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"
37 #include "inet/common.h"
39 #include "inet/ip_ire.h"
41 #include "afs/afs_args.h"
42 #include "afs/afs_osi.h"
43 #ifdef RX_KERNEL_TRACE
44 #include "rx_kcommon.h"
46 #if (defined(AFS_AUX_ENV) || defined(AFS_AIX_ENV))
50 #undef RXDEBUG /* turn off debugging */
52 #if defined(AFS_SGI_ENV)
53 #include "sys/debug.h"
62 #endif /* AFS_OSF_ENV */
64 #include "afs/sysincludes.h"
65 #include "afsincludes.h"
68 #include "rx_kmutex.h"
69 #include "rx_kernel.h"
73 #include "rx_globals.h"
75 #define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
76 #define AFSOP_STOP_AFS 211 /* Stop AFS process */
77 #define AFSOP_STOP_BKG 212 /* Stop BKG process */
79 extern afs_int32 afs_termState;
81 #include "sys/lockl.h"
82 #include "sys/lock_def.h"
83 #endif /* AFS_AIX41_ENV */
84 # include "rxgen_consts.h"
86 # include <sys/types.h>
92 # include <afs/afsutil.h>
93 # include <WINNT\afsreg.h>
95 # include <sys/socket.h>
96 # include <sys/file.h>
98 # include <sys/stat.h>
99 # include <netinet/in.h>
100 # include <sys/time.h>
103 # include "rx_user.h"
104 # include "rx_clock.h"
105 # include "rx_queue.h"
106 # include "rx_globals.h"
107 # include "rx_trace.h"
108 # include <afs/rxgen_consts.h>
111 int (*registerProgram) () = 0;
112 int (*swapNameProgram) () = 0;
114 /* Local static routines */
115 static void rxi_DestroyConnectionNoLock(register struct rx_connection *conn);
116 #ifdef RX_ENABLE_LOCKS
117 static void rxi_SetAcksInTransmitQueue(register struct rx_call *call);
120 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
122 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
123 afs_int32 rxi_start_in_error;
125 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
128 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
129 * currently allocated within rx. This number is used to allocate the
130 * memory required to return the statistics when queried.
133 static unsigned int rxi_rpc_peer_stat_cnt;
136 * rxi_rpc_process_stat_cnt counts the total number of local process stat
137 * structures currently allocated within rx. The number is used to allocate
138 * the memory required to return the statistics when queried.
141 static unsigned int rxi_rpc_process_stat_cnt;
143 #if !defined(offsetof)
144 #include <stddef.h> /* for definition of offsetof() */
147 #ifdef AFS_PTHREAD_ENV
151 * Use procedural initialization of mutexes/condition variables
155 extern pthread_mutex_t rx_stats_mutex;
156 extern pthread_mutex_t des_init_mutex;
157 extern pthread_mutex_t des_random_mutex;
158 extern pthread_mutex_t rx_clock_mutex;
159 extern pthread_mutex_t rxi_connCacheMutex;
160 extern pthread_mutex_t rx_event_mutex;
161 extern pthread_mutex_t osi_malloc_mutex;
162 extern pthread_mutex_t event_handler_mutex;
163 extern pthread_mutex_t listener_mutex;
164 extern pthread_mutex_t rx_if_init_mutex;
165 extern pthread_mutex_t rx_if_mutex;
166 extern pthread_mutex_t rxkad_client_uid_mutex;
167 extern pthread_mutex_t rxkad_random_mutex;
169 extern pthread_cond_t rx_event_handler_cond;
170 extern pthread_cond_t rx_listener_cond;
172 static pthread_mutex_t epoch_mutex;
173 static pthread_mutex_t rx_init_mutex;
174 static pthread_mutex_t rx_debug_mutex;
177 rxi_InitPthread(void)
179 assert(pthread_mutex_init(&rx_clock_mutex, (const pthread_mutexattr_t *)0)
181 assert(pthread_mutex_init(&rx_stats_mutex, (const pthread_mutexattr_t *)0)
183 assert(pthread_mutex_init
184 (&rxi_connCacheMutex, (const pthread_mutexattr_t *)0) == 0);
185 assert(pthread_mutex_init(&rx_init_mutex, (const pthread_mutexattr_t *)0)
187 assert(pthread_mutex_init(&epoch_mutex, (const pthread_mutexattr_t *)0) ==
189 assert(pthread_mutex_init(&rx_event_mutex, (const pthread_mutexattr_t *)0)
191 assert(pthread_mutex_init(&des_init_mutex, (const pthread_mutexattr_t *)0)
193 assert(pthread_mutex_init
194 (&des_random_mutex, (const pthread_mutexattr_t *)0) == 0);
195 assert(pthread_mutex_init
196 (&osi_malloc_mutex, (const pthread_mutexattr_t *)0) == 0);
197 assert(pthread_mutex_init
198 (&event_handler_mutex, (const pthread_mutexattr_t *)0) == 0);
199 assert(pthread_mutex_init(&listener_mutex, (const pthread_mutexattr_t *)0)
201 assert(pthread_mutex_init
202 (&rx_if_init_mutex, (const pthread_mutexattr_t *)0) == 0);
203 assert(pthread_mutex_init(&rx_if_mutex, (const pthread_mutexattr_t *)0) ==
205 assert(pthread_mutex_init
206 (&rxkad_client_uid_mutex, (const pthread_mutexattr_t *)0) == 0);
207 assert(pthread_mutex_init
208 (&rxkad_random_mutex, (const pthread_mutexattr_t *)0) == 0);
209 assert(pthread_mutex_init(&rx_debug_mutex, (const pthread_mutexattr_t *)0)
212 assert(pthread_cond_init
213 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
214 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
216 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
217 assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
219 rxkad_global_stats_init();
222 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
223 #define INIT_PTHREAD_LOCKS \
224 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
226 * The rx_stats_mutex mutex protects the following global variables:
231 * rxi_lowConnRefCount
232 * rxi_lowPeerRefCount
241 #define INIT_PTHREAD_LOCKS
245 /* Variables for handling the minProcs implementation. availProcs gives the
246 * number of threads available in the pool at this moment (not counting dudes
247 * executing right now). totalMin gives the total number of procs required
248 * for handling all minProcs requests. minDeficit is a dynamic variable
249 * tracking the # of procs required to satisfy all of the remaining minProcs
251 * For fine grain locking to work, the quota check and the reservation of
252 * a server thread has to come while rxi_availProcs and rxi_minDeficit
253 * are locked. To this end, the code has been modified under #ifdef
254 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
255 * same time. A new function, ReturnToServerPool() returns the allocation.
257 * A call can be on several queue's (but only one at a time). When
258 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
259 * that no one else is touching the queue. To this end, we store the address
260 * of the queue lock in the call structure (under the call lock) when we
261 * put the call on a queue, and we clear the call_queue_lock when the
262 * call is removed from a queue (once the call lock has been obtained).
263 * This allows rxi_ResetCall to safely synchronize with others wishing
264 * to manipulate the queue.
267 #ifdef RX_ENABLE_LOCKS
268 static afs_kmutex_t rx_rpc_stats;
269 void rxi_StartUnlocked();
272 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
273 ** pretty good that the next packet coming in is from the same connection
274 ** as the last packet, since we're send multiple packets in a transmit window.
276 struct rx_connection *rxLastConn = 0;
278 #ifdef RX_ENABLE_LOCKS
279 /* The locking hierarchy for rx fine grain locking is composed of these
282 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
283 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
284 * call->lock - locks call data fields.
285 * These are independent of each other:
286 * rx_freeCallQueue_lock
291 * serverQueueEntry->lock
293 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
294 * peer->lock - locks peer data fields.
295 * conn_data_lock - that more than one thread is not updating a conn data
296 * field at the same time.
304 * Do we need a lock to protect the peer field in the conn structure?
305 * conn->peer was previously a constant for all intents and so has no
306 * lock protecting this field. The multihomed client delta introduced
307 * a RX code change : change the peer field in the connection structure
308 * to that remote inetrface from which the last packet for this
309 * connection was sent out. This may become an issue if further changes
312 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
313 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
315 /* rxdb_fileID is used to identify the lock location, along with line#. */
316 static int rxdb_fileID = RXDB_FILE_RX;
317 #endif /* RX_LOCKS_DB */
318 #else /* RX_ENABLE_LOCKS */
319 #define SET_CALL_QUEUE_LOCK(C, L)
320 #define CLEAR_CALL_QUEUE_LOCK(C)
321 #endif /* RX_ENABLE_LOCKS */
322 struct rx_serverQueueEntry *rx_waitForPacket = 0;
323 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
325 /* ------------Exported Interfaces------------- */
327 /* This function allows rxkad to set the epoch to a suitably random number
328 * which rx_NewConnection will use in the future. The principle purpose is to
329 * get rxnull connections to use the same epoch as the rxkad connections do, at
330 * least once the first rxkad connection is established. This is important now
331 * that the host/port addresses aren't used in FindConnection: the uniqueness
332 * of epoch/cid matters and the start time won't do. */
334 #ifdef AFS_PTHREAD_ENV
336 * This mutex protects the following global variables:
340 #define LOCK_EPOCH assert(pthread_mutex_lock(&epoch_mutex)==0)
341 #define UNLOCK_EPOCH assert(pthread_mutex_unlock(&epoch_mutex)==0)
345 #endif /* AFS_PTHREAD_ENV */
348 rx_SetEpoch(afs_uint32 epoch)
355 /* Initialize rx. A port number may be mentioned, in which case this
356 * becomes the default port number for any service installed later.
357 * If 0 is provided for the port number, a random port will be chosen
358 * by the kernel. Whether this will ever overlap anything in
359 * /etc/services is anybody's guess... Returns 0 on success, -1 on
364 int rxinit_status = 1;
365 #ifdef AFS_PTHREAD_ENV
367 * This mutex protects the following global variables:
371 #define LOCK_RX_INIT assert(pthread_mutex_lock(&rx_init_mutex)==0)
372 #define UNLOCK_RX_INIT assert(pthread_mutex_unlock(&rx_init_mutex)==0)
375 #define UNLOCK_RX_INIT
379 rx_InitHost(u_int host, u_int port)
386 char *htable, *ptable;
389 #if defined(AFS_DJGPP_ENV) && !defined(DEBUG)
390 __djgpp_set_quiet_socket(1);
397 if (rxinit_status == 0) {
398 tmp_status = rxinit_status;
400 return tmp_status; /* Already started; return previous error code. */
406 if (afs_winsockInit() < 0)
412 * Initialize anything necessary to provide a non-premptive threading
415 rxi_InitializeThreadSupport();
418 /* Allocate and initialize a socket for client and perhaps server
421 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
422 if (rx_socket == OSI_NULLSOCKET) {
426 #ifdef RX_ENABLE_LOCKS
429 #endif /* RX_LOCKS_DB */
430 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
431 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
432 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
433 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
434 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
436 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
438 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
440 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
442 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
444 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
446 #if defined(KERNEL) && defined(AFS_HPUX110_ENV)
448 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
449 #endif /* KERNEL && AFS_HPUX110_ENV */
450 #endif /* RX_ENABLE_LOCKS */
453 rx_connDeadTime = 12;
454 rx_tranquil = 0; /* reset flag */
455 memset((char *)&rx_stats, 0, sizeof(struct rx_stats));
457 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
458 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
459 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
460 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
461 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
462 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
464 /* Malloc up a bunch of packets & buffers */
466 queue_Init(&rx_freePacketQueue);
467 rxi_NeedMorePackets = FALSE;
468 #ifdef RX_ENABLE_TSFPQ
469 rx_nPackets = 0; /* in TSFPQ version, rx_nPackets is managed by rxi_MorePackets* */
470 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
471 #else /* RX_ENABLE_TSFPQ */
472 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
473 rxi_MorePackets(rx_nPackets);
474 #endif /* RX_ENABLE_TSFPQ */
481 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
482 tv.tv_sec = clock_now.sec;
483 tv.tv_usec = clock_now.usec;
484 srand((unsigned int)tv.tv_usec);
491 #if defined(KERNEL) && !defined(UKERNEL)
492 /* Really, this should never happen in a real kernel */
495 struct sockaddr_in addr;
496 int addrlen = sizeof(addr);
497 if (getsockname((int)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
501 rx_port = addr.sin_port;
504 rx_stats.minRtt.sec = 9999999;
506 rx_SetEpoch(tv.tv_sec | 0x80000000);
508 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
509 * will provide a randomer value. */
511 MUTEX_ENTER(&rx_stats_mutex);
512 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
513 MUTEX_EXIT(&rx_stats_mutex);
514 /* *Slightly* random start time for the cid. This is just to help
515 * out with the hashing function at the peer */
516 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
517 rx_connHashTable = (struct rx_connection **)htable;
518 rx_peerHashTable = (struct rx_peer **)ptable;
520 rx_lastAckDelay.sec = 0;
521 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
522 rx_hardAckDelay.sec = 0;
523 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
524 rx_softAckDelay.sec = 0;
525 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
527 rxevent_Init(20, rxi_ReScheduleEvents);
529 /* Initialize various global queues */
530 queue_Init(&rx_idleServerQueue);
531 queue_Init(&rx_incomingCallQueue);
532 queue_Init(&rx_freeCallQueue);
534 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
535 /* Initialize our list of usable IP addresses. */
539 /* Start listener process (exact function is dependent on the
540 * implementation environment--kernel or user space) */
544 tmp_status = rxinit_status = 0;
552 return rx_InitHost(htonl(INADDR_ANY), port);
555 /* called with unincremented nRequestsRunning to see if it is OK to start
556 * a new thread in this service. Could be "no" for two reasons: over the
557 * max quota, or would prevent others from reaching their min quota.
559 #ifdef RX_ENABLE_LOCKS
560 /* This verion of QuotaOK reserves quota if it's ok while the
561 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
564 QuotaOK(register struct rx_service *aservice)
566 /* check if over max quota */
567 if (aservice->nRequestsRunning >= aservice->maxProcs) {
571 /* under min quota, we're OK */
572 /* otherwise, can use only if there are enough to allow everyone
573 * to go to their min quota after this guy starts.
575 MUTEX_ENTER(&rx_stats_mutex);
576 if ((aservice->nRequestsRunning < aservice->minProcs)
577 || (rxi_availProcs > rxi_minDeficit)) {
578 aservice->nRequestsRunning++;
579 /* just started call in minProcs pool, need fewer to maintain
581 if (aservice->nRequestsRunning <= aservice->minProcs)
584 MUTEX_EXIT(&rx_stats_mutex);
587 MUTEX_EXIT(&rx_stats_mutex);
593 ReturnToServerPool(register struct rx_service *aservice)
595 aservice->nRequestsRunning--;
596 MUTEX_ENTER(&rx_stats_mutex);
597 if (aservice->nRequestsRunning < aservice->minProcs)
600 MUTEX_EXIT(&rx_stats_mutex);
603 #else /* RX_ENABLE_LOCKS */
605 QuotaOK(register struct rx_service *aservice)
608 /* under min quota, we're OK */
609 if (aservice->nRequestsRunning < aservice->minProcs)
612 /* check if over max quota */
613 if (aservice->nRequestsRunning >= aservice->maxProcs)
616 /* otherwise, can use only if there are enough to allow everyone
617 * to go to their min quota after this guy starts.
619 if (rxi_availProcs > rxi_minDeficit)
623 #endif /* RX_ENABLE_LOCKS */
626 /* Called by rx_StartServer to start up lwp's to service calls.
627 NExistingProcs gives the number of procs already existing, and which
628 therefore needn't be created. */
630 rxi_StartServerProcs(int nExistingProcs)
632 register struct rx_service *service;
637 /* For each service, reserve N processes, where N is the "minimum"
638 * number of processes that MUST be able to execute a request in parallel,
639 * at any time, for that process. Also compute the maximum difference
640 * between any service's maximum number of processes that can run
641 * (i.e. the maximum number that ever will be run, and a guarantee
642 * that this number will run if other services aren't running), and its
643 * minimum number. The result is the extra number of processes that
644 * we need in order to provide the latter guarantee */
645 for (i = 0; i < RX_MAX_SERVICES; i++) {
647 service = rx_services[i];
648 if (service == (struct rx_service *)0)
650 nProcs += service->minProcs;
651 diff = service->maxProcs - service->minProcs;
655 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
656 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
657 for (i = 0; i < nProcs; i++) {
658 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
664 /* This routine is only required on Windows */
666 rx_StartClientThread(void)
668 #ifdef AFS_PTHREAD_ENV
670 pid = pthread_self();
671 #endif /* AFS_PTHREAD_ENV */
673 #endif /* AFS_NT40_ENV */
675 /* This routine must be called if any services are exported. If the
676 * donateMe flag is set, the calling process is donated to the server
679 rx_StartServer(int donateMe)
681 register struct rx_service *service;
687 /* Start server processes, if necessary (exact function is dependent
688 * on the implementation environment--kernel or user space). DonateMe
689 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
690 * case, one less new proc will be created rx_StartServerProcs.
692 rxi_StartServerProcs(donateMe);
694 /* count up the # of threads in minProcs, and add set the min deficit to
695 * be that value, too.
697 for (i = 0; i < RX_MAX_SERVICES; i++) {
698 service = rx_services[i];
699 if (service == (struct rx_service *)0)
701 MUTEX_ENTER(&rx_stats_mutex);
702 rxi_totalMin += service->minProcs;
703 /* below works even if a thread is running, since minDeficit would
704 * still have been decremented and later re-incremented.
706 rxi_minDeficit += service->minProcs;
707 MUTEX_EXIT(&rx_stats_mutex);
710 /* Turn on reaping of idle server connections */
711 rxi_ReapConnections();
720 #ifdef AFS_PTHREAD_ENV
722 pid = (pid_t) pthread_self();
723 #else /* AFS_PTHREAD_ENV */
725 LWP_CurrentProcess(&pid);
726 #endif /* AFS_PTHREAD_ENV */
728 sprintf(name, "srv_%d", ++nProcs);
730 (*registerProgram) (pid, name);
732 #endif /* AFS_NT40_ENV */
733 rx_ServerProc(NULL); /* Never returns */
735 #ifdef RX_ENABLE_TSFPQ
736 /* no use leaving packets around in this thread's local queue if
737 * it isn't getting donated to the server thread pool.
739 rxi_FlushLocalPacketsTSFPQ();
740 #endif /* RX_ENABLE_TSFPQ */
744 /* Create a new client connection to the specified service, using the
745 * specified security object to implement the security model for this
747 struct rx_connection *
748 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
749 register struct rx_securityClass *securityObject,
750 int serviceSecurityIndex)
754 register struct rx_connection *conn;
759 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", ntohl(shost), ntohs(sport), sservice, securityObject, serviceSecurityIndex));
761 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
762 * the case of kmem_alloc? */
763 conn = rxi_AllocConnection();
764 #ifdef RX_ENABLE_LOCKS
765 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
766 MUTEX_INIT(&conn->conn_data_lock, "conn call lock", MUTEX_DEFAULT, 0);
767 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
770 MUTEX_ENTER(&rx_connHashTable_lock);
771 cid = (rx_nextCid += RX_MAXCALLS);
772 conn->type = RX_CLIENT_CONNECTION;
774 conn->epoch = rx_epoch;
775 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
776 conn->serviceId = sservice;
777 conn->securityObject = securityObject;
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;
788 for (i = 0; i < RX_MAXCALLS; i++) {
789 conn->twind[i] = rx_initSendWindow;
790 conn->rwind[i] = rx_initReceiveWindow;
793 RXS_NewConnection(securityObject, conn);
795 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
797 conn->refCount++; /* no lock required since only this thread knows... */
798 conn->next = rx_connHashTable[hashindex];
799 rx_connHashTable[hashindex] = conn;
800 rx_MutexIncrement(rx_stats.nClientConns, rx_stats_mutex);
801 MUTEX_EXIT(&rx_connHashTable_lock);
807 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
809 /* The idea is to set the dead time to a value that allows several
810 * keepalives to be dropped without timing out the connection. */
811 conn->secondsUntilDead = MAX(seconds, 6);
812 conn->secondsUntilPing = conn->secondsUntilDead / 6;
815 int rxi_lowPeerRefCount = 0;
816 int rxi_lowConnRefCount = 0;
819 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
820 * NOTE: must not be called with rx_connHashTable_lock held.
823 rxi_CleanupConnection(struct rx_connection *conn)
825 /* Notify the service exporter, if requested, that this connection
826 * is being destroyed */
827 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
828 (*conn->service->destroyConnProc) (conn);
830 /* Notify the security module that this connection is being destroyed */
831 RXS_DestroyConnection(conn->securityObject, conn);
833 /* If this is the last connection using the rx_peer struct, set its
834 * idle time to now. rxi_ReapConnections will reap it if it's still
835 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
837 MUTEX_ENTER(&rx_peerHashTable_lock);
838 if (conn->peer->refCount < 2) {
839 conn->peer->idleWhen = clock_Sec();
840 if (conn->peer->refCount < 1) {
841 conn->peer->refCount = 1;
842 MUTEX_ENTER(&rx_stats_mutex);
843 rxi_lowPeerRefCount++;
844 MUTEX_EXIT(&rx_stats_mutex);
847 conn->peer->refCount--;
848 MUTEX_EXIT(&rx_peerHashTable_lock);
850 if (conn->type == RX_SERVER_CONNECTION)
851 rx_MutexDecrement(rx_stats.nServerConns, rx_stats_mutex);
853 rx_MutexDecrement(rx_stats.nClientConns, rx_stats_mutex);
855 if (conn->specific) {
857 for (i = 0; i < conn->nSpecific; i++) {
858 if (conn->specific[i] && rxi_keyCreate_destructor[i])
859 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
860 conn->specific[i] = NULL;
862 free(conn->specific);
864 conn->specific = NULL;
868 MUTEX_DESTROY(&conn->conn_call_lock);
869 MUTEX_DESTROY(&conn->conn_data_lock);
870 CV_DESTROY(&conn->conn_call_cv);
872 rxi_FreeConnection(conn);
875 /* Destroy the specified connection */
877 rxi_DestroyConnection(register struct rx_connection *conn)
879 MUTEX_ENTER(&rx_connHashTable_lock);
880 rxi_DestroyConnectionNoLock(conn);
881 /* conn should be at the head of the cleanup list */
882 if (conn == rx_connCleanup_list) {
883 rx_connCleanup_list = rx_connCleanup_list->next;
884 MUTEX_EXIT(&rx_connHashTable_lock);
885 rxi_CleanupConnection(conn);
887 #ifdef RX_ENABLE_LOCKS
889 MUTEX_EXIT(&rx_connHashTable_lock);
891 #endif /* RX_ENABLE_LOCKS */
895 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
897 register struct rx_connection **conn_ptr;
898 register int havecalls = 0;
899 struct rx_packet *packet;
906 MUTEX_ENTER(&conn->conn_data_lock);
907 if (conn->refCount > 0)
910 MUTEX_ENTER(&rx_stats_mutex);
911 rxi_lowConnRefCount++;
912 MUTEX_EXIT(&rx_stats_mutex);
915 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
916 /* Busy; wait till the last guy before proceeding */
917 MUTEX_EXIT(&conn->conn_data_lock);
922 /* If the client previously called rx_NewCall, but it is still
923 * waiting, treat this as a running call, and wait to destroy the
924 * connection later when the call completes. */
925 if ((conn->type == RX_CLIENT_CONNECTION)
926 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
927 conn->flags |= RX_CONN_DESTROY_ME;
928 MUTEX_EXIT(&conn->conn_data_lock);
932 MUTEX_EXIT(&conn->conn_data_lock);
934 /* Check for extant references to this connection */
935 for (i = 0; i < RX_MAXCALLS; i++) {
936 register struct rx_call *call = conn->call[i];
939 if (conn->type == RX_CLIENT_CONNECTION) {
940 MUTEX_ENTER(&call->lock);
941 if (call->delayedAckEvent) {
942 /* Push the final acknowledgment out now--there
943 * won't be a subsequent call to acknowledge the
944 * last reply packets */
945 rxevent_Cancel(call->delayedAckEvent, call,
946 RX_CALL_REFCOUNT_DELAY);
947 if (call->state == RX_STATE_PRECALL
948 || call->state == RX_STATE_ACTIVE) {
949 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
951 rxi_AckAll(NULL, call, 0);
954 MUTEX_EXIT(&call->lock);
958 #ifdef RX_ENABLE_LOCKS
960 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
961 MUTEX_EXIT(&conn->conn_data_lock);
963 /* Someone is accessing a packet right now. */
967 #endif /* RX_ENABLE_LOCKS */
970 /* Don't destroy the connection if there are any call
971 * structures still in use */
972 MUTEX_ENTER(&conn->conn_data_lock);
973 conn->flags |= RX_CONN_DESTROY_ME;
974 MUTEX_EXIT(&conn->conn_data_lock);
979 if (conn->delayedAbortEvent) {
980 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
981 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
983 MUTEX_ENTER(&conn->conn_data_lock);
984 rxi_SendConnectionAbort(conn, packet, 0, 1);
985 MUTEX_EXIT(&conn->conn_data_lock);
986 rxi_FreePacket(packet);
990 /* Remove from connection hash table before proceeding */
992 &rx_connHashTable[CONN_HASH
993 (peer->host, peer->port, conn->cid, conn->epoch,
995 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
996 if (*conn_ptr == conn) {
997 *conn_ptr = conn->next;
1001 /* if the conn that we are destroying was the last connection, then we
1002 * clear rxLastConn as well */
1003 if (rxLastConn == conn)
1006 /* Make sure the connection is completely reset before deleting it. */
1007 /* get rid of pending events that could zap us later */
1008 if (conn->challengeEvent)
1009 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1010 if (conn->checkReachEvent)
1011 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1013 /* Add the connection to the list of destroyed connections that
1014 * need to be cleaned up. This is necessary to avoid deadlocks
1015 * in the routines we call to inform others that this connection is
1016 * being destroyed. */
1017 conn->next = rx_connCleanup_list;
1018 rx_connCleanup_list = conn;
1021 /* Externally available version */
1023 rx_DestroyConnection(register struct rx_connection *conn)
1028 rxi_DestroyConnection(conn);
1033 rx_GetConnection(register struct rx_connection *conn)
1038 MUTEX_ENTER(&conn->conn_data_lock);
1040 MUTEX_EXIT(&conn->conn_data_lock);
1044 /* Wait for the transmit queue to no longer be busy.
1045 * requires the call->lock to be held */
1046 static void rxi_WaitforTQBusy(struct rx_call *call) {
1047 while (call->flags & RX_CALL_TQ_BUSY) {
1048 call->flags |= RX_CALL_TQ_WAIT;
1050 #ifdef RX_ENABLE_LOCKS
1051 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1052 CV_WAIT(&call->cv_tq, &call->lock);
1053 #else /* RX_ENABLE_LOCKS */
1054 osi_rxSleep(&call->tq);
1055 #endif /* RX_ENABLE_LOCKS */
1057 if (call->tqWaiters == 0) {
1058 call->flags &= ~RX_CALL_TQ_WAIT;
1062 /* Start a new rx remote procedure call, on the specified connection.
1063 * If wait is set to 1, wait for a free call channel; otherwise return
1064 * 0. Maxtime gives the maximum number of seconds this call may take,
1065 * after rx_NewCall returns. After this time interval, a call to any
1066 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1067 * For fine grain locking, we hold the conn_call_lock in order to
1068 * to ensure that we don't get signalle after we found a call in an active
1069 * state and before we go to sleep.
1072 rx_NewCall(register struct rx_connection *conn)
1075 register struct rx_call *call;
1076 struct clock queueTime;
1080 dpf(("rx_NewCall(conn %x)\n", conn));
1083 clock_GetTime(&queueTime);
1084 MUTEX_ENTER(&conn->conn_call_lock);
1087 * Check if there are others waiting for a new call.
1088 * If so, let them go first to avoid starving them.
1089 * This is a fairly simple scheme, and might not be
1090 * a complete solution for large numbers of waiters.
1092 * makeCallWaiters keeps track of the number of
1093 * threads waiting to make calls and the
1094 * RX_CONN_MAKECALL_WAITING flag bit is used to
1095 * indicate that there are indeed calls waiting.
1096 * The flag is set when the waiter is incremented.
1097 * It is only cleared in rx_EndCall when
1098 * makeCallWaiters is 0. This prevents us from
1099 * accidently destroying the connection while it
1100 * is potentially about to be used.
1102 MUTEX_ENTER(&conn->conn_data_lock);
1103 if (conn->makeCallWaiters) {
1104 conn->flags |= RX_CONN_MAKECALL_WAITING;
1105 conn->makeCallWaiters++;
1106 MUTEX_EXIT(&conn->conn_data_lock);
1108 #ifdef RX_ENABLE_LOCKS
1109 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1113 MUTEX_ENTER(&conn->conn_data_lock);
1114 conn->makeCallWaiters--;
1116 MUTEX_EXIT(&conn->conn_data_lock);
1119 for (i = 0; i < RX_MAXCALLS; i++) {
1120 call = conn->call[i];
1122 MUTEX_ENTER(&call->lock);
1123 if (call->state == RX_STATE_DALLY) {
1124 rxi_ResetCall(call, 0);
1125 (*call->callNumber)++;
1128 MUTEX_EXIT(&call->lock);
1130 call = rxi_NewCall(conn, i);
1134 if (i < RX_MAXCALLS) {
1137 MUTEX_ENTER(&conn->conn_data_lock);
1138 conn->flags |= RX_CONN_MAKECALL_WAITING;
1139 conn->makeCallWaiters++;
1140 MUTEX_EXIT(&conn->conn_data_lock);
1142 #ifdef RX_ENABLE_LOCKS
1143 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1147 MUTEX_ENTER(&conn->conn_data_lock);
1148 conn->makeCallWaiters--;
1149 MUTEX_EXIT(&conn->conn_data_lock);
1152 * Wake up anyone else who might be giving us a chance to
1153 * run (see code above that avoids resource starvation).
1155 #ifdef RX_ENABLE_LOCKS
1156 CV_BROADCAST(&conn->conn_call_cv);
1161 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1163 /* Client is initially in send mode */
1164 call->state = RX_STATE_ACTIVE;
1165 call->error = conn->error;
1167 call->mode = RX_MODE_ERROR;
1169 call->mode = RX_MODE_SENDING;
1171 /* remember start time for call in case we have hard dead time limit */
1172 call->queueTime = queueTime;
1173 clock_GetTime(&call->startTime);
1174 hzero(call->bytesSent);
1175 hzero(call->bytesRcvd);
1177 /* Turn on busy protocol. */
1178 rxi_KeepAliveOn(call);
1180 MUTEX_EXIT(&call->lock);
1181 MUTEX_EXIT(&conn->conn_call_lock);
1184 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1185 /* Now, if TQ wasn't cleared earlier, do it now. */
1186 MUTEX_ENTER(&call->lock);
1187 rxi_WaitforTQBusy(call);
1188 if (call->flags & RX_CALL_TQ_CLEARME) {
1189 rxi_ClearTransmitQueue(call, 0);
1190 queue_Init(&call->tq);
1192 MUTEX_EXIT(&call->lock);
1193 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1195 dpf(("rx_NewCall(call %x)\n", call));
1200 rxi_HasActiveCalls(register struct rx_connection *aconn)
1203 register struct rx_call *tcall;
1207 for (i = 0; i < RX_MAXCALLS; i++) {
1208 if ((tcall = aconn->call[i])) {
1209 if ((tcall->state == RX_STATE_ACTIVE)
1210 || (tcall->state == RX_STATE_PRECALL)) {
1221 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1222 register afs_int32 * aint32s)
1225 register struct rx_call *tcall;
1229 for (i = 0; i < RX_MAXCALLS; i++) {
1230 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1231 aint32s[i] = aconn->callNumber[i] + 1;
1233 aint32s[i] = aconn->callNumber[i];
1240 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1241 register afs_int32 * aint32s)
1244 register struct rx_call *tcall;
1248 for (i = 0; i < RX_MAXCALLS; i++) {
1249 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1250 aconn->callNumber[i] = aint32s[i] - 1;
1252 aconn->callNumber[i] = aint32s[i];
1258 /* Advertise a new service. A service is named locally by a UDP port
1259 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1262 char *serviceName; Name for identification purposes (e.g. the
1263 service name might be used for probing for
1266 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1267 char *serviceName, struct rx_securityClass **securityObjects,
1268 int nSecurityObjects,
1269 afs_int32(*serviceProc) (struct rx_call * acall))
1271 osi_socket socket = OSI_NULLSOCKET;
1272 register struct rx_service *tservice;
1278 if (serviceId == 0) {
1280 "rx_NewService: service id for service %s is not non-zero.\n",
1287 "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",
1295 tservice = rxi_AllocService();
1297 for (i = 0; i < RX_MAX_SERVICES; i++) {
1298 register struct rx_service *service = rx_services[i];
1300 if (port == service->servicePort && host == service->serviceHost) {
1301 if (service->serviceId == serviceId) {
1302 /* The identical service has already been
1303 * installed; if the caller was intending to
1304 * change the security classes used by this
1305 * service, he/she loses. */
1307 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1308 serviceName, serviceId, service->serviceName);
1310 rxi_FreeService(tservice);
1313 /* Different service, same port: re-use the socket
1314 * which is bound to the same port */
1315 socket = service->socket;
1318 if (socket == OSI_NULLSOCKET) {
1319 /* If we don't already have a socket (from another
1320 * service on same port) get a new one */
1321 socket = rxi_GetHostUDPSocket(host, port);
1322 if (socket == OSI_NULLSOCKET) {
1324 rxi_FreeService(tservice);
1329 service->socket = socket;
1330 service->serviceHost = host;
1331 service->servicePort = port;
1332 service->serviceId = serviceId;
1333 service->serviceName = serviceName;
1334 service->nSecurityObjects = nSecurityObjects;
1335 service->securityObjects = securityObjects;
1336 service->minProcs = 0;
1337 service->maxProcs = 1;
1338 service->idleDeadTime = 60;
1339 service->idleDeadErr = 0;
1340 service->connDeadTime = rx_connDeadTime;
1341 service->executeRequestProc = serviceProc;
1342 service->checkReach = 0;
1343 rx_services[i] = service; /* not visible until now */
1349 rxi_FreeService(tservice);
1350 (osi_Msg "rx_NewService: cannot support > %d services\n",
1355 /* Set configuration options for all of a service's security objects */
1358 rx_SetSecurityConfiguration(struct rx_service *service,
1359 rx_securityConfigVariables type,
1363 for (i = 0; i<service->nSecurityObjects; i++) {
1364 if (service->securityObjects[i]) {
1365 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1373 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1374 struct rx_securityClass **securityObjects, int nSecurityObjects,
1375 afs_int32(*serviceProc) (struct rx_call * acall))
1377 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1380 /* Generic request processing loop. This routine should be called
1381 * by the implementation dependent rx_ServerProc. If socketp is
1382 * non-null, it will be set to the file descriptor that this thread
1383 * is now listening on. If socketp is null, this routine will never
1386 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1388 register struct rx_call *call;
1389 register afs_int32 code;
1390 register struct rx_service *tservice = NULL;
1397 call = rx_GetCall(threadID, tservice, socketp);
1398 if (socketp && *socketp != OSI_NULLSOCKET) {
1399 /* We are now a listener thread */
1404 /* if server is restarting( typically smooth shutdown) then do not
1405 * allow any new calls.
1408 if (rx_tranquil && (call != NULL)) {
1412 MUTEX_ENTER(&call->lock);
1414 rxi_CallError(call, RX_RESTARTING);
1415 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1417 MUTEX_EXIT(&call->lock);
1421 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1422 #ifdef RX_ENABLE_LOCKS
1424 #endif /* RX_ENABLE_LOCKS */
1425 afs_termState = AFSOP_STOP_AFS;
1426 afs_osi_Wakeup(&afs_termState);
1427 #ifdef RX_ENABLE_LOCKS
1429 #endif /* RX_ENABLE_LOCKS */
1434 tservice = call->conn->service;
1436 if (tservice->beforeProc)
1437 (*tservice->beforeProc) (call);
1439 code = call->conn->service->executeRequestProc(call);
1441 if (tservice->afterProc)
1442 (*tservice->afterProc) (call, code);
1444 rx_EndCall(call, code);
1445 MUTEX_ENTER(&rx_stats_mutex);
1447 MUTEX_EXIT(&rx_stats_mutex);
1453 rx_WakeupServerProcs(void)
1455 struct rx_serverQueueEntry *np, *tqp;
1459 MUTEX_ENTER(&rx_serverPool_lock);
1461 #ifdef RX_ENABLE_LOCKS
1462 if (rx_waitForPacket)
1463 CV_BROADCAST(&rx_waitForPacket->cv);
1464 #else /* RX_ENABLE_LOCKS */
1465 if (rx_waitForPacket)
1466 osi_rxWakeup(rx_waitForPacket);
1467 #endif /* RX_ENABLE_LOCKS */
1468 MUTEX_ENTER(&freeSQEList_lock);
1469 for (np = rx_FreeSQEList; np; np = tqp) {
1470 tqp = *(struct rx_serverQueueEntry **)np;
1471 #ifdef RX_ENABLE_LOCKS
1472 CV_BROADCAST(&np->cv);
1473 #else /* RX_ENABLE_LOCKS */
1475 #endif /* RX_ENABLE_LOCKS */
1477 MUTEX_EXIT(&freeSQEList_lock);
1478 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1479 #ifdef RX_ENABLE_LOCKS
1480 CV_BROADCAST(&np->cv);
1481 #else /* RX_ENABLE_LOCKS */
1483 #endif /* RX_ENABLE_LOCKS */
1485 MUTEX_EXIT(&rx_serverPool_lock);
1490 * One thing that seems to happen is that all the server threads get
1491 * tied up on some empty or slow call, and then a whole bunch of calls
1492 * arrive at once, using up the packet pool, so now there are more
1493 * empty calls. The most critical resources here are server threads
1494 * and the free packet pool. The "doreclaim" code seems to help in
1495 * general. I think that eventually we arrive in this state: there
1496 * are lots of pending calls which do have all their packets present,
1497 * so they won't be reclaimed, are multi-packet calls, so they won't
1498 * be scheduled until later, and thus are tying up most of the free
1499 * packet pool for a very long time.
1501 * 1. schedule multi-packet calls if all the packets are present.
1502 * Probably CPU-bound operation, useful to return packets to pool.
1503 * Do what if there is a full window, but the last packet isn't here?
1504 * 3. preserve one thread which *only* runs "best" calls, otherwise
1505 * it sleeps and waits for that type of call.
1506 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1507 * the current dataquota business is badly broken. The quota isn't adjusted
1508 * to reflect how many packets are presently queued for a running call.
1509 * So, when we schedule a queued call with a full window of packets queued
1510 * up for it, that *should* free up a window full of packets for other 2d-class
1511 * calls to be able to use from the packet pool. But it doesn't.
1513 * NB. Most of the time, this code doesn't run -- since idle server threads
1514 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1515 * as a new call arrives.
1517 /* Sleep until a call arrives. Returns a pointer to the call, ready
1518 * for an rx_Read. */
1519 #ifdef RX_ENABLE_LOCKS
1521 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1523 struct rx_serverQueueEntry *sq;
1524 register struct rx_call *call = (struct rx_call *)0;
1525 struct rx_service *service = NULL;
1528 MUTEX_ENTER(&freeSQEList_lock);
1530 if ((sq = rx_FreeSQEList)) {
1531 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1532 MUTEX_EXIT(&freeSQEList_lock);
1533 } else { /* otherwise allocate a new one and return that */
1534 MUTEX_EXIT(&freeSQEList_lock);
1535 sq = (struct rx_serverQueueEntry *)
1536 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1537 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1538 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1541 MUTEX_ENTER(&rx_serverPool_lock);
1542 if (cur_service != NULL) {
1543 ReturnToServerPool(cur_service);
1546 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1547 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1549 /* Scan for eligible incoming calls. A call is not eligible
1550 * if the maximum number of calls for its service type are
1551 * already executing */
1552 /* One thread will process calls FCFS (to prevent starvation),
1553 * while the other threads may run ahead looking for calls which
1554 * have all their input data available immediately. This helps
1555 * keep threads from blocking, waiting for data from the client. */
1556 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1557 service = tcall->conn->service;
1558 if (!QuotaOK(service)) {
1561 if (tno == rxi_fcfs_thread_num
1562 || !tcall->queue_item_header.next) {
1563 /* If we're the fcfs thread , then we'll just use
1564 * this call. If we haven't been able to find an optimal
1565 * choice, and we're at the end of the list, then use a
1566 * 2d choice if one has been identified. Otherwise... */
1567 call = (choice2 ? choice2 : tcall);
1568 service = call->conn->service;
1569 } else if (!queue_IsEmpty(&tcall->rq)) {
1570 struct rx_packet *rp;
1571 rp = queue_First(&tcall->rq, rx_packet);
1572 if (rp->header.seq == 1) {
1574 || (rp->header.flags & RX_LAST_PACKET)) {
1576 } else if (rxi_2dchoice && !choice2
1577 && !(tcall->flags & RX_CALL_CLEARED)
1578 && (tcall->rprev > rxi_HardAckRate)) {
1587 ReturnToServerPool(service);
1594 MUTEX_EXIT(&rx_serverPool_lock);
1595 MUTEX_ENTER(&call->lock);
1597 if (call->flags & RX_CALL_WAIT_PROC) {
1598 call->flags &= ~RX_CALL_WAIT_PROC;
1599 MUTEX_ENTER(&rx_stats_mutex);
1601 MUTEX_EXIT(&rx_stats_mutex);
1604 if (call->state != RX_STATE_PRECALL || call->error) {
1605 MUTEX_EXIT(&call->lock);
1606 MUTEX_ENTER(&rx_serverPool_lock);
1607 ReturnToServerPool(service);
1612 if (queue_IsEmpty(&call->rq)
1613 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1614 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1616 CLEAR_CALL_QUEUE_LOCK(call);
1619 /* If there are no eligible incoming calls, add this process
1620 * to the idle server queue, to wait for one */
1624 *socketp = OSI_NULLSOCKET;
1626 sq->socketp = socketp;
1627 queue_Append(&rx_idleServerQueue, sq);
1628 #ifndef AFS_AIX41_ENV
1629 rx_waitForPacket = sq;
1631 rx_waitingForPacket = sq;
1632 #endif /* AFS_AIX41_ENV */
1634 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1636 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1637 MUTEX_EXIT(&rx_serverPool_lock);
1638 return (struct rx_call *)0;
1641 } while (!(call = sq->newcall)
1642 && !(socketp && *socketp != OSI_NULLSOCKET));
1643 MUTEX_EXIT(&rx_serverPool_lock);
1645 MUTEX_ENTER(&call->lock);
1651 MUTEX_ENTER(&freeSQEList_lock);
1652 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1653 rx_FreeSQEList = sq;
1654 MUTEX_EXIT(&freeSQEList_lock);
1657 clock_GetTime(&call->startTime);
1658 call->state = RX_STATE_ACTIVE;
1659 call->mode = RX_MODE_RECEIVING;
1660 #ifdef RX_KERNEL_TRACE
1661 if (ICL_SETACTIVE(afs_iclSetp)) {
1662 int glockOwner = ISAFS_GLOCK();
1665 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1666 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1673 rxi_calltrace(RX_CALL_START, call);
1674 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1675 call->conn->service->servicePort, call->conn->service->serviceId,
1678 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1679 MUTEX_EXIT(&call->lock);
1681 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1686 #else /* RX_ENABLE_LOCKS */
1688 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1690 struct rx_serverQueueEntry *sq;
1691 register struct rx_call *call = (struct rx_call *)0, *choice2;
1692 struct rx_service *service = NULL;
1696 MUTEX_ENTER(&freeSQEList_lock);
1698 if ((sq = rx_FreeSQEList)) {
1699 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1700 MUTEX_EXIT(&freeSQEList_lock);
1701 } else { /* otherwise allocate a new one and return that */
1702 MUTEX_EXIT(&freeSQEList_lock);
1703 sq = (struct rx_serverQueueEntry *)
1704 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1705 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1706 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1708 MUTEX_ENTER(&sq->lock);
1710 if (cur_service != NULL) {
1711 cur_service->nRequestsRunning--;
1712 if (cur_service->nRequestsRunning < cur_service->minProcs)
1716 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1717 register struct rx_call *tcall, *ncall;
1718 /* Scan for eligible incoming calls. A call is not eligible
1719 * if the maximum number of calls for its service type are
1720 * already executing */
1721 /* One thread will process calls FCFS (to prevent starvation),
1722 * while the other threads may run ahead looking for calls which
1723 * have all their input data available immediately. This helps
1724 * keep threads from blocking, waiting for data from the client. */
1725 choice2 = (struct rx_call *)0;
1726 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1727 service = tcall->conn->service;
1728 if (QuotaOK(service)) {
1729 if (tno == rxi_fcfs_thread_num
1730 || !tcall->queue_item_header.next) {
1731 /* If we're the fcfs thread, then we'll just use
1732 * this call. If we haven't been able to find an optimal
1733 * choice, and we're at the end of the list, then use a
1734 * 2d choice if one has been identified. Otherwise... */
1735 call = (choice2 ? choice2 : tcall);
1736 service = call->conn->service;
1737 } else if (!queue_IsEmpty(&tcall->rq)) {
1738 struct rx_packet *rp;
1739 rp = queue_First(&tcall->rq, rx_packet);
1740 if (rp->header.seq == 1
1742 || (rp->header.flags & RX_LAST_PACKET))) {
1744 } else if (rxi_2dchoice && !choice2
1745 && !(tcall->flags & RX_CALL_CLEARED)
1746 && (tcall->rprev > rxi_HardAckRate)) {
1759 /* we can't schedule a call if there's no data!!! */
1760 /* send an ack if there's no data, if we're missing the
1761 * first packet, or we're missing something between first
1762 * and last -- there's a "hole" in the incoming data. */
1763 if (queue_IsEmpty(&call->rq)
1764 || queue_First(&call->rq, rx_packet)->header.seq != 1
1765 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1766 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1768 call->flags &= (~RX_CALL_WAIT_PROC);
1769 service->nRequestsRunning++;
1770 /* just started call in minProcs pool, need fewer to maintain
1772 if (service->nRequestsRunning <= service->minProcs)
1776 /* MUTEX_EXIT(&call->lock); */
1778 /* If there are no eligible incoming calls, add this process
1779 * to the idle server queue, to wait for one */
1782 *socketp = OSI_NULLSOCKET;
1784 sq->socketp = socketp;
1785 queue_Append(&rx_idleServerQueue, sq);
1789 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1791 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1792 return (struct rx_call *)0;
1795 } while (!(call = sq->newcall)
1796 && !(socketp && *socketp != OSI_NULLSOCKET));
1798 MUTEX_EXIT(&sq->lock);
1800 MUTEX_ENTER(&freeSQEList_lock);
1801 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1802 rx_FreeSQEList = sq;
1803 MUTEX_EXIT(&freeSQEList_lock);
1806 clock_GetTime(&call->startTime);
1807 call->state = RX_STATE_ACTIVE;
1808 call->mode = RX_MODE_RECEIVING;
1809 #ifdef RX_KERNEL_TRACE
1810 if (ICL_SETACTIVE(afs_iclSetp)) {
1811 int glockOwner = ISAFS_GLOCK();
1814 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1815 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1822 rxi_calltrace(RX_CALL_START, call);
1823 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1824 call->conn->service->servicePort, call->conn->service->serviceId,
1827 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1834 #endif /* RX_ENABLE_LOCKS */
1838 /* Establish a procedure to be called when a packet arrives for a
1839 * call. This routine will be called at most once after each call,
1840 * and will also be called if there is an error condition on the or
1841 * the call is complete. Used by multi rx to build a selection
1842 * function which determines which of several calls is likely to be a
1843 * good one to read from.
1844 * NOTE: the way this is currently implemented it is probably only a
1845 * good idea to (1) use it immediately after a newcall (clients only)
1846 * and (2) only use it once. Other uses currently void your warranty
1849 rx_SetArrivalProc(register struct rx_call *call,
1850 register void (*proc) (register struct rx_call * call,
1852 register int index),
1853 register void * handle, register int arg)
1855 call->arrivalProc = proc;
1856 call->arrivalProcHandle = handle;
1857 call->arrivalProcArg = arg;
1860 /* Call is finished (possibly prematurely). Return rc to the peer, if
1861 * appropriate, and return the final error code from the conversation
1865 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1867 register struct rx_connection *conn = call->conn;
1868 register struct rx_service *service;
1874 dpf(("rx_EndCall(call %x rc %d error %d abortCode %d)\n", call, rc, call->error, call->abortCode));
1877 MUTEX_ENTER(&call->lock);
1879 if (rc == 0 && call->error == 0) {
1880 call->abortCode = 0;
1881 call->abortCount = 0;
1884 call->arrivalProc = (void (*)())0;
1885 if (rc && call->error == 0) {
1886 rxi_CallError(call, rc);
1887 /* Send an abort message to the peer if this error code has
1888 * only just been set. If it was set previously, assume the
1889 * peer has already been sent the error code or will request it
1891 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1893 if (conn->type == RX_SERVER_CONNECTION) {
1894 /* Make sure reply or at least dummy reply is sent */
1895 if (call->mode == RX_MODE_RECEIVING) {
1896 rxi_WriteProc(call, 0, 0);
1898 if (call->mode == RX_MODE_SENDING) {
1899 rxi_FlushWrite(call);
1901 service = conn->service;
1902 rxi_calltrace(RX_CALL_END, call);
1903 /* Call goes to hold state until reply packets are acknowledged */
1904 if (call->tfirst + call->nSoftAcked < call->tnext) {
1905 call->state = RX_STATE_HOLD;
1907 call->state = RX_STATE_DALLY;
1908 rxi_ClearTransmitQueue(call, 0);
1909 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1910 rxevent_Cancel(call->keepAliveEvent, call,
1911 RX_CALL_REFCOUNT_ALIVE);
1913 } else { /* Client connection */
1915 /* Make sure server receives input packets, in the case where
1916 * no reply arguments are expected */
1917 if ((call->mode == RX_MODE_SENDING)
1918 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1919 (void)rxi_ReadProc(call, &dummy, 1);
1922 /* If we had an outstanding delayed ack, be nice to the server
1923 * and force-send it now.
1925 if (call->delayedAckEvent) {
1926 rxevent_Cancel(call->delayedAckEvent, call,
1927 RX_CALL_REFCOUNT_DELAY);
1928 call->delayedAckEvent = NULL;
1929 rxi_SendDelayedAck(NULL, call, NULL);
1932 /* We need to release the call lock since it's lower than the
1933 * conn_call_lock and we don't want to hold the conn_call_lock
1934 * over the rx_ReadProc call. The conn_call_lock needs to be held
1935 * here for the case where rx_NewCall is perusing the calls on
1936 * the connection structure. We don't want to signal until
1937 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1938 * have checked this call, found it active and by the time it
1939 * goes to sleep, will have missed the signal.
1941 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
1942 * there are threads waiting to use the conn object.
1944 MUTEX_EXIT(&call->lock);
1945 MUTEX_ENTER(&conn->conn_call_lock);
1946 MUTEX_ENTER(&call->lock);
1947 MUTEX_ENTER(&conn->conn_data_lock);
1948 conn->flags |= RX_CONN_BUSY;
1949 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1950 if (conn->makeCallWaiters == 0)
1951 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1952 MUTEX_EXIT(&conn->conn_data_lock);
1953 #ifdef RX_ENABLE_LOCKS
1954 CV_BROADCAST(&conn->conn_call_cv);
1959 #ifdef RX_ENABLE_LOCKS
1961 MUTEX_EXIT(&conn->conn_data_lock);
1963 #endif /* RX_ENABLE_LOCKS */
1964 call->state = RX_STATE_DALLY;
1966 error = call->error;
1968 /* currentPacket, nLeft, and NFree must be zeroed here, because
1969 * ResetCall cannot: ResetCall may be called at splnet(), in the
1970 * kernel version, and may interrupt the macros rx_Read or
1971 * rx_Write, which run at normal priority for efficiency. */
1972 if (call->currentPacket) {
1973 queue_Prepend(&call->iovq, call->currentPacket);
1974 call->currentPacket = (struct rx_packet *)0;
1977 call->nLeft = call->nFree = call->curlen = 0;
1979 /* Free any packets from the last call to ReadvProc/WritevProc */
1980 rxi_FreePackets(0, &call->iovq);
1982 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1983 MUTEX_EXIT(&call->lock);
1984 if (conn->type == RX_CLIENT_CONNECTION) {
1985 MUTEX_EXIT(&conn->conn_call_lock);
1986 conn->flags &= ~RX_CONN_BUSY;
1990 * Map errors to the local host's errno.h format.
1992 error = ntoh_syserr_conv(error);
1996 #if !defined(KERNEL)
1998 /* Call this routine when shutting down a server or client (especially
1999 * clients). This will allow Rx to gracefully garbage collect server
2000 * connections, and reduce the number of retries that a server might
2001 * make to a dead client.
2002 * This is not quite right, since some calls may still be ongoing and
2003 * we can't lock them to destroy them. */
2007 register struct rx_connection **conn_ptr, **conn_end;
2011 if (rxinit_status == 1) {
2013 return; /* Already shutdown. */
2015 rxi_DeleteCachedConnections();
2016 if (rx_connHashTable) {
2017 MUTEX_ENTER(&rx_connHashTable_lock);
2018 for (conn_ptr = &rx_connHashTable[0], conn_end =
2019 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2021 struct rx_connection *conn, *next;
2022 for (conn = *conn_ptr; conn; conn = next) {
2024 if (conn->type == RX_CLIENT_CONNECTION) {
2025 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2027 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2028 #ifdef RX_ENABLE_LOCKS
2029 rxi_DestroyConnectionNoLock(conn);
2030 #else /* RX_ENABLE_LOCKS */
2031 rxi_DestroyConnection(conn);
2032 #endif /* RX_ENABLE_LOCKS */
2036 #ifdef RX_ENABLE_LOCKS
2037 while (rx_connCleanup_list) {
2038 struct rx_connection *conn;
2039 conn = rx_connCleanup_list;
2040 rx_connCleanup_list = rx_connCleanup_list->next;
2041 MUTEX_EXIT(&rx_connHashTable_lock);
2042 rxi_CleanupConnection(conn);
2043 MUTEX_ENTER(&rx_connHashTable_lock);
2045 MUTEX_EXIT(&rx_connHashTable_lock);
2046 #endif /* RX_ENABLE_LOCKS */
2051 afs_winsockCleanup();
2059 /* if we wakeup packet waiter too often, can get in loop with two
2060 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2062 rxi_PacketsUnWait(void)
2064 if (!rx_waitingForPackets) {
2068 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2069 return; /* still over quota */
2072 rx_waitingForPackets = 0;
2073 #ifdef RX_ENABLE_LOCKS
2074 CV_BROADCAST(&rx_waitingForPackets_cv);
2076 osi_rxWakeup(&rx_waitingForPackets);
2082 /* ------------------Internal interfaces------------------------- */
2084 /* Return this process's service structure for the
2085 * specified socket and service */
2087 rxi_FindService(register osi_socket socket, register u_short serviceId)
2089 register struct rx_service **sp;
2090 for (sp = &rx_services[0]; *sp; sp++) {
2091 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2097 /* Allocate a call structure, for the indicated channel of the
2098 * supplied connection. The mode and state of the call must be set by
2099 * the caller. Returns the call with mutex locked. */
2101 rxi_NewCall(register struct rx_connection *conn, register int channel)
2103 register struct rx_call *call;
2104 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2105 register struct rx_call *cp; /* Call pointer temp */
2106 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2107 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2109 dpf(("rxi_NewCall(conn %x, channel %d)\n", conn, channel));
2111 /* Grab an existing call structure, or allocate a new one.
2112 * Existing call structures are assumed to have been left reset by
2114 MUTEX_ENTER(&rx_freeCallQueue_lock);
2116 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2118 * EXCEPT that the TQ might not yet be cleared out.
2119 * Skip over those with in-use TQs.
2122 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2123 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2129 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2130 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2131 call = queue_First(&rx_freeCallQueue, rx_call);
2132 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2134 rx_MutexDecrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2135 MUTEX_EXIT(&rx_freeCallQueue_lock);
2136 MUTEX_ENTER(&call->lock);
2137 CLEAR_CALL_QUEUE_LOCK(call);
2138 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2139 /* Now, if TQ wasn't cleared earlier, do it now. */
2140 if (call->flags & RX_CALL_TQ_CLEARME) {
2141 rxi_ClearTransmitQueue(call, 0);
2142 queue_Init(&call->tq);
2144 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2145 /* Bind the call to its connection structure */
2147 rxi_ResetCall(call, 1);
2149 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2151 MUTEX_EXIT(&rx_freeCallQueue_lock);
2152 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2153 MUTEX_ENTER(&call->lock);
2154 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2155 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2156 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2158 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2159 /* Initialize once-only items */
2160 queue_Init(&call->tq);
2161 queue_Init(&call->rq);
2162 queue_Init(&call->iovq);
2163 /* Bind the call to its connection structure (prereq for reset) */
2165 rxi_ResetCall(call, 1);
2167 call->channel = channel;
2168 call->callNumber = &conn->callNumber[channel];
2169 call->rwind = conn->rwind[channel];
2170 call->twind = conn->twind[channel];
2171 /* Note that the next expected call number is retained (in
2172 * conn->callNumber[i]), even if we reallocate the call structure
2174 conn->call[channel] = call;
2175 /* if the channel's never been used (== 0), we should start at 1, otherwise
2176 * the call number is valid from the last time this channel was used */
2177 if (*call->callNumber == 0)
2178 *call->callNumber = 1;
2183 /* A call has been inactive long enough that so we can throw away
2184 * state, including the call structure, which is placed on the call
2186 * Call is locked upon entry.
2187 * haveCTLock set if called from rxi_ReapConnections
2189 #ifdef RX_ENABLE_LOCKS
2191 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2192 #else /* RX_ENABLE_LOCKS */
2194 rxi_FreeCall(register struct rx_call *call)
2195 #endif /* RX_ENABLE_LOCKS */
2197 register int channel = call->channel;
2198 register struct rx_connection *conn = call->conn;
2201 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2202 (*call->callNumber)++;
2203 rxi_ResetCall(call, 0);
2204 call->conn->call[channel] = (struct rx_call *)0;
2206 MUTEX_ENTER(&rx_freeCallQueue_lock);
2207 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2208 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2209 /* A call may be free even though its transmit queue is still in use.
2210 * Since we search the call list from head to tail, put busy calls at
2211 * the head of the list, and idle calls at the tail.
2213 if (call->flags & RX_CALL_TQ_BUSY)
2214 queue_Prepend(&rx_freeCallQueue, call);
2216 queue_Append(&rx_freeCallQueue, call);
2217 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2218 queue_Append(&rx_freeCallQueue, call);
2219 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2220 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2221 MUTEX_EXIT(&rx_freeCallQueue_lock);
2223 /* Destroy the connection if it was previously slated for
2224 * destruction, i.e. the Rx client code previously called
2225 * rx_DestroyConnection (client connections), or
2226 * rxi_ReapConnections called the same routine (server
2227 * connections). Only do this, however, if there are no
2228 * outstanding calls. Note that for fine grain locking, there appears
2229 * to be a deadlock in that rxi_FreeCall has a call locked and
2230 * DestroyConnectionNoLock locks each call in the conn. But note a
2231 * few lines up where we have removed this call from the conn.
2232 * If someone else destroys a connection, they either have no
2233 * call lock held or are going through this section of code.
2235 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2236 MUTEX_ENTER(&conn->conn_data_lock);
2238 MUTEX_EXIT(&conn->conn_data_lock);
2239 #ifdef RX_ENABLE_LOCKS
2241 rxi_DestroyConnectionNoLock(conn);
2243 rxi_DestroyConnection(conn);
2244 #else /* RX_ENABLE_LOCKS */
2245 rxi_DestroyConnection(conn);
2246 #endif /* RX_ENABLE_LOCKS */
2250 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2252 rxi_Alloc(register size_t size)
2256 rx_MutexAdd1Increment2(rxi_Allocsize, (afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2257 p = (char *)osi_Alloc(size);
2260 osi_Panic("rxi_Alloc error");
2266 rxi_Free(void *addr, register size_t size)
2268 rx_MutexAdd1Decrement2(rxi_Allocsize, -(afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2269 osi_Free(addr, size);
2273 rxi_SetPeerMtu(register afs_uint32 host, register afs_uint32 port, int mtu)
2275 struct rx_peer **peer_ptr, **peer_end;
2278 MUTEX_ENTER(&rx_peerHashTable_lock);
2280 for (peer_ptr = &rx_peerHashTable[0], peer_end =
2281 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
2283 struct rx_peer *peer, *next;
2284 for (peer = *peer_ptr; peer; peer = next) {
2286 if (host == peer->host) {
2287 MUTEX_ENTER(&peer->peer_lock);
2288 peer->ifMTU=MIN(mtu, peer->ifMTU);
2289 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2290 MUTEX_EXIT(&peer->peer_lock);
2295 struct rx_peer *peer, *next;
2296 hashIndex = PEER_HASH(host, port);
2297 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2298 if ((peer->host == host) && (peer->port == port)) {
2299 MUTEX_ENTER(&peer->peer_lock);
2300 peer->ifMTU=MIN(mtu, peer->ifMTU);
2301 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2302 MUTEX_EXIT(&peer->peer_lock);
2306 MUTEX_EXIT(&rx_peerHashTable_lock);
2309 /* Find the peer process represented by the supplied (host,port)
2310 * combination. If there is no appropriate active peer structure, a
2311 * new one will be allocated and initialized
2312 * The origPeer, if set, is a pointer to a peer structure on which the
2313 * refcount will be be decremented. This is used to replace the peer
2314 * structure hanging off a connection structure */
2316 rxi_FindPeer(register afs_uint32 host, register u_short port,
2317 struct rx_peer *origPeer, int create)
2319 register struct rx_peer *pp;
2321 hashIndex = PEER_HASH(host, port);
2322 MUTEX_ENTER(&rx_peerHashTable_lock);
2323 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2324 if ((pp->host == host) && (pp->port == port))
2329 pp = rxi_AllocPeer(); /* This bzero's *pp */
2330 pp->host = host; /* set here or in InitPeerParams is zero */
2332 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2333 queue_Init(&pp->congestionQueue);
2334 queue_Init(&pp->rpcStats);
2335 pp->next = rx_peerHashTable[hashIndex];
2336 rx_peerHashTable[hashIndex] = pp;
2337 rxi_InitPeerParams(pp);
2338 rx_MutexIncrement(rx_stats.nPeerStructs, rx_stats_mutex);
2345 origPeer->refCount--;
2346 MUTEX_EXIT(&rx_peerHashTable_lock);
2351 /* Find the connection at (host, port) started at epoch, and with the
2352 * given connection id. Creates the server connection if necessary.
2353 * The type specifies whether a client connection or a server
2354 * connection is desired. In both cases, (host, port) specify the
2355 * peer's (host, pair) pair. Client connections are not made
2356 * automatically by this routine. The parameter socket gives the
2357 * socket descriptor on which the packet was received. This is used,
2358 * in the case of server connections, to check that *new* connections
2359 * come via a valid (port, serviceId). Finally, the securityIndex
2360 * parameter must match the existing index for the connection. If a
2361 * server connection is created, it will be created using the supplied
2362 * index, if the index is valid for this service */
2363 struct rx_connection *
2364 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2365 register u_short port, u_short serviceId, afs_uint32 cid,
2366 afs_uint32 epoch, int type, u_int securityIndex)
2368 int hashindex, flag, i;
2369 register struct rx_connection *conn;
2370 hashindex = CONN_HASH(host, port, cid, epoch, type);
2371 MUTEX_ENTER(&rx_connHashTable_lock);
2372 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2373 rx_connHashTable[hashindex],
2376 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2377 && (epoch == conn->epoch)) {
2378 register struct rx_peer *pp = conn->peer;
2379 if (securityIndex != conn->securityIndex) {
2380 /* this isn't supposed to happen, but someone could forge a packet
2381 * like this, and there seems to be some CM bug that makes this
2382 * happen from time to time -- in which case, the fileserver
2384 MUTEX_EXIT(&rx_connHashTable_lock);
2385 return (struct rx_connection *)0;
2387 if (pp->host == host && pp->port == port)
2389 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2391 /* So what happens when it's a callback connection? */
2392 if ( /*type == RX_CLIENT_CONNECTION && */
2393 (conn->epoch & 0x80000000))
2397 /* the connection rxLastConn that was used the last time is not the
2398 ** one we are looking for now. Hence, start searching in the hash */
2400 conn = rx_connHashTable[hashindex];
2405 struct rx_service *service;
2406 if (type == RX_CLIENT_CONNECTION) {
2407 MUTEX_EXIT(&rx_connHashTable_lock);
2408 return (struct rx_connection *)0;
2410 service = rxi_FindService(socket, serviceId);
2411 if (!service || (securityIndex >= service->nSecurityObjects)
2412 || (service->securityObjects[securityIndex] == 0)) {
2413 MUTEX_EXIT(&rx_connHashTable_lock);
2414 return (struct rx_connection *)0;
2416 conn = rxi_AllocConnection(); /* This bzero's the connection */
2417 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2418 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2419 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2420 conn->next = rx_connHashTable[hashindex];
2421 rx_connHashTable[hashindex] = conn;
2422 conn->peer = rxi_FindPeer(host, port, 0, 1);
2423 conn->type = RX_SERVER_CONNECTION;
2424 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2425 conn->epoch = epoch;
2426 conn->cid = cid & RX_CIDMASK;
2427 /* conn->serial = conn->lastSerial = 0; */
2428 /* conn->timeout = 0; */
2429 conn->ackRate = RX_FAST_ACK_RATE;
2430 conn->service = service;
2431 conn->serviceId = serviceId;
2432 conn->securityIndex = securityIndex;
2433 conn->securityObject = service->securityObjects[securityIndex];
2434 conn->nSpecific = 0;
2435 conn->specific = NULL;
2436 rx_SetConnDeadTime(conn, service->connDeadTime);
2437 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2438 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2439 for (i = 0; i < RX_MAXCALLS; i++) {
2440 conn->twind[i] = rx_initSendWindow;
2441 conn->rwind[i] = rx_initReceiveWindow;
2443 /* Notify security object of the new connection */
2444 RXS_NewConnection(conn->securityObject, conn);
2445 /* XXXX Connection timeout? */
2446 if (service->newConnProc)
2447 (*service->newConnProc) (conn);
2448 rx_MutexIncrement(rx_stats.nServerConns, rx_stats_mutex);
2451 MUTEX_ENTER(&conn->conn_data_lock);
2453 MUTEX_EXIT(&conn->conn_data_lock);
2455 rxLastConn = conn; /* store this connection as the last conn used */
2456 MUTEX_EXIT(&rx_connHashTable_lock);
2460 /* There are two packet tracing routines available for testing and monitoring
2461 * Rx. One is called just after every packet is received and the other is
2462 * called just before every packet is sent. Received packets, have had their
2463 * headers decoded, and packets to be sent have not yet had their headers
2464 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2465 * containing the network address. Both can be modified. The return value, if
2466 * non-zero, indicates that the packet should be dropped. */
2468 int (*rx_justReceived) () = 0;
2469 int (*rx_almostSent) () = 0;
2471 /* A packet has been received off the interface. Np is the packet, socket is
2472 * the socket number it was received from (useful in determining which service
2473 * this packet corresponds to), and (host, port) reflect the host,port of the
2474 * sender. This call returns the packet to the caller if it is finished with
2475 * it, rather than de-allocating it, just as a small performance hack */
2478 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2479 afs_uint32 host, u_short port, int *tnop,
2480 struct rx_call **newcallp)
2482 register struct rx_call *call;
2483 register struct rx_connection *conn;
2485 afs_uint32 currentCallNumber;
2491 struct rx_packet *tnp;
2494 /* We don't print out the packet until now because (1) the time may not be
2495 * accurate enough until now in the lwp implementation (rx_Listener only gets
2496 * the time after the packet is read) and (2) from a protocol point of view,
2497 * this is the first time the packet has been seen */
2498 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2499 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2500 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2501 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2502 np->header.epoch, np->header.cid, np->header.callNumber,
2503 np->header.seq, np->header.flags, np));
2506 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2507 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2510 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2511 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2514 /* If an input tracer function is defined, call it with the packet and
2515 * network address. Note this function may modify its arguments. */
2516 if (rx_justReceived) {
2517 struct sockaddr_in addr;
2519 addr.sin_family = AF_INET;
2520 addr.sin_port = port;
2521 addr.sin_addr.s_addr = host;
2522 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2523 addr.sin_len = sizeof(addr);
2524 #endif /* AFS_OSF_ENV */
2525 drop = (*rx_justReceived) (np, &addr);
2526 /* drop packet if return value is non-zero */
2529 port = addr.sin_port; /* in case fcn changed addr */
2530 host = addr.sin_addr.s_addr;
2534 /* If packet was not sent by the client, then *we* must be the client */
2535 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2536 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2538 /* Find the connection (or fabricate one, if we're the server & if
2539 * necessary) associated with this packet */
2541 rxi_FindConnection(socket, host, port, np->header.serviceId,
2542 np->header.cid, np->header.epoch, type,
2543 np->header.securityIndex);
2546 /* If no connection found or fabricated, just ignore the packet.
2547 * (An argument could be made for sending an abort packet for
2552 MUTEX_ENTER(&conn->conn_data_lock);
2553 if (conn->maxSerial < np->header.serial)
2554 conn->maxSerial = np->header.serial;
2555 MUTEX_EXIT(&conn->conn_data_lock);
2557 /* If the connection is in an error state, send an abort packet and ignore
2558 * the incoming packet */
2560 /* Don't respond to an abort packet--we don't want loops! */
2561 MUTEX_ENTER(&conn->conn_data_lock);
2562 if (np->header.type != RX_PACKET_TYPE_ABORT)
2563 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2565 MUTEX_EXIT(&conn->conn_data_lock);
2569 /* Check for connection-only requests (i.e. not call specific). */
2570 if (np->header.callNumber == 0) {
2571 switch (np->header.type) {
2572 case RX_PACKET_TYPE_ABORT: {
2573 /* What if the supplied error is zero? */
2574 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2575 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2576 rxi_ConnectionError(conn, errcode);
2577 MUTEX_ENTER(&conn->conn_data_lock);
2579 MUTEX_EXIT(&conn->conn_data_lock);
2582 case RX_PACKET_TYPE_CHALLENGE:
2583 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2584 MUTEX_ENTER(&conn->conn_data_lock);
2586 MUTEX_EXIT(&conn->conn_data_lock);
2588 case RX_PACKET_TYPE_RESPONSE:
2589 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2590 MUTEX_ENTER(&conn->conn_data_lock);
2592 MUTEX_EXIT(&conn->conn_data_lock);
2594 case RX_PACKET_TYPE_PARAMS:
2595 case RX_PACKET_TYPE_PARAMS + 1:
2596 case RX_PACKET_TYPE_PARAMS + 2:
2597 /* ignore these packet types for now */
2598 MUTEX_ENTER(&conn->conn_data_lock);
2600 MUTEX_EXIT(&conn->conn_data_lock);
2605 /* Should not reach here, unless the peer is broken: send an
2607 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2608 MUTEX_ENTER(&conn->conn_data_lock);
2609 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2611 MUTEX_EXIT(&conn->conn_data_lock);
2616 channel = np->header.cid & RX_CHANNELMASK;
2617 call = conn->call[channel];
2618 #ifdef RX_ENABLE_LOCKS
2620 MUTEX_ENTER(&call->lock);
2621 /* Test to see if call struct is still attached to conn. */
2622 if (call != conn->call[channel]) {
2624 MUTEX_EXIT(&call->lock);
2625 if (type == RX_SERVER_CONNECTION) {
2626 call = conn->call[channel];
2627 /* If we started with no call attached and there is one now,
2628 * another thread is also running this routine and has gotten
2629 * the connection channel. We should drop this packet in the tests
2630 * below. If there was a call on this connection and it's now
2631 * gone, then we'll be making a new call below.
2632 * If there was previously a call and it's now different then
2633 * the old call was freed and another thread running this routine
2634 * has created a call on this channel. One of these two threads
2635 * has a packet for the old call and the code below handles those
2639 MUTEX_ENTER(&call->lock);
2641 /* This packet can't be for this call. If the new call address is
2642 * 0 then no call is running on this channel. If there is a call
2643 * then, since this is a client connection we're getting data for
2644 * it must be for the previous call.
2646 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2647 MUTEX_ENTER(&conn->conn_data_lock);
2649 MUTEX_EXIT(&conn->conn_data_lock);
2654 currentCallNumber = conn->callNumber[channel];
2656 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2657 if (np->header.callNumber < currentCallNumber) {
2658 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2659 #ifdef RX_ENABLE_LOCKS
2661 MUTEX_EXIT(&call->lock);
2663 MUTEX_ENTER(&conn->conn_data_lock);
2665 MUTEX_EXIT(&conn->conn_data_lock);
2669 MUTEX_ENTER(&conn->conn_call_lock);
2670 call = rxi_NewCall(conn, channel);
2671 MUTEX_EXIT(&conn->conn_call_lock);
2672 *call->callNumber = np->header.callNumber;
2673 if (np->header.callNumber == 0)
2674 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));
2676 call->state = RX_STATE_PRECALL;
2677 clock_GetTime(&call->queueTime);
2678 hzero(call->bytesSent);
2679 hzero(call->bytesRcvd);
2681 * If the number of queued calls exceeds the overload
2682 * threshold then abort this call.
2684 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2685 struct rx_packet *tp;
2687 rxi_CallError(call, rx_BusyError);
2688 tp = rxi_SendCallAbort(call, np, 1, 0);
2689 MUTEX_EXIT(&call->lock);
2690 MUTEX_ENTER(&conn->conn_data_lock);
2692 MUTEX_EXIT(&conn->conn_data_lock);
2693 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2696 rxi_KeepAliveOn(call);
2697 } else if (np->header.callNumber != currentCallNumber) {
2698 /* Wait until the transmit queue is idle before deciding
2699 * whether to reset the current call. Chances are that the
2700 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2703 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2704 while ((call->state == RX_STATE_ACTIVE)
2705 && (call->flags & RX_CALL_TQ_BUSY)) {
2706 call->flags |= RX_CALL_TQ_WAIT;
2708 #ifdef RX_ENABLE_LOCKS
2709 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2710 CV_WAIT(&call->cv_tq, &call->lock);
2711 #else /* RX_ENABLE_LOCKS */
2712 osi_rxSleep(&call->tq);
2713 #endif /* RX_ENABLE_LOCKS */
2715 if (call->tqWaiters == 0)
2716 call->flags &= ~RX_CALL_TQ_WAIT;
2718 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2719 /* If the new call cannot be taken right now send a busy and set
2720 * the error condition in this call, so that it terminates as
2721 * quickly as possible */
2722 if (call->state == RX_STATE_ACTIVE) {
2723 struct rx_packet *tp;
2725 rxi_CallError(call, RX_CALL_DEAD);
2726 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2728 MUTEX_EXIT(&call->lock);
2729 MUTEX_ENTER(&conn->conn_data_lock);
2731 MUTEX_EXIT(&conn->conn_data_lock);
2734 rxi_ResetCall(call, 0);
2735 *call->callNumber = np->header.callNumber;
2736 if (np->header.callNumber == 0)
2737 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));
2739 call->state = RX_STATE_PRECALL;
2740 clock_GetTime(&call->queueTime);
2741 hzero(call->bytesSent);
2742 hzero(call->bytesRcvd);
2744 * If the number of queued calls exceeds the overload
2745 * threshold then abort this call.
2747 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2748 struct rx_packet *tp;
2750 rxi_CallError(call, rx_BusyError);
2751 tp = rxi_SendCallAbort(call, np, 1, 0);
2752 MUTEX_EXIT(&call->lock);
2753 MUTEX_ENTER(&conn->conn_data_lock);
2755 MUTEX_EXIT(&conn->conn_data_lock);
2756 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2759 rxi_KeepAliveOn(call);
2761 /* Continuing call; do nothing here. */
2763 } else { /* we're the client */
2764 /* Ignore all incoming acknowledgements for calls in DALLY state */
2765 if (call && (call->state == RX_STATE_DALLY)
2766 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2767 rx_MutexIncrement(rx_stats.ignorePacketDally, rx_stats_mutex);
2768 #ifdef RX_ENABLE_LOCKS
2770 MUTEX_EXIT(&call->lock);
2773 MUTEX_ENTER(&conn->conn_data_lock);
2775 MUTEX_EXIT(&conn->conn_data_lock);
2779 /* Ignore anything that's not relevant to the current call. If there
2780 * isn't a current call, then no packet is relevant. */
2781 if (!call || (np->header.callNumber != currentCallNumber)) {
2782 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2783 #ifdef RX_ENABLE_LOCKS
2785 MUTEX_EXIT(&call->lock);
2788 MUTEX_ENTER(&conn->conn_data_lock);
2790 MUTEX_EXIT(&conn->conn_data_lock);
2793 /* If the service security object index stamped in the packet does not
2794 * match the connection's security index, ignore the packet */
2795 if (np->header.securityIndex != conn->securityIndex) {
2796 #ifdef RX_ENABLE_LOCKS
2797 MUTEX_EXIT(&call->lock);
2799 MUTEX_ENTER(&conn->conn_data_lock);
2801 MUTEX_EXIT(&conn->conn_data_lock);
2805 /* If we're receiving the response, then all transmit packets are
2806 * implicitly acknowledged. Get rid of them. */
2807 if (np->header.type == RX_PACKET_TYPE_DATA) {
2808 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2809 /* XXX Hack. Because we must release the global rx lock when
2810 * sending packets (osi_NetSend) we drop all acks while we're
2811 * traversing the tq in rxi_Start sending packets out because
2812 * packets may move to the freePacketQueue as result of being here!
2813 * So we drop these packets until we're safely out of the
2814 * traversing. Really ugly!
2815 * For fine grain RX locking, we set the acked field in the
2816 * packets and let rxi_Start remove them from the transmit queue.
2818 if (call->flags & RX_CALL_TQ_BUSY) {
2819 #ifdef RX_ENABLE_LOCKS
2820 rxi_SetAcksInTransmitQueue(call);
2823 return np; /* xmitting; drop packet */
2826 rxi_ClearTransmitQueue(call, 0);
2828 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2829 rxi_ClearTransmitQueue(call, 0);
2830 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2832 if (np->header.type == RX_PACKET_TYPE_ACK) {
2833 /* now check to see if this is an ack packet acknowledging that the
2834 * server actually *lost* some hard-acked data. If this happens we
2835 * ignore this packet, as it may indicate that the server restarted in
2836 * the middle of a call. It is also possible that this is an old ack
2837 * packet. We don't abort the connection in this case, because this
2838 * *might* just be an old ack packet. The right way to detect a server
2839 * restart in the midst of a call is to notice that the server epoch
2841 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2842 * XXX unacknowledged. I think that this is off-by-one, but
2843 * XXX I don't dare change it just yet, since it will
2844 * XXX interact badly with the server-restart detection
2845 * XXX code in receiveackpacket. */
2846 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2847 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2848 MUTEX_EXIT(&call->lock);
2849 MUTEX_ENTER(&conn->conn_data_lock);
2851 MUTEX_EXIT(&conn->conn_data_lock);
2855 } /* else not a data packet */
2858 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2859 /* Set remote user defined status from packet */
2860 call->remoteStatus = np->header.userStatus;
2862 /* Note the gap between the expected next packet and the actual
2863 * packet that arrived, when the new packet has a smaller serial number
2864 * than expected. Rioses frequently reorder packets all by themselves,
2865 * so this will be quite important with very large window sizes.
2866 * Skew is checked against 0 here to avoid any dependence on the type of
2867 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2869 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2870 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2871 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2873 MUTEX_ENTER(&conn->conn_data_lock);
2874 skew = conn->lastSerial - np->header.serial;
2875 conn->lastSerial = np->header.serial;
2876 MUTEX_EXIT(&conn->conn_data_lock);
2878 register struct rx_peer *peer;
2880 if (skew > peer->inPacketSkew) {
2881 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2883 peer->inPacketSkew = skew;
2887 /* Now do packet type-specific processing */
2888 switch (np->header.type) {
2889 case RX_PACKET_TYPE_DATA:
2890 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2893 case RX_PACKET_TYPE_ACK:
2894 /* Respond immediately to ack packets requesting acknowledgement
2896 if (np->header.flags & RX_REQUEST_ACK) {
2898 (void)rxi_SendCallAbort(call, 0, 1, 0);
2900 (void)rxi_SendAck(call, 0, np->header.serial,
2901 RX_ACK_PING_RESPONSE, 1);
2903 np = rxi_ReceiveAckPacket(call, np, 1);
2905 case RX_PACKET_TYPE_ABORT: {
2906 /* An abort packet: reset the call, passing the error up to the user. */
2907 /* What if error is zero? */
2908 /* What if the error is -1? the application will treat it as a timeout. */
2909 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
2910 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
2911 rxi_CallError(call, errdata);
2912 MUTEX_EXIT(&call->lock);
2913 MUTEX_ENTER(&conn->conn_data_lock);
2915 MUTEX_EXIT(&conn->conn_data_lock);
2916 return np; /* xmitting; drop packet */
2918 case RX_PACKET_TYPE_BUSY:
2921 case RX_PACKET_TYPE_ACKALL:
2922 /* All packets acknowledged, so we can drop all packets previously
2923 * readied for sending */
2924 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2925 /* XXX Hack. We because we can't release the global rx lock when
2926 * sending packets (osi_NetSend) we drop all ack pkts while we're
2927 * traversing the tq in rxi_Start sending packets out because
2928 * packets may move to the freePacketQueue as result of being
2929 * here! So we drop these packets until we're safely out of the
2930 * traversing. Really ugly!
2931 * For fine grain RX locking, we set the acked field in the packets
2932 * and let rxi_Start remove the packets from the transmit queue.
2934 if (call->flags & RX_CALL_TQ_BUSY) {
2935 #ifdef RX_ENABLE_LOCKS
2936 rxi_SetAcksInTransmitQueue(call);
2938 #else /* RX_ENABLE_LOCKS */
2939 MUTEX_EXIT(&call->lock);
2940 MUTEX_ENTER(&conn->conn_data_lock);
2942 MUTEX_EXIT(&conn->conn_data_lock);
2943 return np; /* xmitting; drop packet */
2944 #endif /* RX_ENABLE_LOCKS */
2946 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2947 rxi_ClearTransmitQueue(call, 0);
2948 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
2951 /* Should not reach here, unless the peer is broken: send an abort
2953 rxi_CallError(call, RX_PROTOCOL_ERROR);
2954 np = rxi_SendCallAbort(call, np, 1, 0);
2957 /* Note when this last legitimate packet was received, for keep-alive
2958 * processing. Note, we delay getting the time until now in the hope that
2959 * the packet will be delivered to the user before any get time is required
2960 * (if not, then the time won't actually be re-evaluated here). */
2961 call->lastReceiveTime = clock_Sec();
2962 MUTEX_EXIT(&call->lock);
2963 MUTEX_ENTER(&conn->conn_data_lock);
2965 MUTEX_EXIT(&conn->conn_data_lock);
2969 /* return true if this is an "interesting" connection from the point of view
2970 of someone trying to debug the system */
2972 rxi_IsConnInteresting(struct rx_connection *aconn)
2975 register struct rx_call *tcall;
2977 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2979 for (i = 0; i < RX_MAXCALLS; i++) {
2980 tcall = aconn->call[i];
2982 if ((tcall->state == RX_STATE_PRECALL)
2983 || (tcall->state == RX_STATE_ACTIVE))
2985 if ((tcall->mode == RX_MODE_SENDING)
2986 || (tcall->mode == RX_MODE_RECEIVING))
2994 /* if this is one of the last few packets AND it wouldn't be used by the
2995 receiving call to immediately satisfy a read request, then drop it on
2996 the floor, since accepting it might prevent a lock-holding thread from
2997 making progress in its reading. If a call has been cleared while in
2998 the precall state then ignore all subsequent packets until the call
2999 is assigned to a thread. */
3002 TooLow(struct rx_packet *ap, struct rx_call *acall)
3005 MUTEX_ENTER(&rx_stats_mutex);
3006 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3007 && (acall->state == RX_STATE_PRECALL))
3008 || ((rx_nFreePackets < rxi_dataQuota + 2)
3009 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3010 && (acall->flags & RX_CALL_READER_WAIT)))) {
3013 MUTEX_EXIT(&rx_stats_mutex);
3019 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
3020 struct rx_call *acall)
3022 struct rx_call *call = acall;
3023 struct clock when, now;
3026 MUTEX_ENTER(&conn->conn_data_lock);
3027 conn->checkReachEvent = NULL;
3028 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3031 MUTEX_EXIT(&conn->conn_data_lock);
3035 MUTEX_ENTER(&conn->conn_call_lock);
3036 MUTEX_ENTER(&conn->conn_data_lock);
3037 for (i = 0; i < RX_MAXCALLS; i++) {
3038 struct rx_call *tc = conn->call[i];
3039 if (tc && tc->state == RX_STATE_PRECALL) {
3045 /* Indicate that rxi_CheckReachEvent is no longer running by
3046 * clearing the flag. Must be atomic under conn_data_lock to
3047 * avoid a new call slipping by: rxi_CheckConnReach holds
3048 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3050 conn->flags &= ~RX_CONN_ATTACHWAIT;
3051 MUTEX_EXIT(&conn->conn_data_lock);
3052 MUTEX_EXIT(&conn->conn_call_lock);
3057 MUTEX_ENTER(&call->lock);
3058 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3060 MUTEX_EXIT(&call->lock);
3062 clock_GetTime(&now);
3064 when.sec += RX_CHECKREACH_TIMEOUT;
3065 MUTEX_ENTER(&conn->conn_data_lock);
3066 if (!conn->checkReachEvent) {
3068 conn->checkReachEvent =
3069 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3072 MUTEX_EXIT(&conn->conn_data_lock);
3078 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3080 struct rx_service *service = conn->service;
3081 struct rx_peer *peer = conn->peer;
3082 afs_uint32 now, lastReach;
3084 if (service->checkReach == 0)
3088 MUTEX_ENTER(&peer->peer_lock);
3089 lastReach = peer->lastReachTime;
3090 MUTEX_EXIT(&peer->peer_lock);
3091 if (now - lastReach < RX_CHECKREACH_TTL)
3094 MUTEX_ENTER(&conn->conn_data_lock);
3095 if (conn->flags & RX_CONN_ATTACHWAIT) {
3096 MUTEX_EXIT(&conn->conn_data_lock);
3099 conn->flags |= RX_CONN_ATTACHWAIT;
3100 MUTEX_EXIT(&conn->conn_data_lock);
3101 if (!conn->checkReachEvent)
3102 rxi_CheckReachEvent(NULL, conn, call);
3107 /* try to attach call, if authentication is complete */
3109 TryAttach(register struct rx_call *acall, register osi_socket socket,
3110 register int *tnop, register struct rx_call **newcallp,
3113 struct rx_connection *conn = acall->conn;
3115 if (conn->type == RX_SERVER_CONNECTION
3116 && acall->state == RX_STATE_PRECALL) {
3117 /* Don't attach until we have any req'd. authentication. */
3118 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3119 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3120 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3121 /* Note: this does not necessarily succeed; there
3122 * may not any proc available
3125 rxi_ChallengeOn(acall->conn);
3130 /* A data packet has been received off the interface. This packet is
3131 * appropriate to the call (the call is in the right state, etc.). This
3132 * routine can return a packet to the caller, for re-use */
3135 rxi_ReceiveDataPacket(register struct rx_call *call,
3136 register struct rx_packet *np, int istack,
3137 osi_socket socket, afs_uint32 host, u_short port,
3138 int *tnop, struct rx_call **newcallp)
3140 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3144 afs_uint32 seq, serial, flags;
3146 struct rx_packet *tnp;
3147 struct clock when, now;
3148 rx_MutexIncrement(rx_stats.dataPacketsRead, rx_stats_mutex);
3151 /* If there are no packet buffers, drop this new packet, unless we can find
3152 * packet buffers from inactive calls */
3154 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3155 MUTEX_ENTER(&rx_freePktQ_lock);
3156 rxi_NeedMorePackets = TRUE;
3157 MUTEX_EXIT(&rx_freePktQ_lock);
3158 rx_MutexIncrement(rx_stats.noPacketBuffersOnRead, rx_stats_mutex);
3159 call->rprev = np->header.serial;
3160 rxi_calltrace(RX_TRACE_DROP, call);
3161 dpf(("packet %x dropped on receipt - quota problems", np));
3163 rxi_ClearReceiveQueue(call);
3164 clock_GetTime(&now);
3166 clock_Add(&when, &rx_softAckDelay);
3167 if (!call->delayedAckEvent
3168 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3169 rxevent_Cancel(call->delayedAckEvent, call,
3170 RX_CALL_REFCOUNT_DELAY);
3171 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3172 call->delayedAckEvent =
3173 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3175 /* we've damaged this call already, might as well do it in. */
3181 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3182 * packet is one of several packets transmitted as a single
3183 * datagram. Do not send any soft or hard acks until all packets
3184 * in a jumbogram have been processed. Send negative acks right away.
3186 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3187 /* tnp is non-null when there are more packets in the
3188 * current jumbo gram */
3195 seq = np->header.seq;
3196 serial = np->header.serial;
3197 flags = np->header.flags;
3199 /* If the call is in an error state, send an abort message */
3201 return rxi_SendCallAbort(call, np, istack, 0);
3203 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3204 * AFS 3.5 jumbogram. */
3205 if (flags & RX_JUMBO_PACKET) {
3206 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3211 if (np->header.spare != 0) {
3212 MUTEX_ENTER(&call->conn->conn_data_lock);
3213 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3214 MUTEX_EXIT(&call->conn->conn_data_lock);
3217 /* The usual case is that this is the expected next packet */
3218 if (seq == call->rnext) {
3220 /* Check to make sure it is not a duplicate of one already queued */
3221 if (queue_IsNotEmpty(&call->rq)
3222 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3223 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3224 dpf(("packet %x dropped on receipt - duplicate", np));
3225 rxevent_Cancel(call->delayedAckEvent, call,
3226 RX_CALL_REFCOUNT_DELAY);
3227 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3233 /* It's the next packet. Stick it on the receive queue
3234 * for this call. Set newPackets to make sure we wake
3235 * the reader once all packets have been processed */
3236 queue_Prepend(&call->rq, np);
3238 np = NULL; /* We can't use this anymore */
3241 /* If an ack is requested then set a flag to make sure we
3242 * send an acknowledgement for this packet */
3243 if (flags & RX_REQUEST_ACK) {
3244 ackNeeded = RX_ACK_REQUESTED;
3247 /* Keep track of whether we have received the last packet */
3248 if (flags & RX_LAST_PACKET) {
3249 call->flags |= RX_CALL_HAVE_LAST;
3253 /* Check whether we have all of the packets for this call */
3254 if (call->flags & RX_CALL_HAVE_LAST) {
3255 afs_uint32 tseq; /* temporary sequence number */
3256 struct rx_packet *tp; /* Temporary packet pointer */
3257 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3259 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3260 if (tseq != tp->header.seq)
3262 if (tp->header.flags & RX_LAST_PACKET) {
3263 call->flags |= RX_CALL_RECEIVE_DONE;
3270 /* Provide asynchronous notification for those who want it
3271 * (e.g. multi rx) */
3272 if (call->arrivalProc) {
3273 (*call->arrivalProc) (call, call->arrivalProcHandle,
3274 call->arrivalProcArg);
3275 call->arrivalProc = (void (*)())0;
3278 /* Update last packet received */
3281 /* If there is no server process serving this call, grab
3282 * one, if available. We only need to do this once. If a
3283 * server thread is available, this thread becomes a server
3284 * thread and the server thread becomes a listener thread. */
3286 TryAttach(call, socket, tnop, newcallp, 0);
3289 /* This is not the expected next packet. */
3291 /* Determine whether this is a new or old packet, and if it's
3292 * a new one, whether it fits into the current receive window.
3293 * Also figure out whether the packet was delivered in sequence.
3294 * We use the prev variable to determine whether the new packet
3295 * is the successor of its immediate predecessor in the
3296 * receive queue, and the missing flag to determine whether
3297 * any of this packets predecessors are missing. */
3299 afs_uint32 prev; /* "Previous packet" sequence number */
3300 struct rx_packet *tp; /* Temporary packet pointer */
3301 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3302 int missing; /* Are any predecessors missing? */
3304 /* If the new packet's sequence number has been sent to the
3305 * application already, then this is a duplicate */
3306 if (seq < call->rnext) {
3307 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3308 rxevent_Cancel(call->delayedAckEvent, call,
3309 RX_CALL_REFCOUNT_DELAY);
3310 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3316 /* If the sequence number is greater than what can be
3317 * accomodated by the current window, then send a negative
3318 * acknowledge and drop the packet */
3319 if ((call->rnext + call->rwind) <= seq) {
3320 rxevent_Cancel(call->delayedAckEvent, call,
3321 RX_CALL_REFCOUNT_DELAY);
3322 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3329 /* Look for the packet in the queue of old received packets */
3330 for (prev = call->rnext - 1, missing =
3331 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3332 /*Check for duplicate packet */
3333 if (seq == tp->header.seq) {
3334 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3335 rxevent_Cancel(call->delayedAckEvent, call,
3336 RX_CALL_REFCOUNT_DELAY);
3337 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3343 /* If we find a higher sequence packet, break out and
3344 * insert the new packet here. */
3345 if (seq < tp->header.seq)
3347 /* Check for missing packet */
3348 if (tp->header.seq != prev + 1) {
3352 prev = tp->header.seq;
3355 /* Keep track of whether we have received the last packet. */
3356 if (flags & RX_LAST_PACKET) {
3357 call->flags |= RX_CALL_HAVE_LAST;
3360 /* It's within the window: add it to the the receive queue.
3361 * tp is left by the previous loop either pointing at the
3362 * packet before which to insert the new packet, or at the
3363 * queue head if the queue is empty or the packet should be
3365 queue_InsertBefore(tp, np);
3369 /* Check whether we have all of the packets for this call */
3370 if ((call->flags & RX_CALL_HAVE_LAST)
3371 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3372 afs_uint32 tseq; /* temporary sequence number */
3375 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3376 if (tseq != tp->header.seq)
3378 if (tp->header.flags & RX_LAST_PACKET) {
3379 call->flags |= RX_CALL_RECEIVE_DONE;
3386 /* We need to send an ack of the packet is out of sequence,
3387 * or if an ack was requested by the peer. */
3388 if (seq != prev + 1 || missing) {
3389 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3390 } else if (flags & RX_REQUEST_ACK) {
3391 ackNeeded = RX_ACK_REQUESTED;
3394 /* Acknowledge the last packet for each call */
3395 if (flags & RX_LAST_PACKET) {
3406 * If the receiver is waiting for an iovec, fill the iovec
3407 * using the data from the receive queue */
3408 if (call->flags & RX_CALL_IOVEC_WAIT) {
3409 didHardAck = rxi_FillReadVec(call, serial);
3410 /* the call may have been aborted */
3419 /* Wakeup the reader if any */
3420 if ((call->flags & RX_CALL_READER_WAIT)
3421 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3422 || (call->iovNext >= call->iovMax)
3423 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3424 call->flags &= ~RX_CALL_READER_WAIT;
3425 #ifdef RX_ENABLE_LOCKS
3426 CV_BROADCAST(&call->cv_rq);
3428 osi_rxWakeup(&call->rq);
3434 * Send an ack when requested by the peer, or once every
3435 * rxi_SoftAckRate packets until the last packet has been
3436 * received. Always send a soft ack for the last packet in
3437 * the server's reply. */
3439 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3440 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3441 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3442 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3443 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3444 } else if (call->nSoftAcks) {
3445 clock_GetTime(&now);
3447 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3448 clock_Add(&when, &rx_lastAckDelay);
3450 clock_Add(&when, &rx_softAckDelay);
3452 if (!call->delayedAckEvent
3453 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3454 rxevent_Cancel(call->delayedAckEvent, call,
3455 RX_CALL_REFCOUNT_DELAY);
3456 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3457 call->delayedAckEvent =
3458 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3460 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3461 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3468 static void rxi_ComputeRate();
3472 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3474 struct rx_peer *peer = conn->peer;
3476 MUTEX_ENTER(&peer->peer_lock);
3477 peer->lastReachTime = clock_Sec();
3478 MUTEX_EXIT(&peer->peer_lock);
3480 MUTEX_ENTER(&conn->conn_data_lock);
3481 if (conn->flags & RX_CONN_ATTACHWAIT) {
3484 conn->flags &= ~RX_CONN_ATTACHWAIT;
3485 MUTEX_EXIT(&conn->conn_data_lock);
3487 for (i = 0; i < RX_MAXCALLS; i++) {
3488 struct rx_call *call = conn->call[i];
3491 MUTEX_ENTER(&call->lock);
3492 /* tnop can be null if newcallp is null */
3493 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3495 MUTEX_EXIT(&call->lock);
3499 MUTEX_EXIT(&conn->conn_data_lock);
3503 rx_ack_reason(int reason)
3506 case RX_ACK_REQUESTED:
3508 case RX_ACK_DUPLICATE:
3510 case RX_ACK_OUT_OF_SEQUENCE:
3512 case RX_ACK_EXCEEDS_WINDOW:
3514 case RX_ACK_NOSPACE:
3518 case RX_ACK_PING_RESPONSE:
3530 /* rxi_ComputePeerNetStats
3532 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3533 * estimates (like RTT and throughput) based on ack packets. Caller
3534 * must ensure that the packet in question is the right one (i.e.
3535 * serial number matches).
3538 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3539 struct rx_ackPacket *ap, struct rx_packet *np)
3541 struct rx_peer *peer = call->conn->peer;
3543 /* Use RTT if not delayed by client. */
3544 if (ap->reason != RX_ACK_DELAY)
3545 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3547 rxi_ComputeRate(peer, call, p, np, ap->reason);
3551 /* The real smarts of the whole thing. */
3553 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3556 struct rx_ackPacket *ap;
3558 register struct rx_packet *tp;
3559 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3560 register struct rx_connection *conn = call->conn;
3561 struct rx_peer *peer = conn->peer;
3564 /* because there are CM's that are bogus, sending weird values for this. */
3565 afs_uint32 skew = 0;
3570 int newAckCount = 0;
3571 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3572 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3574 rx_MutexIncrement(rx_stats.ackPacketsRead, rx_stats_mutex);
3575 ap = (struct rx_ackPacket *)rx_DataOf(np);
3576 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3578 return np; /* truncated ack packet */
3580 /* depends on ack packet struct */
3581 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3582 first = ntohl(ap->firstPacket);
3583 serial = ntohl(ap->serial);
3584 /* temporarily disabled -- needs to degrade over time
3585 * skew = ntohs(ap->maxSkew); */
3587 /* Ignore ack packets received out of order */
3588 if (first < call->tfirst) {
3592 if (np->header.flags & RX_SLOW_START_OK) {
3593 call->flags |= RX_CALL_SLOW_START_OK;
3596 if (ap->reason == RX_ACK_PING_RESPONSE)
3597 rxi_UpdatePeerReach(conn, call);
3601 if (rxdebug_active) {
3605 len = _snprintf(msg, sizeof(msg),
3606 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3607 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3608 ntohl(ap->serial), ntohl(ap->previousPacket),
3609 (unsigned int)np->header.seq, (unsigned int)skew,
3610 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3614 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3615 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3619 OutputDebugString(msg);
3621 #else /* AFS_NT40_ENV */
3624 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3625 ap->reason, ntohl(ap->previousPacket),
3626 (unsigned int)np->header.seq, (unsigned int)serial,
3627 (unsigned int)skew, ntohl(ap->firstPacket));
3630 for (offset = 0; offset < nAcks; offset++)
3631 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3636 #endif /* AFS_NT40_ENV */
3639 /* Update the outgoing packet skew value to the latest value of
3640 * the peer's incoming packet skew value. The ack packet, of
3641 * course, could arrive out of order, but that won't affect things
3643 MUTEX_ENTER(&peer->peer_lock);
3644 peer->outPacketSkew = skew;
3646 /* Check for packets that no longer need to be transmitted, and
3647 * discard them. This only applies to packets positively
3648 * acknowledged as having been sent to the peer's upper level.
3649 * All other packets must be retained. So only packets with
3650 * sequence numbers < ap->firstPacket are candidates. */
3651 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3652 if (tp->header.seq >= first)
3654 call->tfirst = tp->header.seq + 1;
3656 && (tp->header.serial == serial || tp->firstSerial == serial))
3657 rxi_ComputePeerNetStats(call, tp, ap, np);
3658 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3661 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3662 /* XXX Hack. Because we have to release the global rx lock when sending
3663 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3664 * in rxi_Start sending packets out because packets may move to the
3665 * freePacketQueue as result of being here! So we drop these packets until
3666 * we're safely out of the traversing. Really ugly!
3667 * To make it even uglier, if we're using fine grain locking, we can
3668 * set the ack bits in the packets and have rxi_Start remove the packets
3669 * when it's done transmitting.
3671 if (call->flags & RX_CALL_TQ_BUSY) {
3672 #ifdef RX_ENABLE_LOCKS
3673 tp->flags |= RX_PKTFLAG_ACKED;
3674 call->flags |= RX_CALL_TQ_SOME_ACKED;
3675 #else /* RX_ENABLE_LOCKS */
3677 #endif /* RX_ENABLE_LOCKS */
3679 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3682 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3687 /* Give rate detector a chance to respond to ping requests */
3688 if (ap->reason == RX_ACK_PING_RESPONSE) {
3689 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3693 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3695 /* Now go through explicit acks/nacks and record the results in
3696 * the waiting packets. These are packets that can't be released
3697 * yet, even with a positive acknowledge. This positive
3698 * acknowledge only means the packet has been received by the
3699 * peer, not that it will be retained long enough to be sent to
3700 * the peer's upper level. In addition, reset the transmit timers
3701 * of any missing packets (those packets that must be missing
3702 * because this packet was out of sequence) */
3704 call->nSoftAcked = 0;
3705 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3706 /* Update round trip time if the ack was stimulated on receipt
3708 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3709 #ifdef RX_ENABLE_LOCKS
3710 if (tp->header.seq >= first)
3711 #endif /* RX_ENABLE_LOCKS */
3712 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3714 && (tp->header.serial == serial || tp->firstSerial == serial))
3715 rxi_ComputePeerNetStats(call, tp, ap, np);
3717 /* Set the acknowledge flag per packet based on the
3718 * information in the ack packet. An acknowlegded packet can
3719 * be downgraded when the server has discarded a packet it
3720 * soacked previously, or when an ack packet is received
3721 * out of sequence. */
3722 if (tp->header.seq < first) {
3723 /* Implicit ack information */
3724 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3727 tp->flags |= RX_PKTFLAG_ACKED;
3728 } else if (tp->header.seq < first + nAcks) {
3729 /* Explicit ack information: set it in the packet appropriately */
3730 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3731 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3733 tp->flags |= RX_PKTFLAG_ACKED;
3740 } else /* RX_ACK_TYPE_NACK */ {
3741 tp->flags &= ~RX_PKTFLAG_ACKED;
3745 tp->flags &= ~RX_PKTFLAG_ACKED;
3749 /* If packet isn't yet acked, and it has been transmitted at least
3750 * once, reset retransmit time using latest timeout
3751 * ie, this should readjust the retransmit timer for all outstanding
3752 * packets... So we don't just retransmit when we should know better*/
3754 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3755 tp->retryTime = tp->timeSent;
3756 clock_Add(&tp->retryTime, &peer->timeout);
3757 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3758 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3762 /* If the window has been extended by this acknowledge packet,
3763 * then wakeup a sender waiting in alloc for window space, or try
3764 * sending packets now, if he's been sitting on packets due to
3765 * lack of window space */
3766 if (call->tnext < (call->tfirst + call->twind)) {
3767 #ifdef RX_ENABLE_LOCKS
3768 CV_SIGNAL(&call->cv_twind);
3770 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3771 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3772 osi_rxWakeup(&call->twind);
3775 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3776 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3780 /* if the ack packet has a receivelen field hanging off it,
3781 * update our state */
3782 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3785 /* If the ack packet has a "recommended" size that is less than
3786 * what I am using now, reduce my size to match */
3787 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3788 (int)sizeof(afs_int32), &tSize);
3789 tSize = (afs_uint32) ntohl(tSize);
3790 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3792 /* Get the maximum packet size to send to this peer */
3793 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
3795 tSize = (afs_uint32) ntohl(tSize);
3796 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3797 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3799 /* sanity check - peer might have restarted with different params.
3800 * If peer says "send less", dammit, send less... Peer should never
3801 * be unable to accept packets of the size that prior AFS versions would
3802 * send without asking. */
3803 if (peer->maxMTU != tSize) {
3804 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
3806 peer->maxMTU = tSize;
3807 peer->MTU = MIN(tSize, peer->MTU);
3808 call->MTU = MIN(call->MTU, tSize);
3811 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3814 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3815 (int)sizeof(afs_int32), &tSize);
3816 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3817 if (tSize < call->twind) { /* smaller than our send */
3818 call->twind = tSize; /* window, we must send less... */
3819 call->ssthresh = MIN(call->twind, call->ssthresh);
3820 call->conn->twind[call->channel] = call->twind;
3823 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3824 * network MTU confused with the loopback MTU. Calculate the
3825 * maximum MTU here for use in the slow start code below.
3827 maxMTU = peer->maxMTU;
3828 /* Did peer restart with older RX version? */
3829 if (peer->maxDgramPackets > 1) {
3830 peer->maxDgramPackets = 1;
3832 } else if (np->length >=
3833 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3836 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3837 sizeof(afs_int32), &tSize);
3838 tSize = (afs_uint32) ntohl(tSize);
3840 * As of AFS 3.5 we set the send window to match the receive window.
3842 if (tSize < call->twind) {
3843 call->twind = tSize;
3844 call->conn->twind[call->channel] = call->twind;
3845 call->ssthresh = MIN(call->twind, call->ssthresh);
3846 } else if (tSize > call->twind) {
3847 call->twind = tSize;
3848 call->conn->twind[call->channel] = call->twind;
3852 * As of AFS 3.5, a jumbogram is more than one fixed size
3853 * packet transmitted in a single UDP datagram. If the remote
3854 * MTU is smaller than our local MTU then never send a datagram
3855 * larger than the natural MTU.
3858 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3859 sizeof(afs_int32), &tSize);
3860 maxDgramPackets = (afs_uint32) ntohl(tSize);
3861 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3863 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
3864 maxDgramPackets = MIN(maxDgramPackets, tSize);
3865 if (maxDgramPackets > 1) {
3866 peer->maxDgramPackets = maxDgramPackets;
3867 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3869 peer->maxDgramPackets = 1;
3870 call->MTU = peer->natMTU;
3872 } else if (peer->maxDgramPackets > 1) {
3873 /* Restarted with lower version of RX */
3874 peer->maxDgramPackets = 1;
3876 } else if (peer->maxDgramPackets > 1
3877 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3878 /* Restarted with lower version of RX */
3879 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3880 peer->natMTU = OLD_MAX_PACKET_SIZE;
3881 peer->MTU = OLD_MAX_PACKET_SIZE;
3882 peer->maxDgramPackets = 1;
3883 peer->nDgramPackets = 1;
3885 call->MTU = OLD_MAX_PACKET_SIZE;
3890 * Calculate how many datagrams were successfully received after
3891 * the first missing packet and adjust the negative ack counter
3896 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3897 if (call->nNacks < nNacked) {
3898 call->nNacks = nNacked;
3901 call->nAcks += newAckCount;
3905 if (call->flags & RX_CALL_FAST_RECOVER) {
3907 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3909 call->flags &= ~RX_CALL_FAST_RECOVER;
3910 call->cwind = call->nextCwind;
3911 call->nextCwind = 0;
3914 call->nCwindAcks = 0;
3915 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3916 /* Three negative acks in a row trigger congestion recovery */
3917 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3918 MUTEX_EXIT(&peer->peer_lock);
3919 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3920 /* someone else is waiting to start recovery */
3923 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3924 rxi_WaitforTQBusy(call);
3925 MUTEX_ENTER(&peer->peer_lock);
3926 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3927 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3928 call->flags |= RX_CALL_FAST_RECOVER;
3929 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3931 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3932 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3933 call->nextCwind = call->ssthresh;
3936 peer->MTU = call->MTU;
3937 peer->cwind = call->nextCwind;
3938 peer->nDgramPackets = call->nDgramPackets;
3940 call->congestSeq = peer->congestSeq;
3941 /* Reset the resend times on the packets that were nacked
3942 * so we will retransmit as soon as the window permits*/
3943 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3945 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3946 clock_Zero(&tp->retryTime);
3948 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3953 /* If cwind is smaller than ssthresh, then increase
3954 * the window one packet for each ack we receive (exponential
3956 * If cwind is greater than or equal to ssthresh then increase
3957 * the congestion window by one packet for each cwind acks we
3958 * receive (linear growth). */
3959 if (call->cwind < call->ssthresh) {
3961 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3962 call->nCwindAcks = 0;
3964 call->nCwindAcks += newAckCount;
3965 if (call->nCwindAcks >= call->cwind) {
3966 call->nCwindAcks = 0;
3967 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3971 * If we have received several acknowledgements in a row then
3972 * it is time to increase the size of our datagrams
3974 if ((int)call->nAcks > rx_nDgramThreshold) {
3975 if (peer->maxDgramPackets > 1) {
3976 if (call->nDgramPackets < peer->maxDgramPackets) {
3977 call->nDgramPackets++;
3979 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3980 } else if (call->MTU < peer->maxMTU) {
3981 call->MTU += peer->natMTU;
3982 call->MTU = MIN(call->MTU, peer->maxMTU);
3988 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3990 /* Servers need to hold the call until all response packets have
3991 * been acknowledged. Soft acks are good enough since clients
3992 * are not allowed to clear their receive queues. */
3993 if (call->state == RX_STATE_HOLD
3994 && call->tfirst + call->nSoftAcked >= call->tnext) {
3995 call->state = RX_STATE_DALLY;
3996 rxi_ClearTransmitQueue(call, 0);
3997 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
3998 } else if (!queue_IsEmpty(&call->tq)) {
3999 rxi_Start(0, call, 0, istack);
4004 /* Received a response to a challenge packet */
4006 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
4007 register struct rx_packet *np, int istack)
4011 /* Ignore the packet if we're the client */
4012 if (conn->type == RX_CLIENT_CONNECTION)
4015 /* If already authenticated, ignore the packet (it's probably a retry) */
4016 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4019 /* Otherwise, have the security object evaluate the response packet */
4020 error = RXS_CheckResponse(conn->securityObject, conn, np);
4022 /* If the response is invalid, reset the connection, sending
4023 * an abort to the peer */
4027 rxi_ConnectionError(conn, error);
4028 MUTEX_ENTER(&conn->conn_data_lock);
4029 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4030 MUTEX_EXIT(&conn->conn_data_lock);
4033 /* If the response is valid, any calls waiting to attach
4034 * servers can now do so */
4037 for (i = 0; i < RX_MAXCALLS; i++) {
4038 struct rx_call *call = conn->call[i];
4040 MUTEX_ENTER(&call->lock);
4041 if (call->state == RX_STATE_PRECALL)
4042 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4043 /* tnop can be null if newcallp is null */
4044 MUTEX_EXIT(&call->lock);
4048 /* Update the peer reachability information, just in case
4049 * some calls went into attach-wait while we were waiting
4050 * for authentication..
4052 rxi_UpdatePeerReach(conn, NULL);
4057 /* A client has received an authentication challenge: the security
4058 * object is asked to cough up a respectable response packet to send
4059 * back to the server. The server is responsible for retrying the
4060 * challenge if it fails to get a response. */
4063 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
4064 register struct rx_packet *np, int istack)
4068 /* Ignore the challenge if we're the server */
4069 if (conn->type == RX_SERVER_CONNECTION)
4072 /* Ignore the challenge if the connection is otherwise idle; someone's
4073 * trying to use us as an oracle. */
4074 if (!rxi_HasActiveCalls(conn))
4077 /* Send the security object the challenge packet. It is expected to fill
4078 * in the response. */
4079 error = RXS_GetResponse(conn->securityObject, conn, np);
4081 /* If the security object is unable to return a valid response, reset the
4082 * connection and send an abort to the peer. Otherwise send the response
4083 * packet to the peer connection. */
4085 rxi_ConnectionError(conn, error);
4086 MUTEX_ENTER(&conn->conn_data_lock);
4087 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4088 MUTEX_EXIT(&conn->conn_data_lock);
4090 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4091 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4097 /* Find an available server process to service the current request in
4098 * the given call structure. If one isn't available, queue up this
4099 * call so it eventually gets one */
4101 rxi_AttachServerProc(register struct rx_call *call,
4102 register osi_socket socket, register int *tnop,
4103 register struct rx_call **newcallp)
4105 register struct rx_serverQueueEntry *sq;
4106 register struct rx_service *service = call->conn->service;
4107 register int haveQuota = 0;
4109 /* May already be attached */
4110 if (call->state == RX_STATE_ACTIVE)
4113 MUTEX_ENTER(&rx_serverPool_lock);
4115 haveQuota = QuotaOK(service);
4116 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4117 /* If there are no processes available to service this call,
4118 * put the call on the incoming call queue (unless it's
4119 * already on the queue).
4121 #ifdef RX_ENABLE_LOCKS
4123 ReturnToServerPool(service);
4124 #endif /* RX_ENABLE_LOCKS */
4126 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4127 call->flags |= RX_CALL_WAIT_PROC;
4128 MUTEX_ENTER(&rx_stats_mutex);
4131 MUTEX_EXIT(&rx_stats_mutex);
4132 rxi_calltrace(RX_CALL_ARRIVAL, call);
4133 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4134 queue_Append(&rx_incomingCallQueue, call);
4137 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4139 /* If hot threads are enabled, and both newcallp and sq->socketp
4140 * are non-null, then this thread will process the call, and the
4141 * idle server thread will start listening on this threads socket.
4144 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4147 *sq->socketp = socket;
4148 clock_GetTime(&call->startTime);
4149 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4153 if (call->flags & RX_CALL_WAIT_PROC) {
4154 /* Conservative: I don't think this should happen */
4155 call->flags &= ~RX_CALL_WAIT_PROC;
4156 if (queue_IsOnQueue(call)) {
4158 MUTEX_ENTER(&rx_stats_mutex);
4160 MUTEX_EXIT(&rx_stats_mutex);
4163 call->state = RX_STATE_ACTIVE;
4164 call->mode = RX_MODE_RECEIVING;
4165 #ifdef RX_KERNEL_TRACE
4167 int glockOwner = ISAFS_GLOCK();
4170 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4171 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4177 if (call->flags & RX_CALL_CLEARED) {
4178 /* send an ack now to start the packet flow up again */
4179 call->flags &= ~RX_CALL_CLEARED;
4180 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4182 #ifdef RX_ENABLE_LOCKS
4185 service->nRequestsRunning++;
4186 if (service->nRequestsRunning <= service->minProcs)
4192 MUTEX_EXIT(&rx_serverPool_lock);
4195 /* Delay the sending of an acknowledge event for a short while, while
4196 * a new call is being prepared (in the case of a client) or a reply
4197 * is being prepared (in the case of a server). Rather than sending
4198 * an ack packet, an ACKALL packet is sent. */
4200 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4202 #ifdef RX_ENABLE_LOCKS
4204 MUTEX_ENTER(&call->lock);
4205 call->delayedAckEvent = NULL;
4206 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4208 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4209 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4211 MUTEX_EXIT(&call->lock);
4212 #else /* RX_ENABLE_LOCKS */
4214 call->delayedAckEvent = NULL;
4215 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4216 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4217 #endif /* RX_ENABLE_LOCKS */
4221 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4224 #ifdef RX_ENABLE_LOCKS
4226 MUTEX_ENTER(&call->lock);
4227 if (event == call->delayedAckEvent)
4228 call->delayedAckEvent = NULL;
4229 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4231 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4233 MUTEX_EXIT(&call->lock);
4234 #else /* RX_ENABLE_LOCKS */
4236 call->delayedAckEvent = NULL;
4237 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4238 #endif /* RX_ENABLE_LOCKS */
4242 #ifdef RX_ENABLE_LOCKS
4243 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4244 * clearing them out.
4247 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4249 register struct rx_packet *p, *tp;
4252 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4253 p->flags |= RX_PKTFLAG_ACKED;
4257 call->flags |= RX_CALL_TQ_CLEARME;
4258 call->flags |= RX_CALL_TQ_SOME_ACKED;
4261 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4262 call->tfirst = call->tnext;
4263 call->nSoftAcked = 0;
4265 if (call->flags & RX_CALL_FAST_RECOVER) {
4266 call->flags &= ~RX_CALL_FAST_RECOVER;
4267 call->cwind = call->nextCwind;
4268 call->nextCwind = 0;
4271 CV_SIGNAL(&call->cv_twind);
4273 #endif /* RX_ENABLE_LOCKS */
4275 /* Clear out the transmit queue for the current call (all packets have
4276 * been received by peer) */
4278 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4280 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4281 register struct rx_packet *p, *tp;
4283 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4285 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4286 p->flags |= RX_PKTFLAG_ACKED;
4290 call->flags |= RX_CALL_TQ_CLEARME;
4291 call->flags |= RX_CALL_TQ_SOME_ACKED;
4294 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4295 rxi_FreePackets(0, &call->tq);
4296 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4297 call->flags &= ~RX_CALL_TQ_CLEARME;
4299 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4301 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4302 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4303 call->nSoftAcked = 0;
4305 if (call->flags & RX_CALL_FAST_RECOVER) {
4306 call->flags &= ~RX_CALL_FAST_RECOVER;
4307 call->cwind = call->nextCwind;
4309 #ifdef RX_ENABLE_LOCKS
4310 CV_SIGNAL(&call->cv_twind);
4312 osi_rxWakeup(&call->twind);
4317 rxi_ClearReceiveQueue(register struct rx_call *call)
4319 if (queue_IsNotEmpty(&call->rq)) {
4320 rx_packetReclaims += rxi_FreePackets(0, &call->rq);
4321 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4323 if (call->state == RX_STATE_PRECALL) {
4324 call->flags |= RX_CALL_CLEARED;
4328 /* Send an abort packet for the specified call */
4330 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4331 int istack, int force)
4334 struct clock when, now;
4339 /* Clients should never delay abort messages */
4340 if (rx_IsClientConn(call->conn))
4343 if (call->abortCode != call->error) {
4344 call->abortCode = call->error;
4345 call->abortCount = 0;
4348 if (force || rxi_callAbortThreshhold == 0
4349 || call->abortCount < rxi_callAbortThreshhold) {
4350 if (call->delayedAbortEvent) {
4351 rxevent_Cancel(call->delayedAbortEvent, call,
4352 RX_CALL_REFCOUNT_ABORT);
4354 error = htonl(call->error);
4357 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4358 (char *)&error, sizeof(error), istack);
4359 } else if (!call->delayedAbortEvent) {
4360 clock_GetTime(&now);
4362 clock_Addmsec(&when, rxi_callAbortDelay);
4363 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4364 call->delayedAbortEvent =
4365 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4370 /* Send an abort packet for the specified connection. Packet is an
4371 * optional pointer to a packet that can be used to send the abort.
4372 * Once the number of abort messages reaches the threshhold, an
4373 * event is scheduled to send the abort. Setting the force flag
4374 * overrides sending delayed abort messages.
4376 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4377 * to send the abort packet.
4380 rxi_SendConnectionAbort(register struct rx_connection *conn,
4381 struct rx_packet *packet, int istack, int force)
4384 struct clock when, now;
4389 /* Clients should never delay abort messages */
4390 if (rx_IsClientConn(conn))
4393 if (force || rxi_connAbortThreshhold == 0
4394 || conn->abortCount < rxi_connAbortThreshhold) {
4395 if (conn->delayedAbortEvent) {
4396 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4398 error = htonl(conn->error);
4400 MUTEX_EXIT(&conn->conn_data_lock);
4402 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4403 RX_PACKET_TYPE_ABORT, (char *)&error,
4404 sizeof(error), istack);
4405 MUTEX_ENTER(&conn->conn_data_lock);
4406 } else if (!conn->delayedAbortEvent) {
4407 clock_GetTime(&now);
4409 clock_Addmsec(&when, rxi_connAbortDelay);
4410 conn->delayedAbortEvent =
4411 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4416 /* Associate an error all of the calls owned by a connection. Called
4417 * with error non-zero. This is only for really fatal things, like
4418 * bad authentication responses. The connection itself is set in
4419 * error at this point, so that future packets received will be
4422 rxi_ConnectionError(register struct rx_connection *conn,
4423 register afs_int32 error)
4428 dpf(("rxi_ConnectionError conn %x error %d", conn, error));
4430 MUTEX_ENTER(&conn->conn_data_lock);
4431 if (conn->challengeEvent)
4432 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4433 if (conn->checkReachEvent) {
4434 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4435 conn->checkReachEvent = 0;
4436 conn->flags &= ~RX_CONN_ATTACHWAIT;
4439 MUTEX_EXIT(&conn->conn_data_lock);
4440 for (i = 0; i < RX_MAXCALLS; i++) {
4441 struct rx_call *call = conn->call[i];
4443 MUTEX_ENTER(&call->lock);
4444 rxi_CallError(call, error);
4445 MUTEX_EXIT(&call->lock);
4448 conn->error = error;
4449 rx_MutexIncrement(rx_stats.fatalErrors, rx_stats_mutex);
4454 rxi_CallError(register struct rx_call *call, afs_int32 error)
4456 dpf(("rxi_CallError call %x error %d call->error %d", call, error, call->error));
4458 error = call->error;
4460 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4461 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4462 rxi_ResetCall(call, 0);
4465 rxi_ResetCall(call, 0);
4467 call->error = error;
4468 call->mode = RX_MODE_ERROR;
4471 /* Reset various fields in a call structure, and wakeup waiting
4472 * processes. Some fields aren't changed: state & mode are not
4473 * touched (these must be set by the caller), and bufptr, nLeft, and
4474 * nFree are not reset, since these fields are manipulated by
4475 * unprotected macros, and may only be reset by non-interrupting code.
4478 /* this code requires that call->conn be set properly as a pre-condition. */
4479 #endif /* ADAPT_WINDOW */
4482 rxi_ResetCall(register struct rx_call *call, register int newcall)
4485 register struct rx_peer *peer;
4486 struct rx_packet *packet;
4488 dpf(("rxi_ResetCall(call %x, newcall %d)\n", call, newcall));
4490 /* Notify anyone who is waiting for asynchronous packet arrival */
4491 if (call->arrivalProc) {
4492 (*call->arrivalProc) (call, call->arrivalProcHandle,
4493 call->arrivalProcArg);
4494 call->arrivalProc = (void (*)())0;
4497 if (call->delayedAbortEvent) {
4498 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4499 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4501 rxi_SendCallAbort(call, packet, 0, 1);
4502 rxi_FreePacket(packet);
4507 * Update the peer with the congestion information in this call
4508 * so other calls on this connection can pick up where this call
4509 * left off. If the congestion sequence numbers don't match then
4510 * another call experienced a retransmission.
4512 peer = call->conn->peer;
4513 MUTEX_ENTER(&peer->peer_lock);
4515 if (call->congestSeq == peer->congestSeq) {
4516 peer->cwind = MAX(peer->cwind, call->cwind);
4517 peer->MTU = MAX(peer->MTU, call->MTU);
4518 peer->nDgramPackets =
4519 MAX(peer->nDgramPackets, call->nDgramPackets);
4522 call->abortCode = 0;
4523 call->abortCount = 0;
4525 if (peer->maxDgramPackets > 1) {
4526 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4528 call->MTU = peer->MTU;
4530 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4531 call->ssthresh = rx_maxSendWindow;
4532 call->nDgramPackets = peer->nDgramPackets;
4533 call->congestSeq = peer->congestSeq;
4534 MUTEX_EXIT(&peer->peer_lock);
4536 flags = call->flags;
4537 rxi_ClearReceiveQueue(call);
4538 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4539 if (flags & RX_CALL_TQ_BUSY) {
4540 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4541 call->flags |= (flags & RX_CALL_TQ_WAIT);
4543 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4545 rxi_ClearTransmitQueue(call, 0);
4546 queue_Init(&call->tq);
4547 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4548 dpf(("rcall %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4551 while (call->tqWaiters) {
4552 #ifdef RX_ENABLE_LOCKS
4553 CV_BROADCAST(&call->cv_tq);
4554 #else /* RX_ENABLE_LOCKS */
4555 osi_rxWakeup(&call->tq);
4556 #endif /* RX_ENABLE_LOCKS */
4560 queue_Init(&call->rq);
4562 call->twind = call->conn->twind[call->channel];
4563 call->rwind = call->conn->rwind[call->channel];
4564 call->nSoftAcked = 0;
4565 call->nextCwind = 0;
4568 call->nCwindAcks = 0;
4569 call->nSoftAcks = 0;
4570 call->nHardAcks = 0;
4572 call->tfirst = call->rnext = call->tnext = 1;
4574 call->lastAcked = 0;
4575 call->localStatus = call->remoteStatus = 0;
4577 if (flags & RX_CALL_READER_WAIT) {
4578 #ifdef RX_ENABLE_LOCKS
4579 CV_BROADCAST(&call->cv_rq);
4581 osi_rxWakeup(&call->rq);
4584 if (flags & RX_CALL_WAIT_PACKETS) {
4585 MUTEX_ENTER(&rx_freePktQ_lock);
4586 rxi_PacketsUnWait(); /* XXX */
4587 MUTEX_EXIT(&rx_freePktQ_lock);
4589 #ifdef RX_ENABLE_LOCKS
4590 CV_SIGNAL(&call->cv_twind);
4592 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4593 osi_rxWakeup(&call->twind);
4596 #ifdef RX_ENABLE_LOCKS
4597 /* The following ensures that we don't mess with any queue while some
4598 * other thread might also be doing so. The call_queue_lock field is
4599 * is only modified under the call lock. If the call is in the process
4600 * of being removed from a queue, the call is not locked until the
4601 * the queue lock is dropped and only then is the call_queue_lock field
4602 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4603 * Note that any other routine which removes a call from a queue has to
4604 * obtain the queue lock before examing the queue and removing the call.
4606 if (call->call_queue_lock) {
4607 MUTEX_ENTER(call->call_queue_lock);
4608 if (queue_IsOnQueue(call)) {
4610 if (flags & RX_CALL_WAIT_PROC) {
4611 MUTEX_ENTER(&rx_stats_mutex);
4613 MUTEX_EXIT(&rx_stats_mutex);
4616 MUTEX_EXIT(call->call_queue_lock);
4617 CLEAR_CALL_QUEUE_LOCK(call);
4619 #else /* RX_ENABLE_LOCKS */
4620 if (queue_IsOnQueue(call)) {
4622 if (flags & RX_CALL_WAIT_PROC)
4625 #endif /* RX_ENABLE_LOCKS */
4627 rxi_KeepAliveOff(call);
4628 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4631 /* Send an acknowledge for the indicated packet (seq,serial) of the
4632 * indicated call, for the indicated reason (reason). This
4633 * acknowledge will specifically acknowledge receiving the packet, and
4634 * will also specify which other packets for this call have been
4635 * received. This routine returns the packet that was used to the
4636 * caller. The caller is responsible for freeing it or re-using it.
4637 * This acknowledgement also returns the highest sequence number
4638 * actually read out by the higher level to the sender; the sender
4639 * promises to keep around packets that have not been read by the
4640 * higher level yet (unless, of course, the sender decides to abort
4641 * the call altogether). Any of p, seq, serial, pflags, or reason may
4642 * be set to zero without ill effect. That is, if they are zero, they
4643 * will not convey any information.
4644 * NOW there is a trailer field, after the ack where it will safely be
4645 * ignored by mundanes, which indicates the maximum size packet this
4646 * host can swallow. */
4648 register struct rx_packet *optionalPacket; use to send ack (or null)
4649 int seq; Sequence number of the packet we are acking
4650 int serial; Serial number of the packet
4651 int pflags; Flags field from packet header
4652 int reason; Reason an acknowledge was prompted
4656 rxi_SendAck(register struct rx_call *call,
4657 register struct rx_packet *optionalPacket, int serial, int reason,
4660 struct rx_ackPacket *ap;
4661 register struct rx_packet *rqp;
4662 register struct rx_packet *nxp; /* For queue_Scan */
4663 register struct rx_packet *p;
4666 #ifdef RX_ENABLE_TSFPQ
4667 struct rx_ts_info_t * rx_ts_info;
4671 * Open the receive window once a thread starts reading packets
4673 if (call->rnext > 1) {
4674 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
4677 call->nHardAcks = 0;
4678 call->nSoftAcks = 0;
4679 if (call->rnext > call->lastAcked)
4680 call->lastAcked = call->rnext;
4684 rx_computelen(p, p->length); /* reset length, you never know */
4685 } /* where that's been... */
4686 #ifdef RX_ENABLE_TSFPQ
4688 RX_TS_INFO_GET(rx_ts_info);
4689 if ((p = rx_ts_info->local_special_packet)) {
4690 rx_computelen(p, p->length);
4691 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4692 rx_ts_info->local_special_packet = p;
4693 } else { /* We won't send the ack, but don't panic. */
4694 return optionalPacket;
4698 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4699 /* We won't send the ack, but don't panic. */
4700 return optionalPacket;
4705 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4708 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4709 #ifndef RX_ENABLE_TSFPQ
4710 if (!optionalPacket)
4713 return optionalPacket;
4715 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4716 if (rx_Contiguous(p) < templ) {
4717 #ifndef RX_ENABLE_TSFPQ
4718 if (!optionalPacket)
4721 return optionalPacket;
4726 /* MTUXXX failing to send an ack is very serious. We should */
4727 /* try as hard as possible to send even a partial ack; it's */
4728 /* better than nothing. */
4729 ap = (struct rx_ackPacket *)rx_DataOf(p);
4730 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4731 ap->reason = reason;
4733 /* The skew computation used to be bogus, I think it's better now. */
4734 /* We should start paying attention to skew. XXX */
4735 ap->serial = htonl(serial);
4736 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4738 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4739 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4741 /* No fear of running out of ack packet here because there can only be at most
4742 * one window full of unacknowledged packets. The window size must be constrained
4743 * to be less than the maximum ack size, of course. Also, an ack should always
4744 * fit into a single packet -- it should not ever be fragmented. */
4745 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4746 if (!rqp || !call->rq.next
4747 || (rqp->header.seq > (call->rnext + call->rwind))) {
4748 #ifndef RX_ENABLE_TSFPQ
4749 if (!optionalPacket)
4752 rxi_CallError(call, RX_CALL_DEAD);
4753 return optionalPacket;
4756 while (rqp->header.seq > call->rnext + offset)
4757 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4758 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4760 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4761 #ifndef RX_ENABLE_TSFPQ
4762 if (!optionalPacket)
4765 rxi_CallError(call, RX_CALL_DEAD);
4766 return optionalPacket;
4771 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4773 /* these are new for AFS 3.3 */
4774 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4775 templ = htonl(templ);
4776 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4777 templ = htonl(call->conn->peer->ifMTU);
4778 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4779 sizeof(afs_int32), &templ);
4781 /* new for AFS 3.4 */
4782 templ = htonl(call->rwind);
4783 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4784 sizeof(afs_int32), &templ);
4786 /* new for AFS 3.5 */
4787 templ = htonl(call->conn->peer->ifDgramPackets);
4788 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4789 sizeof(afs_int32), &templ);
4791 p->header.serviceId = call->conn->serviceId;
4792 p->header.cid = (call->conn->cid | call->channel);
4793 p->header.callNumber = *call->callNumber;
4795 p->header.securityIndex = call->conn->securityIndex;
4796 p->header.epoch = call->conn->epoch;
4797 p->header.type = RX_PACKET_TYPE_ACK;
4798 p->header.flags = RX_SLOW_START_OK;
4799 if (reason == RX_ACK_PING) {
4800 p->header.flags |= RX_REQUEST_ACK;
4802 clock_GetTime(&call->pingRequestTime);
4805 if (call->conn->type == RX_CLIENT_CONNECTION)
4806 p->header.flags |= RX_CLIENT_INITIATED;
4810 if (rxdebug_active) {
4814 len = _snprintf(msg, sizeof(msg),
4815 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4816 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4817 ntohl(ap->serial), ntohl(ap->previousPacket),
4818 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
4819 ap->nAcks, ntohs(ap->bufferSpace) );
4823 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
4824 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4828 OutputDebugString(msg);
4830 #else /* AFS_NT40_ENV */
4832 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
4833 ap->reason, ntohl(ap->previousPacket),
4834 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4836 for (offset = 0; offset < ap->nAcks; offset++)
4837 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4842 #endif /* AFS_NT40_ENV */
4845 register int i, nbytes = p->length;
4847 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4848 if (nbytes <= p->wirevec[i].iov_len) {
4849 register int savelen, saven;
4851 savelen = p->wirevec[i].iov_len;
4853 p->wirevec[i].iov_len = nbytes;
4855 rxi_Send(call, p, istack);
4856 p->wirevec[i].iov_len = savelen;
4860 nbytes -= p->wirevec[i].iov_len;
4863 rx_MutexIncrement(rx_stats.ackPacketsSent, rx_stats_mutex);
4864 #ifndef RX_ENABLE_TSFPQ
4865 if (!optionalPacket)
4868 return optionalPacket; /* Return packet for re-use by caller */
4871 /* Send all of the packets in the list in single datagram */
4873 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4874 int istack, int moreFlag, struct clock *now,
4875 struct clock *retryTime, int resending)
4880 struct rx_connection *conn = call->conn;
4881 struct rx_peer *peer = conn->peer;
4883 MUTEX_ENTER(&peer->peer_lock);
4886 peer->reSends += len;
4887 rx_MutexIncrement(rx_stats.dataPacketsSent, rx_stats_mutex);
4888 MUTEX_EXIT(&peer->peer_lock);
4890 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4894 /* Set the packet flags and schedule the resend events */
4895 /* Only request an ack for the last packet in the list */
4896 for (i = 0; i < len; i++) {
4897 list[i]->retryTime = *retryTime;
4898 if (list[i]->header.serial) {
4899 /* Exponentially backoff retry times */
4900 if (list[i]->backoff < MAXBACKOFF) {
4901 /* so it can't stay == 0 */
4902 list[i]->backoff = (list[i]->backoff << 1) + 1;
4905 clock_Addmsec(&(list[i]->retryTime),
4906 ((afs_uint32) list[i]->backoff) << 8);
4909 /* Wait a little extra for the ack on the last packet */
4910 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4911 clock_Addmsec(&(list[i]->retryTime), 400);
4914 /* Record the time sent */
4915 list[i]->timeSent = *now;
4917 /* Ask for an ack on retransmitted packets, on every other packet
4918 * if the peer doesn't support slow start. Ask for an ack on every
4919 * packet until the congestion window reaches the ack rate. */
4920 if (list[i]->header.serial) {
4922 rx_MutexIncrement(rx_stats.dataPacketsReSent, rx_stats_mutex);
4924 /* improved RTO calculation- not Karn */
4925 list[i]->firstSent = *now;
4926 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4927 || (!(call->flags & RX_CALL_SLOW_START_OK)
4928 && (list[i]->header.seq & 1)))) {
4933 MUTEX_ENTER(&peer->peer_lock);
4937 rx_MutexIncrement(rx_stats.dataPacketsSent, rx_stats_mutex);
4938 MUTEX_EXIT(&peer->peer_lock);
4940 /* Tag this packet as not being the last in this group,
4941 * for the receiver's benefit */
4942 if (i < len - 1 || moreFlag) {
4943 list[i]->header.flags |= RX_MORE_PACKETS;
4946 /* Install the new retransmit time for the packet, and
4947 * record the time sent */
4948 list[i]->timeSent = *now;
4952 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4955 /* Since we're about to send a data packet to the peer, it's
4956 * safe to nuke any scheduled end-of-packets ack */
4957 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4959 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4960 MUTEX_EXIT(&call->lock);
4962 rxi_SendPacketList(call, conn, list, len, istack);
4964 rxi_SendPacket(call, conn, list[0], istack);
4966 MUTEX_ENTER(&call->lock);
4967 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4969 /* Update last send time for this call (for keep-alive
4970 * processing), and for the connection (so that we can discover
4971 * idle connections) */
4972 call->lastSendData = conn->lastSendTime = call->lastSendTime = clock_Sec();
4975 /* When sending packets we need to follow these rules:
4976 * 1. Never send more than maxDgramPackets in a jumbogram.
4977 * 2. Never send a packet with more than two iovecs in a jumbogram.
4978 * 3. Never send a retransmitted packet in a jumbogram.
4979 * 4. Never send more than cwind/4 packets in a jumbogram
4980 * We always keep the last list we should have sent so we
4981 * can set the RX_MORE_PACKETS flags correctly.
4984 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4985 int istack, struct clock *now, struct clock *retryTime,
4988 int i, cnt, lastCnt = 0;
4989 struct rx_packet **listP, **lastP = 0;
4990 struct rx_peer *peer = call->conn->peer;
4991 int morePackets = 0;
4993 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4994 /* Does the current packet force us to flush the current list? */
4996 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4997 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4999 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
5001 /* If the call enters an error state stop sending, or if
5002 * we entered congestion recovery mode, stop sending */
5003 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5011 /* Add the current packet to the list if it hasn't been acked.
5012 * Otherwise adjust the list pointer to skip the current packet. */
5013 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5015 /* Do we need to flush the list? */
5016 if (cnt >= (int)peer->maxDgramPackets
5017 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5018 || list[i]->header.serial
5019 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5021 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5022 retryTime, resending);
5023 /* If the call enters an error state stop sending, or if
5024 * we entered congestion recovery mode, stop sending */
5026 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5031 listP = &list[i + 1];
5036 osi_Panic("rxi_SendList error");
5038 listP = &list[i + 1];
5042 /* Send the whole list when the call is in receive mode, when
5043 * the call is in eof mode, when we are in fast recovery mode,
5044 * and when we have the last packet */
5045 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5046 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5047 || (call->flags & RX_CALL_FAST_RECOVER)) {
5048 /* Check for the case where the current list contains
5049 * an acked packet. Since we always send retransmissions
5050 * in a separate packet, we only need to check the first
5051 * packet in the list */
5052 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5056 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5057 retryTime, resending);
5058 /* If the call enters an error state stop sending, or if
5059 * we entered congestion recovery mode, stop sending */
5060 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5064 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5067 } else if (lastCnt > 0) {
5068 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5073 #ifdef RX_ENABLE_LOCKS
5074 /* Call rxi_Start, below, but with the call lock held. */
5076 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
5077 void *arg1, int istack)
5079 MUTEX_ENTER(&call->lock);
5080 rxi_Start(event, call, arg1, istack);
5081 MUTEX_EXIT(&call->lock);
5083 #endif /* RX_ENABLE_LOCKS */
5085 /* This routine is called when new packets are readied for
5086 * transmission and when retransmission may be necessary, or when the
5087 * transmission window or burst count are favourable. This should be
5088 * better optimized for new packets, the usual case, now that we've
5089 * got rid of queues of send packets. XXXXXXXXXXX */
5091 rxi_Start(struct rxevent *event, register struct rx_call *call,
5092 void *arg1, int istack)
5094 struct rx_packet *p;
5095 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
5096 struct rx_peer *peer = call->conn->peer;
5097 struct clock now, usenow, retryTime;
5101 struct rx_packet **xmitList;
5104 /* If rxi_Start is being called as a result of a resend event,
5105 * then make sure that the event pointer is removed from the call
5106 * structure, since there is no longer a per-call retransmission
5108 if (event && event == call->resendEvent) {
5109 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5110 call->resendEvent = NULL;
5112 if (queue_IsEmpty(&call->tq)) {
5116 /* Timeouts trigger congestion recovery */
5117 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5118 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5119 /* someone else is waiting to start recovery */
5122 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5123 rxi_WaitforTQBusy(call);
5124 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5125 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5126 call->flags |= RX_CALL_FAST_RECOVER;
5127 if (peer->maxDgramPackets > 1) {
5128 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5130 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5132 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5133 call->nDgramPackets = 1;
5135 call->nextCwind = 1;
5138 MUTEX_ENTER(&peer->peer_lock);
5139 peer->MTU = call->MTU;
5140 peer->cwind = call->cwind;
5141 peer->nDgramPackets = 1;
5143 call->congestSeq = peer->congestSeq;
5144 MUTEX_EXIT(&peer->peer_lock);
5145 /* Clear retry times on packets. Otherwise, it's possible for
5146 * some packets in the queue to force resends at rates faster
5147 * than recovery rates.
5149 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5150 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5151 clock_Zero(&p->retryTime);
5156 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5157 rx_MutexIncrement(rx_tq_debug.rxi_start_in_error, rx_stats_mutex);
5162 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5163 /* Get clock to compute the re-transmit time for any packets
5164 * in this burst. Note, if we back off, it's reasonable to
5165 * back off all of the packets in the same manner, even if
5166 * some of them have been retransmitted more times than more
5168 * Do a dance to avoid blocking after setting now. */
5169 clock_Zero(&retryTime);
5170 MUTEX_ENTER(&peer->peer_lock);
5171 clock_Add(&retryTime, &peer->timeout);
5172 MUTEX_EXIT(&peer->peer_lock);
5173 clock_GetTime(&now);
5174 clock_Add(&retryTime, &now);
5176 /* Send (or resend) any packets that need it, subject to
5177 * window restrictions and congestion burst control
5178 * restrictions. Ask for an ack on the last packet sent in
5179 * this burst. For now, we're relying upon the window being
5180 * considerably bigger than the largest number of packets that
5181 * are typically sent at once by one initial call to
5182 * rxi_Start. This is probably bogus (perhaps we should ask
5183 * for an ack when we're half way through the current
5184 * window?). Also, for non file transfer applications, this
5185 * may end up asking for an ack for every packet. Bogus. XXXX
5188 * But check whether we're here recursively, and let the other guy
5191 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5192 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5193 call->flags |= RX_CALL_TQ_BUSY;
5195 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5197 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5198 call->flags &= ~RX_CALL_NEED_START;
5199 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5201 maxXmitPackets = MIN(call->twind, call->cwind);
5202 xmitList = (struct rx_packet **)
5203 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5204 if (xmitList == NULL)
5205 osi_Panic("rxi_Start, failed to allocate xmit list");
5206 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5207 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5208 /* We shouldn't be sending packets if a thread is waiting
5209 * to initiate congestion recovery */
5213 && (call->flags & RX_CALL_FAST_RECOVER)) {
5214 /* Only send one packet during fast recovery */
5217 if ((p->flags & RX_PKTFLAG_FREE)
5218 || (!queue_IsEnd(&call->tq, nxp)
5219 && (nxp->flags & RX_PKTFLAG_FREE))
5220 || (p == (struct rx_packet *)&rx_freePacketQueue)
5221 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5222 osi_Panic("rxi_Start: xmit queue clobbered");
5224 if (p->flags & RX_PKTFLAG_ACKED) {
5225 /* Since we may block, don't trust this */
5226 usenow.sec = usenow.usec = 0;
5227 rx_MutexIncrement(rx_stats.ignoreAckedPacket, rx_stats_mutex);
5228 continue; /* Ignore this packet if it has been acknowledged */
5231 /* Turn off all flags except these ones, which are the same
5232 * on each transmission */
5233 p->header.flags &= RX_PRESET_FLAGS;
5235 if (p->header.seq >=
5236 call->tfirst + MIN((int)call->twind,
5237 (int)(call->nSoftAcked +
5239 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5240 /* Note: if we're waiting for more window space, we can
5241 * still send retransmits; hence we don't return here, but
5242 * break out to schedule a retransmit event */
5243 dpf(("call %d waiting for window",
5244 *(call->callNumber)));
5248 /* Transmit the packet if it needs to be sent. */
5249 if (!clock_Lt(&now, &p->retryTime)) {
5250 if (nXmitPackets == maxXmitPackets) {
5251 rxi_SendXmitList(call, xmitList, nXmitPackets,
5252 istack, &now, &retryTime,
5254 osi_Free(xmitList, maxXmitPackets *
5255 sizeof(struct rx_packet *));
5258 xmitList[nXmitPackets++] = p;
5262 /* xmitList now hold pointers to all of the packets that are
5263 * ready to send. Now we loop to send the packets */
5264 if (nXmitPackets > 0) {
5265 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5266 &now, &retryTime, resending);
5269 maxXmitPackets * sizeof(struct rx_packet *));
5271 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5273 * TQ references no longer protected by this flag; they must remain
5274 * protected by the global lock.
5276 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5277 call->flags &= ~RX_CALL_TQ_BUSY;
5278 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5279 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5280 #ifdef RX_ENABLE_LOCKS
5281 osirx_AssertMine(&call->lock, "rxi_Start start");
5282 CV_BROADCAST(&call->cv_tq);
5283 #else /* RX_ENABLE_LOCKS */
5284 osi_rxWakeup(&call->tq);
5285 #endif /* RX_ENABLE_LOCKS */
5290 /* We went into the error state while sending packets. Now is
5291 * the time to reset the call. This will also inform the using
5292 * process that the call is in an error state.
5294 rx_MutexIncrement(rx_tq_debug.rxi_start_aborted, rx_stats_mutex);
5295 call->flags &= ~RX_CALL_TQ_BUSY;
5296 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5297 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5298 #ifdef RX_ENABLE_LOCKS
5299 osirx_AssertMine(&call->lock, "rxi_Start middle");
5300 CV_BROADCAST(&call->cv_tq);
5301 #else /* RX_ENABLE_LOCKS */
5302 osi_rxWakeup(&call->tq);
5303 #endif /* RX_ENABLE_LOCKS */
5305 rxi_CallError(call, call->error);
5308 #ifdef RX_ENABLE_LOCKS
5309 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5310 register int missing;
5311 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5312 /* Some packets have received acks. If they all have, we can clear
5313 * the transmit queue.
5316 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5317 if (p->header.seq < call->tfirst
5318 && (p->flags & RX_PKTFLAG_ACKED)) {
5325 call->flags |= RX_CALL_TQ_CLEARME;
5327 #endif /* RX_ENABLE_LOCKS */
5328 /* Don't bother doing retransmits if the TQ is cleared. */
5329 if (call->flags & RX_CALL_TQ_CLEARME) {
5330 rxi_ClearTransmitQueue(call, 1);
5332 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5335 /* Always post a resend event, if there is anything in the
5336 * queue, and resend is possible. There should be at least
5337 * one unacknowledged packet in the queue ... otherwise none
5338 * of these packets should be on the queue in the first place.
5340 if (call->resendEvent) {
5341 /* Cancel the existing event and post a new one */
5342 rxevent_Cancel(call->resendEvent, call,
5343 RX_CALL_REFCOUNT_RESEND);
5346 /* The retry time is the retry time on the first unacknowledged
5347 * packet inside the current window */
5349 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5350 /* Don't set timers for packets outside the window */
5351 if (p->header.seq >= call->tfirst + call->twind) {
5355 if (!(p->flags & RX_PKTFLAG_ACKED)
5356 && !clock_IsZero(&p->retryTime)) {
5358 retryTime = p->retryTime;
5363 /* Post a new event to re-run rxi_Start when retries may be needed */
5364 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5365 #ifdef RX_ENABLE_LOCKS
5366 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5368 rxevent_PostNow2(&retryTime, &usenow,
5370 (void *)call, 0, istack);
5371 #else /* RX_ENABLE_LOCKS */
5373 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5374 (void *)call, 0, istack);
5375 #endif /* RX_ENABLE_LOCKS */
5378 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5379 } while (call->flags & RX_CALL_NEED_START);
5381 * TQ references no longer protected by this flag; they must remain
5382 * protected by the global lock.
5384 call->flags &= ~RX_CALL_TQ_BUSY;
5385 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5386 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5387 #ifdef RX_ENABLE_LOCKS
5388 osirx_AssertMine(&call->lock, "rxi_Start end");
5389 CV_BROADCAST(&call->cv_tq);
5390 #else /* RX_ENABLE_LOCKS */
5391 osi_rxWakeup(&call->tq);
5392 #endif /* RX_ENABLE_LOCKS */
5395 call->flags |= RX_CALL_NEED_START;
5397 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5399 if (call->resendEvent) {
5400 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5405 /* Also adjusts the keep alive parameters for the call, to reflect
5406 * that we have just sent a packet (so keep alives aren't sent
5409 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5412 register struct rx_connection *conn = call->conn;
5414 /* Stamp each packet with the user supplied status */
5415 p->header.userStatus = call->localStatus;
5417 /* Allow the security object controlling this call's security to
5418 * make any last-minute changes to the packet */
5419 RXS_SendPacket(conn->securityObject, call, p);
5421 /* Since we're about to send SOME sort of packet to the peer, it's
5422 * safe to nuke any scheduled end-of-packets ack */
5423 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5425 /* Actually send the packet, filling in more connection-specific fields */
5426 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5427 MUTEX_EXIT(&call->lock);
5428 rxi_SendPacket(call, conn, p, istack);
5429 MUTEX_ENTER(&call->lock);
5430 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5432 /* Update last send time for this call (for keep-alive
5433 * processing), and for the connection (so that we can discover
5434 * idle connections) */
5435 conn->lastSendTime = call->lastSendTime = clock_Sec();
5436 /* Don't count keepalives here, so idleness can be tracked. */
5437 if (p->header.type != RX_PACKET_TYPE_ACK)
5438 call->lastSendData = call->lastSendTime;
5442 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5443 * that things are fine. Also called periodically to guarantee that nothing
5444 * falls through the cracks (e.g. (error + dally) connections have keepalive
5445 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5447 * haveCTLock Set if calling from rxi_ReapConnections
5449 #ifdef RX_ENABLE_LOCKS
5451 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5452 #else /* RX_ENABLE_LOCKS */
5454 rxi_CheckCall(register struct rx_call *call)
5455 #endif /* RX_ENABLE_LOCKS */
5457 register struct rx_connection *conn = call->conn;
5459 afs_uint32 deadTime;
5461 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5462 if (call->flags & RX_CALL_TQ_BUSY) {
5463 /* Call is active and will be reset by rxi_Start if it's
5464 * in an error state.
5469 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5471 (((afs_uint32) conn->secondsUntilDead << 10) +
5472 ((afs_uint32) conn->peer->rtt >> 3) +
5473 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5475 /* These are computed to the second (+- 1 second). But that's
5476 * good enough for these values, which should be a significant
5477 * number of seconds. */
5478 if (now > (call->lastReceiveTime + deadTime)) {
5479 if (call->state == RX_STATE_ACTIVE) {
5481 #if defined(KERNEL) && defined(AFS_SUN57_ENV)
5483 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5484 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
5485 ip_stack_t *ipst = ns->netstack_ip;
5487 ire = ire_cache_lookup(call->conn->peer->host
5488 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
5490 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
5492 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5499 if (ire && ire->ire_max_frag > 0)
5500 rxi_SetPeerMtu(call->conn->peer->host, 0, ire->ire_max_frag);
5501 #if defined(GLOBAL_NETSTACKID)
5505 #endif /* ADAPT_PMTU */
5506 rxi_CallError(call, RX_CALL_DEAD);
5509 #ifdef RX_ENABLE_LOCKS
5510 /* Cancel pending events */
5511 rxevent_Cancel(call->delayedAckEvent, call,
5512 RX_CALL_REFCOUNT_DELAY);
5513 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5514 rxevent_Cancel(call->keepAliveEvent, call,
5515 RX_CALL_REFCOUNT_ALIVE);
5516 if (call->refCount == 0) {
5517 rxi_FreeCall(call, haveCTLock);
5521 #else /* RX_ENABLE_LOCKS */
5524 #endif /* RX_ENABLE_LOCKS */
5526 /* Non-active calls are destroyed if they are not responding
5527 * to pings; active calls are simply flagged in error, so the
5528 * attached process can die reasonably gracefully. */
5530 /* see if we have a non-activity timeout */
5531 if (call->startWait && conn->idleDeadTime
5532 && ((call->startWait + conn->idleDeadTime) < now)) {
5533 if (call->state == RX_STATE_ACTIVE) {
5534 rxi_CallError(call, RX_CALL_TIMEOUT);
5538 if (call->lastSendData && conn->idleDeadTime && (conn->idleDeadErr != 0)
5539 && ((call->lastSendData + conn->idleDeadTime) < now)) {
5540 if (call->state == RX_STATE_ACTIVE) {
5541 rxi_CallError(call, conn->idleDeadErr);
5545 /* see if we have a hard timeout */
5546 if (conn->hardDeadTime
5547 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5548 if (call->state == RX_STATE_ACTIVE)
5549 rxi_CallError(call, RX_CALL_TIMEOUT);
5556 /* When a call is in progress, this routine is called occasionally to
5557 * make sure that some traffic has arrived (or been sent to) the peer.
5558 * If nothing has arrived in a reasonable amount of time, the call is
5559 * declared dead; if nothing has been sent for a while, we send a
5560 * keep-alive packet (if we're actually trying to keep the call alive)
5563 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5566 struct rx_connection *conn;
5569 MUTEX_ENTER(&call->lock);
5570 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5571 if (event == call->keepAliveEvent)
5572 call->keepAliveEvent = NULL;
5575 #ifdef RX_ENABLE_LOCKS
5576 if (rxi_CheckCall(call, 0)) {
5577 MUTEX_EXIT(&call->lock);
5580 #else /* RX_ENABLE_LOCKS */
5581 if (rxi_CheckCall(call))
5583 #endif /* RX_ENABLE_LOCKS */
5585 /* Don't try to keep alive dallying calls */
5586 if (call->state == RX_STATE_DALLY) {
5587 MUTEX_EXIT(&call->lock);
5592 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5593 /* Don't try to send keepalives if there is unacknowledged data */
5594 /* the rexmit code should be good enough, this little hack
5595 * doesn't quite work XXX */
5596 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5598 rxi_ScheduleKeepAliveEvent(call);
5599 MUTEX_EXIT(&call->lock);
5604 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5606 if (!call->keepAliveEvent) {
5607 struct clock when, now;
5608 clock_GetTime(&now);
5610 when.sec += call->conn->secondsUntilPing;
5611 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5612 call->keepAliveEvent =
5613 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
5617 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5619 rxi_KeepAliveOn(register struct rx_call *call)
5621 /* Pretend last packet received was received now--i.e. if another
5622 * packet isn't received within the keep alive time, then the call
5623 * will die; Initialize last send time to the current time--even
5624 * if a packet hasn't been sent yet. This will guarantee that a
5625 * keep-alive is sent within the ping time */
5626 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5627 rxi_ScheduleKeepAliveEvent(call);
5630 /* This routine is called to send connection abort messages
5631 * that have been delayed to throttle looping clients. */
5633 rxi_SendDelayedConnAbort(struct rxevent *event,
5634 register struct rx_connection *conn, char *dummy)
5637 struct rx_packet *packet;
5639 MUTEX_ENTER(&conn->conn_data_lock);
5640 conn->delayedAbortEvent = NULL;
5641 error = htonl(conn->error);
5643 MUTEX_EXIT(&conn->conn_data_lock);
5644 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5647 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5648 RX_PACKET_TYPE_ABORT, (char *)&error,
5650 rxi_FreePacket(packet);
5654 /* This routine is called to send call abort messages
5655 * that have been delayed to throttle looping clients. */
5657 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5661 struct rx_packet *packet;
5663 MUTEX_ENTER(&call->lock);
5664 call->delayedAbortEvent = NULL;
5665 error = htonl(call->error);
5667 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5670 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5671 (char *)&error, sizeof(error), 0);
5672 rxi_FreePacket(packet);
5674 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5675 MUTEX_EXIT(&call->lock);
5678 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5679 * seconds) to ask the client to authenticate itself. The routine
5680 * issues a challenge to the client, which is obtained from the
5681 * security object associated with the connection */
5683 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5684 void *arg1, int tries)
5686 conn->challengeEvent = NULL;
5687 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5688 register struct rx_packet *packet;
5689 struct clock when, now;
5692 /* We've failed to authenticate for too long.
5693 * Reset any calls waiting for authentication;
5694 * they are all in RX_STATE_PRECALL.
5698 MUTEX_ENTER(&conn->conn_call_lock);
5699 for (i = 0; i < RX_MAXCALLS; i++) {
5700 struct rx_call *call = conn->call[i];
5702 MUTEX_ENTER(&call->lock);
5703 if (call->state == RX_STATE_PRECALL) {
5704 rxi_CallError(call, RX_CALL_DEAD);
5705 rxi_SendCallAbort(call, NULL, 0, 0);
5707 MUTEX_EXIT(&call->lock);
5710 MUTEX_EXIT(&conn->conn_call_lock);
5714 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5716 /* If there's no packet available, do this later. */
5717 RXS_GetChallenge(conn->securityObject, conn, packet);
5718 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5719 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5720 rxi_FreePacket(packet);
5722 clock_GetTime(&now);
5724 when.sec += RX_CHALLENGE_TIMEOUT;
5725 conn->challengeEvent =
5726 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
5731 /* Call this routine to start requesting the client to authenticate
5732 * itself. This will continue until authentication is established,
5733 * the call times out, or an invalid response is returned. The
5734 * security object associated with the connection is asked to create
5735 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5736 * defined earlier. */
5738 rxi_ChallengeOn(register struct rx_connection *conn)
5740 if (!conn->challengeEvent) {
5741 RXS_CreateChallenge(conn->securityObject, conn);
5742 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5747 /* Compute round trip time of the packet provided, in *rttp.
5750 /* rxi_ComputeRoundTripTime is called with peer locked. */
5751 /* sentp and/or peer may be null */
5753 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5754 register struct clock *sentp,
5755 register struct rx_peer *peer)
5757 struct clock thisRtt, *rttp = &thisRtt;
5759 register int rtt_timeout;
5761 clock_GetTime(rttp);
5763 if (clock_Lt(rttp, sentp)) {
5765 return; /* somebody set the clock back, don't count this time. */
5767 clock_Sub(rttp, sentp);
5768 MUTEX_ENTER(&rx_stats_mutex);
5769 if (clock_Lt(rttp, &rx_stats.minRtt))
5770 rx_stats.minRtt = *rttp;
5771 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5772 if (rttp->sec > 60) {
5773 MUTEX_EXIT(&rx_stats_mutex);
5774 return; /* somebody set the clock ahead */
5776 rx_stats.maxRtt = *rttp;
5778 clock_Add(&rx_stats.totalRtt, rttp);
5779 rx_stats.nRttSamples++;
5780 MUTEX_EXIT(&rx_stats_mutex);
5782 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5784 /* Apply VanJacobson round-trip estimations */
5789 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5790 * srtt is stored as fixed point with 3 bits after the binary
5791 * point (i.e., scaled by 8). The following magic is
5792 * equivalent to the smoothing algorithm in rfc793 with an
5793 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5794 * srtt*8 = srtt*8 + rtt - srtt
5795 * srtt = srtt + rtt/8 - srtt/8
5798 delta = MSEC(rttp) - (peer->rtt >> 3);
5802 * We accumulate a smoothed rtt variance (actually, a smoothed
5803 * mean difference), then set the retransmit timer to smoothed
5804 * rtt + 4 times the smoothed variance (was 2x in van's original
5805 * paper, but 4x works better for me, and apparently for him as
5807 * rttvar is stored as
5808 * fixed point with 2 bits after the binary point (scaled by
5809 * 4). The following is equivalent to rfc793 smoothing with
5810 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5811 * replaces rfc793's wired-in beta.
5812 * dev*4 = dev*4 + (|actual - expected| - dev)
5818 delta -= (peer->rtt_dev >> 2);
5819 peer->rtt_dev += delta;
5821 /* I don't have a stored RTT so I start with this value. Since I'm
5822 * probably just starting a call, and will be pushing more data down
5823 * this, I expect congestion to increase rapidly. So I fudge a
5824 * little, and I set deviance to half the rtt. In practice,
5825 * deviance tends to approach something a little less than
5826 * half the smoothed rtt. */
5827 peer->rtt = (MSEC(rttp) << 3) + 8;
5828 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5830 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5831 * the other of these connections is usually in a user process, and can
5832 * be switched and/or swapped out. So on fast, reliable networks, the
5833 * timeout would otherwise be too short.
5835 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5836 clock_Zero(&(peer->timeout));
5837 clock_Addmsec(&(peer->timeout), rtt_timeout);
5839 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)));
5843 /* Find all server connections that have not been active for a long time, and
5846 rxi_ReapConnections(void)
5848 struct clock now, when;
5849 clock_GetTime(&now);
5851 /* Find server connection structures that haven't been used for
5852 * greater than rx_idleConnectionTime */
5854 struct rx_connection **conn_ptr, **conn_end;
5855 int i, havecalls = 0;
5856 MUTEX_ENTER(&rx_connHashTable_lock);
5857 for (conn_ptr = &rx_connHashTable[0], conn_end =
5858 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5860 struct rx_connection *conn, *next;
5861 struct rx_call *call;
5865 for (conn = *conn_ptr; conn; conn = next) {
5866 /* XXX -- Shouldn't the connection be locked? */
5869 for (i = 0; i < RX_MAXCALLS; i++) {
5870 call = conn->call[i];
5873 MUTEX_ENTER(&call->lock);
5874 #ifdef RX_ENABLE_LOCKS
5875 result = rxi_CheckCall(call, 1);
5876 #else /* RX_ENABLE_LOCKS */
5877 result = rxi_CheckCall(call);
5878 #endif /* RX_ENABLE_LOCKS */
5879 MUTEX_EXIT(&call->lock);
5881 /* If CheckCall freed the call, it might
5882 * have destroyed the connection as well,
5883 * which screws up the linked lists.
5889 if (conn->type == RX_SERVER_CONNECTION) {
5890 /* This only actually destroys the connection if
5891 * there are no outstanding calls */
5892 MUTEX_ENTER(&conn->conn_data_lock);
5893 if (!havecalls && !conn->refCount
5894 && ((conn->lastSendTime + rx_idleConnectionTime) <
5896 conn->refCount++; /* it will be decr in rx_DestroyConn */
5897 MUTEX_EXIT(&conn->conn_data_lock);
5898 #ifdef RX_ENABLE_LOCKS
5899 rxi_DestroyConnectionNoLock(conn);
5900 #else /* RX_ENABLE_LOCKS */
5901 rxi_DestroyConnection(conn);
5902 #endif /* RX_ENABLE_LOCKS */
5904 #ifdef RX_ENABLE_LOCKS
5906 MUTEX_EXIT(&conn->conn_data_lock);
5908 #endif /* RX_ENABLE_LOCKS */
5912 #ifdef RX_ENABLE_LOCKS
5913 while (rx_connCleanup_list) {
5914 struct rx_connection *conn;
5915 conn = rx_connCleanup_list;
5916 rx_connCleanup_list = rx_connCleanup_list->next;
5917 MUTEX_EXIT(&rx_connHashTable_lock);
5918 rxi_CleanupConnection(conn);
5919 MUTEX_ENTER(&rx_connHashTable_lock);
5921 MUTEX_EXIT(&rx_connHashTable_lock);
5922 #endif /* RX_ENABLE_LOCKS */
5925 /* Find any peer structures that haven't been used (haven't had an
5926 * associated connection) for greater than rx_idlePeerTime */
5928 struct rx_peer **peer_ptr, **peer_end;
5930 MUTEX_ENTER(&rx_rpc_stats);
5931 MUTEX_ENTER(&rx_peerHashTable_lock);
5932 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5933 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5935 struct rx_peer *peer, *next, *prev;
5936 for (prev = peer = *peer_ptr; peer; peer = next) {
5938 code = MUTEX_TRYENTER(&peer->peer_lock);
5939 if ((code) && (peer->refCount == 0)
5940 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5941 rx_interface_stat_p rpc_stat, nrpc_stat;
5943 MUTEX_EXIT(&peer->peer_lock);
5944 MUTEX_DESTROY(&peer->peer_lock);
5946 (&peer->rpcStats, rpc_stat, nrpc_stat,
5947 rx_interface_stat)) {
5948 unsigned int num_funcs;
5951 queue_Remove(&rpc_stat->queue_header);
5952 queue_Remove(&rpc_stat->all_peers);
5953 num_funcs = rpc_stat->stats[0].func_total;
5955 sizeof(rx_interface_stat_t) +
5956 rpc_stat->stats[0].func_total *
5957 sizeof(rx_function_entry_v1_t);
5959 rxi_Free(rpc_stat, space);
5960 rxi_rpc_peer_stat_cnt -= num_funcs;
5963 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
5964 if (peer == *peer_ptr) {
5971 MUTEX_EXIT(&peer->peer_lock);
5977 MUTEX_EXIT(&rx_peerHashTable_lock);
5978 MUTEX_EXIT(&rx_rpc_stats);
5981 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5982 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5983 * GC, just below. Really, we shouldn't have to keep moving packets from
5984 * one place to another, but instead ought to always know if we can
5985 * afford to hold onto a packet in its particular use. */
5986 MUTEX_ENTER(&rx_freePktQ_lock);
5987 if (rx_waitingForPackets) {
5988 rx_waitingForPackets = 0;
5989 #ifdef RX_ENABLE_LOCKS
5990 CV_BROADCAST(&rx_waitingForPackets_cv);
5992 osi_rxWakeup(&rx_waitingForPackets);
5995 MUTEX_EXIT(&rx_freePktQ_lock);
5998 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5999 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
6003 /* rxs_Release - This isn't strictly necessary but, since the macro name from
6004 * rx.h is sort of strange this is better. This is called with a security
6005 * object before it is discarded. Each connection using a security object has
6006 * its own refcount to the object so it won't actually be freed until the last
6007 * connection is destroyed.
6009 * This is the only rxs module call. A hold could also be written but no one
6013 rxs_Release(struct rx_securityClass *aobj)
6015 return RXS_Close(aobj);
6019 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
6020 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
6021 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
6022 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
6024 /* Adjust our estimate of the transmission rate to this peer, given
6025 * that the packet p was just acked. We can adjust peer->timeout and
6026 * call->twind. Pragmatically, this is called
6027 * only with packets of maximal length.
6028 * Called with peer and call locked.
6032 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
6033 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
6035 afs_int32 xferSize, xferMs;
6036 register afs_int32 minTime;
6039 /* Count down packets */
6040 if (peer->rateFlag > 0)
6042 /* Do nothing until we're enabled */
6043 if (peer->rateFlag != 0)
6048 /* Count only when the ack seems legitimate */
6049 switch (ackReason) {
6050 case RX_ACK_REQUESTED:
6052 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6056 case RX_ACK_PING_RESPONSE:
6057 if (p) /* want the response to ping-request, not data send */
6059 clock_GetTime(&newTO);
6060 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6061 clock_Sub(&newTO, &call->pingRequestTime);
6062 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6066 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6073 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));
6075 /* Track only packets that are big enough. */
6076 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6080 /* absorb RTT data (in milliseconds) for these big packets */
6081 if (peer->smRtt == 0) {
6082 peer->smRtt = xferMs;
6084 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6089 if (peer->countDown) {
6093 peer->countDown = 10; /* recalculate only every so often */
6095 /* In practice, we can measure only the RTT for full packets,
6096 * because of the way Rx acks the data that it receives. (If it's
6097 * smaller than a full packet, it often gets implicitly acked
6098 * either by the call response (from a server) or by the next call
6099 * (from a client), and either case confuses transmission times
6100 * with processing times.) Therefore, replace the above
6101 * more-sophisticated processing with a simpler version, where the
6102 * smoothed RTT is kept for full-size packets, and the time to
6103 * transmit a windowful of full-size packets is simply RTT *
6104 * windowSize. Again, we take two steps:
6105 - ensure the timeout is large enough for a single packet's RTT;
6106 - ensure that the window is small enough to fit in the desired timeout.*/
6108 /* First, the timeout check. */
6109 minTime = peer->smRtt;
6110 /* Get a reasonable estimate for a timeout period */
6112 newTO.sec = minTime / 1000;
6113 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6115 /* Increase the timeout period so that we can always do at least
6116 * one packet exchange */
6117 if (clock_Gt(&newTO, &peer->timeout)) {
6119 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));
6121 peer->timeout = newTO;
6124 /* Now, get an estimate for the transmit window size. */
6125 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6126 /* Now, convert to the number of full packets that could fit in a
6127 * reasonable fraction of that interval */
6128 minTime /= (peer->smRtt << 1);
6129 xferSize = minTime; /* (make a copy) */
6131 /* Now clamp the size to reasonable bounds. */
6134 else if (minTime > rx_Window)
6135 minTime = rx_Window;
6136 /* if (minTime != peer->maxWindow) {
6137 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6138 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6139 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6141 peer->maxWindow = minTime;
6142 elide... call->twind = minTime;
6146 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6147 * Discern this by calculating the timeout necessary for rx_Window
6149 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6150 /* calculate estimate for transmission interval in milliseconds */
6151 minTime = rx_Window * peer->smRtt;
6152 if (minTime < 1000) {
6153 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
6154 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6155 peer->timeout.usec, peer->smRtt, peer->packetSize));
6157 newTO.sec = 0; /* cut back on timeout by half a second */
6158 newTO.usec = 500000;
6159 clock_Sub(&peer->timeout, &newTO);
6164 } /* end of rxi_ComputeRate */
6165 #endif /* ADAPT_WINDOW */
6173 #define TRACE_OPTION_DEBUGLOG 4
6181 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6182 0, KEY_QUERY_VALUE, &parmKey);
6183 if (code != ERROR_SUCCESS)
6186 dummyLen = sizeof(TraceOption);
6187 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6188 (BYTE *) &TraceOption, &dummyLen);
6189 if (code == ERROR_SUCCESS) {
6190 rxdebug_active = (TraceOption & TRACE_OPTION_DEBUGLOG) ? 1 : 0;
6192 RegCloseKey (parmKey);
6193 #endif /* AFS_NT40_ENV */
6198 rx_DebugOnOff(int on)
6200 rxdebug_active = on;
6202 #endif /* AFS_NT40_ENV */
6205 /* Don't call this debugging routine directly; use dpf */
6207 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
6208 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
6216 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6219 len = _snprintf(msg, sizeof(msg)-2,
6220 tformat, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10,
6221 a11, a12, a13, a14, a15);
6223 if (msg[len-1] != '\n') {
6227 OutputDebugString(msg);
6232 clock_GetTime(&now);
6233 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
6234 (unsigned int)now.usec / 1000);
6235 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
6242 * This function is used to process the rx_stats structure that is local
6243 * to a process as well as an rx_stats structure received from a remote
6244 * process (via rxdebug). Therefore, it needs to do minimal version
6248 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
6249 afs_int32 freePackets, char version)
6253 if (size != sizeof(struct rx_stats)) {
6255 "Unexpected size of stats structure: was %d, expected %d\n",
6256 size, sizeof(struct rx_stats));
6259 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6262 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6263 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6264 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6265 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6266 s->specialPktAllocFailures);
6268 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6269 s->receivePktAllocFailures, s->sendPktAllocFailures,
6270 s->specialPktAllocFailures);
6274 " greedy %d, " "bogusReads %d (last from host %x), "
6275 "noPackets %d, " "noBuffers %d, " "selects %d, "
6276 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6277 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6278 s->selects, s->sendSelects);
6280 fprintf(file, " packets read: ");
6281 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6282 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6284 fprintf(file, "\n");
6287 " other read counters: data %d, " "ack %d, " "dup %d "
6288 "spurious %d " "dally %d\n", s->dataPacketsRead,
6289 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6290 s->ignorePacketDally);
6292 fprintf(file, " packets sent: ");
6293 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6294 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6296 fprintf(file, "\n");
6299 " other send counters: ack %d, " "data %d (not resends), "
6300 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6301 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6302 s->dataPacketsPushed, s->ignoreAckedPacket);
6305 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6306 s->netSendFailures, (int)s->fatalErrors);
6308 if (s->nRttSamples) {
6309 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6310 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6312 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6313 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6317 " %d server connections, " "%d client connections, "
6318 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6319 s->nServerConns, s->nClientConns, s->nPeerStructs,
6320 s->nCallStructs, s->nFreeCallStructs);
6322 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6323 fprintf(file, " %d clock updates\n", clock_nUpdates);
6328 /* for backward compatibility */
6330 rx_PrintStats(FILE * file)
6332 MUTEX_ENTER(&rx_stats_mutex);
6333 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6335 MUTEX_EXIT(&rx_stats_mutex);
6339 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6341 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6342 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6343 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6346 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6347 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6348 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6351 " Packet size %d, " "max in packet skew %d, "
6352 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6353 (int)peer->outPacketSkew);
6356 #ifdef AFS_PTHREAD_ENV
6358 * This mutex protects the following static variables:
6362 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6363 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6365 #define LOCK_RX_DEBUG
6366 #define UNLOCK_RX_DEBUG
6367 #endif /* AFS_PTHREAD_ENV */
6370 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6371 u_char type, void *inputData, size_t inputLength,
6372 void *outputData, size_t outputLength)
6374 static afs_int32 counter = 100;
6375 time_t waitTime, waitCount, startTime;
6376 struct rx_header theader;
6378 register afs_int32 code;
6379 struct timeval tv_now, tv_wake, tv_delta;
6380 struct sockaddr_in taddr, faddr;
6385 startTime = time(0);
6391 tp = &tbuffer[sizeof(struct rx_header)];
6392 taddr.sin_family = AF_INET;
6393 taddr.sin_port = remotePort;
6394 taddr.sin_addr.s_addr = remoteAddr;
6395 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6396 taddr.sin_len = sizeof(struct sockaddr_in);
6399 memset(&theader, 0, sizeof(theader));
6400 theader.epoch = htonl(999);
6402 theader.callNumber = htonl(counter);
6405 theader.type = type;
6406 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6407 theader.serviceId = 0;
6409 memcpy(tbuffer, &theader, sizeof(theader));
6410 memcpy(tp, inputData, inputLength);
6412 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6413 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6415 /* see if there's a packet available */
6416 gettimeofday(&tv_wake,0);
6417 tv_wake.tv_sec += waitTime;
6420 FD_SET(socket, &imask);
6421 tv_delta.tv_sec = tv_wake.tv_sec;
6422 tv_delta.tv_usec = tv_wake.tv_usec;
6423 gettimeofday(&tv_now, 0);
6425 if (tv_delta.tv_usec < tv_now.tv_usec) {
6427 tv_delta.tv_usec += 1000000;
6430 tv_delta.tv_usec -= tv_now.tv_usec;
6432 if (tv_delta.tv_sec < tv_now.tv_sec) {
6436 tv_delta.tv_sec -= tv_now.tv_sec;
6438 code = select(socket + 1, &imask, 0, 0, &tv_delta);
6439 if (code == 1 && FD_ISSET(socket, &imask)) {
6440 /* now receive a packet */
6441 faddrLen = sizeof(struct sockaddr_in);
6443 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6444 (struct sockaddr *)&faddr, &faddrLen);
6447 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6448 if (counter == ntohl(theader.callNumber))
6456 /* see if we've timed out */
6464 code -= sizeof(struct rx_header);
6465 if (code > outputLength)
6466 code = outputLength;
6467 memcpy(outputData, tp, code);
6472 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6473 afs_uint16 remotePort, struct rx_debugStats * stat,
6474 afs_uint32 * supportedValues)
6476 struct rx_debugIn in;
6479 *supportedValues = 0;
6480 in.type = htonl(RX_DEBUGI_GETSTATS);
6483 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6484 &in, sizeof(in), stat, sizeof(*stat));
6487 * If the call was successful, fixup the version and indicate
6488 * what contents of the stat structure are valid.
6489 * Also do net to host conversion of fields here.
6493 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6494 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6496 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6497 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6499 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6500 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6502 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6503 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6505 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6506 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6508 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6509 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6511 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6512 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6514 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6515 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6518 stat->nFreePackets = ntohl(stat->nFreePackets);
6519 stat->packetReclaims = ntohl(stat->packetReclaims);
6520 stat->callsExecuted = ntohl(stat->callsExecuted);
6521 stat->nWaiting = ntohl(stat->nWaiting);
6522 stat->idleThreads = ntohl(stat->idleThreads);
6529 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6530 afs_uint16 remotePort, struct rx_stats * stat,
6531 afs_uint32 * supportedValues)
6533 struct rx_debugIn in;
6534 afs_int32 *lp = (afs_int32 *) stat;
6539 * supportedValues is currently unused, but added to allow future
6540 * versioning of this function.
6543 *supportedValues = 0;
6544 in.type = htonl(RX_DEBUGI_RXSTATS);
6546 memset(stat, 0, sizeof(*stat));
6548 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6549 &in, sizeof(in), stat, sizeof(*stat));
6554 * Do net to host conversion here
6557 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6566 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6567 afs_uint16 remotePort, size_t version_length,
6571 return MakeDebugCall(socket, remoteAddr, remotePort,
6572 RX_PACKET_TYPE_VERSION, a, 1, version,
6577 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6578 afs_uint16 remotePort, afs_int32 * nextConnection,
6579 int allConnections, afs_uint32 debugSupportedValues,
6580 struct rx_debugConn * conn,
6581 afs_uint32 * supportedValues)
6583 struct rx_debugIn in;
6588 * supportedValues is currently unused, but added to allow future
6589 * versioning of this function.
6592 *supportedValues = 0;
6593 if (allConnections) {
6594 in.type = htonl(RX_DEBUGI_GETALLCONN);
6596 in.type = htonl(RX_DEBUGI_GETCONN);
6598 in.index = htonl(*nextConnection);
6599 memset(conn, 0, sizeof(*conn));
6601 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6602 &in, sizeof(in), conn, sizeof(*conn));
6605 *nextConnection += 1;
6608 * Convert old connection format to new structure.
6611 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6612 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6613 #define MOVEvL(a) (conn->a = vL->a)
6615 /* any old or unrecognized version... */
6616 for (i = 0; i < RX_MAXCALLS; i++) {
6617 MOVEvL(callState[i]);
6618 MOVEvL(callMode[i]);
6619 MOVEvL(callFlags[i]);
6620 MOVEvL(callOther[i]);
6622 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6623 MOVEvL(secStats.type);
6624 MOVEvL(secStats.level);
6625 MOVEvL(secStats.flags);
6626 MOVEvL(secStats.expires);
6627 MOVEvL(secStats.packetsReceived);
6628 MOVEvL(secStats.packetsSent);
6629 MOVEvL(secStats.bytesReceived);
6630 MOVEvL(secStats.bytesSent);
6635 * Do net to host conversion here
6637 * I don't convert host or port since we are most likely
6638 * going to want these in NBO.
6640 conn->cid = ntohl(conn->cid);
6641 conn->serial = ntohl(conn->serial);
6642 for (i = 0; i < RX_MAXCALLS; i++) {
6643 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6645 conn->error = ntohl(conn->error);
6646 conn->secStats.flags = ntohl(conn->secStats.flags);
6647 conn->secStats.expires = ntohl(conn->secStats.expires);
6648 conn->secStats.packetsReceived =
6649 ntohl(conn->secStats.packetsReceived);
6650 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6651 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6652 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6653 conn->epoch = ntohl(conn->epoch);
6654 conn->natMTU = ntohl(conn->natMTU);
6661 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6662 afs_uint16 remotePort, afs_int32 * nextPeer,
6663 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6664 afs_uint32 * supportedValues)
6666 struct rx_debugIn in;
6670 * supportedValues is currently unused, but added to allow future
6671 * versioning of this function.
6674 *supportedValues = 0;
6675 in.type = htonl(RX_DEBUGI_GETPEER);
6676 in.index = htonl(*nextPeer);
6677 memset(peer, 0, sizeof(*peer));
6679 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6680 &in, sizeof(in), peer, sizeof(*peer));
6686 * Do net to host conversion here
6688 * I don't convert host or port since we are most likely
6689 * going to want these in NBO.
6691 peer->ifMTU = ntohs(peer->ifMTU);
6692 peer->idleWhen = ntohl(peer->idleWhen);
6693 peer->refCount = ntohs(peer->refCount);
6694 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6695 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6696 peer->rtt = ntohl(peer->rtt);
6697 peer->rtt_dev = ntohl(peer->rtt_dev);
6698 peer->timeout.sec = ntohl(peer->timeout.sec);
6699 peer->timeout.usec = ntohl(peer->timeout.usec);
6700 peer->nSent = ntohl(peer->nSent);
6701 peer->reSends = ntohl(peer->reSends);
6702 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6703 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6704 peer->rateFlag = ntohl(peer->rateFlag);
6705 peer->natMTU = ntohs(peer->natMTU);
6706 peer->maxMTU = ntohs(peer->maxMTU);
6707 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6708 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6709 peer->MTU = ntohs(peer->MTU);
6710 peer->cwind = ntohs(peer->cwind);
6711 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6712 peer->congestSeq = ntohs(peer->congestSeq);
6713 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6714 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6715 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6716 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6721 #endif /* RXDEBUG */
6726 struct rx_serverQueueEntry *np;
6729 register struct rx_call *call;
6730 register struct rx_serverQueueEntry *sq;
6734 if (rxinit_status == 1) {
6736 return; /* Already shutdown. */
6740 #ifndef AFS_PTHREAD_ENV
6741 FD_ZERO(&rx_selectMask);
6742 #endif /* AFS_PTHREAD_ENV */
6743 rxi_dataQuota = RX_MAX_QUOTA;
6744 #ifndef AFS_PTHREAD_ENV
6746 #endif /* AFS_PTHREAD_ENV */
6749 #ifndef AFS_PTHREAD_ENV
6750 #ifndef AFS_USE_GETTIMEOFDAY
6752 #endif /* AFS_USE_GETTIMEOFDAY */
6753 #endif /* AFS_PTHREAD_ENV */
6755 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6756 call = queue_First(&rx_freeCallQueue, rx_call);
6758 rxi_Free(call, sizeof(struct rx_call));
6761 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6762 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6768 struct rx_peer **peer_ptr, **peer_end;
6769 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6770 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6772 struct rx_peer *peer, *next;
6773 for (peer = *peer_ptr; peer; peer = next) {
6774 rx_interface_stat_p rpc_stat, nrpc_stat;
6777 (&peer->rpcStats, rpc_stat, nrpc_stat,
6778 rx_interface_stat)) {
6779 unsigned int num_funcs;
6782 queue_Remove(&rpc_stat->queue_header);
6783 queue_Remove(&rpc_stat->all_peers);
6784 num_funcs = rpc_stat->stats[0].func_total;
6786 sizeof(rx_interface_stat_t) +
6787 rpc_stat->stats[0].func_total *
6788 sizeof(rx_function_entry_v1_t);
6790 rxi_Free(rpc_stat, space);
6791 MUTEX_ENTER(&rx_rpc_stats);
6792 rxi_rpc_peer_stat_cnt -= num_funcs;
6793 MUTEX_EXIT(&rx_rpc_stats);
6797 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6801 for (i = 0; i < RX_MAX_SERVICES; i++) {
6803 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6805 for (i = 0; i < rx_hashTableSize; i++) {
6806 register struct rx_connection *tc, *ntc;
6807 MUTEX_ENTER(&rx_connHashTable_lock);
6808 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6810 for (j = 0; j < RX_MAXCALLS; j++) {
6812 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6815 rxi_Free(tc, sizeof(*tc));
6817 MUTEX_EXIT(&rx_connHashTable_lock);
6820 MUTEX_ENTER(&freeSQEList_lock);
6822 while ((np = rx_FreeSQEList)) {
6823 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6824 MUTEX_DESTROY(&np->lock);
6825 rxi_Free(np, sizeof(*np));
6828 MUTEX_EXIT(&freeSQEList_lock);
6829 MUTEX_DESTROY(&freeSQEList_lock);
6830 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6831 MUTEX_DESTROY(&rx_connHashTable_lock);
6832 MUTEX_DESTROY(&rx_peerHashTable_lock);
6833 MUTEX_DESTROY(&rx_serverPool_lock);
6835 osi_Free(rx_connHashTable,
6836 rx_hashTableSize * sizeof(struct rx_connection *));
6837 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6839 UNPIN(rx_connHashTable,
6840 rx_hashTableSize * sizeof(struct rx_connection *));
6841 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6843 rxi_FreeAllPackets();
6845 MUTEX_ENTER(&rx_stats_mutex);
6846 rxi_dataQuota = RX_MAX_QUOTA;
6847 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6848 MUTEX_EXIT(&rx_stats_mutex);
6854 #ifdef RX_ENABLE_LOCKS
6856 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6858 if (!MUTEX_ISMINE(lockaddr))
6859 osi_Panic("Lock not held: %s", msg);
6861 #endif /* RX_ENABLE_LOCKS */
6866 * Routines to implement connection specific data.
6870 rx_KeyCreate(rx_destructor_t rtn)
6873 MUTEX_ENTER(&rxi_keyCreate_lock);
6874 key = rxi_keyCreate_counter++;
6875 rxi_keyCreate_destructor = (rx_destructor_t *)
6876 realloc((void *)rxi_keyCreate_destructor,
6877 (key + 1) * sizeof(rx_destructor_t));
6878 rxi_keyCreate_destructor[key] = rtn;
6879 MUTEX_EXIT(&rxi_keyCreate_lock);
6884 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6887 MUTEX_ENTER(&conn->conn_data_lock);
6888 if (!conn->specific) {
6889 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6890 for (i = 0; i < key; i++)
6891 conn->specific[i] = NULL;
6892 conn->nSpecific = key + 1;
6893 conn->specific[key] = ptr;
6894 } else if (key >= conn->nSpecific) {
6895 conn->specific = (void **)
6896 realloc(conn->specific, (key + 1) * sizeof(void *));
6897 for (i = conn->nSpecific; i < key; i++)
6898 conn->specific[i] = NULL;
6899 conn->nSpecific = key + 1;
6900 conn->specific[key] = ptr;
6902 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6903 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6904 conn->specific[key] = ptr;
6906 MUTEX_EXIT(&conn->conn_data_lock);
6910 rx_GetSpecific(struct rx_connection *conn, int key)
6913 MUTEX_ENTER(&conn->conn_data_lock);
6914 if (key >= conn->nSpecific)
6917 ptr = conn->specific[key];
6918 MUTEX_EXIT(&conn->conn_data_lock);
6922 #endif /* !KERNEL */
6925 * processStats is a queue used to store the statistics for the local
6926 * process. Its contents are similar to the contents of the rpcStats
6927 * queue on a rx_peer structure, but the actual data stored within
6928 * this queue contains totals across the lifetime of the process (assuming
6929 * the stats have not been reset) - unlike the per peer structures
6930 * which can come and go based upon the peer lifetime.
6933 static struct rx_queue processStats = { &processStats, &processStats };
6936 * peerStats is a queue used to store the statistics for all peer structs.
6937 * Its contents are the union of all the peer rpcStats queues.
6940 static struct rx_queue peerStats = { &peerStats, &peerStats };
6943 * rxi_monitor_processStats is used to turn process wide stat collection
6947 static int rxi_monitor_processStats = 0;
6950 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6953 static int rxi_monitor_peerStats = 0;
6956 * rxi_AddRpcStat - given all of the information for a particular rpc
6957 * call, create (if needed) and update the stat totals for the rpc.
6961 * IN stats - the queue of stats that will be updated with the new value
6963 * IN rxInterface - a unique number that identifies the rpc interface
6965 * IN currentFunc - the index of the function being invoked
6967 * IN totalFunc - the total number of functions in this interface
6969 * IN queueTime - the amount of time this function waited for a thread
6971 * IN execTime - the amount of time this function invocation took to execute
6973 * IN bytesSent - the number bytes sent by this invocation
6975 * IN bytesRcvd - the number bytes received by this invocation
6977 * IN isServer - if true, this invocation was made to a server
6979 * IN remoteHost - the ip address of the remote host
6981 * IN remotePort - the port of the remote host
6983 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6985 * INOUT counter - if a new stats structure is allocated, the counter will
6986 * be updated with the new number of allocated stat structures
6994 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6995 afs_uint32 currentFunc, afs_uint32 totalFunc,
6996 struct clock *queueTime, struct clock *execTime,
6997 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6998 afs_uint32 remoteHost, afs_uint32 remotePort,
6999 int addToPeerList, unsigned int *counter)
7002 rx_interface_stat_p rpc_stat, nrpc_stat;
7005 * See if there's already a structure for this interface
7008 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7009 if ((rpc_stat->stats[0].interfaceId == rxInterface)
7010 && (rpc_stat->stats[0].remote_is_server == isServer))
7015 * Didn't find a match so allocate a new structure and add it to the
7019 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
7020 || (rpc_stat->stats[0].interfaceId != rxInterface)
7021 || (rpc_stat->stats[0].remote_is_server != isServer)) {
7026 sizeof(rx_interface_stat_t) +
7027 totalFunc * sizeof(rx_function_entry_v1_t);
7029 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
7030 if (rpc_stat == NULL) {
7034 *counter += totalFunc;
7035 for (i = 0; i < totalFunc; i++) {
7036 rpc_stat->stats[i].remote_peer = remoteHost;
7037 rpc_stat->stats[i].remote_port = remotePort;
7038 rpc_stat->stats[i].remote_is_server = isServer;
7039 rpc_stat->stats[i].interfaceId = rxInterface;
7040 rpc_stat->stats[i].func_total = totalFunc;
7041 rpc_stat->stats[i].func_index = i;
7042 hzero(rpc_stat->stats[i].invocations);
7043 hzero(rpc_stat->stats[i].bytes_sent);
7044 hzero(rpc_stat->stats[i].bytes_rcvd);
7045 rpc_stat->stats[i].queue_time_sum.sec = 0;
7046 rpc_stat->stats[i].queue_time_sum.usec = 0;
7047 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7048 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7049 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7050 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7051 rpc_stat->stats[i].queue_time_max.sec = 0;
7052 rpc_stat->stats[i].queue_time_max.usec = 0;
7053 rpc_stat->stats[i].execution_time_sum.sec = 0;
7054 rpc_stat->stats[i].execution_time_sum.usec = 0;
7055 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7056 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7057 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7058 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7059 rpc_stat->stats[i].execution_time_max.sec = 0;
7060 rpc_stat->stats[i].execution_time_max.usec = 0;
7062 queue_Prepend(stats, rpc_stat);
7063 if (addToPeerList) {
7064 queue_Prepend(&peerStats, &rpc_stat->all_peers);
7069 * Increment the stats for this function
7072 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
7073 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
7074 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
7075 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
7076 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7077 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7078 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7080 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7081 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7083 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7084 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7086 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7087 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7089 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7090 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7098 * rx_IncrementTimeAndCount - increment the times and count for a particular
7103 * IN peer - the peer who invoked the rpc
7105 * IN rxInterface - a unique number that identifies the rpc interface
7107 * IN currentFunc - the index of the function being invoked
7109 * IN totalFunc - the total number of functions in this interface
7111 * IN queueTime - the amount of time this function waited for a thread
7113 * IN execTime - the amount of time this function invocation took to execute
7115 * IN bytesSent - the number bytes sent by this invocation
7117 * IN bytesRcvd - the number bytes received by this invocation
7119 * IN isServer - if true, this invocation was made to a server
7127 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7128 afs_uint32 currentFunc, afs_uint32 totalFunc,
7129 struct clock *queueTime, struct clock *execTime,
7130 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7134 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
7137 MUTEX_ENTER(&rx_rpc_stats);
7138 MUTEX_ENTER(&peer->peer_lock);
7140 if (rxi_monitor_peerStats) {
7141 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7142 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7143 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7146 if (rxi_monitor_processStats) {
7147 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7148 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7149 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7152 MUTEX_EXIT(&peer->peer_lock);
7153 MUTEX_EXIT(&rx_rpc_stats);
7158 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7162 * IN callerVersion - the rpc stat version of the caller.
7164 * IN count - the number of entries to marshall.
7166 * IN stats - pointer to stats to be marshalled.
7168 * OUT ptr - Where to store the marshalled data.
7175 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7176 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7182 * We only support the first version
7184 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7185 *(ptr++) = stats->remote_peer;
7186 *(ptr++) = stats->remote_port;
7187 *(ptr++) = stats->remote_is_server;
7188 *(ptr++) = stats->interfaceId;
7189 *(ptr++) = stats->func_total;
7190 *(ptr++) = stats->func_index;
7191 *(ptr++) = hgethi(stats->invocations);
7192 *(ptr++) = hgetlo(stats->invocations);
7193 *(ptr++) = hgethi(stats->bytes_sent);
7194 *(ptr++) = hgetlo(stats->bytes_sent);
7195 *(ptr++) = hgethi(stats->bytes_rcvd);
7196 *(ptr++) = hgetlo(stats->bytes_rcvd);
7197 *(ptr++) = stats->queue_time_sum.sec;
7198 *(ptr++) = stats->queue_time_sum.usec;
7199 *(ptr++) = stats->queue_time_sum_sqr.sec;
7200 *(ptr++) = stats->queue_time_sum_sqr.usec;
7201 *(ptr++) = stats->queue_time_min.sec;
7202 *(ptr++) = stats->queue_time_min.usec;
7203 *(ptr++) = stats->queue_time_max.sec;
7204 *(ptr++) = stats->queue_time_max.usec;
7205 *(ptr++) = stats->execution_time_sum.sec;
7206 *(ptr++) = stats->execution_time_sum.usec;
7207 *(ptr++) = stats->execution_time_sum_sqr.sec;
7208 *(ptr++) = stats->execution_time_sum_sqr.usec;
7209 *(ptr++) = stats->execution_time_min.sec;
7210 *(ptr++) = stats->execution_time_min.usec;
7211 *(ptr++) = stats->execution_time_max.sec;
7212 *(ptr++) = stats->execution_time_max.usec;
7218 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7223 * IN callerVersion - the rpc stat version of the caller
7225 * OUT myVersion - the rpc stat version of this function
7227 * OUT clock_sec - local time seconds
7229 * OUT clock_usec - local time microseconds
7231 * OUT allocSize - the number of bytes allocated to contain stats
7233 * OUT statCount - the number stats retrieved from this process.
7235 * OUT stats - the actual stats retrieved from this process.
7239 * Returns void. If successful, stats will != NULL.
7243 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7244 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7245 size_t * allocSize, afs_uint32 * statCount,
7246 afs_uint32 ** stats)
7256 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7259 * Check to see if stats are enabled
7262 MUTEX_ENTER(&rx_rpc_stats);
7263 if (!rxi_monitor_processStats) {
7264 MUTEX_EXIT(&rx_rpc_stats);
7268 clock_GetTime(&now);
7269 *clock_sec = now.sec;
7270 *clock_usec = now.usec;
7273 * Allocate the space based upon the caller version
7275 * If the client is at an older version than we are,
7276 * we return the statistic data in the older data format, but
7277 * we still return our version number so the client knows we
7278 * are maintaining more data than it can retrieve.
7281 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7282 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7283 *statCount = rxi_rpc_process_stat_cnt;
7286 * This can't happen yet, but in the future version changes
7287 * can be handled by adding additional code here
7291 if (space > (size_t) 0) {
7293 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7296 rx_interface_stat_p rpc_stat, nrpc_stat;
7300 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7302 * Copy the data based upon the caller version
7304 rx_MarshallProcessRPCStats(callerVersion,
7305 rpc_stat->stats[0].func_total,
7306 rpc_stat->stats, &ptr);
7312 MUTEX_EXIT(&rx_rpc_stats);
7317 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7321 * IN callerVersion - the rpc stat version of the caller
7323 * OUT myVersion - the rpc stat version of this function
7325 * OUT clock_sec - local time seconds
7327 * OUT clock_usec - local time microseconds
7329 * OUT allocSize - the number of bytes allocated to contain stats
7331 * OUT statCount - the number of stats retrieved from the individual
7334 * OUT stats - the actual stats retrieved from the individual peer structures.
7338 * Returns void. If successful, stats will != NULL.
7342 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7343 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7344 size_t * allocSize, afs_uint32 * statCount,
7345 afs_uint32 ** stats)
7355 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7358 * Check to see if stats are enabled
7361 MUTEX_ENTER(&rx_rpc_stats);
7362 if (!rxi_monitor_peerStats) {
7363 MUTEX_EXIT(&rx_rpc_stats);
7367 clock_GetTime(&now);
7368 *clock_sec = now.sec;
7369 *clock_usec = now.usec;
7372 * Allocate the space based upon the caller version
7374 * If the client is at an older version than we are,
7375 * we return the statistic data in the older data format, but
7376 * we still return our version number so the client knows we
7377 * are maintaining more data than it can retrieve.
7380 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7381 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7382 *statCount = rxi_rpc_peer_stat_cnt;
7385 * This can't happen yet, but in the future version changes
7386 * can be handled by adding additional code here
7390 if (space > (size_t) 0) {
7392 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7395 rx_interface_stat_p rpc_stat, nrpc_stat;
7399 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7401 * We have to fix the offset of rpc_stat since we are
7402 * keeping this structure on two rx_queues. The rx_queue
7403 * package assumes that the rx_queue member is the first
7404 * member of the structure. That is, rx_queue assumes that
7405 * any one item is only on one queue at a time. We are
7406 * breaking that assumption and so we have to do a little
7407 * math to fix our pointers.
7410 fix_offset = (char *)rpc_stat;
7411 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7412 rpc_stat = (rx_interface_stat_p) fix_offset;
7415 * Copy the data based upon the caller version
7417 rx_MarshallProcessRPCStats(callerVersion,
7418 rpc_stat->stats[0].func_total,
7419 rpc_stat->stats, &ptr);
7425 MUTEX_EXIT(&rx_rpc_stats);
7430 * rx_FreeRPCStats - free memory allocated by
7431 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7435 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7436 * rx_RetrievePeerRPCStats
7438 * IN allocSize - the number of bytes in stats.
7446 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7448 rxi_Free(stats, allocSize);
7452 * rx_queryProcessRPCStats - see if process rpc stat collection is
7453 * currently enabled.
7459 * Returns 0 if stats are not enabled != 0 otherwise
7463 rx_queryProcessRPCStats(void)
7466 MUTEX_ENTER(&rx_rpc_stats);
7467 rc = rxi_monitor_processStats;
7468 MUTEX_EXIT(&rx_rpc_stats);
7473 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7479 * Returns 0 if stats are not enabled != 0 otherwise
7483 rx_queryPeerRPCStats(void)
7486 MUTEX_ENTER(&rx_rpc_stats);
7487 rc = rxi_monitor_peerStats;
7488 MUTEX_EXIT(&rx_rpc_stats);
7493 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7503 rx_enableProcessRPCStats(void)
7505 MUTEX_ENTER(&rx_rpc_stats);
7506 rx_enable_stats = 1;
7507 rxi_monitor_processStats = 1;
7508 MUTEX_EXIT(&rx_rpc_stats);
7512 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7522 rx_enablePeerRPCStats(void)
7524 MUTEX_ENTER(&rx_rpc_stats);
7525 rx_enable_stats = 1;
7526 rxi_monitor_peerStats = 1;
7527 MUTEX_EXIT(&rx_rpc_stats);
7531 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7541 rx_disableProcessRPCStats(void)
7543 rx_interface_stat_p rpc_stat, nrpc_stat;
7546 MUTEX_ENTER(&rx_rpc_stats);
7549 * Turn off process statistics and if peer stats is also off, turn
7553 rxi_monitor_processStats = 0;
7554 if (rxi_monitor_peerStats == 0) {
7555 rx_enable_stats = 0;
7558 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7559 unsigned int num_funcs = 0;
7562 queue_Remove(rpc_stat);
7563 num_funcs = rpc_stat->stats[0].func_total;
7565 sizeof(rx_interface_stat_t) +
7566 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7568 rxi_Free(rpc_stat, space);
7569 rxi_rpc_process_stat_cnt -= num_funcs;
7571 MUTEX_EXIT(&rx_rpc_stats);
7575 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7585 rx_disablePeerRPCStats(void)
7587 struct rx_peer **peer_ptr, **peer_end;
7590 MUTEX_ENTER(&rx_rpc_stats);
7593 * Turn off peer statistics and if process stats is also off, turn
7597 rxi_monitor_peerStats = 0;
7598 if (rxi_monitor_processStats == 0) {
7599 rx_enable_stats = 0;
7602 MUTEX_ENTER(&rx_peerHashTable_lock);
7603 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7604 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7606 struct rx_peer *peer, *next, *prev;
7607 for (prev = peer = *peer_ptr; peer; peer = next) {
7609 code = MUTEX_TRYENTER(&peer->peer_lock);
7611 rx_interface_stat_p rpc_stat, nrpc_stat;
7614 (&peer->rpcStats, rpc_stat, nrpc_stat,
7615 rx_interface_stat)) {
7616 unsigned int num_funcs = 0;
7619 queue_Remove(&rpc_stat->queue_header);
7620 queue_Remove(&rpc_stat->all_peers);
7621 num_funcs = rpc_stat->stats[0].func_total;
7623 sizeof(rx_interface_stat_t) +
7624 rpc_stat->stats[0].func_total *
7625 sizeof(rx_function_entry_v1_t);
7627 rxi_Free(rpc_stat, space);
7628 rxi_rpc_peer_stat_cnt -= num_funcs;
7630 MUTEX_EXIT(&peer->peer_lock);
7631 if (prev == *peer_ptr) {
7641 MUTEX_EXIT(&rx_peerHashTable_lock);
7642 MUTEX_EXIT(&rx_rpc_stats);
7646 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7651 * IN clearFlag - flag indicating which stats to clear
7659 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7661 rx_interface_stat_p rpc_stat, nrpc_stat;
7663 MUTEX_ENTER(&rx_rpc_stats);
7665 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7666 unsigned int num_funcs = 0, i;
7667 num_funcs = rpc_stat->stats[0].func_total;
7668 for (i = 0; i < num_funcs; i++) {
7669 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7670 hzero(rpc_stat->stats[i].invocations);
7672 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7673 hzero(rpc_stat->stats[i].bytes_sent);
7675 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7676 hzero(rpc_stat->stats[i].bytes_rcvd);
7678 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7679 rpc_stat->stats[i].queue_time_sum.sec = 0;
7680 rpc_stat->stats[i].queue_time_sum.usec = 0;
7682 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7683 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7684 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7686 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7687 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7688 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7690 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7691 rpc_stat->stats[i].queue_time_max.sec = 0;
7692 rpc_stat->stats[i].queue_time_max.usec = 0;
7694 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7695 rpc_stat->stats[i].execution_time_sum.sec = 0;
7696 rpc_stat->stats[i].execution_time_sum.usec = 0;
7698 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7699 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7700 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7702 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7703 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7704 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7706 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7707 rpc_stat->stats[i].execution_time_max.sec = 0;
7708 rpc_stat->stats[i].execution_time_max.usec = 0;
7713 MUTEX_EXIT(&rx_rpc_stats);
7717 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7722 * IN clearFlag - flag indicating which stats to clear
7730 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7732 rx_interface_stat_p rpc_stat, nrpc_stat;
7734 MUTEX_ENTER(&rx_rpc_stats);
7736 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7737 unsigned int num_funcs = 0, i;
7740 * We have to fix the offset of rpc_stat since we are
7741 * keeping this structure on two rx_queues. The rx_queue
7742 * package assumes that the rx_queue member is the first
7743 * member of the structure. That is, rx_queue assumes that
7744 * any one item is only on one queue at a time. We are
7745 * breaking that assumption and so we have to do a little
7746 * math to fix our pointers.
7749 fix_offset = (char *)rpc_stat;
7750 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7751 rpc_stat = (rx_interface_stat_p) fix_offset;
7753 num_funcs = rpc_stat->stats[0].func_total;
7754 for (i = 0; i < num_funcs; i++) {
7755 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7756 hzero(rpc_stat->stats[i].invocations);
7758 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7759 hzero(rpc_stat->stats[i].bytes_sent);
7761 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7762 hzero(rpc_stat->stats[i].bytes_rcvd);
7764 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7765 rpc_stat->stats[i].queue_time_sum.sec = 0;
7766 rpc_stat->stats[i].queue_time_sum.usec = 0;
7768 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7769 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7770 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7772 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7773 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7774 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7776 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7777 rpc_stat->stats[i].queue_time_max.sec = 0;
7778 rpc_stat->stats[i].queue_time_max.usec = 0;
7780 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7781 rpc_stat->stats[i].execution_time_sum.sec = 0;
7782 rpc_stat->stats[i].execution_time_sum.usec = 0;
7784 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7785 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7786 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7788 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7789 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7790 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7792 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7793 rpc_stat->stats[i].execution_time_max.sec = 0;
7794 rpc_stat->stats[i].execution_time_max.usec = 0;
7799 MUTEX_EXIT(&rx_rpc_stats);
7803 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7804 * is authorized to enable/disable/clear RX statistics.
7806 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7809 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7811 rxi_rxstat_userok = proc;
7815 rx_RxStatUserOk(struct rx_call *call)
7817 if (!rxi_rxstat_userok)
7819 return rxi_rxstat_userok(call);
7824 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
7825 * function in the MSVC runtime DLL (msvcrt.dll).
7827 * Note: the system serializes calls to this function.
7830 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
7831 DWORD reason, /* reason function is being called */
7832 LPVOID reserved) /* reserved for future use */
7835 case DLL_PROCESS_ATTACH:
7836 /* library is being attached to a process */
7840 case DLL_PROCESS_DETACH: