serial = ntohl(ap->serial);
3540 /* temporarily disabled -- needs to degrade over time
3541 * skew = ntohs(ap->maxSkew); */
3543 /* Ignore ack packets received out of order */
3544 if (first < call->tfirst) {
3548 if (np->header.flags & RX_SLOW_START_OK) {
3549 call->flags |= RX_CALL_SLOW_START_OK;
3552 if (ap->reason == RX_ACK_PING_RESPONSE)
3553 rxi_UpdatePeerReach(conn, call);
3557 if (rxdebug_active) {
3561 len = _snprintf(msg, sizeof(msg),
3562 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3563 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3564 ntohl(ap->serial), ntohl(ap->previousPacket),
3565 (unsigned int)np->header.seq, (unsigned int)skew,
3566 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3570 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3571 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3575 OutputDebugString(msg);
3577 #else /* AFS_NT40_ENV */
3580 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3581 ap->reason, ntohl(ap->previousPacket),
3582 (unsigned int)np->header.seq, (unsigned int)serial,
3583 (unsigned int)skew, ntohl(ap->firstPacket));
3586 for (offset = 0; offset < nAcks; offset++)
3587 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3592 #endif /* AFS_NT40_ENV */
3595 /* Update the outgoing packet skew value to the latest value of
3596 * the peer's incoming packet skew value. The ack packet, of
3597 * course, could arrive out of order, but that won't affect things
3599 MUTEX_ENTER(&peer->peer_lock);
3600 peer->outPacketSkew = skew;
3602 /* Check for packets that no longer need to be transmitted, and
3603 * discard them. This only applies to packets positively
3604 * acknowledged as having been sent to the peer's upper level.
3605 * All other packets must be retained. So only packets with
3606 * sequence numbers < ap->firstPacket are candidates. */
3607 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3608 if (tp->header.seq >= first)
3610 call->tfirst = tp->header.seq + 1;
3612 && (tp->header.serial == serial || tp->firstSerial == serial))
3613 rxi_ComputePeerNetStats(call, tp, ap, np);
3614 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3615 /* XXX Hack. Because we have to release the global rx lock when sending
3616 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3617 * in rxi_Start sending packets out because packets may move to the
3618 * freePacketQueue as result of being here! So we drop these packets until
3619 * we're safely out of the traversing. Really ugly!
3620 * To make it even uglier, if we're using fine grain locking, we can
3621 * set the ack bits in the packets and have rxi_Start remove the packets
3622 * when it's done transmitting.
3624 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3627 if (call->flags & RX_CALL_TQ_BUSY) {
3628 #ifdef RX_ENABLE_LOCKS
3629 tp->flags |= RX_PKTFLAG_ACKED;
3630 call->flags |= RX_CALL_TQ_SOME_ACKED;
3631 #else /* RX_ENABLE_LOCKS */
3633 #endif /* RX_ENABLE_LOCKS */
3635 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3638 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3643 /* Give rate detector a chance to respond to ping requests */
3644 if (ap->reason == RX_ACK_PING_RESPONSE) {
3645 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3649 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3651 /* Now go through explicit acks/nacks and record the results in
3652 * the waiting packets. These are packets that can't be released
3653 * yet, even with a positive acknowledge. This positive
3654 * acknowledge only means the packet has been received by the
3655 * peer, not that it will be retained long enough to be sent to
3656 * the peer's upper level. In addition, reset the transmit timers
3657 * of any missing packets (those packets that must be missing
3658 * because this packet was out of sequence) */
3660 call->nSoftAcked = 0;
3661 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3662 /* Update round trip time if the ack was stimulated on receipt
3664 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3665 #ifdef RX_ENABLE_LOCKS
3666 if (tp->header.seq >= first)
3667 #endif /* RX_ENABLE_LOCKS */
3668 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3670 && (tp->header.serial == serial || tp->firstSerial == serial))
3671 rxi_ComputePeerNetStats(call, tp, ap, np);
3673 /* Set the acknowledge flag per packet based on the
3674 * information in the ack packet. An acknowlegded packet can
3675 * be downgraded when the server has discarded a packet it
3676 * soacked previously, or when an ack packet is received
3677 * out of sequence. */
3678 if (tp->header.seq < first) {
3679 /* Implicit ack information */
3680 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3683 tp->flags |= RX_PKTFLAG_ACKED;
3684 } else if (tp->header.seq < first + nAcks) {
3685 /* Explicit ack information: set it in the packet appropriately */
3686 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3687 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3689 tp->flags |= RX_PKTFLAG_ACKED;
3696 } else /* RX_ACK_TYPE_NACK */ {
3697 tp->flags &= ~RX_PKTFLAG_ACKED;
3701 tp->flags &= ~RX_PKTFLAG_ACKED;
3705 /* If packet isn't yet acked, and it has been transmitted at least
3706 * once, reset retransmit time using latest timeout
3707 * ie, this should readjust the retransmit timer for all outstanding
3708 * packets... So we don't just retransmit when we should know better*/
3710 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3711 tp->retryTime = tp->timeSent;
3712 clock_Add(&tp->retryTime, &peer->timeout);
3713 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3714 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3718 /* If the window has been extended by this acknowledge packet,
3719 * then wakeup a sender waiting in alloc for window space, or try
3720 * sending packets now, if he's been sitting on packets due to
3721 * lack of window space */
3722 if (call->tnext < (call->tfirst + call->twind)) {
3723 #ifdef RX_ENABLE_LOCKS
3724 CV_SIGNAL(&call->cv_twind);
3726 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3727 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3728 osi_rxWakeup(&call->twind);
3731 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3732 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3736 /* if the ack packet has a receivelen field hanging off it,
3737 * update our state */
3738 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3741 /* If the ack packet has a "recommended" size that is less than
3742 * what I am using now, reduce my size to match */
3743 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3744 (int)sizeof(afs_int32), &tSize);
3745 tSize = (afs_uint32) ntohl(tSize);
3746 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3748 /* Get the maximum packet size to send to this peer */
3749 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
3751 tSize = (afs_uint32) ntohl(tSize);
3752 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3753 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3755 /* sanity check - peer might have restarted with different params.
3756 * If peer says "send less", dammit, send less... Peer should never
3757 * be unable to accept packets of the size that prior AFS versions would
3758 * send without asking. */
3759 if (peer->maxMTU != tSize) {
3760 peer->maxMTU = tSize;
3761 peer->MTU = MIN(tSize, peer->MTU);
3762 call->MTU = MIN(call->MTU, tSize);
3766 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3769 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3770 (int)sizeof(afs_int32), &tSize);
3771 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3772 if (tSize < call->twind) { /* smaller than our send */
3773 call->twind = tSize; /* window, we must send less... */
3774 call->ssthresh = MIN(call->twind, call->ssthresh);
3777 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3778 * network MTU confused with the loopback MTU. Calculate the
3779 * maximum MTU here for use in the slow start code below.
3781 maxMTU = peer->maxMTU;
3782 /* Did peer restart with older RX version? */
3783 if (peer->maxDgramPackets > 1) {
3784 peer->maxDgramPackets = 1;
3786 } else if (np->length >=
3787 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3790 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3791 sizeof(afs_int32), &tSize);
3792 tSize = (afs_uint32) ntohl(tSize);
3794 * As of AFS 3.5 we set the send window to match the receive window.
3796 if (tSize < call->twind) {
3797 call->twind = tSize;
3798 call->ssthresh = MIN(call->twind, call->ssthresh);
3799 } else if (tSize > call->twind) {
3800 call->twind = tSize;
3804 * As of AFS 3.5, a jumbogram is more than one fixed size
3805 * packet transmitted in a single UDP datagram. If the remote
3806 * MTU is smaller than our local MTU then never send a datagram
3807 * larger than the natural MTU.
3810 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3811 sizeof(afs_int32), &tSize);
3812 maxDgramPackets = (afs_uint32) ntohl(tSize);
3813 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3815 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
3816 maxDgramPackets = MIN(maxDgramPackets, tSize);
3817 if (maxDgramPackets > 1) {
3818 peer->maxDgramPackets = maxDgramPackets;
3819 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3821 peer->maxDgramPackets = 1;
3822 call->MTU = peer->natMTU;
3824 } else if (peer->maxDgramPackets > 1) {
3825 /* Restarted with lower version of RX */
3826 peer->maxDgramPackets = 1;
3828 } else if (peer->maxDgramPackets > 1
3829 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3830 /* Restarted with lower version of RX */
3831 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3832 peer->natMTU = OLD_MAX_PACKET_SIZE;
3833 peer->MTU = OLD_MAX_PACKET_SIZE;
3834 peer->maxDgramPackets = 1;
3835 peer->nDgramPackets = 1;
3837 call->MTU = OLD_MAX_PACKET_SIZE;
3842 * Calculate how many datagrams were successfully received after
3843 * the first missing packet and adjust the negative ack counter
3848 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3849 if (call->nNacks < nNacked) {
3850 call->nNacks = nNacked;
3859 if (call->flags & RX_CALL_FAST_RECOVER) {
3861 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3863 call->flags &= ~RX_CALL_FAST_RECOVER;
3864 call->cwind = call->nextCwind;
3865 call->nextCwind = 0;
3868 call->nCwindAcks = 0;
3869 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3870 /* Three negative acks in a row trigger congestion recovery */
3871 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3872 MUTEX_EXIT(&peer->peer_lock);
3873 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3874 /* someone else is waiting to start recovery */
3877 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3878 while (call->flags & RX_CALL_TQ_BUSY) {
3879 call->flags |= RX_CALL_TQ_WAIT;
3881 #ifdef RX_ENABLE_LOCKS
3882 osirx_AssertMine(&call->lock, "rxi_Start lock2");
3883 CV_WAIT(&call->cv_tq, &call->lock);
3884 #else /* RX_ENABLE_LOCKS */
3885 osi_rxSleep(&call->tq);
3886 #endif /* RX_ENABLE_LOCKS */
3888 if (call->tqWaiters == 0)
3889 call->flags &= ~RX_CALL_TQ_WAIT;
3891 MUTEX_ENTER(&peer->peer_lock);
3892 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3893 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3894 call->flags |= RX_CALL_FAST_RECOVER;
3895 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3897 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3898 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3899 call->nextCwind = call->ssthresh;
3902 peer->MTU = call->MTU;
3903 peer->cwind = call->nextCwind;
3904 peer->nDgramPackets = call->nDgramPackets;
3906 call->congestSeq = peer->congestSeq;
3907 /* Reset the resend times on the packets that were nacked
3908 * so we will retransmit as soon as the window permits*/
3909 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3911 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3912 clock_Zero(&tp->retryTime);
3914 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3919 /* If cwind is smaller than ssthresh, then increase
3920 * the window one packet for each ack we receive (exponential
3922 * If cwind is greater than or equal to ssthresh then increase
3923 * the congestion window by one packet for each cwind acks we
3924 * receive (linear growth). */
3925 if (call->cwind < call->ssthresh) {
3927 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3928 call->nCwindAcks = 0;
3930 call->nCwindAcks += newAckCount;
3931 if (call->nCwindAcks >= call->cwind) {
3932 call->nCwindAcks = 0;
3933 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3937 * If we have received several acknowledgements in a row then
3938 * it is time to increase the size of our datagrams
3940 if ((int)call->nAcks > rx_nDgramThreshold) {
3941 if (peer->maxDgramPackets > 1) {
3942 if (call->nDgramPackets < peer->maxDgramPackets) {
3943 call->nDgramPackets++;
3945 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3946 } else if (call->MTU < peer->maxMTU) {
3947 call->MTU += peer->natMTU;
3948 call->MTU = MIN(call->MTU, peer->maxMTU);
3954 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3956 /* Servers need to hold the call until all response packets have
3957 * been acknowledged. Soft acks are good enough since clients
3958 * are not allowed to clear their receive queues. */
3959 if (call->state == RX_STATE_HOLD
3960 && call->tfirst + call->nSoftAcked >= call->tnext) {
3961 call->state = RX_STATE_DALLY;
3962 rxi_ClearTransmitQueue(call, 0);
3963 } else if (!queue_IsEmpty(&call->tq)) {
3964 rxi_Start(0, call, 0, istack);
3969 /* Received a response to a challenge packet */
3971 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
3972 register struct rx_packet *np, int istack)
3976 /* Ignore the packet if we're the client */
3977 if (conn->type == RX_CLIENT_CONNECTION)
3980 /* If already authenticated, ignore the packet (it's probably a retry) */
3981 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
3984 /* Otherwise, have the security object evaluate the response packet */
3985 error = RXS_CheckResponse(conn->securityObject, conn, np);
3987 /* If the response is invalid, reset the connection, sending
3988 * an abort to the peer */
3992 rxi_ConnectionError(conn, error);
3993 MUTEX_ENTER(&conn->conn_data_lock);
3994 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3995 MUTEX_EXIT(&conn->conn_data_lock);
3998 /* If the response is valid, any calls waiting to attach
3999 * servers can now do so */
4002 for (i = 0; i < RX_MAXCALLS; i++) {
4003 struct rx_call *call = conn->call[i];
4005 MUTEX_ENTER(&call->lock);
4006 if (call->state == RX_STATE_PRECALL)
4007 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4008 /* tnop can be null if newcallp is null */
4009 MUTEX_EXIT(&call->lock);
4013 /* Update the peer reachability information, just in case
4014 * some calls went into attach-wait while we were waiting
4015 * for authentication..
4017 rxi_UpdatePeerReach(conn, NULL);
4022 /* A client has received an authentication challenge: the security
4023 * object is asked to cough up a respectable response packet to send
4024 * back to the server. The server is responsible for retrying the
4025 * challenge if it fails to get a response. */
4028 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
4029 register struct rx_packet *np, int istack)
4033 /* Ignore the challenge if we're the server */
4034 if (conn->type == RX_SERVER_CONNECTION)
4037 /* Ignore the challenge if the connection is otherwise idle; someone's
4038 * trying to use us as an oracle. */
4039 if (!rxi_HasActiveCalls(conn))
4042 /* Send the security object the challenge packet. It is expected to fill
4043 * in the response. */
4044 error = RXS_GetResponse(conn->securityObject, conn, np);
4046 /* If the security object is unable to return a valid response, reset the
4047 * connection and send an abort to the peer. Otherwise send the response
4048 * packet to the peer connection. */
4050 rxi_ConnectionError(conn, error);
4051 MUTEX_ENTER(&conn->conn_data_lock);
4052 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4053 MUTEX_EXIT(&conn->conn_data_lock);
4055 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4056 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4062 /* Find an available server process to service the current request in
4063 * the given call structure. If one isn't available, queue up this
4064 * call so it eventually gets one */
4066 rxi_AttachServerProc(register struct rx_call *call,
4067 register osi_socket socket, register int *tnop,
4068 register struct rx_call **newcallp)
4070 register struct rx_serverQueueEntry *sq;
4071 register struct rx_service *service = call->conn->service;
4072 register int haveQuota = 0;
4074 /* May already be attached */
4075 if (call->state == RX_STATE_ACTIVE)
4078 MUTEX_ENTER(&rx_serverPool_lock);
4080 haveQuota = QuotaOK(service);
4081 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4082 /* If there are no processes available to service this call,
4083 * put the call on the incoming call queue (unless it's
4084 * already on the queue).
4086 #ifdef RX_ENABLE_LOCKS
4088 ReturnToServerPool(service);
4089 #endif /* RX_ENABLE_LOCKS */
4091 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4092 call->flags |= RX_CALL_WAIT_PROC;
4093 MUTEX_ENTER(&rx_stats_mutex);
4096 MUTEX_EXIT(&rx_stats_mutex);
4097 rxi_calltrace(RX_CALL_ARRIVAL, call);
4098 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4099 queue_Append(&rx_incomingCallQueue, call);
4102 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4104 /* If hot threads are enabled, and both newcallp and sq->socketp
4105 * are non-null, then this thread will process the call, and the
4106 * idle server thread will start listening on this threads socket.
4109 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4112 *sq->socketp = socket;
4113 clock_GetTime(&call->startTime);
4114 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4118 if (call->flags & RX_CALL_WAIT_PROC) {
4119 /* Conservative: I don't think this should happen */
4120 call->flags &= ~RX_CALL_WAIT_PROC;
4121 if (queue_IsOnQueue(call)) {
4123 MUTEX_ENTER(&rx_stats_mutex);
4125 MUTEX_EXIT(&rx_stats_mutex);
4128 call->state = RX_STATE_ACTIVE;
4129 call->mode = RX_MODE_RECEIVING;
4130 #ifdef RX_KERNEL_TRACE
4132 int glockOwner = ISAFS_GLOCK();
4135 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4136 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4142 if (call->flags & RX_CALL_CLEARED) {
4143 /* send an ack now to start the packet flow up again */
4144 call->flags &= ~RX_CALL_CLEARED;
4145 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4147 #ifdef RX_ENABLE_LOCKS
4150 service->nRequestsRunning++;
4151 if (service->nRequestsRunning <= service->minProcs)
4157 MUTEX_EXIT(&rx_serverPool_lock);
4160 /* Delay the sending of an acknowledge event for a short while, while
4161 * a new call is being prepared (in the case of a client) or a reply
4162 * is being prepared (in the case of a server). Rather than sending
4163 * an ack packet, an ACKALL packet is sent. */
4165 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4167 #ifdef RX_ENABLE_LOCKS
4169 MUTEX_ENTER(&call->lock);
4170 call->delayedAckEvent = NULL;
4171 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4173 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4174 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4176 MUTEX_EXIT(&call->lock);
4177 #else /* RX_ENABLE_LOCKS */
4179 call->delayedAckEvent = NULL;
4180 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4181 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4182 #endif /* RX_ENABLE_LOCKS */
4186 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4189 #ifdef RX_ENABLE_LOCKS
4191 MUTEX_ENTER(&call->lock);
4192 if (event == call->delayedAckEvent)
4193 call->delayedAckEvent = NULL;
4194 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4196 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4198 MUTEX_EXIT(&call->lock);
4199 #else /* RX_ENABLE_LOCKS */
4201 call->delayedAckEvent = NULL;
4202 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4203 #endif /* RX_ENABLE_LOCKS */
4207 #ifdef RX_ENABLE_LOCKS
4208 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4209 * clearing them out.
4212 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4214 register struct rx_packet *p, *tp;
4217 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4218 p->flags |= RX_PKTFLAG_ACKED;
4222 call->flags |= RX_CALL_TQ_CLEARME;
4223 call->flags |= RX_CALL_TQ_SOME_ACKED;
4226 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4227 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4228 call->tfirst = call->tnext;
4229 call->nSoftAcked = 0;
4231 if (call->flags & RX_CALL_FAST_RECOVER) {
4232 call->flags &= ~RX_CALL_FAST_RECOVER;
4233 call->cwind = call->nextCwind;
4234 call->nextCwind = 0;
4237 CV_SIGNAL(&call->cv_twind);
4239 #endif /* RX_ENABLE_LOCKS */
4241 /* Clear out the transmit queue for the current call (all packets have
4242 * been received by peer) */
4244 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4246 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4247 register struct rx_packet *p, *tp;
4249 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4251 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4252 p->flags |= RX_PKTFLAG_ACKED;
4256 call->flags |= RX_CALL_TQ_CLEARME;
4257 call->flags |= RX_CALL_TQ_SOME_ACKED;
4260 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4261 rxi_FreePackets(0, &call->tq);
4262 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4263 call->flags &= ~RX_CALL_TQ_CLEARME;
4265 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4267 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4268 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4269 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4270 call->nSoftAcked = 0;
4272 if (call->flags & RX_CALL_FAST_RECOVER) {
4273 call->flags &= ~RX_CALL_FAST_RECOVER;
4274 call->cwind = call->nextCwind;
4276 #ifdef RX_ENABLE_LOCKS
4277 CV_SIGNAL(&call->cv_twind);
4279 osi_rxWakeup(&call->twind);
4284 rxi_ClearReceiveQueue(register struct rx_call *call)
4286 if (queue_IsNotEmpty(&call->rq)) {
4287 rx_packetReclaims += rxi_FreePackets(0, &call->rq);
4288 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4290 if (call->state == RX_STATE_PRECALL) {
4291 call->flags |= RX_CALL_CLEARED;
4295 /* Send an abort packet for the specified call */
4297 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4298 int istack, int force)
4306 /* Clients should never delay abort messages */
4307 if (rx_IsClientConn(call->conn))
4310 if (call->abortCode != call->error) {
4311 call->abortCode = call->error;
4312 call->abortCount = 0;
4315 if (force || rxi_callAbortThreshhold == 0
4316 || call->abortCount < rxi_callAbortThreshhold) {
4317 if (call->delayedAbortEvent) {
4318 rxevent_Cancel(call->delayedAbortEvent, call,
4319 RX_CALL_REFCOUNT_ABORT);
4321 error = htonl(call->error);
4324 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4325 (char *)&error, sizeof(error), istack);
4326 } else if (!call->delayedAbortEvent) {
4327 clock_GetTime(&when);
4328 clock_Addmsec(&when, rxi_callAbortDelay);
4329 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4330 call->delayedAbortEvent =
4331 rxevent_Post(&when, rxi_SendDelayedCallAbort, call, 0);
4336 /* Send an abort packet for the specified connection. Packet is an
4337 * optional pointer to a packet that can be used to send the abort.
4338 * Once the number of abort messages reaches the threshhold, an
4339 * event is scheduled to send the abort. Setting the force flag
4340 * overrides sending delayed abort messages.
4342 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4343 * to send the abort packet.
4346 rxi_SendConnectionAbort(register struct rx_connection *conn,
4347 struct rx_packet *packet, int istack, int force)
4355 /* Clients should never delay abort messages */
4356 if (rx_IsClientConn(conn))
4359 if (force || rxi_connAbortThreshhold == 0
4360 || conn->abortCount < rxi_connAbortThreshhold) {
4361 if (conn->delayedAbortEvent) {
4362 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4364 error = htonl(conn->error);
4366 MUTEX_EXIT(&conn->conn_data_lock);
4368 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4369 RX_PACKET_TYPE_ABORT, (char *)&error,
4370 sizeof(error), istack);
4371 MUTEX_ENTER(&conn->conn_data_lock);
4372 } else if (!conn->delayedAbortEvent) {
4373 clock_GetTime(&when);
4374 clock_Addmsec(&when, rxi_connAbortDelay);
4375 conn->delayedAbortEvent =
4376 rxevent_Post(&when, rxi_SendDelayedConnAbort, conn, 0);
4381 /* Associate an error all of the calls owned by a connection. Called
4382 * with error non-zero. This is only for really fatal things, like
4383 * bad authentication responses. The connection itself is set in
4384 * error at this point, so that future packets received will be
4387 rxi_ConnectionError(register struct rx_connection *conn,
4388 register afs_int32 error)
4393 dpf(("rxi_ConnectionError conn %x error %d", conn, error));
4395 MUTEX_ENTER(&conn->conn_data_lock);
4396 if (conn->challengeEvent)
4397 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4398 if (conn->checkReachEvent) {
4399 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4400 conn->checkReachEvent = 0;
4401 conn->flags &= ~RX_CONN_ATTACHWAIT;
4404 MUTEX_EXIT(&conn->conn_data_lock);
4405 for (i = 0; i < RX_MAXCALLS; i++) {
4406 struct rx_call *call = conn->call[i];
4408 MUTEX_ENTER(&call->lock);
4409 rxi_CallError(call, error);
4410 MUTEX_EXIT(&call->lock);
4413 conn->error = error;
4414 MUTEX_ENTER(&rx_stats_mutex);
4415 rx_stats.fatalErrors++;
4416 MUTEX_EXIT(&rx_stats_mutex);
4421 rxi_CallError(register struct rx_call *call, afs_int32 error)
4423 dpf(("rxi_CallError call %x error %d call->error %d", call, error, call->error));
4425 error = call->error;
4427 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4428 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4429 rxi_ResetCall(call, 0);
4432 rxi_ResetCall(call, 0);
4434 call->error = error;
4435 call->mode = RX_MODE_ERROR;
4438 /* Reset various fields in a call structure, and wakeup waiting
4439 * processes. Some fields aren't changed: state & mode are not
4440 * touched (these must be set by the caller), and bufptr, nLeft, and
4441 * nFree are not reset, since these fields are manipulated by
4442 * unprotected macros, and may only be reset by non-interrupting code.
4445 /* this code requires that call->conn be set properly as a pre-condition. */
4446 #endif /* ADAPT_WINDOW */
4449 rxi_ResetCall(register struct rx_call *call, register int newcall)
4452 register struct rx_peer *peer;
4453 struct rx_packet *packet;
4455 /* Notify anyone who is waiting for asynchronous packet arrival */
4456 if (call->arrivalProc) {
4457 (*call->arrivalProc) (call, call->arrivalProcHandle,
4458 call->arrivalProcArg);
4459 call->arrivalProc = (void (*)())0;
4462 if (call->delayedAbortEvent) {
4463 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4464 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4466 rxi_SendCallAbort(call, packet, 0, 1);
4467 rxi_FreePacket(packet);
4472 * Update the peer with the congestion information in this call
4473 * so other calls on this connection can pick up where this call
4474 * left off. If the congestion sequence numbers don't match then
4475 * another call experienced a retransmission.
4477 peer = call->conn->peer;
4478 MUTEX_ENTER(&peer->peer_lock);
4480 if (call->congestSeq == peer->congestSeq) {
4481 peer->cwind = MAX(peer->cwind, call->cwind);
4482 peer->MTU = MAX(peer->MTU, call->MTU);
4483 peer->nDgramPackets =
4484 MAX(peer->nDgramPackets, call->nDgramPackets);
4487 call->abortCode = 0;
4488 call->abortCount = 0;
4490 if (peer->maxDgramPackets > 1) {
4491 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4493 call->MTU = peer->MTU;
4495 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4496 call->ssthresh = rx_maxSendWindow;
4497 call->nDgramPackets = peer->nDgramPackets;
4498 call->congestSeq = peer->congestSeq;
4499 MUTEX_EXIT(&peer->peer_lock);
4501 flags = call->flags;
4502 rxi_ClearReceiveQueue(call);
4503 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4504 if (flags & RX_CALL_TQ_BUSY) {
4505 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4506 call->flags |= (flags & RX_CALL_TQ_WAIT);
4508 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4510 rxi_ClearTransmitQueue(call, 0);
4511 queue_Init(&call->tq);
4512 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4513 dpf(("rcall %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4516 while (call->tqWaiters) {
4517 #ifdef RX_ENABLE_LOCKS
4518 CV_BROADCAST(&call->cv_tq);
4519 #else /* RX_ENABLE_LOCKS */
4520 osi_rxWakeup(&call->tq);
4521 #endif /* RX_ENABLE_LOCKS */
4525 queue_Init(&call->rq);
4527 call->rwind = rx_initReceiveWindow;
4528 call->twind = rx_initSendWindow;
4529 call->nSoftAcked = 0;
4530 call->nextCwind = 0;
4533 call->nCwindAcks = 0;
4534 call->nSoftAcks = 0;
4535 call->nHardAcks = 0;
4537 call->tfirst = call->rnext = call->tnext = 1;
4539 call->lastAcked = 0;
4540 call->localStatus = call->remoteStatus = 0;
4542 if (flags & RX_CALL_READER_WAIT) {
4543 #ifdef RX_ENABLE_LOCKS
4544 CV_BROADCAST(&call->cv_rq);
4546 osi_rxWakeup(&call->rq);
4549 if (flags & RX_CALL_WAIT_PACKETS) {
4550 MUTEX_ENTER(&rx_freePktQ_lock);
4551 rxi_PacketsUnWait(); /* XXX */
4552 MUTEX_EXIT(&rx_freePktQ_lock);
4554 #ifdef RX_ENABLE_LOCKS
4555 CV_SIGNAL(&call->cv_twind);
4557 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4558 osi_rxWakeup(&call->twind);
4561 #ifdef RX_ENABLE_LOCKS
4562 /* The following ensures that we don't mess with any queue while some
4563 * other thread might also be doing so. The call_queue_lock field is
4564 * is only modified under the call lock. If the call is in the process
4565 * of being removed from a queue, the call is not locked until the
4566 * the queue lock is dropped and only then is the call_queue_lock field
4567 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4568 * Note that any other routine which removes a call from a queue has to
4569 * obtain the queue lock before examing the queue and removing the call.
4571 if (call->call_queue_lock) {
4572 MUTEX_ENTER(call->call_queue_lock);
4573 if (queue_IsOnQueue(call)) {
4575 if (flags & RX_CALL_WAIT_PROC) {
4576 MUTEX_ENTER(&rx_stats_mutex);
4578 MUTEX_EXIT(&rx_stats_mutex);
4581 MUTEX_EXIT(call->call_queue_lock);
4582 CLEAR_CALL_QUEUE_LOCK(call);
4584 #else /* RX_ENABLE_LOCKS */
4585 if (queue_IsOnQueue(call)) {
4587 if (flags & RX_CALL_WAIT_PROC)
4590 #endif /* RX_ENABLE_LOCKS */
4592 rxi_KeepAliveOff(call);
4593 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4596 /* Send an acknowledge for the indicated packet (seq,serial) of the
4597 * indicated call, for the indicated reason (reason). This
4598 * acknowledge will specifically acknowledge receiving the packet, and
4599 * will also specify which other packets for this call have been
4600 * received. This routine returns the packet that was used to the
4601 * caller. The caller is responsible for freeing it or re-using it.
4602 * This acknowledgement also returns the highest sequence number
4603 * actually read out by the higher level to the sender; the sender
4604 * promises to keep around packets that have not been read by the
4605 * higher level yet (unless, of course, the sender decides to abort
4606 * the call altogether). Any of p, seq, serial, pflags, or reason may
4607 * be set to zero without ill effect. That is, if they are zero, they
4608 * will not convey any information.
4609 * NOW there is a trailer field, after the ack where it will safely be
4610 * ignored by mundanes, which indicates the maximum size packet this
4611 * host can swallow. */
4613 register struct rx_packet *optionalPacket; use to send ack (or null)
4614 int seq; Sequence number of the packet we are acking
4615 int serial; Serial number of the packet
4616 int pflags; Flags field from packet header
4617 int reason; Reason an acknowledge was prompted
4621 rxi_SendAck(register struct rx_call *call,
4622 register struct rx_packet *optionalPacket, int serial, int reason,
4625 struct rx_ackPacket *ap;
4626 register struct rx_packet *rqp;
4627 register struct rx_packet *nxp; /* For queue_Scan */
4628 register struct rx_packet *p;
4631 #ifdef RX_ENABLE_TSFPQ
4632 struct rx_ts_info_t * rx_ts_info;
4636 * Open the receive window once a thread starts reading packets
4638 if (call->rnext > 1) {
4639 call->rwind = rx_maxReceiveWindow;
4642 call->nHardAcks = 0;
4643 call->nSoftAcks = 0;
4644 if (call->rnext > call->lastAcked)
4645 call->lastAcked = call->rnext;
4649 rx_computelen(p, p->length); /* reset length, you never know */
4650 } /* where that's been... */
4651 #ifdef RX_ENABLE_TSFPQ
4653 RX_TS_INFO_GET(rx_ts_info);
4654 if ((p = rx_ts_info->local_special_packet)) {
4655 rx_computelen(p, p->length);
4656 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4657 rx_ts_info->local_special_packet = p;
4658 } else { /* We won't send the ack, but don't panic. */
4659 return optionalPacket;
4663 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4664 /* We won't send the ack, but don't panic. */
4665 return optionalPacket;
4670 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4673 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4674 #ifndef RX_ENABLE_TSFPQ
4675 if (!optionalPacket)
4678 return optionalPacket;
4680 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4681 if (rx_Contiguous(p) < templ) {
4682 #ifndef RX_ENABLE_TSFPQ
4683 if (!optionalPacket)
4686 return optionalPacket;
4691 /* MTUXXX failing to send an ack is very serious. We should */
4692 /* try as hard as possible to send even a partial ack; it's */
4693 /* better than nothing. */
4694 ap = (struct rx_ackPacket *)rx_DataOf(p);
4695 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4696 ap->reason = reason;
4698 /* The skew computation used to be bogus, I think it's better now. */
4699 /* We should start paying attention to skew. XXX */
4700 ap->serial = htonl(serial);
4701 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4703 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4704 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4706 /* No fear of running out of ack packet here because there can only be at most
4707 * one window full of unacknowledged packets. The window size must be constrained
4708 * to be less than the maximum ack size, of course. Also, an ack should always
4709 * fit into a single packet -- it should not ever be fragmented. */
4710 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4711 if (!rqp || !call->rq.next
4712 || (rqp->header.seq > (call->rnext + call->rwind))) {
4713 #ifndef RX_ENABLE_TSFPQ
4714 if (!optionalPacket)
4717 rxi_CallError(call, RX_CALL_DEAD);
4718 return optionalPacket;
4721 while (rqp->header.seq > call->rnext + offset)
4722 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4723 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4725 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4726 #ifndef RX_ENABLE_TSFPQ
4727 if (!optionalPacket)
4730 rxi_CallError(call, RX_CALL_DEAD);
4731 return optionalPacket;
4736 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4738 /* these are new for AFS 3.3 */
4739 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4740 templ = htonl(templ);
4741 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4742 templ = htonl(call->conn->peer->ifMTU);
4743 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4744 sizeof(afs_int32), &templ);
4746 /* new for AFS 3.4 */
4747 templ = htonl(call->rwind);
4748 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4749 sizeof(afs_int32), &templ);
4751 /* new for AFS 3.5 */
4752 templ = htonl(call->conn->peer->ifDgramPackets);
4753 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4754 sizeof(afs_int32), &templ);
4756 p->header.serviceId = call->conn->serviceId;
4757 p->header.cid = (call->conn->cid | call->channel);
4758 p->header.callNumber = *call->callNumber;
4760 p->header.securityIndex = call->conn->securityIndex;
4761 p->header.epoch = call->conn->epoch;
4762 p->header.type = RX_PACKET_TYPE_ACK;
4763 p->header.flags = RX_SLOW_START_OK;
4764 if (reason == RX_ACK_PING) {
4765 p->header.flags |= RX_REQUEST_ACK;
4767 clock_GetTime(&call->pingRequestTime);
4770 if (call->conn->type == RX_CLIENT_CONNECTION)
4771 p->header.flags |= RX_CLIENT_INITIATED;
4775 if (rxdebug_active) {
4779 len = _snprintf(msg, sizeof(msg),
4780 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4781 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4782 ntohl(ap->serial), ntohl(ap->previousPacket),
4783 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
4784 ap->nAcks, ntohs(ap->bufferSpace) );
4788 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
4789 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4793 OutputDebugString(msg);
4795 #else /* AFS_NT40_ENV */
4797 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
4798 ap->reason, ntohl(ap->previousPacket),
4799 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4801 for (offset = 0; offset < ap->nAcks; offset++)
4802 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4807 #endif /* AFS_NT40_ENV */
4810 register int i, nbytes = p->length;
4812 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4813 if (nbytes <= p->wirevec[i].iov_len) {
4814 register int savelen, saven;
4816 savelen = p->wirevec[i].iov_len;
4818 p->wirevec[i].iov_len = nbytes;
4820 rxi_Send(call, p, istack);
4821 p->wirevec[i].iov_len = savelen;
4825 nbytes -= p->wirevec[i].iov_len;
4828 MUTEX_ENTER(&rx_stats_mutex);
4829 rx_stats.ackPacketsSent++;
4830 MUTEX_EXIT(&rx_stats_mutex);
4831 #ifndef RX_ENABLE_TSFPQ
4832 if (!optionalPacket)
4835 return optionalPacket; /* Return packet for re-use by caller */
4838 /* Send all of the packets in the list in single datagram */
4840 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4841 int istack, int moreFlag, struct clock *now,
4842 struct clock *retryTime, int resending)
4847 struct rx_connection *conn = call->conn;
4848 struct rx_peer *peer = conn->peer;
4850 MUTEX_ENTER(&peer->peer_lock);
4853 peer->reSends += len;
4854 MUTEX_ENTER(&rx_stats_mutex);
4855 rx_stats.dataPacketsSent += len;
4856 MUTEX_EXIT(&rx_stats_mutex);
4857 MUTEX_EXIT(&peer->peer_lock);
4859 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4863 /* Set the packet flags and schedule the resend events */
4864 /* Only request an ack for the last packet in the list */
4865 for (i = 0; i < len; i++) {
4866 list[i]->retryTime = *retryTime;
4867 if (list[i]->header.serial) {
4868 /* Exponentially backoff retry times */
4869 if (list[i]->backoff < MAXBACKOFF) {
4870 /* so it can't stay == 0 */
4871 list[i]->backoff = (list[i]->backoff << 1) + 1;
4874 clock_Addmsec(&(list[i]->retryTime),
4875 ((afs_uint32) list[i]->backoff) << 8);
4878 /* Wait a little extra for the ack on the last packet */
4879 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4880 clock_Addmsec(&(list[i]->retryTime), 400);
4883 /* Record the time sent */
4884 list[i]->timeSent = *now;
4886 /* Ask for an ack on retransmitted packets, on every other packet
4887 * if the peer doesn't support slow start. Ask for an ack on every
4888 * packet until the congestion window reaches the ack rate. */
4889 if (list[i]->header.serial) {
4891 MUTEX_ENTER(&rx_stats_mutex);
4892 rx_stats.dataPacketsReSent++;
4893 MUTEX_EXIT(&rx_stats_mutex);
4895 /* improved RTO calculation- not Karn */
4896 list[i]->firstSent = *now;
4897 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4898 || (!(call->flags & RX_CALL_SLOW_START_OK)
4899 && (list[i]->header.seq & 1)))) {
4904 MUTEX_ENTER(&peer->peer_lock);
4908 MUTEX_ENTER(&rx_stats_mutex);
4909 rx_stats.dataPacketsSent++;
4910 MUTEX_EXIT(&rx_stats_mutex);
4911 MUTEX_EXIT(&peer->peer_lock);
4913 /* Tag this packet as not being the last in this group,
4914 * for the receiver's benefit */
4915 if (i < len - 1 || moreFlag) {
4916 list[i]->header.flags |= RX_MORE_PACKETS;
4919 /* Install the new retransmit time for the packet, and
4920 * record the time sent */
4921 list[i]->timeSent = *now;
4925 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4928 /* Since we're about to send a data packet to the peer, it's
4929 * safe to nuke any scheduled end-of-packets ack */
4930 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4932 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4933 MUTEX_EXIT(&call->lock);
4935 rxi_SendPacketList(call, conn, list, len, istack);
4937 rxi_SendPacket(call, conn, list[0], istack);
4939 MUTEX_ENTER(&call->lock);
4940 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4942 /* Update last send time for this call (for keep-alive
4943 * processing), and for the connection (so that we can discover
4944 * idle connections) */
4945 conn->lastSendTime = call->lastSendTime = clock_Sec();
4948 /* When sending packets we need to follow these rules:
4949 * 1. Never send more than maxDgramPackets in a jumbogram.
4950 * 2. Never send a packet with more than two iovecs in a jumbogram.
4951 * 3. Never send a retransmitted packet in a jumbogram.
4952 * 4. Never send more than cwind/4 packets in a jumbogram
4953 * We always keep the last list we should have sent so we
4954 * can set the RX_MORE_PACKETS flags correctly.
4957 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4958 int istack, struct clock *now, struct clock *retryTime,
4961 int i, cnt, lastCnt = 0;
4962 struct rx_packet **listP, **lastP = 0;
4963 struct rx_peer *peer = call->conn->peer;
4964 int morePackets = 0;
4966 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4967 /* Does the current packet force us to flush the current list? */
4969 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4970 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4972 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
4974 /* If the call enters an error state stop sending, or if
4975 * we entered congestion recovery mode, stop sending */
4976 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4984 /* Add the current packet to the list if it hasn't been acked.
4985 * Otherwise adjust the list pointer to skip the current packet. */
4986 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
4988 /* Do we need to flush the list? */
4989 if (cnt >= (int)peer->maxDgramPackets
4990 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
4991 || list[i]->header.serial
4992 || list[i]->length != RX_JUMBOBUFFERSIZE) {
4994 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
4995 retryTime, resending);
4996 /* If the call enters an error state stop sending, or if
4997 * we entered congestion recovery mode, stop sending */
4999 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5004 listP = &list[i + 1];
5009 osi_Panic("rxi_SendList error");
5011 listP = &list[i + 1];
5015 /* Send the whole list when the call is in receive mode, when
5016 * the call is in eof mode, when we are in fast recovery mode,
5017 * and when we have the last packet */
5018 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5019 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5020 || (call->flags & RX_CALL_FAST_RECOVER)) {
5021 /* Check for the case where the current list contains
5022 * an acked packet. Since we always send retransmissions
5023 * in a separate packet, we only need to check the first
5024 * packet in the list */
5025 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5029 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5030 retryTime, resending);
5031 /* If the call enters an error state stop sending, or if
5032 * we entered congestion recovery mode, stop sending */
5033 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5037 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5040 } else if (lastCnt > 0) {
5041 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5046 #ifdef RX_ENABLE_LOCKS
5047 /* Call rxi_Start, below, but with the call lock held. */
5049 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
5050 void *arg1, int istack)
5052 MUTEX_ENTER(&call->lock);
5053 rxi_Start(event, call, arg1, istack);
5054 MUTEX_EXIT(&call->lock);
5056 #endif /* RX_ENABLE_LOCKS */
5058 /* This routine is called when new packets are readied for
5059 * transmission and when retransmission may be necessary, or when the
5060 * transmission window or burst count are favourable. This should be
5061 * better optimized for new packets, the usual case, now that we've
5062 * got rid of queues of send packets. XXXXXXXXXXX */
5064 rxi_Start(struct rxevent *event, register struct rx_call *call,
5065 void *arg1, int istack)
5067 struct rx_packet *p;
5068 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
5069 struct rx_peer *peer = call->conn->peer;
5070 struct clock now, retryTime;
5074 struct rx_packet **xmitList;
5077 /* If rxi_Start is being called as a result of a resend event,
5078 * then make sure that the event pointer is removed from the call
5079 * structure, since there is no longer a per-call retransmission
5081 if (event && event == call->resendEvent) {
5082 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5083 call->resendEvent = NULL;
5085 if (queue_IsEmpty(&call->tq)) {
5089 /* Timeouts trigger congestion recovery */
5090 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5091 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5092 /* someone else is waiting to start recovery */
5095 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5096 while (call->flags & RX_CALL_TQ_BUSY) {
5097 call->flags |= RX_CALL_TQ_WAIT;
5099 #ifdef RX_ENABLE_LOCKS
5100 osirx_AssertMine(&call->lock, "rxi_Start lock1");
5101 CV_WAIT(&call->cv_tq, &call->lock);
5102 #else /* RX_ENABLE_LOCKS */
5103 osi_rxSleep(&call->tq);
5104 #endif /* RX_ENABLE_LOCKS */
5106 if (call->tqWaiters == 0)
5107 call->flags &= ~RX_CALL_TQ_WAIT;
5109 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5110 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5111 call->flags |= RX_CALL_FAST_RECOVER;
5112 if (peer->maxDgramPackets > 1) {
5113 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5115 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5117 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5118 call->nDgramPackets = 1;
5120 call->nextCwind = 1;
5123 MUTEX_ENTER(&peer->peer_lock);
5124 peer->MTU = call->MTU;
5125 peer->cwind = call->cwind;
5126 peer->nDgramPackets = 1;
5128 call->congestSeq = peer->congestSeq;
5129 MUTEX_EXIT(&peer->peer_lock);
5130 /* Clear retry times on packets. Otherwise, it's possible for
5131 * some packets in the queue to force resends at rates faster
5132 * than recovery rates.
5134 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5135 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5136 clock_Zero(&p->retryTime);
5141 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5142 MUTEX_ENTER(&rx_stats_mutex);
5143 rx_tq_debug.rxi_start_in_error++;
5144 MUTEX_EXIT(&rx_stats_mutex);
5149 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5150 /* Get clock to compute the re-transmit time for any packets
5151 * in this burst. Note, if we back off, it's reasonable to
5152 * back off all of the packets in the same manner, even if
5153 * some of them have been retransmitted more times than more
5154 * recent additions */
5155 clock_GetTime(&now);
5156 retryTime = now; /* initialize before use */
5157 MUTEX_ENTER(&peer->peer_lock);
5158 clock_Add(&retryTime, &peer->timeout);
5159 MUTEX_EXIT(&peer->peer_lock);
5161 /* Send (or resend) any packets that need it, subject to
5162 * window restrictions and congestion burst control
5163 * restrictions. Ask for an ack on the last packet sent in
5164 * this burst. For now, we're relying upon the window being
5165 * considerably bigger than the largest number of packets that
5166 * are typically sent at once by one initial call to
5167 * rxi_Start. This is probably bogus (perhaps we should ask
5168 * for an ack when we're half way through the current
5169 * window?). Also, for non file transfer applications, this
5170 * may end up asking for an ack for every packet. Bogus. XXXX
5173 * But check whether we're here recursively, and let the other guy
5176 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5177 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5178 call->flags |= RX_CALL_TQ_BUSY;
5180 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5182 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5183 call->flags &= ~RX_CALL_NEED_START;
5184 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5186 maxXmitPackets = MIN(call->twind, call->cwind);
5187 xmitList = (struct rx_packet **)
5188 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5189 if (xmitList == NULL)
5190 osi_Panic("rxi_Start, failed to allocate xmit list");
5191 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5192 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5193 /* We shouldn't be sending packets if a thread is waiting
5194 * to initiate congestion recovery */
5198 && (call->flags & RX_CALL_FAST_RECOVER)) {
5199 /* Only send one packet during fast recovery */
5202 if ((p->flags & RX_PKTFLAG_FREE)
5203 || (!queue_IsEnd(&call->tq, nxp)
5204 && (nxp->flags & RX_PKTFLAG_FREE))
5205 || (p == (struct rx_packet *)&rx_freePacketQueue)
5206 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5207 osi_Panic("rxi_Start: xmit queue clobbered");
5209 if (p->flags & RX_PKTFLAG_ACKED) {
5210 MUTEX_ENTER(&rx_stats_mutex);
5211 rx_stats.ignoreAckedPacket++;
5212 MUTEX_EXIT(&rx_stats_mutex);
5213 continue; /* Ignore this packet if it has been acknowledged */
5216 /* Turn off all flags except these ones, which are the same
5217 * on each transmission */
5218 p->header.flags &= RX_PRESET_FLAGS;
5220 if (p->header.seq >=
5221 call->tfirst + MIN((int)call->twind,
5222 (int)(call->nSoftAcked +
5224 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5225 /* Note: if we're waiting for more window space, we can
5226 * still send retransmits; hence we don't return here, but
5227 * break out to schedule a retransmit event */
5228 dpf(("call %d waiting for window",
5229 *(call->callNumber)));
5233 /* Transmit the packet if it needs to be sent. */
5234 if (!clock_Lt(&now, &p->retryTime)) {
5235 if (nXmitPackets == maxXmitPackets) {
5236 rxi_SendXmitList(call, xmitList, nXmitPackets,
5237 istack, &now, &retryTime,
5239 osi_Free(xmitList, maxXmitPackets *
5240 sizeof(struct rx_packet *));
5243 xmitList[nXmitPackets++] = p;
5247 /* xmitList now hold pointers to all of the packets that are
5248 * ready to send. Now we loop to send the packets */
5249 if (nXmitPackets > 0) {
5250 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5251 &now, &retryTime, resending);
5254 maxXmitPackets * sizeof(struct rx_packet *));
5256 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5258 * TQ references no longer protected by this flag; they must remain
5259 * protected by the global lock.
5261 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5262 call->flags &= ~RX_CALL_TQ_BUSY;
5263 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5264 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5265 #ifdef RX_ENABLE_LOCKS
5266 osirx_AssertMine(&call->lock, "rxi_Start start");
5267 CV_BROADCAST(&call->cv_tq);
5268 #else /* RX_ENABLE_LOCKS */
5269 osi_rxWakeup(&call->tq);
5270 #endif /* RX_ENABLE_LOCKS */
5275 /* We went into the error state while sending packets. Now is
5276 * the time to reset the call. This will also inform the using
5277 * process that the call is in an error state.
5279 MUTEX_ENTER(&rx_stats_mutex);
5280 rx_tq_debug.rxi_start_aborted++;
5281 MUTEX_EXIT(&rx_stats_mutex);
5282 call->flags &= ~RX_CALL_TQ_BUSY;
5283 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5284 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5285 #ifdef RX_ENABLE_LOCKS
5286 osirx_AssertMine(&call->lock, "rxi_Start middle");
5287 CV_BROADCAST(&call->cv_tq);
5288 #else /* RX_ENABLE_LOCKS */
5289 osi_rxWakeup(&call->tq);
5290 #endif /* RX_ENABLE_LOCKS */
5292 rxi_CallError(call, call->error);
5295 #ifdef RX_ENABLE_LOCKS
5296 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5297 register int missing;
5298 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5299 /* Some packets have received acks. If they all have, we can clear
5300 * the transmit queue.
5303 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5304 if (p->header.seq < call->tfirst
5305 && (p->flags & RX_PKTFLAG_ACKED)) {
5312 call->flags |= RX_CALL_TQ_CLEARME;
5314 #endif /* RX_ENABLE_LOCKS */
5315 /* Don't bother doing retransmits if the TQ is cleared. */
5316 if (call->flags & RX_CALL_TQ_CLEARME) {
5317 rxi_ClearTransmitQueue(call, 1);
5319 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5322 /* Always post a resend event, if there is anything in the
5323 * queue, and resend is possible. There should be at least
5324 * one unacknowledged packet in the queue ... otherwise none
5325 * of these packets should be on the queue in the first place.
5327 if (call->resendEvent) {
5328 /* Cancel the existing event and post a new one */
5329 rxevent_Cancel(call->resendEvent, call,
5330 RX_CALL_REFCOUNT_RESEND);
5333 /* The retry time is the retry time on the first unacknowledged
5334 * packet inside the current window */
5336 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5337 /* Don't set timers for packets outside the window */
5338 if (p->header.seq >= call->tfirst + call->twind) {
5342 if (!(p->flags & RX_PKTFLAG_ACKED)
5343 && !clock_IsZero(&p->retryTime)) {
5345 retryTime = p->retryTime;
5350 /* Post a new event to re-run rxi_Start when retries may be needed */
5351 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5352 #ifdef RX_ENABLE_LOCKS
5353 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5355 rxevent_Post2(&retryTime, rxi_StartUnlocked,
5356 (void *)call, 0, istack);
5357 #else /* RX_ENABLE_LOCKS */
5359 rxevent_Post2(&retryTime, rxi_Start, (void *)call,
5361 #endif /* RX_ENABLE_LOCKS */
5364 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5365 } while (call->flags & RX_CALL_NEED_START);
5367 * TQ references no longer protected by this flag; they must remain
5368 * protected by the global lock.
5370 call->flags &= ~RX_CALL_TQ_BUSY;
5371 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5372 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5373 #ifdef RX_ENABLE_LOCKS
5374 osirx_AssertMine(&call->lock, "rxi_Start end");
5375 CV_BROADCAST(&call->cv_tq);
5376 #else /* RX_ENABLE_LOCKS */
5377 osi_rxWakeup(&call->tq);
5378 #endif /* RX_ENABLE_LOCKS */
5381 call->flags |= RX_CALL_NEED_START;
5383 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5385 if (call->resendEvent) {
5386 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5391 /* Also adjusts the keep alive parameters for the call, to reflect
5392 * that we have just sent a packet (so keep alives aren't sent
5395 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5398 register struct rx_connection *conn = call->conn;
5400 /* Stamp each packet with the user supplied status */
5401 p->header.userStatus = call->localStatus;
5403 /* Allow the security object controlling this call's security to
5404 * make any last-minute changes to the packet */
5405 RXS_SendPacket(conn->securityObject, call, p);
5407 /* Since we're about to send SOME sort of packet to the peer, it's
5408 * safe to nuke any scheduled end-of-packets ack */
5409 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5411 /* Actually send the packet, filling in more connection-specific fields */
5412 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5413 MUTEX_EXIT(&call->lock);
5414 rxi_SendPacket(call, conn, p, istack);
5415 MUTEX_ENTER(&call->lock);
5416 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5418 /* Update last send time for this call (for keep-alive
5419 * processing), and for the connection (so that we can discover
5420 * idle connections) */
5421 conn->lastSendTime = call->lastSendTime = clock_Sec();
5425 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5426 * that things are fine. Also called periodically to guarantee that nothing
5427 * falls through the cracks (e.g. (error + dally) connections have keepalive
5428 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5430 * haveCTLock Set if calling from rxi_ReapConnections
5432 #ifdef RX_ENABLE_LOCKS
5434 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5435 #else /* RX_ENABLE_LOCKS */
5437 rxi_CheckCall(register struct rx_call *call)
5438 #endif /* RX_ENABLE_LOCKS */
5440 register struct rx_connection *conn = call->conn;
5442 afs_uint32 deadTime;
5444 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5445 if (call->flags & RX_CALL_TQ_BUSY) {
5446 /* Call is active and will be reset by rxi_Start if it's
5447 * in an error state.
5452 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5454 (((afs_uint32) conn->secondsUntilDead << 10) +
5455 ((afs_uint32) conn->peer->rtt >> 3) +
5456 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5458 /* These are computed to the second (+- 1 second). But that's
5459 * good enough for these values, which should be a significant
5460 * number of seconds. */
5461 if (now > (call->lastReceiveTime + deadTime)) {
5462 if (call->state == RX_STATE_ACTIVE) {
5463 rxi_CallError(call, RX_CALL_DEAD);
5466 #ifdef RX_ENABLE_LOCKS
5467 /* Cancel pending events */
5468 rxevent_Cancel(call->delayedAckEvent, call,
5469 RX_CALL_REFCOUNT_DELAY);
5470 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5471 rxevent_Cancel(call->keepAliveEvent, call,
5472 RX_CALL_REFCOUNT_ALIVE);
5473 if (call->refCount == 0) {
5474 rxi_FreeCall(call, haveCTLock);
5478 #else /* RX_ENABLE_LOCKS */
5481 #endif /* RX_ENABLE_LOCKS */
5483 /* Non-active calls are destroyed if they are not responding
5484 * to pings; active calls are simply flagged in error, so the
5485 * attached process can die reasonably gracefully. */
5487 /* see if we have a non-activity timeout */
5488 if (call->startWait && conn->idleDeadTime
5489 && ((call->startWait + conn->idleDeadTime) < now)) {
5490 if (call->state == RX_STATE_ACTIVE) {
5491 rxi_CallError(call, RX_CALL_TIMEOUT);
5495 /* see if we have a hard timeout */
5496 if (conn->hardDeadTime
5497 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5498 if (call->state == RX_STATE_ACTIVE)
5499 rxi_CallError(call, RX_CALL_TIMEOUT);
5506 /* When a call is in progress, this routine is called occasionally to
5507 * make sure that some traffic has arrived (or been sent to) the peer.
5508 * If nothing has arrived in a reasonable amount of time, the call is
5509 * declared dead; if nothing has been sent for a while, we send a
5510 * keep-alive packet (if we're actually trying to keep the call alive)
5513 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5516 struct rx_connection *conn;
5519 MUTEX_ENTER(&call->lock);
5520 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5521 if (event == call->keepAliveEvent)
5522 call->keepAliveEvent = NULL;
5525 #ifdef RX_ENABLE_LOCKS
5526 if (rxi_CheckCall(call, 0)) {
5527 MUTEX_EXIT(&call->lock);
5530 #else /* RX_ENABLE_LOCKS */
5531 if (rxi_CheckCall(call))
5533 #endif /* RX_ENABLE_LOCKS */
5535 /* Don't try to keep alive dallying calls */
5536 if (call->state == RX_STATE_DALLY) {
5537 MUTEX_EXIT(&call->lock);
5542 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5543 /* Don't try to send keepalives if there is unacknowledged data */
5544 /* the rexmit code should be good enough, this little hack
5545 * doesn't quite work XXX */
5546 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5548 rxi_ScheduleKeepAliveEvent(call);
5549 MUTEX_EXIT(&call->lock);
5554 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5556 if (!call->keepAliveEvent) {
5558 clock_GetTime(&when);
5559 when.sec += call->conn->secondsUntilPing;
5560 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5561 call->keepAliveEvent =
5562 rxevent_Post(&when, rxi_KeepAliveEvent, call, 0);
5566 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5568 rxi_KeepAliveOn(register struct rx_call *call)
5570 /* Pretend last packet received was received now--i.e. if another
5571 * packet isn't received within the keep alive time, then the call
5572 * will die; Initialize last send time to the current time--even
5573 * if a packet hasn't been sent yet. This will guarantee that a
5574 * keep-alive is sent within the ping time */
5575 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5576 rxi_ScheduleKeepAliveEvent(call);
5579 /* This routine is called to send connection abort messages
5580 * that have been delayed to throttle looping clients. */
5582 rxi_SendDelayedConnAbort(struct rxevent *event,
5583 register struct rx_connection *conn, char *dummy)
5586 struct rx_packet *packet;
5588 MUTEX_ENTER(&conn->conn_data_lock);
5589 conn->delayedAbortEvent = NULL;
5590 error = htonl(conn->error);
5592 MUTEX_EXIT(&conn->conn_data_lock);
5593 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5596 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5597 RX_PACKET_TYPE_ABORT, (char *)&error,
5599 rxi_FreePacket(packet);
5603 /* This routine is called to send call abort messages
5604 * that have been delayed to throttle looping clients. */
5606 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5610 struct rx_packet *packet;
5612 MUTEX_ENTER(&call->lock);
5613 call->delayedAbortEvent = NULL;
5614 error = htonl(call->error);
5616 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5619 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5620 (char *)&error, sizeof(error), 0);
5621 rxi_FreePacket(packet);
5623 MUTEX_EXIT(&call->lock);
5626 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5627 * seconds) to ask the client to authenticate itself. The routine
5628 * issues a challenge to the client, which is obtained from the
5629 * security object associated with the connection */
5631 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5632 void *arg1, int tries)
5634 conn->challengeEvent = NULL;
5635 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5636 register struct rx_packet *packet;
5640 /* We've failed to authenticate for too long.
5641 * Reset any calls waiting for authentication;
5642 * they are all in RX_STATE_PRECALL.
5646 MUTEX_ENTER(&conn->conn_call_lock);
5647 for (i = 0; i < RX_MAXCALLS; i++) {
5648 struct rx_call *call = conn->call[i];
5650 MUTEX_ENTER(&call->lock);
5651 if (call->state == RX_STATE_PRECALL) {
5652 rxi_CallError(call, RX_CALL_DEAD);
5653 rxi_SendCallAbort(call, NULL, 0, 0);
5655 MUTEX_EXIT(&call->lock);
5658 MUTEX_EXIT(&conn->conn_call_lock);
5662 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5664 /* If there's no packet available, do this later. */
5665 RXS_GetChallenge(conn->securityObject, conn, packet);
5666 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5667 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5668 rxi_FreePacket(packet);
5670 clock_GetTime(&when);
5671 when.sec += RX_CHALLENGE_TIMEOUT;
5672 conn->challengeEvent =
5673 rxevent_Post2(&when, rxi_ChallengeEvent, conn, 0,
5678 /* Call this routine to start requesting the client to authenticate
5679 * itself. This will continue until authentication is established,
5680 * the call times out, or an invalid response is returned. The
5681 * security object associated with the connection is asked to create
5682 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5683 * defined earlier. */
5685 rxi_ChallengeOn(register struct rx_connection *conn)
5687 if (!conn->challengeEvent) {
5688 RXS_CreateChallenge(conn->securityObject, conn);
5689 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5694 /* Compute round trip time of the packet provided, in *rttp.
5697 /* rxi_ComputeRoundTripTime is called with peer locked. */
5698 /* sentp and/or peer may be null */
5700 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5701 register struct clock *sentp,
5702 register struct rx_peer *peer)
5704 struct clock thisRtt, *rttp = &thisRtt;
5706 register int rtt_timeout;
5708 clock_GetTime(rttp);
5710 if (clock_Lt(rttp, sentp)) {
5712 return; /* somebody set the clock back, don't count this time. */
5714 clock_Sub(rttp, sentp);
5715 MUTEX_ENTER(&rx_stats_mutex);
5716 if (clock_Lt(rttp, &rx_stats.minRtt))
5717 rx_stats.minRtt = *rttp;
5718 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5719 if (rttp->sec > 60) {
5720 MUTEX_EXIT(&rx_stats_mutex);
5721 return; /* somebody set the clock ahead */
5723 rx_stats.maxRtt = *rttp;
5725 clock_Add(&rx_stats.totalRtt, rttp);
5726 rx_stats.nRttSamples++;
5727 MUTEX_EXIT(&rx_stats_mutex);
5729 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5731 /* Apply VanJacobson round-trip estimations */
5736 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5737 * srtt is stored as fixed point with 3 bits after the binary
5738 * point (i.e., scaled by 8). The following magic is
5739 * equivalent to the smoothing algorithm in rfc793 with an
5740 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5741 * srtt*8 = srtt*8 + rtt - srtt
5742 * srtt = srtt + rtt/8 - srtt/8
5745 delta = MSEC(rttp) - (peer->rtt >> 3);
5749 * We accumulate a smoothed rtt variance (actually, a smoothed
5750 * mean difference), then set the retransmit timer to smoothed
5751 * rtt + 4 times the smoothed variance (was 2x in van's original
5752 * paper, but 4x works better for me, and apparently for him as
5754 * rttvar is stored as
5755 * fixed point with 2 bits after the binary point (scaled by
5756 * 4). The following is equivalent to rfc793 smoothing with
5757 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5758 * replaces rfc793's wired-in beta.
5759 * dev*4 = dev*4 + (|actual - expected| - dev)
5765 delta -= (peer->rtt_dev >> 2);
5766 peer->rtt_dev += delta;
5768 /* I don't have a stored RTT so I start with this value. Since I'm
5769 * probably just starting a call, and will be pushing more data down
5770 * this, I expect congestion to increase rapidly. So I fudge a
5771 * little, and I set deviance to half the rtt. In practice,
5772 * deviance tends to approach something a little less than
5773 * half the smoothed rtt. */
5774 peer->rtt = (MSEC(rttp) << 3) + 8;
5775 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5777 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5778 * the other of these connections is usually in a user process, and can
5779 * be switched and/or swapped out. So on fast, reliable networks, the
5780 * timeout would otherwise be too short.
5782 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5783 clock_Zero(&(peer->timeout));
5784 clock_Addmsec(&(peer->timeout), rtt_timeout);
5786 dpf(("rxi_ComputeRoundTripTime(rtt=%d ms, srtt=%d ms, rtt_dev=%d ms, timeout=%d.%0.3d sec)\n", MSEC(rttp), peer->rtt >> 3, peer->rtt_dev >> 2, (peer->timeout.sec), (peer->timeout.usec)));
5790 /* Find all server connections that have not been active for a long time, and
5793 rxi_ReapConnections(void)
5796 clock_GetTime(&now);
5798 /* Find server connection structures that haven't been used for
5799 * greater than rx_idleConnectionTime */
5801 struct rx_connection **conn_ptr, **conn_end;
5802 int i, havecalls = 0;
5803 MUTEX_ENTER(&rx_connHashTable_lock);
5804 for (conn_ptr = &rx_connHashTable[0], conn_end =
5805 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5807 struct rx_connection *conn, *next;
5808 struct rx_call *call;
5812 for (conn = *conn_ptr; conn; conn = next) {
5813 /* XXX -- Shouldn't the connection be locked? */
5816 for (i = 0; i < RX_MAXCALLS; i++) {
5817 call = conn->call[i];
5820 MUTEX_ENTER(&call->lock);
5821 #ifdef RX_ENABLE_LOCKS
5822 result = rxi_CheckCall(call, 1);
5823 #else /* RX_ENABLE_LOCKS */
5824 result = rxi_CheckCall(call);
5825 #endif /* RX_ENABLE_LOCKS */
5826 MUTEX_EXIT(&call->lock);
5828 /* If CheckCall freed the call, it might
5829 * have destroyed the connection as well,
5830 * which screws up the linked lists.
5836 if (conn->type == RX_SERVER_CONNECTION) {
5837 /* This only actually destroys the connection if
5838 * there are no outstanding calls */
5839 MUTEX_ENTER(&conn->conn_data_lock);
5840 if (!havecalls && !conn->refCount
5841 && ((conn->lastSendTime + rx_idleConnectionTime) <
5843 conn->refCount++; /* it will be decr in rx_DestroyConn */
5844 MUTEX_EXIT(&conn->conn_data_lock);
5845 #ifdef RX_ENABLE_LOCKS
5846 rxi_DestroyConnectionNoLock(conn);
5847 #else /* RX_ENABLE_LOCKS */
5848 rxi_DestroyConnection(conn);
5849 #endif /* RX_ENABLE_LOCKS */
5851 #ifdef RX_ENABLE_LOCKS
5853 MUTEX_EXIT(&conn->conn_data_lock);
5855 #endif /* RX_ENABLE_LOCKS */
5859 #ifdef RX_ENABLE_LOCKS
5860 while (rx_connCleanup_list) {
5861 struct rx_connection *conn;
5862 conn = rx_connCleanup_list;
5863 rx_connCleanup_list = rx_connCleanup_list->next;
5864 MUTEX_EXIT(&rx_connHashTable_lock);
5865 rxi_CleanupConnection(conn);
5866 MUTEX_ENTER(&rx_connHashTable_lock);
5868 MUTEX_EXIT(&rx_connHashTable_lock);
5869 #endif /* RX_ENABLE_LOCKS */
5872 /* Find any peer structures that haven't been used (haven't had an
5873 * associated connection) for greater than rx_idlePeerTime */
5875 struct rx_peer **peer_ptr, **peer_end;
5877 MUTEX_ENTER(&rx_rpc_stats);
5878 MUTEX_ENTER(&rx_peerHashTable_lock);
5879 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5880 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5882 struct rx_peer *peer, *next, *prev;
5883 for (prev = peer = *peer_ptr; peer; peer = next) {
5885 code = MUTEX_TRYENTER(&peer->peer_lock);
5886 if ((code) && (peer->refCount == 0)
5887 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5888 rx_interface_stat_p rpc_stat, nrpc_stat;
5890 MUTEX_EXIT(&peer->peer_lock);
5891 MUTEX_DESTROY(&peer->peer_lock);
5893 (&peer->rpcStats, rpc_stat, nrpc_stat,
5894 rx_interface_stat)) {
5895 unsigned int num_funcs;
5898 queue_Remove(&rpc_stat->queue_header);
5899 queue_Remove(&rpc_stat->all_peers);
5900 num_funcs = rpc_stat->stats[0].func_total;
5902 sizeof(rx_interface_stat_t) +
5903 rpc_stat->stats[0].func_total *
5904 sizeof(rx_function_entry_v1_t);
5906 rxi_Free(rpc_stat, space);
5907 rxi_rpc_peer_stat_cnt -= num_funcs;
5910 MUTEX_ENTER(&rx_stats_mutex);
5911 rx_stats.nPeerStructs--;
5912 MUTEX_EXIT(&rx_stats_mutex);
5913 if (peer == *peer_ptr) {
5920 MUTEX_EXIT(&peer->peer_lock);
5926 MUTEX_EXIT(&rx_peerHashTable_lock);
5927 MUTEX_EXIT(&rx_rpc_stats);
5930 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5931 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5932 * GC, just below. Really, we shouldn't have to keep moving packets from
5933 * one place to another, but instead ought to always know if we can
5934 * afford to hold onto a packet in its particular use. */
5935 MUTEX_ENTER(&rx_freePktQ_lock);
5936 if (rx_waitingForPackets) {
5937 rx_waitingForPackets = 0;
5938 #ifdef RX_ENABLE_LOCKS
5939 CV_BROADCAST(&rx_waitingForPackets_cv);
5941 osi_rxWakeup(&rx_waitingForPackets);
5944 MUTEX_EXIT(&rx_freePktQ_lock);
5946 now.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5947 rxevent_Post(&now, rxi_ReapConnections, 0, 0);
5951 /* rxs_Release - This isn't strictly necessary but, since the macro name from
5952 * rx.h is sort of strange this is better. This is called with a security
5953 * object before it is discarded. Each connection using a security object has
5954 * its own refcount to the object so it won't actually be freed until the last
5955 * connection is destroyed.
5957 * This is the only rxs module call. A hold could also be written but no one
5961 rxs_Release(struct rx_securityClass *aobj)
5963 return RXS_Close(aobj);
5967 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
5968 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
5969 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
5970 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
5972 /* Adjust our estimate of the transmission rate to this peer, given
5973 * that the packet p was just acked. We can adjust peer->timeout and
5974 * call->twind. Pragmatically, this is called
5975 * only with packets of maximal length.
5976 * Called with peer and call locked.
5980 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
5981 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
5983 afs_int32 xferSize, xferMs;
5984 register afs_int32 minTime;
5987 /* Count down packets */
5988 if (peer->rateFlag > 0)
5990 /* Do nothing until we're enabled */
5991 if (peer->rateFlag != 0)
5996 /* Count only when the ack seems legitimate */
5997 switch (ackReason) {
5998 case RX_ACK_REQUESTED:
6000 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6004 case RX_ACK_PING_RESPONSE:
6005 if (p) /* want the response to ping-request, not data send */
6007 clock_GetTime(&newTO);
6008 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6009 clock_Sub(&newTO, &call->pingRequestTime);
6010 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6014 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6021 dpf(("CONG peer %lx/%u: sample (%s) size %ld, %ld ms (to %lu.%06lu, rtt %u, ps %u)", ntohl(peer->host), ntohs(peer->port), (ackReason == RX_ACK_REQUESTED ? "dataack" : "pingack"), xferSize, xferMs, peer->timeout.sec, peer->timeout.usec, peer->smRtt, peer->ifMTU));
6023 /* Track only packets that are big enough. */
6024 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6028 /* absorb RTT data (in milliseconds) for these big packets */
6029 if (peer->smRtt == 0) {
6030 peer->smRtt = xferMs;
6032 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6037 if (peer->countDown) {
6041 peer->countDown = 10; /* recalculate only every so often */
6043 /* In practice, we can measure only the RTT for full packets,
6044 * because of the way Rx acks the data that it receives. (If it's
6045 * smaller than a full packet, it often gets implicitly acked
6046 * either by the call response (from a server) or by the next call
6047 * (from a client), and either case confuses transmission times
6048 * with processing times.) Therefore, replace the above
6049 * more-sophisticated processing with a simpler version, where the
6050 * smoothed RTT is kept for full-size packets, and the time to
6051 * transmit a windowful of full-size packets is simply RTT *
6052 * windowSize. Again, we take two steps:
6053 - ensure the timeout is large enough for a single packet's RTT;
6054 - ensure that the window is small enough to fit in the desired timeout.*/
6056 /* First, the timeout check. */
6057 minTime = peer->smRtt;
6058 /* Get a reasonable estimate for a timeout period */
6060 newTO.sec = minTime / 1000;
6061 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6063 /* Increase the timeout period so that we can always do at least
6064 * one packet exchange */
6065 if (clock_Gt(&newTO, &peer->timeout)) {
6067 dpf(("CONG peer %lx/%u: timeout %lu.%06lu ==> %lu.%06lu (rtt %u, ps %u)", ntohl(peer->host), ntohs(peer->port), peer->timeout.sec, peer->timeout.usec, newTO.sec, newTO.usec, peer->smRtt, peer->packetSize));
6069 peer->timeout = newTO;
6072 /* Now, get an estimate for the transmit window size. */
6073 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6074 /* Now, convert to the number of full packets that could fit in a
6075 * reasonable fraction of that interval */
6076 minTime /= (peer->smRtt << 1);
6077 xferSize = minTime; /* (make a copy) */
6079 /* Now clamp the size to reasonable bounds. */
6082 else if (minTime > rx_Window)
6083 minTime = rx_Window;
6084 /* if (minTime != peer->maxWindow) {
6085 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6086 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6087 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6089 peer->maxWindow = minTime;
6090 elide... call->twind = minTime;
6094 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6095 * Discern this by calculating the timeout necessary for rx_Window
6097 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6098 /* calculate estimate for transmission interval in milliseconds */
6099 minTime = rx_Window * peer->smRtt;
6100 if (minTime < 1000) {
6101 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
6102 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6103 peer->timeout.usec, peer->smRtt, peer->packetSize));
6105 newTO.sec = 0; /* cut back on timeout by half a second */
6106 newTO.usec = 500000;
6107 clock_Sub(&peer->timeout, &newTO);
6112 } /* end of rxi_ComputeRate */
6113 #endif /* ADAPT_WINDOW */
6121 #define TRACE_OPTION_DEBUGLOG 4
6129 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6130 0, KEY_QUERY_VALUE, &parmKey);
6131 if (code != ERROR_SUCCESS)
6134 dummyLen = sizeof(TraceOption);
6135 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6136 (BYTE *) &TraceOption, &dummyLen);
6137 if (code == ERROR_SUCCESS) {
6138 rxdebug_active = (TraceOption & TRACE_OPTION_DEBUGLOG) ? 1 : 0;
6140 RegCloseKey (parmKey);
6141 #endif /* AFS_NT40_ENV */
6146 rx_DebugOnOff(int on)
6148 rxdebug_active = on;
6150 #endif /* AFS_NT40_ENV */
6153 /* Don't call this debugging routine directly; use dpf */
6155 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
6156 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
6164 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6167 len = _snprintf(msg, sizeof(msg)-2,
6168 tformat, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10,
6169 a11, a12, a13, a14, a15);
6171 if (msg[len-1] != '\n') {
6175 OutputDebugString(msg);
6180 clock_GetTime(&now);
6181 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
6182 (unsigned int)now.usec / 1000);
6183 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
6190 * This function is used to process the rx_stats structure that is local
6191 * to a process as well as an rx_stats structure received from a remote
6192 * process (via rxdebug). Therefore, it needs to do minimal version
6196 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
6197 afs_int32 freePackets, char version)
6201 if (size != sizeof(struct rx_stats)) {
6203 "Unexpected size of stats structure: was %d, expected %d\n",
6204 size, sizeof(struct rx_stats));
6207 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6210 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6211 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6212 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6213 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6214 s->specialPktAllocFailures);
6216 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6217 s->receivePktAllocFailures, s->sendPktAllocFailures,
6218 s->specialPktAllocFailures);
6222 " greedy %d, " "bogusReads %d (last from host %x), "
6223 "noPackets %d, " "noBuffers %d, " "selects %d, "
6224 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6225 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6226 s->selects, s->sendSelects);
6228 fprintf(file, " packets read: ");
6229 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6230 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6232 fprintf(file, "\n");
6235 " other read counters: data %d, " "ack %d, " "dup %d "
6236 "spurious %d " "dally %d\n", s->dataPacketsRead,
6237 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6238 s->ignorePacketDally);
6240 fprintf(file, " packets sent: ");
6241 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6242 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6244 fprintf(file, "\n");
6247 " other send counters: ack %d, " "data %d (not resends), "
6248 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6249 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6250 s->dataPacketsPushed, s->ignoreAckedPacket);
6253 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6254 s->netSendFailures, (int)s->fatalErrors);
6256 if (s->nRttSamples) {
6257 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6258 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6260 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6261 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6265 " %d server connections, " "%d client connections, "
6266 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6267 s->nServerConns, s->nClientConns, s->nPeerStructs,
6268 s->nCallStructs, s->nFreeCallStructs);
6270 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6271 fprintf(file, " %d clock updates\n", clock_nUpdates);
6276 /* for backward compatibility */
6278 rx_PrintStats(FILE * file)
6280 MUTEX_ENTER(&rx_stats_mutex);
6281 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6283 MUTEX_EXIT(&rx_stats_mutex);
6287 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6289 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6290 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6291 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6294 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6295 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6296 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6299 " Packet size %d, " "max in packet skew %d, "
6300 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6301 (int)peer->outPacketSkew);
6304 #ifdef AFS_PTHREAD_ENV
6306 * This mutex protects the following static variables:
6310 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6311 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6313 #define LOCK_RX_DEBUG
6314 #define UNLOCK_RX_DEBUG
6315 #endif /* AFS_PTHREAD_ENV */
6318 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6319 u_char type, void *inputData, size_t inputLength,
6320 void *outputData, size_t outputLength)
6322 static afs_int32 counter = 100;
6324 struct rx_header theader;
6326 register afs_int32 code;
6328 struct sockaddr_in taddr, faddr;
6333 endTime = time(0) + 20; /* try for 20 seconds */
6337 tp = &tbuffer[sizeof(struct rx_header)];
6338 taddr.sin_family = AF_INET;
6339 taddr.sin_port = remotePort;
6340 taddr.sin_addr.s_addr = remoteAddr;
6341 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6342 taddr.sin_len = sizeof(struct sockaddr_in);
6345 memset(&theader, 0, sizeof(theader));
6346 theader.epoch = htonl(999);
6348 theader.callNumber = htonl(counter);
6351 theader.type = type;
6352 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6353 theader.serviceId = 0;
6355 memcpy(tbuffer, &theader, sizeof(theader));
6356 memcpy(tp, inputData, inputLength);
6358 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6359 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6361 /* see if there's a packet available */
6363 FD_SET(socket, &imask);
6366 code = select((int)(socket + 1), &imask, 0, 0, &tv);
6367 if (code == 1 && FD_ISSET(socket, &imask)) {
6368 /* now receive a packet */
6369 faddrLen = sizeof(struct sockaddr_in);
6371 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6372 (struct sockaddr *)&faddr, &faddrLen);
6375 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6376 if (counter == ntohl(theader.callNumber))
6381 /* see if we've timed out */
6382 if (endTime < time(0))
6385 code -= sizeof(struct rx_header);
6386 if (code > outputLength)
6387 code = outputLength;
6388 memcpy(outputData, tp, code);
6393 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6394 afs_uint16 remotePort, struct rx_debugStats * stat,
6395 afs_uint32 * supportedValues)
6397 struct rx_debugIn in;
6400 *supportedValues = 0;
6401 in.type = htonl(RX_DEBUGI_GETSTATS);
6404 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6405 &in, sizeof(in), stat, sizeof(*stat));
6408 * If the call was successful, fixup the version and indicate
6409 * what contents of the stat structure are valid.
6410 * Also do net to host conversion of fields here.
6414 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6415 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6417 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6418 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6420 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6421 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6423 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6424 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6426 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6427 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6429 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6430 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6432 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6433 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6435 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6436 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6439 stat->nFreePackets = ntohl(stat->nFreePackets);
6440 stat->packetReclaims = ntohl(stat->packetReclaims);
6441 stat->callsExecuted = ntohl(stat->callsExecuted);
6442 stat->nWaiting = ntohl(stat->nWaiting);
6443 stat->idleThreads = ntohl(stat->idleThreads);
6450 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6451 afs_uint16 remotePort, struct rx_stats * stat,
6452 afs_uint32 * supportedValues)
6454 struct rx_debugIn in;
6455 afs_int32 *lp = (afs_int32 *) stat;
6460 * supportedValues is currently unused, but added to allow future
6461 * versioning of this function.
6464 *supportedValues = 0;
6465 in.type = htonl(RX_DEBUGI_RXSTATS);
6467 memset(stat, 0, sizeof(*stat));
6469 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6470 &in, sizeof(in), stat, sizeof(*stat));
6475 * Do net to host conversion here
6478 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6487 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6488 afs_uint16 remotePort, size_t version_length,
6492 return MakeDebugCall(socket, remoteAddr, remotePort,
6493 RX_PACKET_TYPE_VERSION, a, 1, version,
6498 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6499 afs_uint16 remotePort, afs_int32 * nextConnection,
6500 int allConnections, afs_uint32 debugSupportedValues,
6501 struct rx_debugConn * conn,
6502 afs_uint32 * supportedValues)
6504 struct rx_debugIn in;
6509 * supportedValues is currently unused, but added to allow future
6510 * versioning of this function.
6513 *supportedValues = 0;
6514 if (allConnections) {
6515 in.type = htonl(RX_DEBUGI_GETALLCONN);
6517 in.type = htonl(RX_DEBUGI_GETCONN);
6519 in.index = htonl(*nextConnection);
6520 memset(conn, 0, sizeof(*conn));
6522 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6523 &in, sizeof(in), conn, sizeof(*conn));
6526 *nextConnection += 1;
6529 * Convert old connection format to new structure.
6532 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6533 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6534 #define MOVEvL(a) (conn->a = vL->a)
6536 /* any old or unrecognized version... */
6537 for (i = 0; i < RX_MAXCALLS; i++) {
6538 MOVEvL(callState[i]);
6539 MOVEvL(callMode[i]);
6540 MOVEvL(callFlags[i]);
6541 MOVEvL(callOther[i]);
6543 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6544 MOVEvL(secStats.type);
6545 MOVEvL(secStats.level);
6546 MOVEvL(secStats.flags);
6547 MOVEvL(secStats.expires);
6548 MOVEvL(secStats.packetsReceived);
6549 MOVEvL(secStats.packetsSent);
6550 MOVEvL(secStats.bytesReceived);
6551 MOVEvL(secStats.bytesSent);
6556 * Do net to host conversion here
6558 * I don't convert host or port since we are most likely
6559 * going to want these in NBO.
6561 conn->cid = ntohl(conn->cid);
6562 conn->serial = ntohl(conn->serial);
6563 for (i = 0; i < RX_MAXCALLS; i++) {
6564 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6566 conn->error = ntohl(conn->error);
6567 conn->secStats.flags = ntohl(conn->secStats.flags);
6568 conn->secStats.expires = ntohl(conn->secStats.expires);
6569 conn->secStats.packetsReceived =
6570 ntohl(conn->secStats.packetsReceived);
6571 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6572 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6573 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6574 conn->epoch = ntohl(conn->epoch);
6575 conn->natMTU = ntohl(conn->natMTU);
6582 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6583 afs_uint16 remotePort, afs_int32 * nextPeer,
6584 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6585 afs_uint32 * supportedValues)
6587 struct rx_debugIn in;
6591 * supportedValues is currently unused, but added to allow future
6592 * versioning of this function.
6595 *supportedValues = 0;
6596 in.type = htonl(RX_DEBUGI_GETPEER);
6597 in.index = htonl(*nextPeer);
6598 memset(peer, 0, sizeof(*peer));
6600 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6601 &in, sizeof(in), peer, sizeof(*peer));
6607 * Do net to host conversion here
6609 * I don't convert host or port since we are most likely
6610 * going to want these in NBO.
6612 peer->ifMTU = ntohs(peer->ifMTU);
6613 peer->idleWhen = ntohl(peer->idleWhen);
6614 peer->refCount = ntohs(peer->refCount);
6615 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6616 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6617 peer->rtt = ntohl(peer->rtt);
6618 peer->rtt_dev = ntohl(peer->rtt_dev);
6619 peer->timeout.sec = ntohl(peer->timeout.sec);
6620 peer->timeout.usec = ntohl(peer->timeout.usec);
6621 peer->nSent = ntohl(peer->nSent);
6622 peer->reSends = ntohl(peer->reSends);
6623 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6624 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6625 peer->rateFlag = ntohl(peer->rateFlag);
6626 peer->natMTU = ntohs(peer->natMTU);
6627 peer->maxMTU = ntohs(peer->maxMTU);
6628 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6629 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6630 peer->MTU = ntohs(peer->MTU);
6631 peer->cwind = ntohs(peer->cwind);
6632 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6633 peer->congestSeq = ntohs(peer->congestSeq);
6634 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6635 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6636 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6637 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6642 #endif /* RXDEBUG */
6647 struct rx_serverQueueEntry *np;
6650 register struct rx_call *call;
6651 register struct rx_serverQueueEntry *sq;
6655 if (rxinit_status == 1) {
6657 return; /* Already shutdown. */
6661 #ifndef AFS_PTHREAD_ENV
6662 FD_ZERO(&rx_selectMask);
6663 #endif /* AFS_PTHREAD_ENV */
6664 rxi_dataQuota = RX_MAX_QUOTA;
6665 #ifndef AFS_PTHREAD_ENV
6667 #endif /* AFS_PTHREAD_ENV */
6670 #ifndef AFS_PTHREAD_ENV
6671 #ifndef AFS_USE_GETTIMEOFDAY
6673 #endif /* AFS_USE_GETTIMEOFDAY */
6674 #endif /* AFS_PTHREAD_ENV */
6676 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6677 call = queue_First(&rx_freeCallQueue, rx_call);
6679 rxi_Free(call, sizeof(struct rx_call));
6682 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6683 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6689 struct rx_peer **peer_ptr, **peer_end;
6690 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6691 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6693 struct rx_peer *peer, *next;
6694 for (peer = *peer_ptr; peer; peer = next) {
6695 rx_interface_stat_p rpc_stat, nrpc_stat;
6698 (&peer->rpcStats, rpc_stat, nrpc_stat,
6699 rx_interface_stat)) {
6700 unsigned int num_funcs;
6703 queue_Remove(&rpc_stat->queue_header);
6704 queue_Remove(&rpc_stat->all_peers);
6705 num_funcs = rpc_stat->stats[0].func_total;
6707 sizeof(rx_interface_stat_t) +
6708 rpc_stat->stats[0].func_total *
6709 sizeof(rx_function_entry_v1_t);
6711 rxi_Free(rpc_stat, space);
6712 MUTEX_ENTER(&rx_rpc_stats);
6713 rxi_rpc_peer_stat_cnt -= num_funcs;
6714 MUTEX_EXIT(&rx_rpc_stats);
6718 MUTEX_ENTER(&rx_stats_mutex);
6719 rx_stats.nPeerStructs--;
6720 MUTEX_EXIT(&rx_stats_mutex);
6724 for (i = 0; i < RX_MAX_SERVICES; i++) {
6726 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6728 for (i = 0; i < rx_hashTableSize; i++) {
6729 register struct rx_connection *tc, *ntc;
6730 MUTEX_ENTER(&rx_connHashTable_lock);
6731 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6733 for (j = 0; j < RX_MAXCALLS; j++) {
6735 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6738 rxi_Free(tc, sizeof(*tc));
6740 MUTEX_EXIT(&rx_connHashTable_lock);
6743 MUTEX_ENTER(&freeSQEList_lock);
6745 while ((np = rx_FreeSQEList)) {
6746 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6747 MUTEX_DESTROY(&np->lock);
6748 rxi_Free(np, sizeof(*np));
6751 MUTEX_EXIT(&freeSQEList_lock);
6752 MUTEX_DESTROY(&freeSQEList_lock);
6753 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6754 MUTEX_DESTROY(&rx_connHashTable_lock);
6755 MUTEX_DESTROY(&rx_peerHashTable_lock);
6756 MUTEX_DESTROY(&rx_serverPool_lock);
6758 osi_Free(rx_connHashTable,
6759 rx_hashTableSize * sizeof(struct rx_connection *));
6760 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6762 UNPIN(rx_connHashTable,
6763 rx_hashTableSize * sizeof(struct rx_connection *));
6764 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6766 rxi_FreeAllPackets();
6768 MUTEX_ENTER(&rx_stats_mutex);
6769 rxi_dataQuota = RX_MAX_QUOTA;
6770 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6771 MUTEX_EXIT(&rx_stats_mutex);
6777 #ifdef RX_ENABLE_LOCKS
6779 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6781 if (!MUTEX_ISMINE(lockaddr))
6782 osi_Panic("Lock not held: %s", msg);
6784 #endif /* RX_ENABLE_LOCKS */
6789 * Routines to implement connection specific data.
6793 rx_KeyCreate(rx_destructor_t rtn)
6796 MUTEX_ENTER(&rxi_keyCreate_lock);
6797 key = rxi_keyCreate_counter++;
6798 rxi_keyCreate_destructor = (rx_destructor_t *)
6799 realloc((void *)rxi_keyCreate_destructor,
6800 (key + 1) * sizeof(rx_destructor_t));
6801 rxi_keyCreate_destructor[key] = rtn;
6802 MUTEX_EXIT(&rxi_keyCreate_lock);
6807 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6810 MUTEX_ENTER(&conn->conn_data_lock);
6811 if (!conn->specific) {
6812 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6813 for (i = 0; i < key; i++)
6814 conn->specific[i] = NULL;
6815 conn->nSpecific = key + 1;
6816 conn->specific[key] = ptr;
6817 } else if (key >= conn->nSpecific) {
6818 conn->specific = (void **)
6819 realloc(conn->specific, (key + 1) * sizeof(void *));
6820 for (i = conn->nSpecific; i < key; i++)
6821 conn->specific[i] = NULL;
6822 conn->nSpecific = key + 1;
6823 conn->specific[key] = ptr;
6825 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6826 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6827 conn->specific[key] = ptr;
6829 MUTEX_EXIT(&conn->conn_data_lock);
6833 rx_GetSpecific(struct rx_connection *conn, int key)
6836 MUTEX_ENTER(&conn->conn_data_lock);
6837 if (key >= conn->nSpecific)
6840 ptr = conn->specific[key];
6841 MUTEX_EXIT(&conn->conn_data_lock);
6845 #endif /* !KERNEL */
6848 * processStats is a queue used to store the statistics for the local
6849 * process. Its contents are similar to the contents of the rpcStats
6850 * queue on a rx_peer structure, but the actual data stored within
6851 * this queue contains totals across the lifetime of the process (assuming
6852 * the stats have not been reset) - unlike the per peer structures
6853 * which can come and go based upon the peer lifetime.
6856 static struct rx_queue processStats = { &processStats, &processStats };
6859 * peerStats is a queue used to store the statistics for all peer structs.
6860 * Its contents are the union of all the peer rpcStats queues.
6863 static struct rx_queue peerStats = { &peerStats, &peerStats };
6866 * rxi_monitor_processStats is used to turn process wide stat collection
6870 static int rxi_monitor_processStats = 0;
6873 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6876 static int rxi_monitor_peerStats = 0;
6879 * rxi_AddRpcStat - given all of the information for a particular rpc
6880 * call, create (if needed) and update the stat totals for the rpc.
6884 * IN stats - the queue of stats that will be updated with the new value
6886 * IN rxInterface - a unique number that identifies the rpc interface
6888 * IN currentFunc - the index of the function being invoked
6890 * IN totalFunc - the total number of functions in this interface
6892 * IN queueTime - the amount of time this function waited for a thread
6894 * IN execTime - the amount of time this function invocation took to execute
6896 * IN bytesSent - the number bytes sent by this invocation
6898 * IN bytesRcvd - the number bytes received by this invocation
6900 * IN isServer - if true, this invocation was made to a server
6902 * IN remoteHost - the ip address of the remote host
6904 * IN remotePort - the port of the remote host
6906 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6908 * INOUT counter - if a new stats structure is allocated, the counter will
6909 * be updated with the new number of allocated stat structures
6917 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6918 afs_uint32 currentFunc, afs_uint32 totalFunc,
6919 struct clock *queueTime, struct clock *execTime,
6920 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6921 afs_uint32 remoteHost, afs_uint32 remotePort,
6922 int addToPeerList, unsigned int *counter)
6925 rx_interface_stat_p rpc_stat, nrpc_stat;
6928 * See if there's already a structure for this interface
6931 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6932 if ((rpc_stat->stats[0].interfaceId == rxInterface)
6933 && (rpc_stat->stats[0].remote_is_server == isServer))
6938 * Didn't find a match so allocate a new structure and add it to the
6942 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
6943 || (rpc_stat->stats[0].interfaceId != rxInterface)
6944 || (rpc_stat->stats[0].remote_is_server != isServer)) {
6949 sizeof(rx_interface_stat_t) +
6950 totalFunc * sizeof(rx_function_entry_v1_t);
6952 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
6953 if (rpc_stat == NULL) {
6957 *counter += totalFunc;
6958 for (i = 0; i < totalFunc; i++) {
6959 rpc_stat->stats[i].remote_peer = remoteHost;
6960 rpc_stat->stats[i].remote_port = remotePort;
6961 rpc_stat->stats[i].remote_is_server = isServer;
6962 rpc_stat->stats[i].interfaceId = rxInterface;
6963 rpc_stat->stats[i].func_total = totalFunc;
6964 rpc_stat->stats[i].func_index = i;
6965 hzero(rpc_stat->stats[i].invocations);
6966 hzero(rpc_stat->stats[i].bytes_sent);
6967 hzero(rpc_stat->stats[i].bytes_rcvd);
6968 rpc_stat->stats[i].queue_time_sum.sec = 0;
6969 rpc_stat->stats[i].queue_time_sum.usec = 0;
6970 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
6971 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
6972 rpc_stat->stats[i].queue_time_min.sec = 9999999;
6973 rpc_stat->stats[i].queue_time_min.usec = 9999999;
6974 rpc_stat->stats[i].queue_time_max.sec = 0;
6975 rpc_stat->stats[i].queue_time_max.usec = 0;
6976 rpc_stat->stats[i].execution_time_sum.sec = 0;
6977 rpc_stat->stats[i].execution_time_sum.usec = 0;
6978 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
6979 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
6980 rpc_stat->stats[i].execution_time_min.sec = 9999999;
6981 rpc_stat->stats[i].execution_time_min.usec = 9999999;
6982 rpc_stat->stats[i].execution_time_max.sec = 0;
6983 rpc_stat->stats[i].execution_time_max.usec = 0;
6985 queue_Prepend(stats, rpc_stat);
6986 if (addToPeerList) {
6987 queue_Prepend(&peerStats, &rpc_stat->all_peers);
6992 * Increment the stats for this function
6995 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
6996 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
6997 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
6998 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
6999 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7000 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7001 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7003 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7004 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7006 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7007 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7009 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7010 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7012 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7013 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7021 * rx_IncrementTimeAndCount - increment the times and count for a particular
7026 * IN peer - the peer who invoked the rpc
7028 * IN rxInterface - a unique number that identifies the rpc interface
7030 * IN currentFunc - the index of the function being invoked
7032 * IN totalFunc - the total number of functions in this interface
7034 * IN queueTime - the amount of time this function waited for a thread
7036 * IN execTime - the amount of time this function invocation took to execute
7038 * IN bytesSent - the number bytes sent by this invocation
7040 * IN bytesRcvd - the number bytes received by this invocation
7042 * IN isServer - if true, this invocation was made to a server
7050 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7051 afs_uint32 currentFunc, afs_uint32 totalFunc,
7052 struct clock *queueTime, struct clock *execTime,
7053 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7057 MUTEX_ENTER(&rx_rpc_stats);
7058 MUTEX_ENTER(&peer->peer_lock);
7060 if (rxi_monitor_peerStats) {
7061 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7062 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7063 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7066 if (rxi_monitor_processStats) {
7067 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7068 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7069 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7072 MUTEX_EXIT(&peer->peer_lock);
7073 MUTEX_EXIT(&rx_rpc_stats);
7078 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7082 * IN callerVersion - the rpc stat version of the caller.
7084 * IN count - the number of entries to marshall.
7086 * IN stats - pointer to stats to be marshalled.
7088 * OUT ptr - Where to store the marshalled data.
7095 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7096 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7102 * We only support the first version
7104 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7105 *(ptr++) = stats->remote_peer;
7106 *(ptr++) = stats->remote_port;
7107 *(ptr++) = stats->remote_is_server;
7108 *(ptr++) = stats->interfaceId;
7109 *(ptr++) = stats->func_total;
7110 *(ptr++) = stats->func_index;
7111 *(ptr++) = hgethi(stats->invocations);
7112 *(ptr++) = hgetlo(stats->invocations);
7113 *(ptr++) = hgethi(stats->bytes_sent);
7114 *(ptr++) = hgetlo(stats->bytes_sent);
7115 *(ptr++) = hgethi(stats->bytes_rcvd);
7116 *(ptr++) = hgetlo(stats->bytes_rcvd);
7117 *(ptr++) = stats->queue_time_sum.sec;
7118 *(ptr++) = stats->queue_time_sum.usec;
7119 *(ptr++) = stats->queue_time_sum_sqr.sec;
7120 *(ptr++) = stats->queue_time_sum_sqr.usec;
7121 *(ptr++) = stats->queue_time_min.sec;
7122 *(ptr++) = stats->queue_time_min.usec;
7123 *(ptr++) = stats->queue_time_max.sec;
7124 *(ptr++) = stats->queue_time_max.usec;
7125 *(ptr++) = stats->execution_time_sum.sec;
7126 *(ptr++) = stats->execution_time_sum.usec;
7127 *(ptr++) = stats->execution_time_sum_sqr.sec;
7128 *(ptr++) = stats->execution_time_sum_sqr.usec;
7129 *(ptr++) = stats->execution_time_min.sec;
7130 *(ptr++) = stats->execution_time_min.usec;
7131 *(ptr++) = stats->execution_time_max.sec;
7132 *(ptr++) = stats->execution_time_max.usec;
7138 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7143 * IN callerVersion - the rpc stat version of the caller
7145 * OUT myVersion - the rpc stat version of this function
7147 * OUT clock_sec - local time seconds
7149 * OUT clock_usec - local time microseconds
7151 * OUT allocSize - the number of bytes allocated to contain stats
7153 * OUT statCount - the number stats retrieved from this process.
7155 * OUT stats - the actual stats retrieved from this process.
7159 * Returns void. If successful, stats will != NULL.
7163 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7164 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7165 size_t * allocSize, afs_uint32 * statCount,
7166 afs_uint32 ** stats)
7176 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7179 * Check to see if stats are enabled
7182 MUTEX_ENTER(&rx_rpc_stats);
7183 if (!rxi_monitor_processStats) {
7184 MUTEX_EXIT(&rx_rpc_stats);
7188 clock_GetTime(&now);
7189 *clock_sec = now.sec;
7190 *clock_usec = now.usec;
7193 * Allocate the space based upon the caller version
7195 * If the client is at an older version than we are,
7196 * we return the statistic data in the older data format, but
7197 * we still return our version number so the client knows we
7198 * are maintaining more data than it can retrieve.
7201 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7202 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7203 *statCount = rxi_rpc_process_stat_cnt;
7206 * This can't happen yet, but in the future version changes
7207 * can be handled by adding additional code here
7211 if (space > (size_t) 0) {
7213 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7216 rx_interface_stat_p rpc_stat, nrpc_stat;
7220 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7222 * Copy the data based upon the caller version
7224 rx_MarshallProcessRPCStats(callerVersion,
7225 rpc_stat->stats[0].func_total,
7226 rpc_stat->stats, &ptr);
7232 MUTEX_EXIT(&rx_rpc_stats);
7237 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7241 * IN callerVersion - the rpc stat version of the caller
7243 * OUT myVersion - the rpc stat version of this function
7245 * OUT clock_sec - local time seconds
7247 * OUT clock_usec - local time microseconds
7249 * OUT allocSize - the number of bytes allocated to contain stats
7251 * OUT statCount - the number of stats retrieved from the individual
7254 * OUT stats - the actual stats retrieved from the individual peer structures.
7258 * Returns void. If successful, stats will != NULL.
7262 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7263 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7264 size_t * allocSize, afs_uint32 * statCount,
7265 afs_uint32 ** stats)
7275 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7278 * Check to see if stats are enabled
7281 MUTEX_ENTER(&rx_rpc_stats);
7282 if (!rxi_monitor_peerStats) {
7283 MUTEX_EXIT(&rx_rpc_stats);
7287 clock_GetTime(&now);
7288 *clock_sec = now.sec;
7289 *clock_usec = now.usec;
7292 * Allocate the space based upon the caller version
7294 * If the client is at an older version than we are,
7295 * we return the statistic data in the older data format, but
7296 * we still return our version number so the client knows we
7297 * are maintaining more data than it can retrieve.
7300 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7301 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7302 *statCount = rxi_rpc_peer_stat_cnt;
7305 * This can't happen yet, but in the future version changes
7306 * can be handled by adding additional code here
7310 if (space > (size_t) 0) {
7312 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7315 rx_interface_stat_p rpc_stat, nrpc_stat;
7319 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7321 * We have to fix the offset of rpc_stat since we are
7322 * keeping this structure on two rx_queues. The rx_queue
7323 * package assumes that the rx_queue member is the first
7324 * member of the structure. That is, rx_queue assumes that
7325 * any one item is only on one queue at a time. We are
7326 * breaking that assumption and so we have to do a little
7327 * math to fix our pointers.
7330 fix_offset = (char *)rpc_stat;
7331 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7332 rpc_stat = (rx_interface_stat_p) fix_offset;
7335 * Copy the data based upon the caller version
7337 rx_MarshallProcessRPCStats(callerVersion,
7338 rpc_stat->stats[0].func_total,
7339 rpc_stat->stats, &ptr);
7345 MUTEX_EXIT(&rx_rpc_stats);
7350 * rx_FreeRPCStats - free memory allocated by
7351 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7355 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7356 * rx_RetrievePeerRPCStats
7358 * IN allocSize - the number of bytes in stats.
7366 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7368 rxi_Free(stats, allocSize);
7372 * rx_queryProcessRPCStats - see if process rpc stat collection is
7373 * currently enabled.
7379 * Returns 0 if stats are not enabled != 0 otherwise
7383 rx_queryProcessRPCStats(void)
7386 MUTEX_ENTER(&rx_rpc_stats);
7387 rc = rxi_monitor_processStats;
7388 MUTEX_EXIT(&rx_rpc_stats);
7393 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7399 * Returns 0 if stats are not enabled != 0 otherwise
7403 rx_queryPeerRPCStats(void)
7406 MUTEX_ENTER(&rx_rpc_stats);
7407 rc = rxi_monitor_peerStats;
7408 MUTEX_EXIT(&rx_rpc_stats);
7413 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7423 rx_enableProcessRPCStats(void)
7425 MUTEX_ENTER(&rx_rpc_stats);
7426 rx_enable_stats = 1;
7427 rxi_monitor_processStats = 1;
7428 MUTEX_EXIT(&rx_rpc_stats);
7432 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7442 rx_enablePeerRPCStats(void)
7444 MUTEX_ENTER(&rx_rpc_stats);
7445 rx_enable_stats = 1;
7446 rxi_monitor_peerStats = 1;
7447 MUTEX_EXIT(&rx_rpc_stats);
7451 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7461 rx_disableProcessRPCStats(void)
7463 rx_interface_stat_p rpc_stat, nrpc_stat;
7466 MUTEX_ENTER(&rx_rpc_stats);
7469 * Turn off process statistics and if peer stats is also off, turn
7473 rxi_monitor_processStats = 0;
7474 if (rxi_monitor_peerStats == 0) {
7475 rx_enable_stats = 0;
7478 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7479 unsigned int num_funcs = 0;
7482 queue_Remove(rpc_stat);
7483 num_funcs = rpc_stat->stats[0].func_total;
7485 sizeof(rx_interface_stat_t) +
7486 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7488 rxi_Free(rpc_stat, space);
7489 rxi_rpc_process_stat_cnt -= num_funcs;
7491 MUTEX_EXIT(&rx_rpc_stats);
7495 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7505 rx_disablePeerRPCStats(void)
7507 struct rx_peer **peer_ptr, **peer_end;
7510 MUTEX_ENTER(&rx_rpc_stats);
7513 * Turn off peer statistics and if process stats is also off, turn
7517 rxi_monitor_peerStats = 0;
7518 if (rxi_monitor_processStats == 0) {
7519 rx_enable_stats = 0;
7522 MUTEX_ENTER(&rx_peerHashTable_lock);
7523 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7524 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7526 struct rx_peer *peer, *next, *prev;
7527 for (prev = peer = *peer_ptr; peer; peer = next) {
7529 code = MUTEX_TRYENTER(&peer->peer_lock);
7531 rx_interface_stat_p rpc_stat, nrpc_stat;
7534 (&peer->rpcStats, rpc_stat, nrpc_stat,
7535 rx_interface_stat)) {
7536 unsigned int num_funcs = 0;
7539 queue_Remove(&rpc_stat->queue_header);
7540 queue_Remove(&rpc_stat->all_peers);
7541 num_funcs = rpc_stat->stats[0].func_total;
7543 sizeof(rx_interface_stat_t) +
7544 rpc_stat->stats[0].func_total *
7545 sizeof(rx_function_entry_v1_t);
7547 rxi_Free(rpc_stat, space);
7548 rxi_rpc_peer_stat_cnt -= num_funcs;
7550 MUTEX_EXIT(&peer->peer_lock);
7551 if (prev == *peer_ptr) {
7561 MUTEX_EXIT(&rx_peerHashTable_lock);
7562 MUTEX_EXIT(&rx_rpc_stats);
7566 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7571 * IN clearFlag - flag indicating which stats to clear
7579 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7581 rx_interface_stat_p rpc_stat, nrpc_stat;
7583 MUTEX_ENTER(&rx_rpc_stats);
7585 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7586 unsigned int num_funcs = 0, i;
7587 num_funcs = rpc_stat->stats[0].func_total;
7588 for (i = 0; i < num_funcs; i++) {
7589 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7590 hzero(rpc_stat->stats[i].invocations);
7592 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7593 hzero(rpc_stat->stats[i].bytes_sent);
7595 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7596 hzero(rpc_stat->stats[i].bytes_rcvd);
7598 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7599 rpc_stat->stats[i].queue_time_sum.sec = 0;
7600 rpc_stat->stats[i].queue_time_sum.usec = 0;
7602 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7603 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7604 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7606 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7607 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7608 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7610 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7611 rpc_stat->stats[i].queue_time_max.sec = 0;
7612 rpc_stat->stats[i].queue_time_max.usec = 0;
7614 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7615 rpc_stat->stats[i].execution_time_sum.sec = 0;
7616 rpc_stat->stats[i].execution_time_sum.usec = 0;
7618 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7619 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7620 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7622 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7623 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7624 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7626 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7627 rpc_stat->stats[i].execution_time_max.sec = 0;
7628 rpc_stat->stats[i].execution_time_max.usec = 0;
7633 MUTEX_EXIT(&rx_rpc_stats);
7637 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7642 * IN clearFlag - flag indicating which stats to clear
7650 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7652 rx_interface_stat_p rpc_stat, nrpc_stat;
7654 MUTEX_ENTER(&rx_rpc_stats);
7656 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7657 unsigned int num_funcs = 0, i;
7660 * We have to fix the offset of rpc_stat since we are
7661 * keeping this structure on two rx_queues. The rx_queue
7662 * package assumes that the rx_queue member is the first
7663 * member of the structure. That is, rx_queue assumes that
7664 * any one item is only on one queue at a time. We are
7665 * breaking that assumption and so we have to do a little
7666 * math to fix our pointers.
7669 fix_offset = (char *)rpc_stat;
7670 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7671 rpc_stat = (rx_interface_stat_p) fix_offset;
7673 num_funcs = rpc_stat->stats[0].func_total;
7674 for (i = 0; i < num_funcs; i++) {
7675 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7676 hzero(rpc_stat->stats[i].invocations);
7678 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7679 hzero(rpc_stat->stats[i].bytes_sent);
7681 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7682 hzero(rpc_stat->stats[i].bytes_rcvd);
7684 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7685 rpc_stat->stats[i].queue_time_sum.sec = 0;
7686 rpc_stat->stats[i].queue_time_sum.usec = 0;
7688 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7689 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7690 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7692 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7693 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7694 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7696 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7697 rpc_stat->stats[i].queue_time_max.sec = 0;
7698 rpc_stat->stats[i].queue_time_max.usec = 0;
7700 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7701 rpc_stat->stats[i].execution_time_sum.sec = 0;
7702 rpc_stat->stats[i].execution_time_sum.usec = 0;
7704 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7705 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7706 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7708 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7709 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7710 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7712 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7713 rpc_stat->stats[i].execution_time_max.sec = 0;
7714 rpc_stat->stats[i].execution_time_max.usec = 0;
7719 MUTEX_EXIT(&rx_rpc_stats);
7723 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7724 * is authorized to enable/disable/clear RX statistics.
7726 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7729 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7731 rxi_rxstat_userok = proc;
7735 rx_RxStatUserOk(struct rx_call *call)
7737 if (!rxi_rxstat_userok)
7739 return rxi_rxstat_userok(call);