2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
14 #include "afs/param.h"
16 #include <afs/param.h>
23 #include "afs/sysincludes.h"
24 #include "afsincludes.h"
30 #include <net/net_globals.h>
31 #endif /* AFS_OSF_ENV */
32 #ifdef AFS_LINUX20_ENV
35 #include "netinet/in.h"
36 #include "afs/afs_args.h"
37 #include "afs/afs_osi.h"
38 #ifdef RX_KERNEL_TRACE
39 #include "rx_kcommon.h"
41 #if (defined(AFS_AUX_ENV) || defined(AFS_AIX_ENV))
45 #undef RXDEBUG /* turn off debugging */
47 #if defined(AFS_SGI_ENV)
48 #include "sys/debug.h"
57 #endif /* AFS_OSF_ENV */
59 #include "afs/sysincludes.h"
60 #include "afsincludes.h"
63 #include "rx_kmutex.h"
64 #include "rx_kernel.h"
68 #include "rx_globals.h"
70 #define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
71 #define AFSOP_STOP_AFS 211 /* Stop AFS process */
72 #define AFSOP_STOP_BKG 212 /* Stop BKG process */
74 extern afs_int32 afs_termState;
76 #include "sys/lockl.h"
77 #include "sys/lock_def.h"
78 #endif /* AFS_AIX41_ENV */
79 # include "rxgen_consts.h"
81 # include <sys/types.h>
87 # include <afs/afsutil.h>
88 # include <WINNT\afsreg.h>
90 # include <sys/socket.h>
91 # include <sys/file.h>
93 # include <sys/stat.h>
94 # include <netinet/in.h>
95 # include <sys/time.h>
99 # include "rx_clock.h"
100 # include "rx_queue.h"
101 # include "rx_globals.h"
102 # include "rx_trace.h"
103 # include <afs/rxgen_consts.h>
106 int (*registerProgram) () = 0;
107 int (*swapNameProgram) () = 0;
109 /* Local static routines */
110 static void rxi_DestroyConnectionNoLock(register struct rx_connection *conn);
111 #ifdef RX_ENABLE_LOCKS
112 static void rxi_SetAcksInTransmitQueue(register struct rx_call *call);
115 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
117 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
118 afs_int32 rxi_start_in_error;
120 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
123 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
124 * currently allocated within rx. This number is used to allocate the
125 * memory required to return the statistics when queried.
128 static unsigned int rxi_rpc_peer_stat_cnt;
131 * rxi_rpc_process_stat_cnt counts the total number of local process stat
132 * structures currently allocated within rx. The number is used to allocate
133 * the memory required to return the statistics when queried.
136 static unsigned int rxi_rpc_process_stat_cnt;
138 #if !defined(offsetof)
139 #include <stddef.h> /* for definition of offsetof() */
142 #ifdef AFS_PTHREAD_ENV
146 * Use procedural initialization of mutexes/condition variables
150 extern pthread_mutex_t rx_stats_mutex;
151 extern pthread_mutex_t des_init_mutex;
152 extern pthread_mutex_t des_random_mutex;
153 extern pthread_mutex_t rx_clock_mutex;
154 extern pthread_mutex_t rxi_connCacheMutex;
155 extern pthread_mutex_t rx_event_mutex;
156 extern pthread_mutex_t osi_malloc_mutex;
157 extern pthread_mutex_t event_handler_mutex;
158 extern pthread_mutex_t listener_mutex;
159 extern pthread_mutex_t rx_if_init_mutex;
160 extern pthread_mutex_t rx_if_mutex;
161 extern pthread_mutex_t rxkad_client_uid_mutex;
162 extern pthread_mutex_t rxkad_random_mutex;
164 extern pthread_cond_t rx_event_handler_cond;
165 extern pthread_cond_t rx_listener_cond;
167 static pthread_mutex_t epoch_mutex;
168 static pthread_mutex_t rx_init_mutex;
169 static pthread_mutex_t rx_debug_mutex;
172 rxi_InitPthread(void)
174 assert(pthread_mutex_init(&rx_clock_mutex, (const pthread_mutexattr_t *)0)
176 assert(pthread_mutex_init(&rx_stats_mutex, (const pthread_mutexattr_t *)0)
178 assert(pthread_mutex_init
179 (&rxi_connCacheMutex, (const pthread_mutexattr_t *)0) == 0);
180 assert(pthread_mutex_init(&rx_init_mutex, (const pthread_mutexattr_t *)0)
182 assert(pthread_mutex_init(&epoch_mutex, (const pthread_mutexattr_t *)0) ==
184 assert(pthread_mutex_init(&rx_event_mutex, (const pthread_mutexattr_t *)0)
186 assert(pthread_mutex_init(&des_init_mutex, (const pthread_mutexattr_t *)0)
188 assert(pthread_mutex_init
189 (&des_random_mutex, (const pthread_mutexattr_t *)0) == 0);
190 assert(pthread_mutex_init
191 (&osi_malloc_mutex, (const pthread_mutexattr_t *)0) == 0);
192 assert(pthread_mutex_init
193 (&event_handler_mutex, (const pthread_mutexattr_t *)0) == 0);
194 assert(pthread_mutex_init(&listener_mutex, (const pthread_mutexattr_t *)0)
196 assert(pthread_mutex_init
197 (&rx_if_init_mutex, (const pthread_mutexattr_t *)0) == 0);
198 assert(pthread_mutex_init(&rx_if_mutex, (const pthread_mutexattr_t *)0) ==
200 assert(pthread_mutex_init
201 (&rxkad_client_uid_mutex, (const pthread_mutexattr_t *)0) == 0);
202 assert(pthread_mutex_init
203 (&rxkad_random_mutex, (const pthread_mutexattr_t *)0) == 0);
204 assert(pthread_mutex_init(&rx_debug_mutex, (const pthread_mutexattr_t *)0)
207 assert(pthread_cond_init
208 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
209 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
211 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
212 assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
214 rxkad_global_stats_init();
217 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
218 #define INIT_PTHREAD_LOCKS \
219 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
221 * The rx_stats_mutex mutex protects the following global variables:
226 * rxi_lowConnRefCount
227 * rxi_lowPeerRefCount
236 #define INIT_PTHREAD_LOCKS
240 /* Variables for handling the minProcs implementation. availProcs gives the
241 * number of threads available in the pool at this moment (not counting dudes
242 * executing right now). totalMin gives the total number of procs required
243 * for handling all minProcs requests. minDeficit is a dynamic variable
244 * tracking the # of procs required to satisfy all of the remaining minProcs
246 * For fine grain locking to work, the quota check and the reservation of
247 * a server thread has to come while rxi_availProcs and rxi_minDeficit
248 * are locked. To this end, the code has been modified under #ifdef
249 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
250 * same time. A new function, ReturnToServerPool() returns the allocation.
252 * A call can be on several queue's (but only one at a time). When
253 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
254 * that no one else is touching the queue. To this end, we store the address
255 * of the queue lock in the call structure (under the call lock) when we
256 * put the call on a queue, and we clear the call_queue_lock when the
257 * call is removed from a queue (once the call lock has been obtained).
258 * This allows rxi_ResetCall to safely synchronize with others wishing
259 * to manipulate the queue.
262 #ifdef RX_ENABLE_LOCKS
263 static afs_kmutex_t rx_rpc_stats;
264 void rxi_StartUnlocked();
267 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
268 ** pretty good that the next packet coming in is from the same connection
269 ** as the last packet, since we're send multiple packets in a transmit window.
271 struct rx_connection *rxLastConn = 0;
273 #ifdef RX_ENABLE_LOCKS
274 /* The locking hierarchy for rx fine grain locking is composed of these
277 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
278 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
279 * call->lock - locks call data fields.
280 * These are independent of each other:
281 * rx_freeCallQueue_lock
286 * serverQueueEntry->lock
288 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
289 * peer->lock - locks peer data fields.
290 * conn_data_lock - that more than one thread is not updating a conn data
291 * field at the same time.
299 * Do we need a lock to protect the peer field in the conn structure?
300 * conn->peer was previously a constant for all intents and so has no
301 * lock protecting this field. The multihomed client delta introduced
302 * a RX code change : change the peer field in the connection structure
303 * to that remote inetrface from which the last packet for this
304 * connection was sent out. This may become an issue if further changes
307 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
308 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
310 /* rxdb_fileID is used to identify the lock location, along with line#. */
311 static int rxdb_fileID = RXDB_FILE_RX;
312 #endif /* RX_LOCKS_DB */
313 #else /* RX_ENABLE_LOCKS */
314 #define SET_CALL_QUEUE_LOCK(C, L)
315 #define CLEAR_CALL_QUEUE_LOCK(C)
316 #endif /* RX_ENABLE_LOCKS */
317 struct rx_serverQueueEntry *rx_waitForPacket = 0;
318 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
320 /* ------------Exported Interfaces------------- */
322 /* This function allows rxkad to set the epoch to a suitably random number
323 * which rx_NewConnection will use in the future. The principle purpose is to
324 * get rxnull connections to use the same epoch as the rxkad connections do, at
325 * least once the first rxkad connection is established. This is important now
326 * that the host/port addresses aren't used in FindConnection: the uniqueness
327 * of epoch/cid matters and the start time won't do. */
329 #ifdef AFS_PTHREAD_ENV
331 * This mutex protects the following global variables:
335 #define LOCK_EPOCH assert(pthread_mutex_lock(&epoch_mutex)==0)
336 #define UNLOCK_EPOCH assert(pthread_mutex_unlock(&epoch_mutex)==0)
340 #endif /* AFS_PTHREAD_ENV */
343 rx_SetEpoch(afs_uint32 epoch)
350 /* Initialize rx. A port number may be mentioned, in which case this
351 * becomes the default port number for any service installed later.
352 * If 0 is provided for the port number, a random port will be chosen
353 * by the kernel. Whether this will ever overlap anything in
354 * /etc/services is anybody's guess... Returns 0 on success, -1 on
356 static int rxinit_status = 1;
357 #ifdef AFS_PTHREAD_ENV
359 * This mutex protects the following global variables:
363 #define LOCK_RX_INIT assert(pthread_mutex_lock(&rx_init_mutex)==0)
364 #define UNLOCK_RX_INIT assert(pthread_mutex_unlock(&rx_init_mutex)==0)
367 #define UNLOCK_RX_INIT
371 rx_InitHost(u_int host, u_int port)
378 char *htable, *ptable;
381 #if defined(AFS_DJGPP_ENV) && !defined(DEBUG)
382 __djgpp_set_quiet_socket(1);
389 if (rxinit_status == 0) {
390 tmp_status = rxinit_status;
392 return tmp_status; /* Already started; return previous error code. */
398 if (afs_winsockInit() < 0)
404 * Initialize anything necessary to provide a non-premptive threading
407 rxi_InitializeThreadSupport();
410 /* Allocate and initialize a socket for client and perhaps server
413 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
414 if (rx_socket == OSI_NULLSOCKET) {
418 #ifdef RX_ENABLE_LOCKS
421 #endif /* RX_LOCKS_DB */
422 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
423 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
424 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
425 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
426 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
428 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
430 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
432 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
434 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
436 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
438 #if defined(KERNEL) && defined(AFS_HPUX110_ENV)
440 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
441 #endif /* KERNEL && AFS_HPUX110_ENV */
442 #endif /* RX_ENABLE_LOCKS */
445 rx_connDeadTime = 12;
446 rx_tranquil = 0; /* reset flag */
447 memset((char *)&rx_stats, 0, sizeof(struct rx_stats));
449 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
450 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
451 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
452 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
453 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
454 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
456 /* Malloc up a bunch of packets & buffers */
458 queue_Init(&rx_freePacketQueue);
459 rxi_NeedMorePackets = FALSE;
460 #ifdef RX_ENABLE_TSFPQ
461 rx_nPackets = 0; /* in TSFPQ version, rx_nPackets is managed by rxi_MorePackets* */
462 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
463 #else /* RX_ENABLE_TSFPQ */
464 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
465 rxi_MorePackets(rx_nPackets);
466 #endif /* RX_ENABLE_TSFPQ */
473 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
474 tv.tv_sec = clock_now.sec;
475 tv.tv_usec = clock_now.usec;
476 srand((unsigned int)tv.tv_usec);
483 #if defined(KERNEL) && !defined(UKERNEL)
484 /* Really, this should never happen in a real kernel */
487 struct sockaddr_in addr;
488 int addrlen = sizeof(addr);
489 if (getsockname((int)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
493 rx_port = addr.sin_port;
496 rx_stats.minRtt.sec = 9999999;
498 rx_SetEpoch(tv.tv_sec | 0x80000000);
500 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
501 * will provide a randomer value. */
503 MUTEX_ENTER(&rx_stats_mutex);
504 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
505 MUTEX_EXIT(&rx_stats_mutex);
506 /* *Slightly* random start time for the cid. This is just to help
507 * out with the hashing function at the peer */
508 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
509 rx_connHashTable = (struct rx_connection **)htable;
510 rx_peerHashTable = (struct rx_peer **)ptable;
512 rx_lastAckDelay.sec = 0;
513 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
514 rx_hardAckDelay.sec = 0;
515 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
516 rx_softAckDelay.sec = 0;
517 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
519 rxevent_Init(20, rxi_ReScheduleEvents);
521 /* Initialize various global queues */
522 queue_Init(&rx_idleServerQueue);
523 queue_Init(&rx_incomingCallQueue);
524 queue_Init(&rx_freeCallQueue);
526 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
527 /* Initialize our list of usable IP addresses. */
531 /* Start listener process (exact function is dependent on the
532 * implementation environment--kernel or user space) */
536 tmp_status = rxinit_status = 0;
544 return rx_InitHost(htonl(INADDR_ANY), port);
547 /* called with unincremented nRequestsRunning to see if it is OK to start
548 * a new thread in this service. Could be "no" for two reasons: over the
549 * max quota, or would prevent others from reaching their min quota.
551 #ifdef RX_ENABLE_LOCKS
552 /* This verion of QuotaOK reserves quota if it's ok while the
553 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
556 QuotaOK(register struct rx_service *aservice)
558 /* check if over max quota */
559 if (aservice->nRequestsRunning >= aservice->maxProcs) {
563 /* under min quota, we're OK */
564 /* otherwise, can use only if there are enough to allow everyone
565 * to go to their min quota after this guy starts.
567 MUTEX_ENTER(&rx_stats_mutex);
568 if ((aservice->nRequestsRunning < aservice->minProcs)
569 || (rxi_availProcs > rxi_minDeficit)) {
570 aservice->nRequestsRunning++;
571 /* just started call in minProcs pool, need fewer to maintain
573 if (aservice->nRequestsRunning <= aservice->minProcs)
576 MUTEX_EXIT(&rx_stats_mutex);
579 MUTEX_EXIT(&rx_stats_mutex);
585 ReturnToServerPool(register struct rx_service *aservice)
587 aservice->nRequestsRunning--;
588 MUTEX_ENTER(&rx_stats_mutex);
589 if (aservice->nRequestsRunning < aservice->minProcs)
592 MUTEX_EXIT(&rx_stats_mutex);
595 #else /* RX_ENABLE_LOCKS */
597 QuotaOK(register struct rx_service *aservice)
600 /* under min quota, we're OK */
601 if (aservice->nRequestsRunning < aservice->minProcs)
604 /* check if over max quota */
605 if (aservice->nRequestsRunning >= aservice->maxProcs)
608 /* otherwise, can use only if there are enough to allow everyone
609 * to go to their min quota after this guy starts.
611 if (rxi_availProcs > rxi_minDeficit)
615 #endif /* RX_ENABLE_LOCKS */
618 /* Called by rx_StartServer to start up lwp's to service calls.
619 NExistingProcs gives the number of procs already existing, and which
620 therefore needn't be created. */
622 rxi_StartServerProcs(int nExistingProcs)
624 register struct rx_service *service;
629 /* For each service, reserve N processes, where N is the "minimum"
630 * number of processes that MUST be able to execute a request in parallel,
631 * at any time, for that process. Also compute the maximum difference
632 * between any service's maximum number of processes that can run
633 * (i.e. the maximum number that ever will be run, and a guarantee
634 * that this number will run if other services aren't running), and its
635 * minimum number. The result is the extra number of processes that
636 * we need in order to provide the latter guarantee */
637 for (i = 0; i < RX_MAX_SERVICES; i++) {
639 service = rx_services[i];
640 if (service == (struct rx_service *)0)
642 nProcs += service->minProcs;
643 diff = service->maxProcs - service->minProcs;
647 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
648 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
649 for (i = 0; i < nProcs; i++) {
650 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
656 /* This routine is only required on Windows */
658 rx_StartClientThread(void)
660 #ifdef AFS_PTHREAD_ENV
662 pid = (int) pthread_self();
663 #endif /* AFS_PTHREAD_ENV */
665 #endif /* AFS_NT40_ENV */
667 /* This routine must be called if any services are exported. If the
668 * donateMe flag is set, the calling process is donated to the server
671 rx_StartServer(int donateMe)
673 register struct rx_service *service;
679 /* Start server processes, if necessary (exact function is dependent
680 * on the implementation environment--kernel or user space). DonateMe
681 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
682 * case, one less new proc will be created rx_StartServerProcs.
684 rxi_StartServerProcs(donateMe);
686 /* count up the # of threads in minProcs, and add set the min deficit to
687 * be that value, too.
689 for (i = 0; i < RX_MAX_SERVICES; i++) {
690 service = rx_services[i];
691 if (service == (struct rx_service *)0)
693 MUTEX_ENTER(&rx_stats_mutex);
694 rxi_totalMin += service->minProcs;
695 /* below works even if a thread is running, since minDeficit would
696 * still have been decremented and later re-incremented.
698 rxi_minDeficit += service->minProcs;
699 MUTEX_EXIT(&rx_stats_mutex);
702 /* Turn on reaping of idle server connections */
703 rxi_ReapConnections();
712 #ifdef AFS_PTHREAD_ENV
714 pid = (pid_t) pthread_self();
715 #else /* AFS_PTHREAD_ENV */
717 LWP_CurrentProcess(&pid);
718 #endif /* AFS_PTHREAD_ENV */
720 sprintf(name, "srv_%d", ++nProcs);
722 (*registerProgram) (pid, name);
724 #endif /* AFS_NT40_ENV */
725 rx_ServerProc(); /* Never returns */
727 #ifdef RX_ENABLE_TSFPQ
728 /* no use leaving packets around in this thread's local queue if
729 * it isn't getting donated to the server thread pool.
731 rxi_FlushLocalPacketsTSFPQ();
732 #endif /* RX_ENABLE_TSFPQ */
736 /* Create a new client connection to the specified service, using the
737 * specified security object to implement the security model for this
739 struct rx_connection *
740 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
741 register struct rx_securityClass *securityObject,
742 int serviceSecurityIndex)
746 register struct rx_connection *conn;
751 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", ntohl(shost), ntohs(sport), sservice, securityObject, serviceSecurityIndex));
753 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
754 * the case of kmem_alloc? */
755 conn = rxi_AllocConnection();
756 #ifdef RX_ENABLE_LOCKS
757 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
758 MUTEX_INIT(&conn->conn_data_lock, "conn call lock", MUTEX_DEFAULT, 0);
759 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
762 MUTEX_ENTER(&rx_connHashTable_lock);
763 cid = (rx_nextCid += RX_MAXCALLS);
764 conn->type = RX_CLIENT_CONNECTION;
766 conn->epoch = rx_epoch;
767 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
768 conn->serviceId = sservice;
769 conn->securityObject = securityObject;
770 conn->securityData = (void *) 0;
771 conn->securityIndex = serviceSecurityIndex;
772 rx_SetConnDeadTime(conn, rx_connDeadTime);
773 conn->ackRate = RX_FAST_ACK_RATE;
775 conn->specific = NULL;
776 conn->challengeEvent = NULL;
777 conn->delayedAbortEvent = NULL;
778 conn->abortCount = 0;
781 RXS_NewConnection(securityObject, conn);
783 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
785 conn->refCount++; /* no lock required since only this thread knows... */
786 conn->next = rx_connHashTable[hashindex];
787 rx_connHashTable[hashindex] = conn;
788 MUTEX_ENTER(&rx_stats_mutex);
789 rx_stats.nClientConns++;
790 MUTEX_EXIT(&rx_stats_mutex);
792 MUTEX_EXIT(&rx_connHashTable_lock);
798 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
800 /* The idea is to set the dead time to a value that allows several
801 * keepalives to be dropped without timing out the connection. */
802 conn->secondsUntilDead = MAX(seconds, 6);
803 conn->secondsUntilPing = conn->secondsUntilDead / 6;
806 int rxi_lowPeerRefCount = 0;
807 int rxi_lowConnRefCount = 0;
810 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
811 * NOTE: must not be called with rx_connHashTable_lock held.
814 rxi_CleanupConnection(struct rx_connection *conn)
816 /* Notify the service exporter, if requested, that this connection
817 * is being destroyed */
818 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
819 (*conn->service->destroyConnProc) (conn);
821 /* Notify the security module that this connection is being destroyed */
822 RXS_DestroyConnection(conn->securityObject, conn);
824 /* If this is the last connection using the rx_peer struct, set its
825 * idle time to now. rxi_ReapConnections will reap it if it's still
826 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
828 MUTEX_ENTER(&rx_peerHashTable_lock);
829 if (conn->peer->refCount < 2) {
830 conn->peer->idleWhen = clock_Sec();
831 if (conn->peer->refCount < 1) {
832 conn->peer->refCount = 1;
833 MUTEX_ENTER(&rx_stats_mutex);
834 rxi_lowPeerRefCount++;
835 MUTEX_EXIT(&rx_stats_mutex);
838 conn->peer->refCount--;
839 MUTEX_EXIT(&rx_peerHashTable_lock);
841 MUTEX_ENTER(&rx_stats_mutex);
842 if (conn->type == RX_SERVER_CONNECTION)
843 rx_stats.nServerConns--;
845 rx_stats.nClientConns--;
846 MUTEX_EXIT(&rx_stats_mutex);
849 if (conn->specific) {
851 for (i = 0; i < conn->nSpecific; i++) {
852 if (conn->specific[i] && rxi_keyCreate_destructor[i])
853 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
854 conn->specific[i] = NULL;
856 free(conn->specific);
858 conn->specific = NULL;
862 MUTEX_DESTROY(&conn->conn_call_lock);
863 MUTEX_DESTROY(&conn->conn_data_lock);
864 CV_DESTROY(&conn->conn_call_cv);
866 rxi_FreeConnection(conn);
869 /* Destroy the specified connection */
871 rxi_DestroyConnection(register struct rx_connection *conn)
873 MUTEX_ENTER(&rx_connHashTable_lock);
874 rxi_DestroyConnectionNoLock(conn);
875 /* conn should be at the head of the cleanup list */
876 if (conn == rx_connCleanup_list) {
877 rx_connCleanup_list = rx_connCleanup_list->next;
878 MUTEX_EXIT(&rx_connHashTable_lock);
879 rxi_CleanupConnection(conn);
881 #ifdef RX_ENABLE_LOCKS
883 MUTEX_EXIT(&rx_connHashTable_lock);
885 #endif /* RX_ENABLE_LOCKS */
889 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
891 register struct rx_connection **conn_ptr;
892 register int havecalls = 0;
893 struct rx_packet *packet;
900 MUTEX_ENTER(&conn->conn_data_lock);
901 if (conn->refCount > 0)
904 MUTEX_ENTER(&rx_stats_mutex);
905 rxi_lowConnRefCount++;
906 MUTEX_EXIT(&rx_stats_mutex);
909 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
910 /* Busy; wait till the last guy before proceeding */
911 MUTEX_EXIT(&conn->conn_data_lock);
916 /* If the client previously called rx_NewCall, but it is still
917 * waiting, treat this as a running call, and wait to destroy the
918 * connection later when the call completes. */
919 if ((conn->type == RX_CLIENT_CONNECTION)
920 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
921 conn->flags |= RX_CONN_DESTROY_ME;
922 MUTEX_EXIT(&conn->conn_data_lock);
926 MUTEX_EXIT(&conn->conn_data_lock);
928 /* Check for extant references to this connection */
929 for (i = 0; i < RX_MAXCALLS; i++) {
930 register struct rx_call *call = conn->call[i];
933 if (conn->type == RX_CLIENT_CONNECTION) {
934 MUTEX_ENTER(&call->lock);
935 if (call->delayedAckEvent) {
936 /* Push the final acknowledgment out now--there
937 * won't be a subsequent call to acknowledge the
938 * last reply packets */
939 rxevent_Cancel(call->delayedAckEvent, call,
940 RX_CALL_REFCOUNT_DELAY);
941 if (call->state == RX_STATE_PRECALL
942 || call->state == RX_STATE_ACTIVE) {
943 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
945 rxi_AckAll(NULL, call, 0);
948 MUTEX_EXIT(&call->lock);
952 #ifdef RX_ENABLE_LOCKS
954 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
955 MUTEX_EXIT(&conn->conn_data_lock);
957 /* Someone is accessing a packet right now. */
961 #endif /* RX_ENABLE_LOCKS */
964 /* Don't destroy the connection if there are any call
965 * structures still in use */
966 MUTEX_ENTER(&conn->conn_data_lock);
967 conn->flags |= RX_CONN_DESTROY_ME;
968 MUTEX_EXIT(&conn->conn_data_lock);
973 if (conn->delayedAbortEvent) {
974 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
975 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
977 MUTEX_ENTER(&conn->conn_data_lock);
978 rxi_SendConnectionAbort(conn, packet, 0, 1);
979 MUTEX_EXIT(&conn->conn_data_lock);
980 rxi_FreePacket(packet);
984 /* Remove from connection hash table before proceeding */
986 &rx_connHashTable[CONN_HASH
987 (peer->host, peer->port, conn->cid, conn->epoch,
989 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
990 if (*conn_ptr == conn) {
991 *conn_ptr = conn->next;
995 /* if the conn that we are destroying was the last connection, then we
996 * clear rxLastConn as well */
997 if (rxLastConn == conn)
1000 /* Make sure the connection is completely reset before deleting it. */
1001 /* get rid of pending events that could zap us later */
1002 if (conn->challengeEvent)
1003 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1004 if (conn->checkReachEvent)
1005 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1007 /* Add the connection to the list of destroyed connections that
1008 * need to be cleaned up. This is necessary to avoid deadlocks
1009 * in the routines we call to inform others that this connection is
1010 * being destroyed. */
1011 conn->next = rx_connCleanup_list;
1012 rx_connCleanup_list = conn;
1015 /* Externally available version */
1017 rx_DestroyConnection(register struct rx_connection *conn)
1022 rxi_DestroyConnection(conn);
1027 rx_GetConnection(register struct rx_connection *conn)
1032 MUTEX_ENTER(&conn->conn_data_lock);
1034 MUTEX_EXIT(&conn->conn_data_lock);
1038 /* Wait for the transmit queue to no longer be busy.
1039 * requires the call->lock to be held */
1040 static void rxi_WaitforTQBusy(struct rx_call *call) {
1041 while (call->flags & RX_CALL_TQ_BUSY) {
1042 call->flags |= RX_CALL_TQ_WAIT;
1044 #ifdef RX_ENABLE_LOCKS
1045 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1046 CV_WAIT(&call->cv_tq, &call->lock);
1047 #else /* RX_ENABLE_LOCKS */
1048 osi_rxSleep(&call->tq);
1049 #endif /* RX_ENABLE_LOCKS */
1051 if (call->tqWaiters == 0) {
1052 call->flags &= ~RX_CALL_TQ_WAIT;
1056 /* Start a new rx remote procedure call, on the specified connection.
1057 * If wait is set to 1, wait for a free call channel; otherwise return
1058 * 0. Maxtime gives the maximum number of seconds this call may take,
1059 * after rx_NewCall returns. After this time interval, a call to any
1060 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1061 * For fine grain locking, we hold the conn_call_lock in order to
1062 * to ensure that we don't get signalle after we found a call in an active
1063 * state and before we go to sleep.
1066 rx_NewCall(register struct rx_connection *conn)
1069 register struct rx_call *call;
1070 struct clock queueTime;
1074 dpf(("rx_NewCall(conn %x)\n", conn));
1077 clock_GetTime(&queueTime);
1078 MUTEX_ENTER(&conn->conn_call_lock);
1081 * Check if there are others waiting for a new call.
1082 * If so, let them go first to avoid starving them.
1083 * This is a fairly simple scheme, and might not be
1084 * a complete solution for large numbers of waiters.
1086 * makeCallWaiters keeps track of the number of
1087 * threads waiting to make calls and the
1088 * RX_CONN_MAKECALL_WAITING flag bit is used to
1089 * indicate that there are indeed calls waiting.
1090 * The flag is set when the waiter is incremented.
1091 * It is only cleared in rx_EndCall when
1092 * makeCallWaiters is 0. This prevents us from
1093 * accidently destroying the connection while it
1094 * is potentially about to be used.
1096 MUTEX_ENTER(&conn->conn_data_lock);
1097 if (conn->makeCallWaiters) {
1098 conn->flags |= RX_CONN_MAKECALL_WAITING;
1099 conn->makeCallWaiters++;
1100 MUTEX_EXIT(&conn->conn_data_lock);
1102 #ifdef RX_ENABLE_LOCKS
1103 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1107 MUTEX_ENTER(&conn->conn_data_lock);
1108 conn->makeCallWaiters--;
1110 MUTEX_EXIT(&conn->conn_data_lock);
1113 for (i = 0; i < RX_MAXCALLS; i++) {
1114 call = conn->call[i];
1116 MUTEX_ENTER(&call->lock);
1117 if (call->state == RX_STATE_DALLY) {
1118 rxi_ResetCall(call, 0);
1119 (*call->callNumber)++;
1122 MUTEX_EXIT(&call->lock);
1124 call = rxi_NewCall(conn, i);
1128 if (i < RX_MAXCALLS) {
1131 MUTEX_ENTER(&conn->conn_data_lock);
1132 conn->flags |= RX_CONN_MAKECALL_WAITING;
1133 conn->makeCallWaiters++;
1134 MUTEX_EXIT(&conn->conn_data_lock);
1136 #ifdef RX_ENABLE_LOCKS
1137 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1141 MUTEX_ENTER(&conn->conn_data_lock);
1142 conn->makeCallWaiters--;
1143 MUTEX_EXIT(&conn->conn_data_lock);
1146 * Wake up anyone else who might be giving us a chance to
1147 * run (see code above that avoids resource starvation).
1149 #ifdef RX_ENABLE_LOCKS
1150 CV_BROADCAST(&conn->conn_call_cv);
1155 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1157 /* Client is initially in send mode */
1158 call->state = RX_STATE_ACTIVE;
1159 call->error = conn->error;
1161 call->mode = RX_MODE_ERROR;
1163 call->mode = RX_MODE_SENDING;
1165 /* remember start time for call in case we have hard dead time limit */
1166 call->queueTime = queueTime;
1167 clock_GetTime(&call->startTime);
1168 hzero(call->bytesSent);
1169 hzero(call->bytesRcvd);
1171 /* Turn on busy protocol. */
1172 rxi_KeepAliveOn(call);
1174 MUTEX_EXIT(&call->lock);
1175 MUTEX_EXIT(&conn->conn_call_lock);
1178 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1179 /* Now, if TQ wasn't cleared earlier, do it now. */
1180 MUTEX_ENTER(&call->lock);
1181 rxi_WaitforTQBusy(call);
1182 if (call->flags & RX_CALL_TQ_CLEARME) {
1183 rxi_ClearTransmitQueue(call, 0);
1184 queue_Init(&call->tq);
1186 MUTEX_EXIT(&call->lock);
1187 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1189 dpf(("rx_NewCall(call %x)\n", call));
1194 rxi_HasActiveCalls(register struct rx_connection *aconn)
1197 register struct rx_call *tcall;
1201 for (i = 0; i < RX_MAXCALLS; i++) {
1202 if ((tcall = aconn->call[i])) {
1203 if ((tcall->state == RX_STATE_ACTIVE)
1204 || (tcall->state == RX_STATE_PRECALL)) {
1215 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1216 register afs_int32 * aint32s)
1219 register struct rx_call *tcall;
1223 for (i = 0; i < RX_MAXCALLS; i++) {
1224 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1225 aint32s[i] = aconn->callNumber[i] + 1;
1227 aint32s[i] = aconn->callNumber[i];
1234 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1235 register afs_int32 * aint32s)
1238 register struct rx_call *tcall;
1242 for (i = 0; i < RX_MAXCALLS; i++) {
1243 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1244 aconn->callNumber[i] = aint32s[i] - 1;
1246 aconn->callNumber[i] = aint32s[i];
1252 /* Advertise a new service. A service is named locally by a UDP port
1253 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1256 char *serviceName; Name for identification purposes (e.g. the
1257 service name might be used for probing for
1260 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1261 char *serviceName, struct rx_securityClass **securityObjects,
1262 int nSecurityObjects,
1263 afs_int32(*serviceProc) (struct rx_call * acall))
1265 osi_socket socket = OSI_NULLSOCKET;
1266 register struct rx_service *tservice;
1272 if (serviceId == 0) {
1274 "rx_NewService: service id for service %s is not non-zero.\n",
1281 "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",
1289 tservice = rxi_AllocService();
1291 for (i = 0; i < RX_MAX_SERVICES; i++) {
1292 register struct rx_service *service = rx_services[i];
1294 if (port == service->servicePort && host == service->serviceHost) {
1295 if (service->serviceId == serviceId) {
1296 /* The identical service has already been
1297 * installed; if the caller was intending to
1298 * change the security classes used by this
1299 * service, he/she loses. */
1301 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1302 serviceName, serviceId, service->serviceName);
1304 rxi_FreeService(tservice);
1307 /* Different service, same port: re-use the socket
1308 * which is bound to the same port */
1309 socket = service->socket;
1312 if (socket == OSI_NULLSOCKET) {
1313 /* If we don't already have a socket (from another
1314 * service on same port) get a new one */
1315 socket = rxi_GetHostUDPSocket(host, port);
1316 if (socket == OSI_NULLSOCKET) {
1318 rxi_FreeService(tservice);
1323 service->socket = socket;
1324 service->serviceHost = host;
1325 service->servicePort = port;
1326 service->serviceId = serviceId;
1327 service->serviceName = serviceName;
1328 service->nSecurityObjects = nSecurityObjects;
1329 service->securityObjects = securityObjects;
1330 service->minProcs = 0;
1331 service->maxProcs = 1;
1332 service->idleDeadTime = 60;
1333 service->connDeadTime = rx_connDeadTime;
1334 service->executeRequestProc = serviceProc;
1335 service->checkReach = 0;
1336 rx_services[i] = service; /* not visible until now */
1342 rxi_FreeService(tservice);
1343 (osi_Msg "rx_NewService: cannot support > %d services\n",
1348 /* Set configuration options for all of a service's security objects */
1351 rx_SetSecurityConfiguration(struct rx_service *service,
1352 rx_securityConfigVariables type,
1356 for (i = 0; i<service->nSecurityObjects; i++) {
1357 if (service->securityObjects[i]) {
1358 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1366 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1367 struct rx_securityClass **securityObjects, int nSecurityObjects,
1368 afs_int32(*serviceProc) (struct rx_call * acall))
1370 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1373 /* Generic request processing loop. This routine should be called
1374 * by the implementation dependent rx_ServerProc. If socketp is
1375 * non-null, it will be set to the file descriptor that this thread
1376 * is now listening on. If socketp is null, this routine will never
1379 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1381 register struct rx_call *call;
1382 register afs_int32 code;
1383 register struct rx_service *tservice = NULL;
1390 call = rx_GetCall(threadID, tservice, socketp);
1391 if (socketp && *socketp != OSI_NULLSOCKET) {
1392 /* We are now a listener thread */
1397 /* if server is restarting( typically smooth shutdown) then do not
1398 * allow any new calls.
1401 if (rx_tranquil && (call != NULL)) {
1405 MUTEX_ENTER(&call->lock);
1407 rxi_CallError(call, RX_RESTARTING);
1408 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1410 MUTEX_EXIT(&call->lock);
1414 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1415 #ifdef RX_ENABLE_LOCKS
1417 #endif /* RX_ENABLE_LOCKS */
1418 afs_termState = AFSOP_STOP_AFS;
1419 afs_osi_Wakeup(&afs_termState);
1420 #ifdef RX_ENABLE_LOCKS
1422 #endif /* RX_ENABLE_LOCKS */
1427 tservice = call->conn->service;
1429 if (tservice->beforeProc)
1430 (*tservice->beforeProc) (call);
1432 code = call->conn->service->executeRequestProc(call);
1434 if (tservice->afterProc)
1435 (*tservice->afterProc) (call, code);
1437 rx_EndCall(call, code);
1438 MUTEX_ENTER(&rx_stats_mutex);
1440 MUTEX_EXIT(&rx_stats_mutex);
1446 rx_WakeupServerProcs(void)
1448 struct rx_serverQueueEntry *np, *tqp;
1452 MUTEX_ENTER(&rx_serverPool_lock);
1454 #ifdef RX_ENABLE_LOCKS
1455 if (rx_waitForPacket)
1456 CV_BROADCAST(&rx_waitForPacket->cv);
1457 #else /* RX_ENABLE_LOCKS */
1458 if (rx_waitForPacket)
1459 osi_rxWakeup(rx_waitForPacket);
1460 #endif /* RX_ENABLE_LOCKS */
1461 MUTEX_ENTER(&freeSQEList_lock);
1462 for (np = rx_FreeSQEList; np; np = tqp) {
1463 tqp = *(struct rx_serverQueueEntry **)np;
1464 #ifdef RX_ENABLE_LOCKS
1465 CV_BROADCAST(&np->cv);
1466 #else /* RX_ENABLE_LOCKS */
1468 #endif /* RX_ENABLE_LOCKS */
1470 MUTEX_EXIT(&freeSQEList_lock);
1471 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1472 #ifdef RX_ENABLE_LOCKS
1473 CV_BROADCAST(&np->cv);
1474 #else /* RX_ENABLE_LOCKS */
1476 #endif /* RX_ENABLE_LOCKS */
1478 MUTEX_EXIT(&rx_serverPool_lock);
1483 * One thing that seems to happen is that all the server threads get
1484 * tied up on some empty or slow call, and then a whole bunch of calls
1485 * arrive at once, using up the packet pool, so now there are more
1486 * empty calls. The most critical resources here are server threads
1487 * and the free packet pool. The "doreclaim" code seems to help in
1488 * general. I think that eventually we arrive in this state: there
1489 * are lots of pending calls which do have all their packets present,
1490 * so they won't be reclaimed, are multi-packet calls, so they won't
1491 * be scheduled until later, and thus are tying up most of the free
1492 * packet pool for a very long time.
1494 * 1. schedule multi-packet calls if all the packets are present.
1495 * Probably CPU-bound operation, useful to return packets to pool.
1496 * Do what if there is a full window, but the last packet isn't here?
1497 * 3. preserve one thread which *only* runs "best" calls, otherwise
1498 * it sleeps and waits for that type of call.
1499 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1500 * the current dataquota business is badly broken. The quota isn't adjusted
1501 * to reflect how many packets are presently queued for a running call.
1502 * So, when we schedule a queued call with a full window of packets queued
1503 * up for it, that *should* free up a window full of packets for other 2d-class
1504 * calls to be able to use from the packet pool. But it doesn't.
1506 * NB. Most of the time, this code doesn't run -- since idle server threads
1507 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1508 * as a new call arrives.
1510 /* Sleep until a call arrives. Returns a pointer to the call, ready
1511 * for an rx_Read. */
1512 #ifdef RX_ENABLE_LOCKS
1514 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1516 struct rx_serverQueueEntry *sq;
1517 register struct rx_call *call = (struct rx_call *)0;
1518 struct rx_service *service = NULL;
1521 MUTEX_ENTER(&freeSQEList_lock);
1523 if ((sq = rx_FreeSQEList)) {
1524 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1525 MUTEX_EXIT(&freeSQEList_lock);
1526 } else { /* otherwise allocate a new one and return that */
1527 MUTEX_EXIT(&freeSQEList_lock);
1528 sq = (struct rx_serverQueueEntry *)
1529 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1530 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1531 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1534 MUTEX_ENTER(&rx_serverPool_lock);
1535 if (cur_service != NULL) {
1536 ReturnToServerPool(cur_service);
1539 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1540 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1542 /* Scan for eligible incoming calls. A call is not eligible
1543 * if the maximum number of calls for its service type are
1544 * already executing */
1545 /* One thread will process calls FCFS (to prevent starvation),
1546 * while the other threads may run ahead looking for calls which
1547 * have all their input data available immediately. This helps
1548 * keep threads from blocking, waiting for data from the client. */
1549 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1550 service = tcall->conn->service;
1551 if (!QuotaOK(service)) {
1554 if (tno == rxi_fcfs_thread_num
1555 || !tcall->queue_item_header.next) {
1556 /* If we're the fcfs thread , then we'll just use
1557 * this call. If we haven't been able to find an optimal
1558 * choice, and we're at the end of the list, then use a
1559 * 2d choice if one has been identified. Otherwise... */
1560 call = (choice2 ? choice2 : tcall);
1561 service = call->conn->service;
1562 } else if (!queue_IsEmpty(&tcall->rq)) {
1563 struct rx_packet *rp;
1564 rp = queue_First(&tcall->rq, rx_packet);
1565 if (rp->header.seq == 1) {
1567 || (rp->header.flags & RX_LAST_PACKET)) {
1569 } else if (rxi_2dchoice && !choice2
1570 && !(tcall->flags & RX_CALL_CLEARED)
1571 && (tcall->rprev > rxi_HardAckRate)) {
1580 ReturnToServerPool(service);
1587 MUTEX_EXIT(&rx_serverPool_lock);
1588 MUTEX_ENTER(&call->lock);
1590 if (call->flags & RX_CALL_WAIT_PROC) {
1591 call->flags &= ~RX_CALL_WAIT_PROC;
1592 MUTEX_ENTER(&rx_stats_mutex);
1594 MUTEX_EXIT(&rx_stats_mutex);
1597 if (call->state != RX_STATE_PRECALL || call->error) {
1598 MUTEX_EXIT(&call->lock);
1599 MUTEX_ENTER(&rx_serverPool_lock);
1600 ReturnToServerPool(service);
1605 if (queue_IsEmpty(&call->rq)
1606 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1607 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1609 CLEAR_CALL_QUEUE_LOCK(call);
1612 /* If there are no eligible incoming calls, add this process
1613 * to the idle server queue, to wait for one */
1617 *socketp = OSI_NULLSOCKET;
1619 sq->socketp = socketp;
1620 queue_Append(&rx_idleServerQueue, sq);
1621 #ifndef AFS_AIX41_ENV
1622 rx_waitForPacket = sq;
1624 rx_waitingForPacket = sq;
1625 #endif /* AFS_AIX41_ENV */
1627 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1629 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1630 MUTEX_EXIT(&rx_serverPool_lock);
1631 return (struct rx_call *)0;
1634 } while (!(call = sq->newcall)
1635 && !(socketp && *socketp != OSI_NULLSOCKET));
1636 MUTEX_EXIT(&rx_serverPool_lock);
1638 MUTEX_ENTER(&call->lock);
1644 MUTEX_ENTER(&freeSQEList_lock);
1645 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1646 rx_FreeSQEList = sq;
1647 MUTEX_EXIT(&freeSQEList_lock);
1650 clock_GetTime(&call->startTime);
1651 call->state = RX_STATE_ACTIVE;
1652 call->mode = RX_MODE_RECEIVING;
1653 #ifdef RX_KERNEL_TRACE
1654 if (ICL_SETACTIVE(afs_iclSetp)) {
1655 int glockOwner = ISAFS_GLOCK();
1658 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1659 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1666 rxi_calltrace(RX_CALL_START, call);
1667 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1668 call->conn->service->servicePort, call->conn->service->serviceId,
1671 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1672 MUTEX_EXIT(&call->lock);
1674 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1679 #else /* RX_ENABLE_LOCKS */
1681 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1683 struct rx_serverQueueEntry *sq;
1684 register struct rx_call *call = (struct rx_call *)0, *choice2;
1685 struct rx_service *service = NULL;
1689 MUTEX_ENTER(&freeSQEList_lock);
1691 if ((sq = rx_FreeSQEList)) {
1692 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1693 MUTEX_EXIT(&freeSQEList_lock);
1694 } else { /* otherwise allocate a new one and return that */
1695 MUTEX_EXIT(&freeSQEList_lock);
1696 sq = (struct rx_serverQueueEntry *)
1697 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1698 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1699 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1701 MUTEX_ENTER(&sq->lock);
1703 if (cur_service != NULL) {
1704 cur_service->nRequestsRunning--;
1705 if (cur_service->nRequestsRunning < cur_service->minProcs)
1709 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1710 register struct rx_call *tcall, *ncall;
1711 /* Scan for eligible incoming calls. A call is not eligible
1712 * if the maximum number of calls for its service type are
1713 * already executing */
1714 /* One thread will process calls FCFS (to prevent starvation),
1715 * while the other threads may run ahead looking for calls which
1716 * have all their input data available immediately. This helps
1717 * keep threads from blocking, waiting for data from the client. */
1718 choice2 = (struct rx_call *)0;
1719 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1720 service = tcall->conn->service;
1721 if (QuotaOK(service)) {
1722 if (tno == rxi_fcfs_thread_num
1723 || !tcall->queue_item_header.next) {
1724 /* If we're the fcfs thread, then we'll just use
1725 * this call. If we haven't been able to find an optimal
1726 * choice, and we're at the end of the list, then use a
1727 * 2d choice if one has been identified. Otherwise... */
1728 call = (choice2 ? choice2 : tcall);
1729 service = call->conn->service;
1730 } else if (!queue_IsEmpty(&tcall->rq)) {
1731 struct rx_packet *rp;
1732 rp = queue_First(&tcall->rq, rx_packet);
1733 if (rp->header.seq == 1
1735 || (rp->header.flags & RX_LAST_PACKET))) {
1737 } else if (rxi_2dchoice && !choice2
1738 && !(tcall->flags & RX_CALL_CLEARED)
1739 && (tcall->rprev > rxi_HardAckRate)) {
1752 /* we can't schedule a call if there's no data!!! */
1753 /* send an ack if there's no data, if we're missing the
1754 * first packet, or we're missing something between first
1755 * and last -- there's a "hole" in the incoming data. */
1756 if (queue_IsEmpty(&call->rq)
1757 || queue_First(&call->rq, rx_packet)->header.seq != 1
1758 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1759 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1761 call->flags &= (~RX_CALL_WAIT_PROC);
1762 service->nRequestsRunning++;
1763 /* just started call in minProcs pool, need fewer to maintain
1765 if (service->nRequestsRunning <= service->minProcs)
1769 /* MUTEX_EXIT(&call->lock); */
1771 /* If there are no eligible incoming calls, add this process
1772 * to the idle server queue, to wait for one */
1775 *socketp = OSI_NULLSOCKET;
1777 sq->socketp = socketp;
1778 queue_Append(&rx_idleServerQueue, sq);
1782 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1784 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1785 return (struct rx_call *)0;
1788 } while (!(call = sq->newcall)
1789 && !(socketp && *socketp != OSI_NULLSOCKET));
1791 MUTEX_EXIT(&sq->lock);
1793 MUTEX_ENTER(&freeSQEList_lock);
1794 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1795 rx_FreeSQEList = sq;
1796 MUTEX_EXIT(&freeSQEList_lock);
1799 clock_GetTime(&call->startTime);
1800 call->state = RX_STATE_ACTIVE;
1801 call->mode = RX_MODE_RECEIVING;
1802 #ifdef RX_KERNEL_TRACE
1803 if (ICL_SETACTIVE(afs_iclSetp)) {
1804 int glockOwner = ISAFS_GLOCK();
1807 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1808 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1815 rxi_calltrace(RX_CALL_START, call);
1816 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1817 call->conn->service->servicePort, call->conn->service->serviceId,
1820 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1827 #endif /* RX_ENABLE_LOCKS */
1831 /* Establish a procedure to be called when a packet arrives for a
1832 * call. This routine will be called at most once after each call,
1833 * and will also be called if there is an error condition on the or
1834 * the call is complete. Used by multi rx to build a selection
1835 * function which determines which of several calls is likely to be a
1836 * good one to read from.
1837 * NOTE: the way this is currently implemented it is probably only a
1838 * good idea to (1) use it immediately after a newcall (clients only)
1839 * and (2) only use it once. Other uses currently void your warranty
1842 rx_SetArrivalProc(register struct rx_call *call,
1843 register void (*proc) (register struct rx_call * call,
1845 register int index),
1846 register void * handle, register int arg)
1848 call->arrivalProc = proc;
1849 call->arrivalProcHandle = handle;
1850 call->arrivalProcArg = arg;
1853 /* Call is finished (possibly prematurely). Return rc to the peer, if
1854 * appropriate, and return the final error code from the conversation
1858 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1860 register struct rx_connection *conn = call->conn;
1861 register struct rx_service *service;
1867 dpf(("rx_EndCall(call %x rc %d error %d abortCode %d)\n", call, rc, call->error, call->abortCode));
1870 MUTEX_ENTER(&call->lock);
1872 if (rc == 0 && call->error == 0) {
1873 call->abortCode = 0;
1874 call->abortCount = 0;
1877 call->arrivalProc = (void (*)())0;
1878 if (rc && call->error == 0) {
1879 rxi_CallError(call, rc);
1880 /* Send an abort message to the peer if this error code has
1881 * only just been set. If it was set previously, assume the
1882 * peer has already been sent the error code or will request it
1884 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1886 if (conn->type == RX_SERVER_CONNECTION) {
1887 /* Make sure reply or at least dummy reply is sent */
1888 if (call->mode == RX_MODE_RECEIVING) {
1889 rxi_WriteProc(call, 0, 0);
1891 if (call->mode == RX_MODE_SENDING) {
1892 rxi_FlushWrite(call);
1894 service = conn->service;
1895 rxi_calltrace(RX_CALL_END, call);
1896 /* Call goes to hold state until reply packets are acknowledged */
1897 if (call->tfirst + call->nSoftAcked < call->tnext) {
1898 call->state = RX_STATE_HOLD;
1900 call->state = RX_STATE_DALLY;
1901 rxi_ClearTransmitQueue(call, 0);
1902 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1903 rxevent_Cancel(call->keepAliveEvent, call,
1904 RX_CALL_REFCOUNT_ALIVE);
1906 } else { /* Client connection */
1908 /* Make sure server receives input packets, in the case where
1909 * no reply arguments are expected */
1910 if ((call->mode == RX_MODE_SENDING)
1911 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1912 (void)rxi_ReadProc(call, &dummy, 1);
1915 /* If we had an outstanding delayed ack, be nice to the server
1916 * and force-send it now.
1918 if (call->delayedAckEvent) {
1919 rxevent_Cancel(call->delayedAckEvent, call,
1920 RX_CALL_REFCOUNT_DELAY);
1921 call->delayedAckEvent = NULL;
1922 rxi_SendDelayedAck(NULL, call, NULL);
1925 /* We need to release the call lock since it's lower than the
1926 * conn_call_lock and we don't want to hold the conn_call_lock
1927 * over the rx_ReadProc call. The conn_call_lock needs to be held
1928 * here for the case where rx_NewCall is perusing the calls on
1929 * the connection structure. We don't want to signal until
1930 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1931 * have checked this call, found it active and by the time it
1932 * goes to sleep, will have missed the signal.
1934 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
1935 * there are threads waiting to use the conn object.
1937 MUTEX_EXIT(&call->lock);
1938 MUTEX_ENTER(&conn->conn_call_lock);
1939 MUTEX_ENTER(&call->lock);
1940 MUTEX_ENTER(&conn->conn_data_lock);
1941 conn->flags |= RX_CONN_BUSY;
1942 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1943 if (conn->makeCallWaiters == 0)
1944 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1945 MUTEX_EXIT(&conn->conn_data_lock);
1946 #ifdef RX_ENABLE_LOCKS
1947 CV_BROADCAST(&conn->conn_call_cv);
1952 #ifdef RX_ENABLE_LOCKS
1954 MUTEX_EXIT(&conn->conn_data_lock);
1956 #endif /* RX_ENABLE_LOCKS */
1957 call->state = RX_STATE_DALLY;
1959 error = call->error;
1961 /* currentPacket, nLeft, and NFree must be zeroed here, because
1962 * ResetCall cannot: ResetCall may be called at splnet(), in the
1963 * kernel version, and may interrupt the macros rx_Read or
1964 * rx_Write, which run at normal priority for efficiency. */
1965 if (call->currentPacket) {
1966 queue_Prepend(&call->iovq, call->currentPacket);
1967 call->currentPacket = (struct rx_packet *)0;
1970 call->nLeft = call->nFree = call->curlen = 0;
1972 /* Free any packets from the last call to ReadvProc/WritevProc */
1973 rxi_FreePackets(0, &call->iovq);
1975 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1976 MUTEX_EXIT(&call->lock);
1977 if (conn->type == RX_CLIENT_CONNECTION) {
1978 MUTEX_EXIT(&conn->conn_call_lock);
1979 conn->flags &= ~RX_CONN_BUSY;
1983 * Map errors to the local host's errno.h format.
1985 error = ntoh_syserr_conv(error);
1989 #if !defined(KERNEL)
1991 /* Call this routine when shutting down a server or client (especially
1992 * clients). This will allow Rx to gracefully garbage collect server
1993 * connections, and reduce the number of retries that a server might
1994 * make to a dead client.
1995 * This is not quite right, since some calls may still be ongoing and
1996 * we can't lock them to destroy them. */
2000 register struct rx_connection **conn_ptr, **conn_end;
2004 if (rxinit_status == 1) {
2006 return; /* Already shutdown. */
2008 rxi_DeleteCachedConnections();
2009 if (rx_connHashTable) {
2010 MUTEX_ENTER(&rx_connHashTable_lock);
2011 for (conn_ptr = &rx_connHashTable[0], conn_end =
2012 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2014 struct rx_connection *conn, *next;
2015 for (conn = *conn_ptr; conn; conn = next) {
2017 if (conn->type == RX_CLIENT_CONNECTION) {
2018 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2020 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2021 #ifdef RX_ENABLE_LOCKS
2022 rxi_DestroyConnectionNoLock(conn);
2023 #else /* RX_ENABLE_LOCKS */
2024 rxi_DestroyConnection(conn);
2025 #endif /* RX_ENABLE_LOCKS */
2029 #ifdef RX_ENABLE_LOCKS
2030 while (rx_connCleanup_list) {
2031 struct rx_connection *conn;
2032 conn = rx_connCleanup_list;
2033 rx_connCleanup_list = rx_connCleanup_list->next;
2034 MUTEX_EXIT(&rx_connHashTable_lock);
2035 rxi_CleanupConnection(conn);
2036 MUTEX_ENTER(&rx_connHashTable_lock);
2038 MUTEX_EXIT(&rx_connHashTable_lock);
2039 #endif /* RX_ENABLE_LOCKS */
2044 afs_winsockCleanup();
2052 /* if we wakeup packet waiter too often, can get in loop with two
2053 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2055 rxi_PacketsUnWait(void)
2057 if (!rx_waitingForPackets) {
2061 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2062 return; /* still over quota */
2065 rx_waitingForPackets = 0;
2066 #ifdef RX_ENABLE_LOCKS
2067 CV_BROADCAST(&rx_waitingForPackets_cv);
2069 osi_rxWakeup(&rx_waitingForPackets);
2075 /* ------------------Internal interfaces------------------------- */
2077 /* Return this process's service structure for the
2078 * specified socket and service */
2080 rxi_FindService(register osi_socket socket, register u_short serviceId)
2082 register struct rx_service **sp;
2083 for (sp = &rx_services[0]; *sp; sp++) {
2084 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2090 /* Allocate a call structure, for the indicated channel of the
2091 * supplied connection. The mode and state of the call must be set by
2092 * the caller. Returns the call with mutex locked. */
2094 rxi_NewCall(register struct rx_connection *conn, register int channel)
2096 register struct rx_call *call;
2097 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2098 register struct rx_call *cp; /* Call pointer temp */
2099 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2100 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2102 dpf(("rxi_NewCall(conn %x, channel %d)\n", conn, channel));
2104 /* Grab an existing call structure, or allocate a new one.
2105 * Existing call structures are assumed to have been left reset by
2107 MUTEX_ENTER(&rx_freeCallQueue_lock);
2109 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2111 * EXCEPT that the TQ might not yet be cleared out.
2112 * Skip over those with in-use TQs.
2115 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2116 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2122 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2123 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2124 call = queue_First(&rx_freeCallQueue, rx_call);
2125 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2127 MUTEX_ENTER(&rx_stats_mutex);
2128 rx_stats.nFreeCallStructs--;
2129 MUTEX_EXIT(&rx_stats_mutex);
2130 MUTEX_EXIT(&rx_freeCallQueue_lock);
2131 MUTEX_ENTER(&call->lock);
2132 CLEAR_CALL_QUEUE_LOCK(call);
2133 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2134 /* Now, if TQ wasn't cleared earlier, do it now. */
2135 if (call->flags & RX_CALL_TQ_CLEARME) {
2136 rxi_ClearTransmitQueue(call, 0);
2137 queue_Init(&call->tq);
2139 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2140 /* Bind the call to its connection structure */
2142 rxi_ResetCall(call, 1);
2144 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2146 MUTEX_EXIT(&rx_freeCallQueue_lock);
2147 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2148 MUTEX_ENTER(&call->lock);
2149 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2150 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2151 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2153 MUTEX_ENTER(&rx_stats_mutex);
2154 rx_stats.nCallStructs++;
2155 MUTEX_EXIT(&rx_stats_mutex);
2156 /* Initialize once-only items */
2157 queue_Init(&call->tq);
2158 queue_Init(&call->rq);
2159 queue_Init(&call->iovq);
2160 /* Bind the call to its connection structure (prereq for reset) */
2162 rxi_ResetCall(call, 1);
2164 call->channel = channel;
2165 call->callNumber = &conn->callNumber[channel];
2166 /* Note that the next expected call number is retained (in
2167 * conn->callNumber[i]), even if we reallocate the call structure
2169 conn->call[channel] = call;
2170 /* if the channel's never been used (== 0), we should start at 1, otherwise
2171 * the call number is valid from the last time this channel was used */
2172 if (*call->callNumber == 0)
2173 *call->callNumber = 1;
2178 /* A call has been inactive long enough that so we can throw away
2179 * state, including the call structure, which is placed on the call
2181 * Call is locked upon entry.
2182 * haveCTLock set if called from rxi_ReapConnections
2184 #ifdef RX_ENABLE_LOCKS
2186 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2187 #else /* RX_ENABLE_LOCKS */
2189 rxi_FreeCall(register struct rx_call *call)
2190 #endif /* RX_ENABLE_LOCKS */
2192 register int channel = call->channel;
2193 register struct rx_connection *conn = call->conn;
2196 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2197 (*call->callNumber)++;
2198 rxi_ResetCall(call, 0);
2199 call->conn->call[channel] = (struct rx_call *)0;
2201 MUTEX_ENTER(&rx_freeCallQueue_lock);
2202 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2203 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2204 /* A call may be free even though its transmit queue is still in use.
2205 * Since we search the call list from head to tail, put busy calls at
2206 * the head of the list, and idle calls at the tail.
2208 if (call->flags & RX_CALL_TQ_BUSY)
2209 queue_Prepend(&rx_freeCallQueue, call);
2211 queue_Append(&rx_freeCallQueue, call);
2212 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2213 queue_Append(&rx_freeCallQueue, call);
2214 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2215 MUTEX_ENTER(&rx_stats_mutex);
2216 rx_stats.nFreeCallStructs++;
2217 MUTEX_EXIT(&rx_stats_mutex);
2219 MUTEX_EXIT(&rx_freeCallQueue_lock);
2221 /* Destroy the connection if it was previously slated for
2222 * destruction, i.e. the Rx client code previously called
2223 * rx_DestroyConnection (client connections), or
2224 * rxi_ReapConnections called the same routine (server
2225 * connections). Only do this, however, if there are no
2226 * outstanding calls. Note that for fine grain locking, there appears
2227 * to be a deadlock in that rxi_FreeCall has a call locked and
2228 * DestroyConnectionNoLock locks each call in the conn. But note a
2229 * few lines up where we have removed this call from the conn.
2230 * If someone else destroys a connection, they either have no
2231 * call lock held or are going through this section of code.
2233 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2234 MUTEX_ENTER(&conn->conn_data_lock);
2236 MUTEX_EXIT(&conn->conn_data_lock);
2237 #ifdef RX_ENABLE_LOCKS
2239 rxi_DestroyConnectionNoLock(conn);
2241 rxi_DestroyConnection(conn);
2242 #else /* RX_ENABLE_LOCKS */
2243 rxi_DestroyConnection(conn);
2244 #endif /* RX_ENABLE_LOCKS */
2248 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2250 rxi_Alloc(register size_t size)
2254 MUTEX_ENTER(&rx_stats_mutex);
2256 rxi_Allocsize += (afs_int32)size;
2257 MUTEX_EXIT(&rx_stats_mutex);
2259 p = (char *)osi_Alloc(size);
2262 osi_Panic("rxi_Alloc error");
2268 rxi_Free(void *addr, register size_t size)
2270 MUTEX_ENTER(&rx_stats_mutex);
2272 rxi_Allocsize -= (afs_int32)size;
2273 MUTEX_EXIT(&rx_stats_mutex);
2275 osi_Free(addr, size);
2278 /* Find the peer process represented by the supplied (host,port)
2279 * combination. If there is no appropriate active peer structure, a
2280 * new one will be allocated and initialized
2281 * The origPeer, if set, is a pointer to a peer structure on which the
2282 * refcount will be be decremented. This is used to replace the peer
2283 * structure hanging off a connection structure */
2285 rxi_FindPeer(register afs_uint32 host, register u_short port,
2286 struct rx_peer *origPeer, int create)
2288 register struct rx_peer *pp;
2290 hashIndex = PEER_HASH(host, port);
2291 MUTEX_ENTER(&rx_peerHashTable_lock);
2292 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2293 if ((pp->host == host) && (pp->port == port))
2298 pp = rxi_AllocPeer(); /* This bzero's *pp */
2299 pp->host = host; /* set here or in InitPeerParams is zero */
2301 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2302 queue_Init(&pp->congestionQueue);
2303 queue_Init(&pp->rpcStats);
2304 pp->next = rx_peerHashTable[hashIndex];
2305 rx_peerHashTable[hashIndex] = pp;
2306 rxi_InitPeerParams(pp);
2307 MUTEX_ENTER(&rx_stats_mutex);
2308 rx_stats.nPeerStructs++;
2309 MUTEX_EXIT(&rx_stats_mutex);
2316 origPeer->refCount--;
2317 MUTEX_EXIT(&rx_peerHashTable_lock);
2322 /* Find the connection at (host, port) started at epoch, and with the
2323 * given connection id. Creates the server connection if necessary.
2324 * The type specifies whether a client connection or a server
2325 * connection is desired. In both cases, (host, port) specify the
2326 * peer's (host, pair) pair. Client connections are not made
2327 * automatically by this routine. The parameter socket gives the
2328 * socket descriptor on which the packet was received. This is used,
2329 * in the case of server connections, to check that *new* connections
2330 * come via a valid (port, serviceId). Finally, the securityIndex
2331 * parameter must match the existing index for the connection. If a
2332 * server connection is created, it will be created using the supplied
2333 * index, if the index is valid for this service */
2334 struct rx_connection *
2335 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2336 register u_short port, u_short serviceId, afs_uint32 cid,
2337 afs_uint32 epoch, int type, u_int securityIndex)
2339 int hashindex, flag;
2340 register struct rx_connection *conn;
2341 hashindex = CONN_HASH(host, port, cid, epoch, type);
2342 MUTEX_ENTER(&rx_connHashTable_lock);
2343 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2344 rx_connHashTable[hashindex],
2347 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2348 && (epoch == conn->epoch)) {
2349 register struct rx_peer *pp = conn->peer;
2350 if (securityIndex != conn->securityIndex) {
2351 /* this isn't supposed to happen, but someone could forge a packet
2352 * like this, and there seems to be some CM bug that makes this
2353 * happen from time to time -- in which case, the fileserver
2355 MUTEX_EXIT(&rx_connHashTable_lock);
2356 return (struct rx_connection *)0;
2358 if (pp->host == host && pp->port == port)
2360 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2362 /* So what happens when it's a callback connection? */
2363 if ( /*type == RX_CLIENT_CONNECTION && */
2364 (conn->epoch & 0x80000000))
2368 /* the connection rxLastConn that was used the last time is not the
2369 ** one we are looking for now. Hence, start searching in the hash */
2371 conn = rx_connHashTable[hashindex];
2376 struct rx_service *service;
2377 if (type == RX_CLIENT_CONNECTION) {
2378 MUTEX_EXIT(&rx_connHashTable_lock);
2379 return (struct rx_connection *)0;
2381 service = rxi_FindService(socket, serviceId);
2382 if (!service || (securityIndex >= service->nSecurityObjects)
2383 || (service->securityObjects[securityIndex] == 0)) {
2384 MUTEX_EXIT(&rx_connHashTable_lock);
2385 return (struct rx_connection *)0;
2387 conn = rxi_AllocConnection(); /* This bzero's the connection */
2388 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2389 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2390 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2391 conn->next = rx_connHashTable[hashindex];
2392 rx_connHashTable[hashindex] = conn;
2393 conn->peer = rxi_FindPeer(host, port, 0, 1);
2394 conn->type = RX_SERVER_CONNECTION;
2395 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2396 conn->epoch = epoch;
2397 conn->cid = cid & RX_CIDMASK;
2398 /* conn->serial = conn->lastSerial = 0; */
2399 /* conn->timeout = 0; */
2400 conn->ackRate = RX_FAST_ACK_RATE;
2401 conn->service = service;
2402 conn->serviceId = serviceId;
2403 conn->securityIndex = securityIndex;
2404 conn->securityObject = service->securityObjects[securityIndex];
2405 conn->nSpecific = 0;
2406 conn->specific = NULL;
2407 rx_SetConnDeadTime(conn, service->connDeadTime);
2408 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2409 /* Notify security object of the new connection */
2410 RXS_NewConnection(conn->securityObject, conn);
2411 /* XXXX Connection timeout? */
2412 if (service->newConnProc)
2413 (*service->newConnProc) (conn);
2414 MUTEX_ENTER(&rx_stats_mutex);
2415 rx_stats.nServerConns++;
2416 MUTEX_EXIT(&rx_stats_mutex);
2419 MUTEX_ENTER(&conn->conn_data_lock);
2421 MUTEX_EXIT(&conn->conn_data_lock);
2423 rxLastConn = conn; /* store this connection as the last conn used */
2424 MUTEX_EXIT(&rx_connHashTable_lock);
2428 /* There are two packet tracing routines available for testing and monitoring
2429 * Rx. One is called just after every packet is received and the other is
2430 * called just before every packet is sent. Received packets, have had their
2431 * headers decoded, and packets to be sent have not yet had their headers
2432 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2433 * containing the network address. Both can be modified. The return value, if
2434 * non-zero, indicates that the packet should be dropped. */
2436 int (*rx_justReceived) () = 0;
2437 int (*rx_almostSent) () = 0;
2439 /* A packet has been received off the interface. Np is the packet, socket is
2440 * the socket number it was received from (useful in determining which service
2441 * this packet corresponds to), and (host, port) reflect the host,port of the
2442 * sender. This call returns the packet to the caller if it is finished with
2443 * it, rather than de-allocating it, just as a small performance hack */
2446 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2447 afs_uint32 host, u_short port, int *tnop,
2448 struct rx_call **newcallp)
2450 register struct rx_call *call;
2451 register struct rx_connection *conn;
2453 afs_uint32 currentCallNumber;
2459 struct rx_packet *tnp;
2462 /* We don't print out the packet until now because (1) the time may not be
2463 * accurate enough until now in the lwp implementation (rx_Listener only gets
2464 * the time after the packet is read) and (2) from a protocol point of view,
2465 * this is the first time the packet has been seen */
2466 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2467 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2468 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2469 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2470 np->header.epoch, np->header.cid, np->header.callNumber,
2471 np->header.seq, np->header.flags, np));
2474 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2475 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2478 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2479 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2482 /* If an input tracer function is defined, call it with the packet and
2483 * network address. Note this function may modify its arguments. */
2484 if (rx_justReceived) {
2485 struct sockaddr_in addr;
2487 addr.sin_family = AF_INET;
2488 addr.sin_port = port;
2489 addr.sin_addr.s_addr = host;
2490 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2491 addr.sin_len = sizeof(addr);
2492 #endif /* AFS_OSF_ENV */
2493 drop = (*rx_justReceived) (np, &addr);
2494 /* drop packet if return value is non-zero */
2497 port = addr.sin_port; /* in case fcn changed addr */
2498 host = addr.sin_addr.s_addr;
2502 /* If packet was not sent by the client, then *we* must be the client */
2503 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2504 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2506 /* Find the connection (or fabricate one, if we're the server & if
2507 * necessary) associated with this packet */
2509 rxi_FindConnection(socket, host, port, np->header.serviceId,
2510 np->header.cid, np->header.epoch, type,
2511 np->header.securityIndex);
2514 /* If no connection found or fabricated, just ignore the packet.
2515 * (An argument could be made for sending an abort packet for
2520 MUTEX_ENTER(&conn->conn_data_lock);
2521 if (conn->maxSerial < np->header.serial)
2522 conn->maxSerial = np->header.serial;
2523 MUTEX_EXIT(&conn->conn_data_lock);
2525 /* If the connection is in an error state, send an abort packet and ignore
2526 * the incoming packet */
2528 /* Don't respond to an abort packet--we don't want loops! */
2529 MUTEX_ENTER(&conn->conn_data_lock);
2530 if (np->header.type != RX_PACKET_TYPE_ABORT)
2531 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2533 MUTEX_EXIT(&conn->conn_data_lock);
2537 /* Check for connection-only requests (i.e. not call specific). */
2538 if (np->header.callNumber == 0) {
2539 switch (np->header.type) {
2540 case RX_PACKET_TYPE_ABORT: {
2541 /* What if the supplied error is zero? */
2542 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2543 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2544 rxi_ConnectionError(conn, errcode);
2545 MUTEX_ENTER(&conn->conn_data_lock);
2547 MUTEX_EXIT(&conn->conn_data_lock);
2550 case RX_PACKET_TYPE_CHALLENGE:
2551 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2552 MUTEX_ENTER(&conn->conn_data_lock);
2554 MUTEX_EXIT(&conn->conn_data_lock);
2556 case RX_PACKET_TYPE_RESPONSE:
2557 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2558 MUTEX_ENTER(&conn->conn_data_lock);
2560 MUTEX_EXIT(&conn->conn_data_lock);
2562 case RX_PACKET_TYPE_PARAMS:
2563 case RX_PACKET_TYPE_PARAMS + 1:
2564 case RX_PACKET_TYPE_PARAMS + 2:
2565 /* ignore these packet types for now */
2566 MUTEX_ENTER(&conn->conn_data_lock);
2568 MUTEX_EXIT(&conn->conn_data_lock);
2573 /* Should not reach here, unless the peer is broken: send an
2575 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2576 MUTEX_ENTER(&conn->conn_data_lock);
2577 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2579 MUTEX_EXIT(&conn->conn_data_lock);
2584 channel = np->header.cid & RX_CHANNELMASK;
2585 call = conn->call[channel];
2586 #ifdef RX_ENABLE_LOCKS
2588 MUTEX_ENTER(&call->lock);
2589 /* Test to see if call struct is still attached to conn. */
2590 if (call != conn->call[channel]) {
2592 MUTEX_EXIT(&call->lock);
2593 if (type == RX_SERVER_CONNECTION) {
2594 call = conn->call[channel];
2595 /* If we started with no call attached and there is one now,
2596 * another thread is also running this routine and has gotten
2597 * the connection channel. We should drop this packet in the tests
2598 * below. If there was a call on this connection and it's now
2599 * gone, then we'll be making a new call below.
2600 * If there was previously a call and it's now different then
2601 * the old call was freed and another thread running this routine
2602 * has created a call on this channel. One of these two threads
2603 * has a packet for the old call and the code below handles those
2607 MUTEX_ENTER(&call->lock);
2609 /* This packet can't be for this call. If the new call address is
2610 * 0 then no call is running on this channel. If there is a call
2611 * then, since this is a client connection we're getting data for
2612 * it must be for the previous call.
2614 MUTEX_ENTER(&rx_stats_mutex);
2615 rx_stats.spuriousPacketsRead++;
2616 MUTEX_EXIT(&rx_stats_mutex);
2617 MUTEX_ENTER(&conn->conn_data_lock);
2619 MUTEX_EXIT(&conn->conn_data_lock);
2624 currentCallNumber = conn->callNumber[channel];
2626 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2627 if (np->header.callNumber < currentCallNumber) {
2628 MUTEX_ENTER(&rx_stats_mutex);
2629 rx_stats.spuriousPacketsRead++;
2630 MUTEX_EXIT(&rx_stats_mutex);
2631 #ifdef RX_ENABLE_LOCKS
2633 MUTEX_EXIT(&call->lock);
2635 MUTEX_ENTER(&conn->conn_data_lock);
2637 MUTEX_EXIT(&conn->conn_data_lock);
2641 MUTEX_ENTER(&conn->conn_call_lock);
2642 call = rxi_NewCall(conn, channel);
2643 MUTEX_EXIT(&conn->conn_call_lock);
2644 *call->callNumber = np->header.callNumber;
2645 if (np->header.callNumber == 0)
2646 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));
2648 call->state = RX_STATE_PRECALL;
2649 clock_GetTime(&call->queueTime);
2650 hzero(call->bytesSent);
2651 hzero(call->bytesRcvd);
2653 * If the number of queued calls exceeds the overload
2654 * threshold then abort this call.
2656 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2657 struct rx_packet *tp;
2659 rxi_CallError(call, rx_BusyError);
2660 tp = rxi_SendCallAbort(call, np, 1, 0);
2661 MUTEX_EXIT(&call->lock);
2662 MUTEX_ENTER(&conn->conn_data_lock);
2664 MUTEX_EXIT(&conn->conn_data_lock);
2665 MUTEX_ENTER(&rx_stats_mutex);
2667 MUTEX_EXIT(&rx_stats_mutex);
2670 rxi_KeepAliveOn(call);
2671 } else if (np->header.callNumber != currentCallNumber) {
2672 /* Wait until the transmit queue is idle before deciding
2673 * whether to reset the current call. Chances are that the
2674 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2677 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2678 while ((call->state == RX_STATE_ACTIVE)
2679 && (call->flags & RX_CALL_TQ_BUSY)) {
2680 call->flags |= RX_CALL_TQ_WAIT;
2682 #ifdef RX_ENABLE_LOCKS
2683 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2684 CV_WAIT(&call->cv_tq, &call->lock);
2685 #else /* RX_ENABLE_LOCKS */
2686 osi_rxSleep(&call->tq);
2687 #endif /* RX_ENABLE_LOCKS */
2689 if (call->tqWaiters == 0)
2690 call->flags &= ~RX_CALL_TQ_WAIT;
2692 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2693 /* If the new call cannot be taken right now send a busy and set
2694 * the error condition in this call, so that it terminates as
2695 * quickly as possible */
2696 if (call->state == RX_STATE_ACTIVE) {
2697 struct rx_packet *tp;
2699 rxi_CallError(call, RX_CALL_DEAD);
2700 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2702 MUTEX_EXIT(&call->lock);
2703 MUTEX_ENTER(&conn->conn_data_lock);
2705 MUTEX_EXIT(&conn->conn_data_lock);
2708 rxi_ResetCall(call, 0);
2709 *call->callNumber = np->header.callNumber;
2710 if (np->header.callNumber == 0)
2711 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));
2713 call->state = RX_STATE_PRECALL;
2714 clock_GetTime(&call->queueTime);
2715 hzero(call->bytesSent);
2716 hzero(call->bytesRcvd);
2718 * If the number of queued calls exceeds the overload
2719 * threshold then abort this call.
2721 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2722 struct rx_packet *tp;
2724 rxi_CallError(call, rx_BusyError);
2725 tp = rxi_SendCallAbort(call, np, 1, 0);
2726 MUTEX_EXIT(&call->lock);
2727 MUTEX_ENTER(&conn->conn_data_lock);
2729 MUTEX_EXIT(&conn->conn_data_lock);
2730 MUTEX_ENTER(&rx_stats_mutex);
2732 MUTEX_EXIT(&rx_stats_mutex);
2735 rxi_KeepAliveOn(call);
2737 /* Continuing call; do nothing here. */
2739 } else { /* we're the client */
2740 /* Ignore all incoming acknowledgements for calls in DALLY state */
2741 if (call && (call->state == RX_STATE_DALLY)
2742 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2743 MUTEX_ENTER(&rx_stats_mutex);
2744 rx_stats.ignorePacketDally++;
2745 MUTEX_EXIT(&rx_stats_mutex);
2746 #ifdef RX_ENABLE_LOCKS
2748 MUTEX_EXIT(&call->lock);
2751 MUTEX_ENTER(&conn->conn_data_lock);
2753 MUTEX_EXIT(&conn->conn_data_lock);
2757 /* Ignore anything that's not relevant to the current call. If there
2758 * isn't a current call, then no packet is relevant. */
2759 if (!call || (np->header.callNumber != currentCallNumber)) {
2760 MUTEX_ENTER(&rx_stats_mutex);
2761 rx_stats.spuriousPacketsRead++;
2762 MUTEX_EXIT(&rx_stats_mutex);
2763 #ifdef RX_ENABLE_LOCKS
2765 MUTEX_EXIT(&call->lock);
2768 MUTEX_ENTER(&conn->conn_data_lock);
2770 MUTEX_EXIT(&conn->conn_data_lock);
2773 /* If the service security object index stamped in the packet does not
2774 * match the connection's security index, ignore the packet */
2775 if (np->header.securityIndex != conn->securityIndex) {
2776 #ifdef RX_ENABLE_LOCKS
2777 MUTEX_EXIT(&call->lock);
2779 MUTEX_ENTER(&conn->conn_data_lock);
2781 MUTEX_EXIT(&conn->conn_data_lock);
2785 /* If we're receiving the response, then all transmit packets are
2786 * implicitly acknowledged. Get rid of them. */
2787 if (np->header.type == RX_PACKET_TYPE_DATA) {
2788 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2789 /* XXX Hack. Because we must release the global rx lock when
2790 * sending packets (osi_NetSend) we drop all acks while we're
2791 * traversing the tq in rxi_Start sending packets out because
2792 * packets may move to the freePacketQueue as result of being here!
2793 * So we drop these packets until we're safely out of the
2794 * traversing. Really ugly!
2795 * For fine grain RX locking, we set the acked field in the
2796 * packets and let rxi_Start remove them from the transmit queue.
2798 if (call->flags & RX_CALL_TQ_BUSY) {
2799 #ifdef RX_ENABLE_LOCKS
2800 rxi_SetAcksInTransmitQueue(call);
2803 return np; /* xmitting; drop packet */
2806 rxi_ClearTransmitQueue(call, 0);
2808 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2809 rxi_ClearTransmitQueue(call, 0);
2810 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2812 if (np->header.type == RX_PACKET_TYPE_ACK) {
2813 /* now check to see if this is an ack packet acknowledging that the
2814 * server actually *lost* some hard-acked data. If this happens we
2815 * ignore this packet, as it may indicate that the server restarted in
2816 * the middle of a call. It is also possible that this is an old ack
2817 * packet. We don't abort the connection in this case, because this
2818 * *might* just be an old ack packet. The right way to detect a server
2819 * restart in the midst of a call is to notice that the server epoch
2821 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2822 * XXX unacknowledged. I think that this is off-by-one, but
2823 * XXX I don't dare change it just yet, since it will
2824 * XXX interact badly with the server-restart detection
2825 * XXX code in receiveackpacket. */
2826 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2827 MUTEX_ENTER(&rx_stats_mutex);
2828 rx_stats.spuriousPacketsRead++;
2829 MUTEX_EXIT(&rx_stats_mutex);
2830 MUTEX_EXIT(&call->lock);
2831 MUTEX_ENTER(&conn->conn_data_lock);
2833 MUTEX_EXIT(&conn->conn_data_lock);
2837 } /* else not a data packet */
2840 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2841 /* Set remote user defined status from packet */
2842 call->remoteStatus = np->header.userStatus;
2844 /* Note the gap between the expected next packet and the actual
2845 * packet that arrived, when the new packet has a smaller serial number
2846 * than expected. Rioses frequently reorder packets all by themselves,
2847 * so this will be quite important with very large window sizes.
2848 * Skew is checked against 0 here to avoid any dependence on the type of
2849 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2851 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2852 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2853 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2855 MUTEX_ENTER(&conn->conn_data_lock);
2856 skew = conn->lastSerial - np->header.serial;
2857 conn->lastSerial = np->header.serial;
2858 MUTEX_EXIT(&conn->conn_data_lock);
2860 register struct rx_peer *peer;
2862 if (skew > peer->inPacketSkew) {
2863 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2865 peer->inPacketSkew = skew;
2869 /* Now do packet type-specific processing */
2870 switch (np->header.type) {
2871 case RX_PACKET_TYPE_DATA:
2872 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2875 case RX_PACKET_TYPE_ACK:
2876 /* Respond immediately to ack packets requesting acknowledgement
2878 if (np->header.flags & RX_REQUEST_ACK) {
2880 (void)rxi_SendCallAbort(call, 0, 1, 0);
2882 (void)rxi_SendAck(call, 0, np->header.serial,
2883 RX_ACK_PING_RESPONSE, 1);
2885 np = rxi_ReceiveAckPacket(call, np, 1);
2887 case RX_PACKET_TYPE_ABORT: {
2888 /* An abort packet: reset the call, passing the error up to the user. */
2889 /* What if error is zero? */
2890 /* What if the error is -1? the application will treat it as a timeout. */
2891 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
2892 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
2893 rxi_CallError(call, errdata);
2894 MUTEX_EXIT(&call->lock);
2895 MUTEX_ENTER(&conn->conn_data_lock);
2897 MUTEX_EXIT(&conn->conn_data_lock);
2898 return np; /* xmitting; drop packet */
2900 case RX_PACKET_TYPE_BUSY:
2903 case RX_PACKET_TYPE_ACKALL:
2904 /* All packets acknowledged, so we can drop all packets previously
2905 * readied for sending */
2906 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2907 /* XXX Hack. We because we can't release the global rx lock when
2908 * sending packets (osi_NetSend) we drop all ack pkts while we're
2909 * traversing the tq in rxi_Start sending packets out because
2910 * packets may move to the freePacketQueue as result of being
2911 * here! So we drop these packets until we're safely out of the
2912 * traversing. Really ugly!
2913 * For fine grain RX locking, we set the acked field in the packets
2914 * and let rxi_Start remove the packets from the transmit queue.
2916 if (call->flags & RX_CALL_TQ_BUSY) {
2917 #ifdef RX_ENABLE_LOCKS
2918 rxi_SetAcksInTransmitQueue(call);
2920 #else /* RX_ENABLE_LOCKS */
2921 MUTEX_EXIT(&call->lock);
2922 MUTEX_ENTER(&conn->conn_data_lock);
2924 MUTEX_EXIT(&conn->conn_data_lock);
2925 return np; /* xmitting; drop packet */
2926 #endif /* RX_ENABLE_LOCKS */
2928 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2929 rxi_ClearTransmitQueue(call, 0);
2932 /* Should not reach here, unless the peer is broken: send an abort
2934 rxi_CallError(call, RX_PROTOCOL_ERROR);
2935 np = rxi_SendCallAbort(call, np, 1, 0);
2938 /* Note when this last legitimate packet was received, for keep-alive
2939 * processing. Note, we delay getting the time until now in the hope that
2940 * the packet will be delivered to the user before any get time is required
2941 * (if not, then the time won't actually be re-evaluated here). */
2942 call->lastReceiveTime = clock_Sec();
2943 MUTEX_EXIT(&call->lock);
2944 MUTEX_ENTER(&conn->conn_data_lock);
2946 MUTEX_EXIT(&conn->conn_data_lock);
2950 /* return true if this is an "interesting" connection from the point of view
2951 of someone trying to debug the system */
2953 rxi_IsConnInteresting(struct rx_connection *aconn)
2956 register struct rx_call *tcall;
2958 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2960 for (i = 0; i < RX_MAXCALLS; i++) {
2961 tcall = aconn->call[i];
2963 if ((tcall->state == RX_STATE_PRECALL)
2964 || (tcall->state == RX_STATE_ACTIVE))
2966 if ((tcall->mode == RX_MODE_SENDING)
2967 || (tcall->mode == RX_MODE_RECEIVING))
2975 /* if this is one of the last few packets AND it wouldn't be used by the
2976 receiving call to immediately satisfy a read request, then drop it on
2977 the floor, since accepting it might prevent a lock-holding thread from
2978 making progress in its reading. If a call has been cleared while in
2979 the precall state then ignore all subsequent packets until the call
2980 is assigned to a thread. */
2983 TooLow(struct rx_packet *ap, struct rx_call *acall)
2986 MUTEX_ENTER(&rx_stats_mutex);
2987 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
2988 && (acall->state == RX_STATE_PRECALL))
2989 || ((rx_nFreePackets < rxi_dataQuota + 2)
2990 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
2991 && (acall->flags & RX_CALL_READER_WAIT)))) {
2994 MUTEX_EXIT(&rx_stats_mutex);
3000 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
3001 struct rx_call *acall)
3003 struct rx_call *call = acall;
3007 MUTEX_ENTER(&conn->conn_data_lock);
3008 conn->checkReachEvent = NULL;
3009 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3012 MUTEX_EXIT(&conn->conn_data_lock);
3016 MUTEX_ENTER(&conn->conn_call_lock);
3017 MUTEX_ENTER(&conn->conn_data_lock);
3018 for (i = 0; i < RX_MAXCALLS; i++) {
3019 struct rx_call *tc = conn->call[i];
3020 if (tc && tc->state == RX_STATE_PRECALL) {
3026 /* Indicate that rxi_CheckReachEvent is no longer running by
3027 * clearing the flag. Must be atomic under conn_data_lock to
3028 * avoid a new call slipping by: rxi_CheckConnReach holds
3029 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3031 conn->flags &= ~RX_CONN_ATTACHWAIT;
3032 MUTEX_EXIT(&conn->conn_data_lock);
3033 MUTEX_EXIT(&conn->conn_call_lock);
3038 MUTEX_ENTER(&call->lock);
3039 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3041 MUTEX_EXIT(&call->lock);
3043 clock_GetTime(&when);
3044 when.sec += RX_CHECKREACH_TIMEOUT;
3045 MUTEX_ENTER(&conn->conn_data_lock);
3046 if (!conn->checkReachEvent) {
3048 conn->checkReachEvent =
3049 rxevent_Post(&when, rxi_CheckReachEvent, conn, NULL);
3051 MUTEX_EXIT(&conn->conn_data_lock);
3057 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3059 struct rx_service *service = conn->service;
3060 struct rx_peer *peer = conn->peer;
3061 afs_uint32 now, lastReach;
3063 if (service->checkReach == 0)
3067 MUTEX_ENTER(&peer->peer_lock);
3068 lastReach = peer->lastReachTime;
3069 MUTEX_EXIT(&peer->peer_lock);
3070 if (now - lastReach < RX_CHECKREACH_TTL)
3073 MUTEX_ENTER(&conn->conn_data_lock);
3074 if (conn->flags & RX_CONN_ATTACHWAIT) {
3075 MUTEX_EXIT(&conn->conn_data_lock);
3078 conn->flags |= RX_CONN_ATTACHWAIT;
3079 MUTEX_EXIT(&conn->conn_data_lock);
3080 if (!conn->checkReachEvent)
3081 rxi_CheckReachEvent(NULL, conn, call);
3086 /* try to attach call, if authentication is complete */
3088 TryAttach(register struct rx_call *acall, register osi_socket socket,
3089 register int *tnop, register struct rx_call **newcallp,
3092 struct rx_connection *conn = acall->conn;
3094 if (conn->type == RX_SERVER_CONNECTION
3095 && acall->state == RX_STATE_PRECALL) {
3096 /* Don't attach until we have any req'd. authentication. */
3097 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3098 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3099 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3100 /* Note: this does not necessarily succeed; there
3101 * may not any proc available
3104 rxi_ChallengeOn(acall->conn);
3109 /* A data packet has been received off the interface. This packet is
3110 * appropriate to the call (the call is in the right state, etc.). This
3111 * routine can return a packet to the caller, for re-use */
3114 rxi_ReceiveDataPacket(register struct rx_call *call,
3115 register struct rx_packet *np, int istack,
3116 osi_socket socket, afs_uint32 host, u_short port,
3117 int *tnop, struct rx_call **newcallp)
3119 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3123 afs_uint32 seq, serial, flags;
3125 struct rx_packet *tnp;
3127 MUTEX_ENTER(&rx_stats_mutex);
3128 rx_stats.dataPacketsRead++;
3129 MUTEX_EXIT(&rx_stats_mutex);
3132 /* If there are no packet buffers, drop this new packet, unless we can find
3133 * packet buffers from inactive calls */
3135 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3136 MUTEX_ENTER(&rx_freePktQ_lock);
3137 rxi_NeedMorePackets = TRUE;
3138 MUTEX_EXIT(&rx_freePktQ_lock);
3139 MUTEX_ENTER(&rx_stats_mutex);
3140 rx_stats.noPacketBuffersOnRead++;
3141 MUTEX_EXIT(&rx_stats_mutex);
3142 call->rprev = np->header.serial;
3143 rxi_calltrace(RX_TRACE_DROP, call);
3144 dpf(("packet %x dropped on receipt - quota problems", np));
3146 rxi_ClearReceiveQueue(call);
3147 clock_GetTime(&when);
3148 clock_Add(&when, &rx_softAckDelay);
3149 if (!call->delayedAckEvent
3150 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3151 rxevent_Cancel(call->delayedAckEvent, call,
3152 RX_CALL_REFCOUNT_DELAY);
3153 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3154 call->delayedAckEvent =
3155 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3157 /* we've damaged this call already, might as well do it in. */
3163 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3164 * packet is one of several packets transmitted as a single
3165 * datagram. Do not send any soft or hard acks until all packets
3166 * in a jumbogram have been processed. Send negative acks right away.
3168 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3169 /* tnp is non-null when there are more packets in the
3170 * current jumbo gram */
3177 seq = np->header.seq;
3178 serial = np->header.serial;
3179 flags = np->header.flags;
3181 /* If the call is in an error state, send an abort message */
3183 return rxi_SendCallAbort(call, np, istack, 0);
3185 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3186 * AFS 3.5 jumbogram. */
3187 if (flags & RX_JUMBO_PACKET) {
3188 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3193 if (np->header.spare != 0) {
3194 MUTEX_ENTER(&call->conn->conn_data_lock);
3195 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3196 MUTEX_EXIT(&call->conn->conn_data_lock);
3199 /* The usual case is that this is the expected next packet */
3200 if (seq == call->rnext) {
3202 /* Check to make sure it is not a duplicate of one already queued */
3203 if (queue_IsNotEmpty(&call->rq)
3204 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3205 MUTEX_ENTER(&rx_stats_mutex);
3206 rx_stats.dupPacketsRead++;
3207 MUTEX_EXIT(&rx_stats_mutex);
3208 dpf(("packet %x dropped on receipt - duplicate", np));
3209 rxevent_Cancel(call->delayedAckEvent, call,
3210 RX_CALL_REFCOUNT_DELAY);
3211 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3217 /* It's the next packet. Stick it on the receive queue
3218 * for this call. Set newPackets to make sure we wake
3219 * the reader once all packets have been processed */
3220 queue_Prepend(&call->rq, np);
3222 np = NULL; /* We can't use this anymore */
3225 /* If an ack is requested then set a flag to make sure we
3226 * send an acknowledgement for this packet */
3227 if (flags & RX_REQUEST_ACK) {
3228 ackNeeded = RX_ACK_REQUESTED;
3231 /* Keep track of whether we have received the last packet */
3232 if (flags & RX_LAST_PACKET) {
3233 call->flags |= RX_CALL_HAVE_LAST;
3237 /* Check whether we have all of the packets for this call */
3238 if (call->flags & RX_CALL_HAVE_LAST) {
3239 afs_uint32 tseq; /* temporary sequence number */
3240 struct rx_packet *tp; /* Temporary packet pointer */
3241 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3243 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3244 if (tseq != tp->header.seq)
3246 if (tp->header.flags & RX_LAST_PACKET) {
3247 call->flags |= RX_CALL_RECEIVE_DONE;
3254 /* Provide asynchronous notification for those who want it
3255 * (e.g. multi rx) */
3256 if (call->arrivalProc) {
3257 (*call->arrivalProc) (call, call->arrivalProcHandle,
3258 call->arrivalProcArg);
3259 call->arrivalProc = (void (*)())0;
3262 /* Update last packet received */
3265 /* If there is no server process serving this call, grab
3266 * one, if available. We only need to do this once. If a
3267 * server thread is available, this thread becomes a server
3268 * thread and the server thread becomes a listener thread. */
3270 TryAttach(call, socket, tnop, newcallp, 0);
3273 /* This is not the expected next packet. */
3275 /* Determine whether this is a new or old packet, and if it's
3276 * a new one, whether it fits into the current receive window.
3277 * Also figure out whether the packet was delivered in sequence.
3278 * We use the prev variable to determine whether the new packet
3279 * is the successor of its immediate predecessor in the
3280 * receive queue, and the missing flag to determine whether
3281 * any of this packets predecessors are missing. */
3283 afs_uint32 prev; /* "Previous packet" sequence number */
3284 struct rx_packet *tp; /* Temporary packet pointer */
3285 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3286 int missing; /* Are any predecessors missing? */
3288 /* If the new packet's sequence number has been sent to the
3289 * application already, then this is a duplicate */
3290 if (seq < call->rnext) {
3291 MUTEX_ENTER(&rx_stats_mutex);
3292 rx_stats.dupPacketsRead++;
3293 MUTEX_EXIT(&rx_stats_mutex);
3294 rxevent_Cancel(call->delayedAckEvent, call,
3295 RX_CALL_REFCOUNT_DELAY);
3296 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3302 /* If the sequence number is greater than what can be
3303 * accomodated by the current window, then send a negative
3304 * acknowledge and drop the packet */
3305 if ((call->rnext + call->rwind) <= seq) {
3306 rxevent_Cancel(call->delayedAckEvent, call,
3307 RX_CALL_REFCOUNT_DELAY);
3308 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3315 /* Look for the packet in the queue of old received packets */
3316 for (prev = call->rnext - 1, missing =
3317 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3318 /*Check for duplicate packet */
3319 if (seq == tp->header.seq) {
3320 MUTEX_ENTER(&rx_stats_mutex);
3321 rx_stats.dupPacketsRead++;
3322 MUTEX_EXIT(&rx_stats_mutex);
3323 rxevent_Cancel(call->delayedAckEvent, call,
3324 RX_CALL_REFCOUNT_DELAY);
3325 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3331 /* If we find a higher sequence packet, break out and
3332 * insert the new packet here. */
3333 if (seq < tp->header.seq)
3335 /* Check for missing packet */
3336 if (tp->header.seq != prev + 1) {
3340 prev = tp->header.seq;
3343 /* Keep track of whether we have received the last packet. */
3344 if (flags & RX_LAST_PACKET) {
3345 call->flags |= RX_CALL_HAVE_LAST;
3348 /* It's within the window: add it to the the receive queue.
3349 * tp is left by the previous loop either pointing at the
3350 * packet before which to insert the new packet, or at the
3351 * queue head if the queue is empty or the packet should be
3353 queue_InsertBefore(tp, np);
3357 /* Check whether we have all of the packets for this call */
3358 if ((call->flags & RX_CALL_HAVE_LAST)
3359 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3360 afs_uint32 tseq; /* temporary sequence number */
3363 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3364 if (tseq != tp->header.seq)
3366 if (tp->header.flags & RX_LAST_PACKET) {
3367 call->flags |= RX_CALL_RECEIVE_DONE;
3374 /* We need to send an ack of the packet is out of sequence,
3375 * or if an ack was requested by the peer. */
3376 if (seq != prev + 1 || missing) {
3377 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3378 } else if (flags & RX_REQUEST_ACK) {
3379 ackNeeded = RX_ACK_REQUESTED;
3382 /* Acknowledge the last packet for each call */
3383 if (flags & RX_LAST_PACKET) {
3394 * If the receiver is waiting for an iovec, fill the iovec
3395 * using the data from the receive queue */
3396 if (call->flags & RX_CALL_IOVEC_WAIT) {
3397 didHardAck = rxi_FillReadVec(call, serial);
3398 /* the call may have been aborted */
3407 /* Wakeup the reader if any */
3408 if ((call->flags & RX_CALL_READER_WAIT)
3409 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3410 || (call->iovNext >= call->iovMax)
3411 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3412 call->flags &= ~RX_CALL_READER_WAIT;
3413 #ifdef RX_ENABLE_LOCKS
3414 CV_BROADCAST(&call->cv_rq);
3416 osi_rxWakeup(&call->rq);
3422 * Send an ack when requested by the peer, or once every
3423 * rxi_SoftAckRate packets until the last packet has been
3424 * received. Always send a soft ack for the last packet in
3425 * the server's reply. */
3427 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3428 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3429 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3430 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3431 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3432 } else if (call->nSoftAcks) {
3433 clock_GetTime(&when);
3434 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3435 clock_Add(&when, &rx_lastAckDelay);
3437 clock_Add(&when, &rx_softAckDelay);
3439 if (!call->delayedAckEvent
3440 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3441 rxevent_Cancel(call->delayedAckEvent, call,
3442 RX_CALL_REFCOUNT_DELAY);
3443 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3444 call->delayedAckEvent =
3445 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3447 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3448 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3455 static void rxi_ComputeRate();
3459 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3461 struct rx_peer *peer = conn->peer;
3463 MUTEX_ENTER(&peer->peer_lock);
3464 peer->lastReachTime = clock_Sec();
3465 MUTEX_EXIT(&peer->peer_lock);
3467 MUTEX_ENTER(&conn->conn_data_lock);
3468 if (conn->flags & RX_CONN_ATTACHWAIT) {
3471 conn->flags &= ~RX_CONN_ATTACHWAIT;
3472 MUTEX_EXIT(&conn->conn_data_lock);
3474 for (i = 0; i < RX_MAXCALLS; i++) {
3475 struct rx_call *call = conn->call[i];
3478 MUTEX_ENTER(&call->lock);
3479 /* tnop can be null if newcallp is null */
3480 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3482 MUTEX_EXIT(&call->lock);
3486 MUTEX_EXIT(&conn->conn_data_lock);
3490 rx_ack_reason(int reason)
3493 case RX_ACK_REQUESTED:
3495 case RX_ACK_DUPLICATE:
3497 case RX_ACK_OUT_OF_SEQUENCE:
3499 case RX_ACK_EXCEEDS_WINDOW:
3501 case RX_ACK_NOSPACE:
3505 case RX_ACK_PING_RESPONSE:
3517 /* rxi_ComputePeerNetStats
3519 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3520 * estimates (like RTT and throughput) based on ack packets. Caller
3521 * must ensure that the packet in question is the right one (i.e.
3522 * serial number matches).
3525 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3526 struct rx_ackPacket *ap, struct rx_packet *np)
3528 struct rx_peer *peer = call->conn->peer;
3530 /* Use RTT if not delayed by client. */
3531 if (ap->reason != RX_ACK_DELAY)
3532 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3534 rxi_ComputeRate(peer, call, p, np, ap->reason);
3538 /* The real smarts of the whole thing. */
3540 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3543 struct rx_ackPacket *ap;
3545 register struct rx_packet *tp;
3546 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3547 register struct rx_connection *conn = call->conn;
3548 struct rx_peer *peer = conn->peer;
3551 /* because there are CM's that are bogus, sending weird values for this. */
3552 afs_uint32 skew = 0;
3557 int newAckCount = 0;
3558 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3559 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3561 MUTEX_ENTER(&rx_stats_mutex);
3562 rx_stats.ackPacketsRead++;
3563 MUTEX_EXIT(&rx_stats_mutex);
3564 ap = (struct rx_ackPacket *)rx_DataOf(np);
3565 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3567 return np; /* truncated ack packet */
3569 /* depends on ack packet struct */
3570 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3571 first = ntohl(ap->firstPacket);
3572 serial = ntohl(ap->serial);
3573 /* temporarily disabled -- needs to degrade over time
3574 * skew = ntohs(ap->maxSkew); */
3576 /* Ignore ack packets received out of order */
3577 if (first < call->tfirst) {
3581 if (np->header.flags & RX_SLOW_START_OK) {
3582 call->flags |= RX_CALL_SLOW_START_OK;
3585 if (ap->reason == RX_ACK_PING_RESPONSE)
3586 rxi_UpdatePeerReach(conn, call);
3590 if (rxdebug_active) {
3594 len = _snprintf(msg, sizeof(msg),
3595 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3596 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3597 ntohl(ap->serial), ntohl(ap->previousPacket),
3598 (unsigned int)np->header.seq, (unsigned int)skew,
3599 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3603 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3604 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3608 OutputDebugString(msg);
3610 #else /* AFS_NT40_ENV */
3613 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3614 ap->reason, ntohl(ap->previousPacket),
3615 (unsigned int)np->header.seq, (unsigned int)serial,
3616 (unsigned int)skew, ntohl(ap->firstPacket));
3619 for (offset = 0; offset < nAcks; offset++)
3620 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3625 #endif /* AFS_NT40_ENV */
3628 /* Update the outgoing packet skew value to the latest value of
3629 * the peer's incoming packet skew value. The ack packet, of
3630 * course, could arrive out of order, but that won't affect things
3632 MUTEX_ENTER(&peer->peer_lock);
3633 peer->outPacketSkew = skew;
3635 /* Check for packets that no longer need to be transmitted, and
3636 * discard them. This only applies to packets positively
3637 * acknowledged as having been sent to the peer's upper level.
3638 * All other packets must be retained. So only packets with
3639 * sequence numbers < ap->firstPacket are candidates. */
3640 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3641 if (tp->header.seq >= first)
3643 call->tfirst = tp->header.seq + 1;
3645 && (tp->header.serial == serial || tp->firstSerial == serial))
3646 rxi_ComputePeerNetStats(call, tp, ap, np);
3647 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3650 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3651 /* XXX Hack. Because we have to release the global rx lock when sending
3652 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3653 * in rxi_Start sending packets out because packets may move to the
3654 * freePacketQueue as result of being here! So we drop these packets until
3655 * we're safely out of the traversing. Really ugly!
3656 * To make it even uglier, if we're using fine grain locking, we can
3657 * set the ack bits in the packets and have rxi_Start remove the packets
3658 * when it's done transmitting.
3660 if (call->flags & RX_CALL_TQ_BUSY) {
3661 #ifdef RX_ENABLE_LOCKS
3662 tp->flags |= RX_PKTFLAG_ACKED;
3663 call->flags |= RX_CALL_TQ_SOME_ACKED;
3664 #else /* RX_ENABLE_LOCKS */
3666 #endif /* RX_ENABLE_LOCKS */
3668 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3671 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3676 /* Give rate detector a chance to respond to ping requests */
3677 if (ap->reason == RX_ACK_PING_RESPONSE) {
3678 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3682 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3684 /* Now go through explicit acks/nacks and record the results in
3685 * the waiting packets. These are packets that can't be released
3686 * yet, even with a positive acknowledge. This positive
3687 * acknowledge only means the packet has been received by the
3688 * peer, not that it will be retained long enough to be sent to
3689 * the peer's upper level. In addition, reset the transmit timers
3690 * of any missing packets (those packets that must be missing
3691 * because this packet was out of sequence) */
3693 call->nSoftAcked = 0;
3694 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3695 /* Update round trip time if the ack was stimulated on receipt
3697 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3698 #ifdef RX_ENABLE_LOCKS
3699 if (tp->header.seq >= first)
3700 #endif /* RX_ENABLE_LOCKS */
3701 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3703 && (tp->header.serial == serial || tp->firstSerial == serial))
3704 rxi_ComputePeerNetStats(call, tp, ap, np);
3706 /* Set the acknowledge flag per packet based on the
3707 * information in the ack packet. An acknowlegded packet can
3708 * be downgraded when the server has discarded a packet it
3709 * soacked previously, or when an ack packet is received
3710 * out of sequence. */
3711 if (tp->header.seq < first) {
3712 /* Implicit ack information */
3713 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3716 tp->flags |= RX_PKTFLAG_ACKED;
3717 } else if (tp->header.seq < first + nAcks) {
3718 /* Explicit ack information: set it in the packet appropriately */
3719 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3720 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3722 tp->flags |= RX_PKTFLAG_ACKED;
3729 } else /* RX_ACK_TYPE_NACK */ {
3730 tp->flags &= ~RX_PKTFLAG_ACKED;
3734 tp->flags &= ~RX_PKTFLAG_ACKED;
3738 /* If packet isn't yet acked, and it has been transmitted at least
3739 * once, reset retransmit time using latest timeout
3740 * ie, this should readjust the retransmit timer for all outstanding
3741 * packets... So we don't just retransmit when we should know better*/
3743 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3744 tp->retryTime = tp->timeSent;
3745 clock_Add(&tp->retryTime, &peer->timeout);
3746 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3747 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3751 /* If the window has been extended by this acknowledge packet,
3752 * then wakeup a sender waiting in alloc for window space, or try
3753 * sending packets now, if he's been sitting on packets due to
3754 * lack of window space */
3755 if (call->tnext < (call->tfirst + call->twind)) {
3756 #ifdef RX_ENABLE_LOCKS
3757 CV_SIGNAL(&call->cv_twind);
3759 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3760 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3761 osi_rxWakeup(&call->twind);
3764 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3765 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3769 /* if the ack packet has a receivelen field hanging off it,
3770 * update our state */
3771 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3774 /* If the ack packet has a "recommended" size that is less than
3775 * what I am using now, reduce my size to match */
3776 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3777 (int)sizeof(afs_int32), &tSize);
3778 tSize = (afs_uint32) ntohl(tSize);
3779 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3781 /* Get the maximum packet size to send to this peer */
3782 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
3784 tSize = (afs_uint32) ntohl(tSize);
3785 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3786 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3788 /* sanity check - peer might have restarted with different params.
3789 * If peer says "send less", dammit, send less... Peer should never
3790 * be unable to accept packets of the size that prior AFS versions would
3791 * send without asking. */
3792 if (peer->maxMTU != tSize) {
3793 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
3795 peer->maxMTU = tSize;
3796 peer->MTU = MIN(tSize, peer->MTU);
3797 call->MTU = MIN(call->MTU, tSize);
3800 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3803 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3804 (int)sizeof(afs_int32), &tSize);
3805 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3806 if (tSize < call->twind) { /* smaller than our send */
3807 call->twind = tSize; /* window, we must send less... */
3808 call->ssthresh = MIN(call->twind, call->ssthresh);
3811 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3812 * network MTU confused with the loopback MTU. Calculate the
3813 * maximum MTU here for use in the slow start code below.
3815 maxMTU = peer->maxMTU;
3816 /* Did peer restart with older RX version? */
3817 if (peer->maxDgramPackets > 1) {
3818 peer->maxDgramPackets = 1;
3820 } else if (np->length >=
3821 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3824 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3825 sizeof(afs_int32), &tSize);
3826 tSize = (afs_uint32) ntohl(tSize);
3828 * As of AFS 3.5 we set the send window to match the receive window.
3830 if (tSize < call->twind) {
3831 call->twind = tSize;
3832 call->ssthresh = MIN(call->twind, call->ssthresh);
3833 } else if (tSize > call->twind) {
3834 call->twind = tSize;
3838 * As of AFS 3.5, a jumbogram is more than one fixed size
3839 * packet transmitted in a single UDP datagram. If the remote
3840 * MTU is smaller than our local MTU then never send a datagram
3841 * larger than the natural MTU.
3844 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3845 sizeof(afs_int32), &tSize);
3846 maxDgramPackets = (afs_uint32) ntohl(tSize);
3847 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3849 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
3850 maxDgramPackets = MIN(maxDgramPackets, tSize);
3851 if (maxDgramPackets > 1) {
3852 peer->maxDgramPackets = maxDgramPackets;
3853 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3855 peer->maxDgramPackets = 1;
3856 call->MTU = peer->natMTU;
3858 } else if (peer->maxDgramPackets > 1) {
3859 /* Restarted with lower version of RX */
3860 peer->maxDgramPackets = 1;
3862 } else if (peer->maxDgramPackets > 1
3863 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3864 /* Restarted with lower version of RX */
3865 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3866 peer->natMTU = OLD_MAX_PACKET_SIZE;
3867 peer->MTU = OLD_MAX_PACKET_SIZE;
3868 peer->maxDgramPackets = 1;
3869 peer->nDgramPackets = 1;
3871 call->MTU = OLD_MAX_PACKET_SIZE;
3876 * Calculate how many datagrams were successfully received after
3877 * the first missing packet and adjust the negative ack counter
3882 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3883 if (call->nNacks < nNacked) {
3884 call->nNacks = nNacked;
3893 if (call->flags & RX_CALL_FAST_RECOVER) {
3895 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3897 call->flags &= ~RX_CALL_FAST_RECOVER;
3898 call->cwind = call->nextCwind;
3899 call->nextCwind = 0;
3902 call->nCwindAcks = 0;
3903 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3904 /* Three negative acks in a row trigger congestion recovery */
3905 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3906 MUTEX_EXIT(&peer->peer_lock);
3907 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3908 /* someone else is waiting to start recovery */
3911 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3912 rxi_WaitforTQBusy(call);
3913 MUTEX_ENTER(&peer->peer_lock);
3914 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3915 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3916 call->flags |= RX_CALL_FAST_RECOVER;
3917 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3919 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3920 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3921 call->nextCwind = call->ssthresh;
3924 peer->MTU = call->MTU;
3925 peer->cwind = call->nextCwind;
3926 peer->nDgramPackets = call->nDgramPackets;
3928 call->congestSeq = peer->congestSeq;
3929 /* Reset the resend times on the packets that were nacked
3930 * so we will retransmit as soon as the window permits*/
3931 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3933 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3934 clock_Zero(&tp->retryTime);
3936 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3941 /* If cwind is smaller than ssthresh, then increase
3942 * the window one packet for each ack we receive (exponential
3944 * If cwind is greater than or equal to ssthresh then increase
3945 * the congestion window by one packet for each cwind acks we
3946 * receive (linear growth). */
3947 if (call->cwind < call->ssthresh) {
3949 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3950 call->nCwindAcks = 0;
3952 call->nCwindAcks += newAckCount;
3953 if (call->nCwindAcks >= call->cwind) {
3954 call->nCwindAcks = 0;
3955 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3959 * If we have received several acknowledgements in a row then
3960 * it is time to increase the size of our datagrams
3962 if ((int)call->nAcks > rx_nDgramThreshold) {
3963 if (peer->maxDgramPackets > 1) {
3964 if (call->nDgramPackets < peer->maxDgramPackets) {
3965 call->nDgramPackets++;
3967 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3968 } else if (call->MTU < peer->maxMTU) {
3969 call->MTU += peer->natMTU;
3970 call->MTU = MIN(call->MTU, peer->maxMTU);
3976 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3978 /* Servers need to hold the call until all response packets have
3979 * been acknowledged. Soft acks are good enough since clients
3980 * are not allowed to clear their receive queues. */
3981 if (call->state == RX_STATE_HOLD
3982 && call->tfirst + call->nSoftAcked >= call->tnext) {
3983 call->state = RX_STATE_DALLY;
3984 rxi_ClearTransmitQueue(call, 0);
3985 } else if (!queue_IsEmpty(&call->tq)) {
3986 rxi_Start(0, call, 0, istack);
3991 /* Received a response to a challenge packet */
3993 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
3994 register struct rx_packet *np, int istack)
3998 /* Ignore the packet if we're the client */
3999 if (conn->type == RX_CLIENT_CONNECTION)
4002 /* If already authenticated, ignore the packet (it's probably a retry) */
4003 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4006 /* Otherwise, have the security object evaluate the response packet */
4007 error = RXS_CheckResponse(conn->securityObject, conn, np);
4009 /* If the response is invalid, reset the connection, sending
4010 * an abort to the peer */
4014 rxi_ConnectionError(conn, error);
4015 MUTEX_ENTER(&conn->conn_data_lock);
4016 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4017 MUTEX_EXIT(&conn->conn_data_lock);
4020 /* If the response is valid, any calls waiting to attach
4021 * servers can now do so */
4024 for (i = 0; i < RX_MAXCALLS; i++) {
4025 struct rx_call *call = conn->call[i];
4027 MUTEX_ENTER(&call->lock);
4028 if (call->state == RX_STATE_PRECALL)
4029 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4030 /* tnop can be null if newcallp is null */
4031 MUTEX_EXIT(&call->lock);
4035 /* Update the peer reachability information, just in case
4036 * some calls went into attach-wait while we were waiting
4037 * for authentication..
4039 rxi_UpdatePeerReach(conn, NULL);
4044 /* A client has received an authentication challenge: the security
4045 * object is asked to cough up a respectable response packet to send
4046 * back to the server. The server is responsible for retrying the
4047 * challenge if it fails to get a response. */
4050 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
4051 register struct rx_packet *np, int istack)
4055 /* Ignore the challenge if we're the server */
4056 if (conn->type == RX_SERVER_CONNECTION)
4059 /* Ignore the challenge if the connection is otherwise idle; someone's
4060 * trying to use us as an oracle. */
4061 if (!rxi_HasActiveCalls(conn))
4064 /* Send the security object the challenge packet. It is expected to fill
4065 * in the response. */
4066 error = RXS_GetResponse(conn->securityObject, conn, np);
4068 /* If the security object is unable to return a valid response, reset the
4069 * connection and send an abort to the peer. Otherwise send the response
4070 * packet to the peer connection. */
4072 rxi_ConnectionError(conn, error);
4073 MUTEX_ENTER(&conn->conn_data_lock);
4074 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4075 MUTEX_EXIT(&conn->conn_data_lock);
4077 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4078 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4084 /* Find an available server process to service the current request in
4085 * the given call structure. If one isn't available, queue up this
4086 * call so it eventually gets one */
4088 rxi_AttachServerProc(register struct rx_call *call,
4089 register osi_socket socket, register int *tnop,
4090 register struct rx_call **newcallp)
4092 register struct rx_serverQueueEntry *sq;
4093 register struct rx_service *service = call->conn->service;
4094 register int haveQuota = 0;
4096 /* May already be attached */
4097 if (call->state == RX_STATE_ACTIVE)
4100 MUTEX_ENTER(&rx_serverPool_lock);
4102 haveQuota = QuotaOK(service);
4103 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4104 /* If there are no processes available to service this call,
4105 * put the call on the incoming call queue (unless it's
4106 * already on the queue).
4108 #ifdef RX_ENABLE_LOCKS
4110 ReturnToServerPool(service);
4111 #endif /* RX_ENABLE_LOCKS */
4113 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4114 call->flags |= RX_CALL_WAIT_PROC;
4115 MUTEX_ENTER(&rx_stats_mutex);
4118 MUTEX_EXIT(&rx_stats_mutex);
4119 rxi_calltrace(RX_CALL_ARRIVAL, call);
4120 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4121 queue_Append(&rx_incomingCallQueue, call);
4124 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4126 /* If hot threads are enabled, and both newcallp and sq->socketp
4127 * are non-null, then this thread will process the call, and the
4128 * idle server thread will start listening on this threads socket.
4131 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4134 *sq->socketp = socket;
4135 clock_GetTime(&call->startTime);
4136 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4140 if (call->flags & RX_CALL_WAIT_PROC) {
4141 /* Conservative: I don't think this should happen */
4142 call->flags &= ~RX_CALL_WAIT_PROC;
4143 if (queue_IsOnQueue(call)) {
4145 MUTEX_ENTER(&rx_stats_mutex);
4147 MUTEX_EXIT(&rx_stats_mutex);
4150 call->state = RX_STATE_ACTIVE;
4151 call->mode = RX_MODE_RECEIVING;
4152 #ifdef RX_KERNEL_TRACE
4154 int glockOwner = ISAFS_GLOCK();
4157 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4158 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4164 if (call->flags & RX_CALL_CLEARED) {
4165 /* send an ack now to start the packet flow up again */
4166 call->flags &= ~RX_CALL_CLEARED;
4167 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4169 #ifdef RX_ENABLE_LOCKS
4172 service->nRequestsRunning++;
4173 if (service->nRequestsRunning <= service->minProcs)
4179 MUTEX_EXIT(&rx_serverPool_lock);
4182 /* Delay the sending of an acknowledge event for a short while, while
4183 * a new call is being prepared (in the case of a client) or a reply
4184 * is being prepared (in the case of a server). Rather than sending
4185 * an ack packet, an ACKALL packet is sent. */
4187 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4189 #ifdef RX_ENABLE_LOCKS
4191 MUTEX_ENTER(&call->lock);
4192 call->delayedAckEvent = NULL;
4193 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4195 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4196 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4198 MUTEX_EXIT(&call->lock);
4199 #else /* RX_ENABLE_LOCKS */
4201 call->delayedAckEvent = NULL;
4202 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4203 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4204 #endif /* RX_ENABLE_LOCKS */
4208 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4211 #ifdef RX_ENABLE_LOCKS
4213 MUTEX_ENTER(&call->lock);
4214 if (event == call->delayedAckEvent)
4215 call->delayedAckEvent = NULL;
4216 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4218 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4220 MUTEX_EXIT(&call->lock);
4221 #else /* RX_ENABLE_LOCKS */
4223 call->delayedAckEvent = NULL;
4224 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4225 #endif /* RX_ENABLE_LOCKS */
4229 #ifdef RX_ENABLE_LOCKS
4230 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4231 * clearing them out.
4234 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4236 register struct rx_packet *p, *tp;
4239 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4240 p->flags |= RX_PKTFLAG_ACKED;
4244 call->flags |= RX_CALL_TQ_CLEARME;
4245 call->flags |= RX_CALL_TQ_SOME_ACKED;
4248 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4249 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4250 call->tfirst = call->tnext;
4251 call->nSoftAcked = 0;
4253 if (call->flags & RX_CALL_FAST_RECOVER) {
4254 call->flags &= ~RX_CALL_FAST_RECOVER;
4255 call->cwind = call->nextCwind;
4256 call->nextCwind = 0;
4259 CV_SIGNAL(&call->cv_twind);
4261 #endif /* RX_ENABLE_LOCKS */
4263 /* Clear out the transmit queue for the current call (all packets have
4264 * been received by peer) */
4266 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4268 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4269 register struct rx_packet *p, *tp;
4271 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4273 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4274 p->flags |= RX_PKTFLAG_ACKED;
4278 call->flags |= RX_CALL_TQ_CLEARME;
4279 call->flags |= RX_CALL_TQ_SOME_ACKED;
4282 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4283 rxi_FreePackets(0, &call->tq);
4284 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4285 call->flags &= ~RX_CALL_TQ_CLEARME;
4287 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4289 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4290 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4291 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4292 call->nSoftAcked = 0;
4294 if (call->flags & RX_CALL_FAST_RECOVER) {
4295 call->flags &= ~RX_CALL_FAST_RECOVER;
4296 call->cwind = call->nextCwind;
4298 #ifdef RX_ENABLE_LOCKS
4299 CV_SIGNAL(&call->cv_twind);
4301 osi_rxWakeup(&call->twind);
4306 rxi_ClearReceiveQueue(register struct rx_call *call)
4308 if (queue_IsNotEmpty(&call->rq)) {
4309 rx_packetReclaims += rxi_FreePackets(0, &call->rq);
4310 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4312 if (call->state == RX_STATE_PRECALL) {
4313 call->flags |= RX_CALL_CLEARED;
4317 /* Send an abort packet for the specified call */
4319 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4320 int istack, int force)
4328 /* Clients should never delay abort messages */
4329 if (rx_IsClientConn(call->conn))
4332 if (call->abortCode != call->error) {
4333 call->abortCode = call->error;
4334 call->abortCount = 0;
4337 if (force || rxi_callAbortThreshhold == 0
4338 || call->abortCount < rxi_callAbortThreshhold) {
4339 if (call->delayedAbortEvent) {
4340 rxevent_Cancel(call->delayedAbortEvent, call,
4341 RX_CALL_REFCOUNT_ABORT);
4343 error = htonl(call->error);
4346 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4347 (char *)&error, sizeof(error), istack);
4348 } else if (!call->delayedAbortEvent) {
4349 clock_GetTime(&when);
4350 clock_Addmsec(&when, rxi_callAbortDelay);
4351 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4352 call->delayedAbortEvent =
4353 rxevent_Post(&when, rxi_SendDelayedCallAbort, call, 0);
4358 /* Send an abort packet for the specified connection. Packet is an
4359 * optional pointer to a packet that can be used to send the abort.
4360 * Once the number of abort messages reaches the threshhold, an
4361 * event is scheduled to send the abort. Setting the force flag
4362 * overrides sending delayed abort messages.
4364 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4365 * to send the abort packet.
4368 rxi_SendConnectionAbort(register struct rx_connection *conn,
4369 struct rx_packet *packet, int istack, int force)
4377 /* Clients should never delay abort messages */
4378 if (rx_IsClientConn(conn))
4381 if (force || rxi_connAbortThreshhold == 0
4382 || conn->abortCount < rxi_connAbortThreshhold) {
4383 if (conn->delayedAbortEvent) {
4384 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4386 error = htonl(conn->error);
4388 MUTEX_EXIT(&conn->conn_data_lock);
4390 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4391 RX_PACKET_TYPE_ABORT, (char *)&error,
4392 sizeof(error), istack);
4393 MUTEX_ENTER(&conn->conn_data_lock);
4394 } else if (!conn->delayedAbortEvent) {
4395 clock_GetTime(&when);
4396 clock_Addmsec(&when, rxi_connAbortDelay);
4397 conn->delayedAbortEvent =
4398 rxevent_Post(&when, rxi_SendDelayedConnAbort, conn, 0);
4403 /* Associate an error all of the calls owned by a connection. Called
4404 * with error non-zero. This is only for really fatal things, like
4405 * bad authentication responses. The connection itself is set in
4406 * error at this point, so that future packets received will be
4409 rxi_ConnectionError(register struct rx_connection *conn,
4410 register afs_int32 error)
4415 dpf(("rxi_ConnectionError conn %x error %d", conn, error));
4417 MUTEX_ENTER(&conn->conn_data_lock);
4418 if (conn->challengeEvent)
4419 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4420 if (conn->checkReachEvent) {
4421 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4422 conn->checkReachEvent = 0;
4423 conn->flags &= ~RX_CONN_ATTACHWAIT;
4426 MUTEX_EXIT(&conn->conn_data_lock);
4427 for (i = 0; i < RX_MAXCALLS; i++) {
4428 struct rx_call *call = conn->call[i];
4430 MUTEX_ENTER(&call->lock);
4431 rxi_CallError(call, error);
4432 MUTEX_EXIT(&call->lock);
4435 conn->error = error;
4436 MUTEX_ENTER(&rx_stats_mutex);
4437 rx_stats.fatalErrors++;
4438 MUTEX_EXIT(&rx_stats_mutex);
4443 rxi_CallError(register struct rx_call *call, afs_int32 error)
4445 dpf(("rxi_CallError call %x error %d call->error %d", call, error, call->error));
4447 error = call->error;
4449 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4450 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4451 rxi_ResetCall(call, 0);
4454 rxi_ResetCall(call, 0);
4456 call->error = error;
4457 call->mode = RX_MODE_ERROR;
4460 /* Reset various fields in a call structure, and wakeup waiting
4461 * processes. Some fields aren't changed: state & mode are not
4462 * touched (these must be set by the caller), and bufptr, nLeft, and
4463 * nFree are not reset, since these fields are manipulated by
4464 * unprotected macros, and may only be reset by non-interrupting code.
4467 /* this code requires that call->conn be set properly as a pre-condition. */
4468 #endif /* ADAPT_WINDOW */
4471 rxi_ResetCall(register struct rx_call *call, register int newcall)
4474 register struct rx_peer *peer;
4475 struct rx_packet *packet;
4477 dpf(("rxi_ResetCall(call %x, newcall %d)\n", call, newcall));
4479 /* Notify anyone who is waiting for asynchronous packet arrival */
4480 if (call->arrivalProc) {
4481 (*call->arrivalProc) (call, call->arrivalProcHandle,
4482 call->arrivalProcArg);
4483 call->arrivalProc = (void (*)())0;
4486 if (call->delayedAbortEvent) {
4487 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4488 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4490 rxi_SendCallAbort(call, packet, 0, 1);
4491 rxi_FreePacket(packet);
4496 * Update the peer with the congestion information in this call
4497 * so other calls on this connection can pick up where this call
4498 * left off. If the congestion sequence numbers don't match then
4499 * another call experienced a retransmission.
4501 peer = call->conn->peer;
4502 MUTEX_ENTER(&peer->peer_lock);
4504 if (call->congestSeq == peer->congestSeq) {
4505 peer->cwind = MAX(peer->cwind, call->cwind);
4506 peer->MTU = MAX(peer->MTU, call->MTU);
4507 peer->nDgramPackets =
4508 MAX(peer->nDgramPackets, call->nDgramPackets);
4511 call->abortCode = 0;
4512 call->abortCount = 0;
4514 if (peer->maxDgramPackets > 1) {
4515 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4517 call->MTU = peer->MTU;
4519 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4520 call->ssthresh = rx_maxSendWindow;
4521 call->nDgramPackets = peer->nDgramPackets;
4522 call->congestSeq = peer->congestSeq;
4523 MUTEX_EXIT(&peer->peer_lock);
4525 flags = call->flags;
4526 rxi_ClearReceiveQueue(call);
4527 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4528 if (flags & RX_CALL_TQ_BUSY) {
4529 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4530 call->flags |= (flags & RX_CALL_TQ_WAIT);
4532 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4534 rxi_ClearTransmitQueue(call, 0);
4535 queue_Init(&call->tq);
4536 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4537 dpf(("rcall %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4540 while (call->tqWaiters) {
4541 #ifdef RX_ENABLE_LOCKS
4542 CV_BROADCAST(&call->cv_tq);
4543 #else /* RX_ENABLE_LOCKS */
4544 osi_rxWakeup(&call->tq);
4545 #endif /* RX_ENABLE_LOCKS */
4549 queue_Init(&call->rq);
4551 call->rwind = rx_initReceiveWindow;
4552 call->twind = rx_initSendWindow;
4553 call->nSoftAcked = 0;
4554 call->nextCwind = 0;
4557 call->nCwindAcks = 0;
4558 call->nSoftAcks = 0;
4559 call->nHardAcks = 0;
4561 call->tfirst = call->rnext = call->tnext = 1;
4563 call->lastAcked = 0;
4564 call->localStatus = call->remoteStatus = 0;
4566 if (flags & RX_CALL_READER_WAIT) {
4567 #ifdef RX_ENABLE_LOCKS
4568 CV_BROADCAST(&call->cv_rq);
4570 osi_rxWakeup(&call->rq);
4573 if (flags & RX_CALL_WAIT_PACKETS) {
4574 MUTEX_ENTER(&rx_freePktQ_lock);
4575 rxi_PacketsUnWait(); /* XXX */
4576 MUTEX_EXIT(&rx_freePktQ_lock);
4578 #ifdef RX_ENABLE_LOCKS
4579 CV_SIGNAL(&call->cv_twind);
4581 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4582 osi_rxWakeup(&call->twind);
4585 #ifdef RX_ENABLE_LOCKS
4586 /* The following ensures that we don't mess with any queue while some
4587 * other thread might also be doing so. The call_queue_lock field is
4588 * is only modified under the call lock. If the call is in the process
4589 * of being removed from a queue, the call is not locked until the
4590 * the queue lock is dropped and only then is the call_queue_lock field
4591 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4592 * Note that any other routine which removes a call from a queue has to
4593 * obtain the queue lock before examing the queue and removing the call.
4595 if (call->call_queue_lock) {
4596 MUTEX_ENTER(call->call_queue_lock);
4597 if (queue_IsOnQueue(call)) {
4599 if (flags & RX_CALL_WAIT_PROC) {
4600 MUTEX_ENTER(&rx_stats_mutex);
4602 MUTEX_EXIT(&rx_stats_mutex);
4605 MUTEX_EXIT(call->call_queue_lock);
4606 CLEAR_CALL_QUEUE_LOCK(call);
4608 #else /* RX_ENABLE_LOCKS */
4609 if (queue_IsOnQueue(call)) {
4611 if (flags & RX_CALL_WAIT_PROC)
4614 #endif /* RX_ENABLE_LOCKS */
4616 rxi_KeepAliveOff(call);
4617 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4620 /* Send an acknowledge for the indicated packet (seq,serial) of the
4621 * indicated call, for the indicated reason (reason). This
4622 * acknowledge will specifically acknowledge receiving the packet, and
4623 * will also specify which other packets for this call have been
4624 * received. This routine returns the packet that was used to the
4625 * caller. The caller is responsible for freeing it or re-using it.
4626 * This acknowledgement also returns the highest sequence number
4627 * actually read out by the higher level to the sender; the sender
4628 * promises to keep around packets that have not been read by the
4629 * higher level yet (unless, of course, the sender decides to abort
4630 * the call altogether). Any of p, seq, serial, pflags, or reason may
4631 * be set to zero without ill effect. That is, if they are zero, they
4632 * will not convey any information.
4633 * NOW there is a trailer field, after the ack where it will safely be
4634 * ignored by mundanes, which indicates the maximum size packet this
4635 * host can swallow. */
4637 register struct rx_packet *optionalPacket; use to send ack (or null)
4638 int seq; Sequence number of the packet we are acking
4639 int serial; Serial number of the packet
4640 int pflags; Flags field from packet header
4641 int reason; Reason an acknowledge was prompted
4645 rxi_SendAck(register struct rx_call *call,
4646 register struct rx_packet *optionalPacket, int serial, int reason,
4649 struct rx_ackPacket *ap;
4650 register struct rx_packet *rqp;
4651 register struct rx_packet *nxp; /* For queue_Scan */
4652 register struct rx_packet *p;
4655 #ifdef RX_ENABLE_TSFPQ
4656 struct rx_ts_info_t * rx_ts_info;
4660 * Open the receive window once a thread starts reading packets
4662 if (call->rnext > 1) {
4663 call->rwind = rx_maxReceiveWindow;
4666 call->nHardAcks = 0;
4667 call->nSoftAcks = 0;
4668 if (call->rnext > call->lastAcked)
4669 call->lastAcked = call->rnext;
4673 rx_computelen(p, p->length); /* reset length, you never know */
4674 } /* where that's been... */
4675 #ifdef RX_ENABLE_TSFPQ
4677 RX_TS_INFO_GET(rx_ts_info);
4678 if ((p = rx_ts_info->local_special_packet)) {
4679 rx_computelen(p, p->length);
4680 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4681 rx_ts_info->local_special_packet = p;
4682 } else { /* We won't send the ack, but don't panic. */
4683 return optionalPacket;
4687 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4688 /* We won't send the ack, but don't panic. */
4689 return optionalPacket;
4694 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4697 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4698 #ifndef RX_ENABLE_TSFPQ
4699 if (!optionalPacket)
4702 return optionalPacket;
4704 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4705 if (rx_Contiguous(p) < templ) {
4706 #ifndef RX_ENABLE_TSFPQ
4707 if (!optionalPacket)
4710 return optionalPacket;
4715 /* MTUXXX failing to send an ack is very serious. We should */
4716 /* try as hard as possible to send even a partial ack; it's */
4717 /* better than nothing. */
4718 ap = (struct rx_ackPacket *)rx_DataOf(p);
4719 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4720 ap->reason = reason;
4722 /* The skew computation used to be bogus, I think it's better now. */
4723 /* We should start paying attention to skew. XXX */
4724 ap->serial = htonl(serial);
4725 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4727 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4728 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4730 /* No fear of running out of ack packet here because there can only be at most
4731 * one window full of unacknowledged packets. The window size must be constrained
4732 * to be less than the maximum ack size, of course. Also, an ack should always
4733 * fit into a single packet -- it should not ever be fragmented. */
4734 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4735 if (!rqp || !call->rq.next
4736 || (rqp->header.seq > (call->rnext + call->rwind))) {
4737 #ifndef RX_ENABLE_TSFPQ
4738 if (!optionalPacket)
4741 rxi_CallError(call, RX_CALL_DEAD);
4742 return optionalPacket;
4745 while (rqp->header.seq > call->rnext + offset)
4746 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4747 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4749 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4750 #ifndef RX_ENABLE_TSFPQ
4751 if (!optionalPacket)
4754 rxi_CallError(call, RX_CALL_DEAD);
4755 return optionalPacket;
4760 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4762 /* these are new for AFS 3.3 */
4763 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4764 templ = htonl(templ);
4765 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4766 templ = htonl(call->conn->peer->ifMTU);
4767 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4768 sizeof(afs_int32), &templ);
4770 /* new for AFS 3.4 */
4771 templ = htonl(call->rwind);
4772 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4773 sizeof(afs_int32), &templ);
4775 /* new for AFS 3.5 */
4776 templ = htonl(call->conn->peer->ifDgramPackets);
4777 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4778 sizeof(afs_int32), &templ);
4780 p->header.serviceId = call->conn->serviceId;
4781 p->header.cid = (call->conn->cid | call->channel);
4782 p->header.callNumber = *call->callNumber;
4784 p->header.securityIndex = call->conn->securityIndex;
4785 p->header.epoch = call->conn->epoch;
4786 p->header.type = RX_PACKET_TYPE_ACK;
4787 p->header.flags = RX_SLOW_START_OK;
4788 if (reason == RX_ACK_PING) {
4789 p->header.flags |= RX_REQUEST_ACK;
4791 clock_GetTime(&call->pingRequestTime);
4794 if (call->conn->type == RX_CLIENT_CONNECTION)
4795 p->header.flags |= RX_CLIENT_INITIATED;
4799 if (rxdebug_active) {
4803 len = _snprintf(msg, sizeof(msg),
4804 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4805 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4806 ntohl(ap->serial), ntohl(ap->previousPacket),
4807 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
4808 ap->nAcks, ntohs(ap->bufferSpace) );
4812 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
4813 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4817 OutputDebugString(msg);
4819 #else /* AFS_NT40_ENV */
4821 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
4822 ap->reason, ntohl(ap->previousPacket),
4823 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4825 for (offset = 0; offset < ap->nAcks; offset++)
4826 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4831 #endif /* AFS_NT40_ENV */
4834 register int i, nbytes = p->length;
4836 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4837 if (nbytes <= p->wirevec[i].iov_len) {
4838 register int savelen, saven;
4840 savelen = p->wirevec[i].iov_len;
4842 p->wirevec[i].iov_len = nbytes;
4844 rxi_Send(call, p, istack);
4845 p->wirevec[i].iov_len = savelen;
4849 nbytes -= p->wirevec[i].iov_len;
4852 MUTEX_ENTER(&rx_stats_mutex);
4853 rx_stats.ackPacketsSent++;
4854 MUTEX_EXIT(&rx_stats_mutex);
4855 #ifndef RX_ENABLE_TSFPQ
4856 if (!optionalPacket)
4859 return optionalPacket; /* Return packet for re-use by caller */
4862 /* Send all of the packets in the list in single datagram */
4864 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4865 int istack, int moreFlag, struct clock *now,
4866 struct clock *retryTime, int resending)
4871 struct rx_connection *conn = call->conn;
4872 struct rx_peer *peer = conn->peer;
4874 MUTEX_ENTER(&peer->peer_lock);
4877 peer->reSends += len;
4878 MUTEX_ENTER(&rx_stats_mutex);
4879 rx_stats.dataPacketsSent += len;
4880 MUTEX_EXIT(&rx_stats_mutex);
4881 MUTEX_EXIT(&peer->peer_lock);
4883 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4887 /* Set the packet flags and schedule the resend events */
4888 /* Only request an ack for the last packet in the list */
4889 for (i = 0; i < len; i++) {
4890 list[i]->retryTime = *retryTime;
4891 if (list[i]->header.serial) {
4892 /* Exponentially backoff retry times */
4893 if (list[i]->backoff < MAXBACKOFF) {
4894 /* so it can't stay == 0 */
4895 list[i]->backoff = (list[i]->backoff << 1) + 1;
4898 clock_Addmsec(&(list[i]->retryTime),
4899 ((afs_uint32) list[i]->backoff) << 8);
4902 /* Wait a little extra for the ack on the last packet */
4903 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4904 clock_Addmsec(&(list[i]->retryTime), 400);
4907 /* Record the time sent */
4908 list[i]->timeSent = *now;
4910 /* Ask for an ack on retransmitted packets, on every other packet
4911 * if the peer doesn't support slow start. Ask for an ack on every
4912 * packet until the congestion window reaches the ack rate. */
4913 if (list[i]->header.serial) {
4915 MUTEX_ENTER(&rx_stats_mutex);
4916 rx_stats.dataPacketsReSent++;
4917 MUTEX_EXIT(&rx_stats_mutex);
4919 /* improved RTO calculation- not Karn */
4920 list[i]->firstSent = *now;
4921 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4922 || (!(call->flags & RX_CALL_SLOW_START_OK)
4923 && (list[i]->header.seq & 1)))) {
4928 MUTEX_ENTER(&peer->peer_lock);
4932 MUTEX_ENTER(&rx_stats_mutex);
4933 rx_stats.dataPacketsSent++;
4934 MUTEX_EXIT(&rx_stats_mutex);
4935 MUTEX_EXIT(&peer->peer_lock);
4937 /* Tag this packet as not being the last in this group,
4938 * for the receiver's benefit */
4939 if (i < len - 1 || moreFlag) {
4940 list[i]->header.flags |= RX_MORE_PACKETS;
4943 /* Install the new retransmit time for the packet, and
4944 * record the time sent */
4945 list[i]->timeSent = *now;
4949 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4952 /* Since we're about to send a data packet to the peer, it's
4953 * safe to nuke any scheduled end-of-packets ack */
4954 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4956 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4957 MUTEX_EXIT(&call->lock);
4959 rxi_SendPacketList(call, conn, list, len, istack);
4961 rxi_SendPacket(call, conn, list[0], istack);
4963 MUTEX_ENTER(&call->lock);
4964 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4966 /* Update last send time for this call (for keep-alive
4967 * processing), and for the connection (so that we can discover
4968 * idle connections) */
4969 conn->lastSendTime = call->lastSendTime = clock_Sec();
4972 /* When sending packets we need to follow these rules:
4973 * 1. Never send more than maxDgramPackets in a jumbogram.
4974 * 2. Never send a packet with more than two iovecs in a jumbogram.
4975 * 3. Never send a retransmitted packet in a jumbogram.
4976 * 4. Never send more than cwind/4 packets in a jumbogram
4977 * We always keep the last list we should have sent so we
4978 * can set the RX_MORE_PACKETS flags correctly.
4981 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4982 int istack, struct clock *now, struct clock *retryTime,
4985 int i, cnt, lastCnt = 0;
4986 struct rx_packet **listP, **lastP = 0;
4987 struct rx_peer *peer = call->conn->peer;
4988 int morePackets = 0;
4990 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4991 /* Does the current packet force us to flush the current list? */
4993 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4994 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4996 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
4998 /* If the call enters an error state stop sending, or if
4999 * we entered congestion recovery mode, stop sending */
5000 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5008 /* Add the current packet to the list if it hasn't been acked.
5009 * Otherwise adjust the list pointer to skip the current packet. */
5010 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5012 /* Do we need to flush the list? */
5013 if (cnt >= (int)peer->maxDgramPackets
5014 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5015 || list[i]->header.serial
5016 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5018 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5019 retryTime, resending);
5020 /* If the call enters an error state stop sending, or if
5021 * we entered congestion recovery mode, stop sending */
5023 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5028 listP = &list[i + 1];
5033 osi_Panic("rxi_SendList error");
5035 listP = &list[i + 1];
5039 /* Send the whole list when the call is in receive mode, when
5040 * the call is in eof mode, when we are in fast recovery mode,
5041 * and when we have the last packet */
5042 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5043 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5044 || (call->flags & RX_CALL_FAST_RECOVER)) {
5045 /* Check for the case where the current list contains
5046 * an acked packet. Since we always send retransmissions
5047 * in a separate packet, we only need to check the first
5048 * packet in the list */
5049 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5053 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5054 retryTime, resending);
5055 /* If the call enters an error state stop sending, or if
5056 * we entered congestion recovery mode, stop sending */
5057 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5061 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5064 } else if (lastCnt > 0) {
5065 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5070 #ifdef RX_ENABLE_LOCKS
5071 /* Call rxi_Start, below, but with the call lock held. */
5073 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
5074 void *arg1, int istack)
5076 MUTEX_ENTER(&call->lock);
5077 rxi_Start(event, call, arg1, istack);
5078 MUTEX_EXIT(&call->lock);
5080 #endif /* RX_ENABLE_LOCKS */
5082 /* This routine is called when new packets are readied for
5083 * transmission and when retransmission may be necessary, or when the
5084 * transmission window or burst count are favourable. This should be
5085 * better optimized for new packets, the usual case, now that we've
5086 * got rid of queues of send packets. XXXXXXXXXXX */
5088 rxi_Start(struct rxevent *event, register struct rx_call *call,
5089 void *arg1, int istack)
5091 struct rx_packet *p;
5092 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
5093 struct rx_peer *peer = call->conn->peer;
5094 struct clock now, retryTime;
5098 struct rx_packet **xmitList;
5101 /* If rxi_Start is being called as a result of a resend event,
5102 * then make sure that the event pointer is removed from the call
5103 * structure, since there is no longer a per-call retransmission
5105 if (event && event == call->resendEvent) {
5106 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5107 call->resendEvent = NULL;
5109 if (queue_IsEmpty(&call->tq)) {
5113 /* Timeouts trigger congestion recovery */
5114 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5115 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5116 /* someone else is waiting to start recovery */
5119 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5120 rxi_WaitforTQBusy(call);
5121 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5122 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5123 call->flags |= RX_CALL_FAST_RECOVER;
5124 if (peer->maxDgramPackets > 1) {
5125 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5127 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5129 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5130 call->nDgramPackets = 1;
5132 call->nextCwind = 1;
5135 MUTEX_ENTER(&peer->peer_lock);
5136 peer->MTU = call->MTU;
5137 peer->cwind = call->cwind;
5138 peer->nDgramPackets = 1;
5140 call->congestSeq = peer->congestSeq;
5141 MUTEX_EXIT(&peer->peer_lock);
5142 /* Clear retry times on packets. Otherwise, it's possible for
5143 * some packets in the queue to force resends at rates faster
5144 * than recovery rates.
5146 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5147 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5148 clock_Zero(&p->retryTime);
5153 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5154 MUTEX_ENTER(&rx_stats_mutex);
5155 rx_tq_debug.rxi_start_in_error++;
5156 MUTEX_EXIT(&rx_stats_mutex);
5161 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5162 /* Get clock to compute the re-transmit time for any packets
5163 * in this burst. Note, if we back off, it's reasonable to
5164 * back off all of the packets in the same manner, even if
5165 * some of them have been retransmitted more times than more
5166 * recent additions */
5167 clock_GetTime(&now);
5168 retryTime = now; /* initialize before use */
5169 MUTEX_ENTER(&peer->peer_lock);
5170 clock_Add(&retryTime, &peer->timeout);
5171 MUTEX_EXIT(&peer->peer_lock);
5173 /* Send (or resend) any packets that need it, subject to
5174 * window restrictions and congestion burst control
5175 * restrictions. Ask for an ack on the last packet sent in
5176 * this burst. For now, we're relying upon the window being
5177 * considerably bigger than the largest number of packets that
5178 * are typically sent at once by one initial call to
5179 * rxi_Start. This is probably bogus (perhaps we should ask
5180 * for an ack when we're half way through the current
5181 * window?). Also, for non file transfer applications, this
5182 * may end up asking for an ack for every packet. Bogus. XXXX
5185 * But check whether we're here recursively, and let the other guy
5188 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5189 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5190 call->flags |= RX_CALL_TQ_BUSY;
5192 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5194 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5195 call->flags &= ~RX_CALL_NEED_START;
5196 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5198 maxXmitPackets = MIN(call->twind, call->cwind);
5199 xmitList = (struct rx_packet **)
5200 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5201 if (xmitList == NULL)
5202 osi_Panic("rxi_Start, failed to allocate xmit list");
5203 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5204 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5205 /* We shouldn't be sending packets if a thread is waiting
5206 * to initiate congestion recovery */
5210 && (call->flags & RX_CALL_FAST_RECOVER)) {
5211 /* Only send one packet during fast recovery */
5214 if ((p->flags & RX_PKTFLAG_FREE)
5215 || (!queue_IsEnd(&call->tq, nxp)
5216 && (nxp->flags & RX_PKTFLAG_FREE))
5217 || (p == (struct rx_packet *)&rx_freePacketQueue)
5218 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5219 osi_Panic("rxi_Start: xmit queue clobbered");
5221 if (p->flags & RX_PKTFLAG_ACKED) {
5222 MUTEX_ENTER(&rx_stats_mutex);
5223 rx_stats.ignoreAckedPacket++;
5224 MUTEX_EXIT(&rx_stats_mutex);
5225 continue; /* Ignore this packet if it has been acknowledged */
5228 /* Turn off all flags except these ones, which are the same
5229 * on each transmission */
5230 p->header.flags &= RX_PRESET_FLAGS;
5232 if (p->header.seq >=
5233 call->tfirst + MIN((int)call->twind,
5234 (int)(call->nSoftAcked +
5236 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5237 /* Note: if we're waiting for more window space, we can
5238 * still send retransmits; hence we don't return here, but
5239 * break out to schedule a retransmit event */
5240 dpf(("call %d waiting for window",
5241 *(call->callNumber)));
5245 /* Transmit the packet if it needs to be sent. */
5246 if (!clock_Lt(&now, &p->retryTime)) {
5247 if (nXmitPackets == maxXmitPackets) {
5248 rxi_SendXmitList(call, xmitList, nXmitPackets,
5249 istack, &now, &retryTime,
5251 osi_Free(xmitList, maxXmitPackets *
5252 sizeof(struct rx_packet *));
5255 xmitList[nXmitPackets++] = p;
5259 /* xmitList now hold pointers to all of the packets that are
5260 * ready to send. Now we loop to send the packets */
5261 if (nXmitPackets > 0) {
5262 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5263 &now, &retryTime, resending);
5266 maxXmitPackets * sizeof(struct rx_packet *));
5268 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5270 * TQ references no longer protected by this flag; they must remain
5271 * protected by the global lock.
5273 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5274 call->flags &= ~RX_CALL_TQ_BUSY;
5275 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5276 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5277 #ifdef RX_ENABLE_LOCKS
5278 osirx_AssertMine(&call->lock, "rxi_Start start");
5279 CV_BROADCAST(&call->cv_tq);
5280 #else /* RX_ENABLE_LOCKS */
5281 osi_rxWakeup(&call->tq);
5282 #endif /* RX_ENABLE_LOCKS */
5287 /* We went into the error state while sending packets. Now is
5288 * the time to reset the call. This will also inform the using
5289 * process that the call is in an error state.
5291 MUTEX_ENTER(&rx_stats_mutex);
5292 rx_tq_debug.rxi_start_aborted++;
5293 MUTEX_EXIT(&rx_stats_mutex);
5294 call->flags &= ~RX_CALL_TQ_BUSY;
5295 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5296 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5297 #ifdef RX_ENABLE_LOCKS
5298 osirx_AssertMine(&call->lock, "rxi_Start middle");
5299 CV_BROADCAST(&call->cv_tq);
5300 #else /* RX_ENABLE_LOCKS */
5301 osi_rxWakeup(&call->tq);
5302 #endif /* RX_ENABLE_LOCKS */
5304 rxi_CallError(call, call->error);
5307 #ifdef RX_ENABLE_LOCKS
5308 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5309 register int missing;
5310 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5311 /* Some packets have received acks. If they all have, we can clear
5312 * the transmit queue.
5315 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5316 if (p->header.seq < call->tfirst
5317 && (p->flags & RX_PKTFLAG_ACKED)) {
5324 call->flags |= RX_CALL_TQ_CLEARME;
5326 #endif /* RX_ENABLE_LOCKS */
5327 /* Don't bother doing retransmits if the TQ is cleared. */
5328 if (call->flags & RX_CALL_TQ_CLEARME) {
5329 rxi_ClearTransmitQueue(call, 1);
5331 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5334 /* Always post a resend event, if there is anything in the
5335 * queue, and resend is possible. There should be at least
5336 * one unacknowledged packet in the queue ... otherwise none
5337 * of these packets should be on the queue in the first place.
5339 if (call->resendEvent) {
5340 /* Cancel the existing event and post a new one */
5341 rxevent_Cancel(call->resendEvent, call,
5342 RX_CALL_REFCOUNT_RESEND);
5345 /* The retry time is the retry time on the first unacknowledged
5346 * packet inside the current window */
5348 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5349 /* Don't set timers for packets outside the window */
5350 if (p->header.seq >= call->tfirst + call->twind) {
5354 if (!(p->flags & RX_PKTFLAG_ACKED)
5355 && !clock_IsZero(&p->retryTime)) {
5357 retryTime = p->retryTime;
5362 /* Post a new event to re-run rxi_Start when retries may be needed */
5363 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5364 #ifdef RX_ENABLE_LOCKS
5365 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5367 rxevent_Post2(&retryTime, rxi_StartUnlocked,
5368 (void *)call, 0, istack);
5369 #else /* RX_ENABLE_LOCKS */
5371 rxevent_Post2(&retryTime, rxi_Start, (void *)call,
5373 #endif /* RX_ENABLE_LOCKS */
5376 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5377 } while (call->flags & RX_CALL_NEED_START);
5379 * TQ references no longer protected by this flag; they must remain
5380 * protected by the global lock.
5382 call->flags &= ~RX_CALL_TQ_BUSY;
5383 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5384 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5385 #ifdef RX_ENABLE_LOCKS
5386 osirx_AssertMine(&call->lock, "rxi_Start end");
5387 CV_BROADCAST(&call->cv_tq);
5388 #else /* RX_ENABLE_LOCKS */
5389 osi_rxWakeup(&call->tq);
5390 #endif /* RX_ENABLE_LOCKS */
5393 call->flags |= RX_CALL_NEED_START;
5395 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5397 if (call->resendEvent) {
5398 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5403 /* Also adjusts the keep alive parameters for the call, to reflect
5404 * that we have just sent a packet (so keep alives aren't sent
5407 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5410 register struct rx_connection *conn = call->conn;
5412 /* Stamp each packet with the user supplied status */
5413 p->header.userStatus = call->localStatus;
5415 /* Allow the security object controlling this call's security to
5416 * make any last-minute changes to the packet */
5417 RXS_SendPacket(conn->securityObject, call, p);
5419 /* Since we're about to send SOME sort of packet to the peer, it's
5420 * safe to nuke any scheduled end-of-packets ack */
5421 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5423 /* Actually send the packet, filling in more connection-specific fields */
5424 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5425 MUTEX_EXIT(&call->lock);
5426 rxi_SendPacket(call, conn, p, istack);
5427 MUTEX_ENTER(&call->lock);
5428 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5430 /* Update last send time for this call (for keep-alive
5431 * processing), and for the connection (so that we can discover
5432 * idle connections) */
5433 conn->lastSendTime = call->lastSendTime = clock_Sec();
5437 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5438 * that things are fine. Also called periodically to guarantee that nothing
5439 * falls through the cracks (e.g. (error + dally) connections have keepalive
5440 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5442 * haveCTLock Set if calling from rxi_ReapConnections
5444 #ifdef RX_ENABLE_LOCKS
5446 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5447 #else /* RX_ENABLE_LOCKS */
5449 rxi_CheckCall(register struct rx_call *call)
5450 #endif /* RX_ENABLE_LOCKS */
5452 register struct rx_connection *conn = call->conn;
5454 afs_uint32 deadTime;
5456 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5457 if (call->flags & RX_CALL_TQ_BUSY) {
5458 /* Call is active and will be reset by rxi_Start if it's
5459 * in an error state.
5464 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5466 (((afs_uint32) conn->secondsUntilDead << 10) +
5467 ((afs_uint32) conn->peer->rtt >> 3) +
5468 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5470 /* These are computed to the second (+- 1 second). But that's
5471 * good enough for these values, which should be a significant
5472 * number of seconds. */
5473 if (now > (call->lastReceiveTime + deadTime)) {
5474 if (call->state == RX_STATE_ACTIVE) {
5475 rxi_CallError(call, RX_CALL_DEAD);
5478 #ifdef RX_ENABLE_LOCKS
5479 /* Cancel pending events */
5480 rxevent_Cancel(call->delayedAckEvent, call,
5481 RX_CALL_REFCOUNT_DELAY);
5482 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5483 rxevent_Cancel(call->keepAliveEvent, call,
5484 RX_CALL_REFCOUNT_ALIVE);
5485 if (call->refCount == 0) {
5486 rxi_FreeCall(call, haveCTLock);
5490 #else /* RX_ENABLE_LOCKS */
5493 #endif /* RX_ENABLE_LOCKS */
5495 /* Non-active calls are destroyed if they are not responding
5496 * to pings; active calls are simply flagged in error, so the
5497 * attached process can die reasonably gracefully. */
5499 /* see if we have a non-activity timeout */
5500 if (call->startWait && conn->idleDeadTime
5501 && ((call->startWait + conn->idleDeadTime) < now)) {
5502 if (call->state == RX_STATE_ACTIVE) {
5503 rxi_CallError(call, RX_CALL_TIMEOUT);
5507 /* see if we have a hard timeout */
5508 if (conn->hardDeadTime
5509 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5510 if (call->state == RX_STATE_ACTIVE)
5511 rxi_CallError(call, RX_CALL_TIMEOUT);
5518 /* When a call is in progress, this routine is called occasionally to
5519 * make sure that some traffic has arrived (or been sent to) the peer.
5520 * If nothing has arrived in a reasonable amount of time, the call is
5521 * declared dead; if nothing has been sent for a while, we send a
5522 * keep-alive packet (if we're actually trying to keep the call alive)
5525 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5528 struct rx_connection *conn;
5531 MUTEX_ENTER(&call->lock);
5532 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5533 if (event == call->keepAliveEvent)
5534 call->keepAliveEvent = NULL;
5537 #ifdef RX_ENABLE_LOCKS
5538 if (rxi_CheckCall(call, 0)) {
5539 MUTEX_EXIT(&call->lock);
5542 #else /* RX_ENABLE_LOCKS */
5543 if (rxi_CheckCall(call))
5545 #endif /* RX_ENABLE_LOCKS */
5547 /* Don't try to keep alive dallying calls */
5548 if (call->state == RX_STATE_DALLY) {
5549 MUTEX_EXIT(&call->lock);
5554 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5555 /* Don't try to send keepalives if there is unacknowledged data */
5556 /* the rexmit code should be good enough, this little hack
5557 * doesn't quite work XXX */
5558 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5560 rxi_ScheduleKeepAliveEvent(call);
5561 MUTEX_EXIT(&call->lock);
5566 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5568 if (!call->keepAliveEvent) {
5570 clock_GetTime(&when);
5571 when.sec += call->conn->secondsUntilPing;
5572 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5573 call->keepAliveEvent =
5574 rxevent_Post(&when, rxi_KeepAliveEvent, call, 0);
5578 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5580 rxi_KeepAliveOn(register struct rx_call *call)
5582 /* Pretend last packet received was received now--i.e. if another
5583 * packet isn't received within the keep alive time, then the call
5584 * will die; Initialize last send time to the current time--even
5585 * if a packet hasn't been sent yet. This will guarantee that a
5586 * keep-alive is sent within the ping time */
5587 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5588 rxi_ScheduleKeepAliveEvent(call);
5591 /* This routine is called to send connection abort messages
5592 * that have been delayed to throttle looping clients. */
5594 rxi_SendDelayedConnAbort(struct rxevent *event,
5595 register struct rx_connection *conn, char *dummy)
5598 struct rx_packet *packet;
5600 MUTEX_ENTER(&conn->conn_data_lock);
5601 conn->delayedAbortEvent = NULL;
5602 error = htonl(conn->error);
5604 MUTEX_EXIT(&conn->conn_data_lock);
5605 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5608 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5609 RX_PACKET_TYPE_ABORT, (char *)&error,
5611 rxi_FreePacket(packet);
5615 /* This routine is called to send call abort messages
5616 * that have been delayed to throttle looping clients. */
5618 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5622 struct rx_packet *packet;
5624 MUTEX_ENTER(&call->lock);
5625 call->delayedAbortEvent = NULL;
5626 error = htonl(call->error);
5628 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5631 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5632 (char *)&error, sizeof(error), 0);
5633 rxi_FreePacket(packet);
5635 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5636 MUTEX_EXIT(&call->lock);
5639 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5640 * seconds) to ask the client to authenticate itself. The routine
5641 * issues a challenge to the client, which is obtained from the
5642 * security object associated with the connection */
5644 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5645 void *arg1, int tries)
5647 conn->challengeEvent = NULL;
5648 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5649 register struct rx_packet *packet;
5653 /* We've failed to authenticate for too long.
5654 * Reset any calls waiting for authentication;
5655 * they are all in RX_STATE_PRECALL.
5659 MUTEX_ENTER(&conn->conn_call_lock);
5660 for (i = 0; i < RX_MAXCALLS; i++) {
5661 struct rx_call *call = conn->call[i];
5663 MUTEX_ENTER(&call->lock);
5664 if (call->state == RX_STATE_PRECALL) {
5665 rxi_CallError(call, RX_CALL_DEAD);
5666 rxi_SendCallAbort(call, NULL, 0, 0);
5668 MUTEX_EXIT(&call->lock);
5671 MUTEX_EXIT(&conn->conn_call_lock);
5675 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5677 /* If there's no packet available, do this later. */
5678 RXS_GetChallenge(conn->securityObject, conn, packet);
5679 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5680 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5681 rxi_FreePacket(packet);
5683 clock_GetTime(&when);
5684 when.sec += RX_CHALLENGE_TIMEOUT;
5685 conn->challengeEvent =
5686 rxevent_Post2(&when, rxi_ChallengeEvent, conn, 0,
5691 /* Call this routine to start requesting the client to authenticate
5692 * itself. This will continue until authentication is established,
5693 * the call times out, or an invalid response is returned. The
5694 * security object associated with the connection is asked to create
5695 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5696 * defined earlier. */
5698 rxi_ChallengeOn(register struct rx_connection *conn)
5700 if (!conn->challengeEvent) {
5701 RXS_CreateChallenge(conn->securityObject, conn);
5702 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5707 /* Compute round trip time of the packet provided, in *rttp.
5710 /* rxi_ComputeRoundTripTime is called with peer locked. */
5711 /* sentp and/or peer may be null */
5713 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5714 register struct clock *sentp,
5715 register struct rx_peer *peer)
5717 struct clock thisRtt, *rttp = &thisRtt;
5719 register int rtt_timeout;
5721 clock_GetTime(rttp);
5723 if (clock_Lt(rttp, sentp)) {
5725 return; /* somebody set the clock back, don't count this time. */
5727 clock_Sub(rttp, sentp);
5728 MUTEX_ENTER(&rx_stats_mutex);
5729 if (clock_Lt(rttp, &rx_stats.minRtt))
5730 rx_stats.minRtt = *rttp;
5731 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5732 if (rttp->sec > 60) {
5733 MUTEX_EXIT(&rx_stats_mutex);
5734 return; /* somebody set the clock ahead */
5736 rx_stats.maxRtt = *rttp;
5738 clock_Add(&rx_stats.totalRtt, rttp);
5739 rx_stats.nRttSamples++;
5740 MUTEX_EXIT(&rx_stats_mutex);
5742 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5744 /* Apply VanJacobson round-trip estimations */
5749 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5750 * srtt is stored as fixed point with 3 bits after the binary
5751 * point (i.e., scaled by 8). The following magic is
5752 * equivalent to the smoothing algorithm in rfc793 with an
5753 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5754 * srtt*8 = srtt*8 + rtt - srtt
5755 * srtt = srtt + rtt/8 - srtt/8
5758 delta = MSEC(rttp) - (peer->rtt >> 3);
5762 * We accumulate a smoothed rtt variance (actually, a smoothed
5763 * mean difference), then set the retransmit timer to smoothed
5764 * rtt + 4 times the smoothed variance (was 2x in van's original
5765 * paper, but 4x works better for me, and apparently for him as
5767 * rttvar is stored as
5768 * fixed point with 2 bits after the binary point (scaled by
5769 * 4). The following is equivalent to rfc793 smoothing with
5770 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5771 * replaces rfc793's wired-in beta.
5772 * dev*4 = dev*4 + (|actual - expected| - dev)
5778 delta -= (peer->rtt_dev >> 2);
5779 peer->rtt_dev += delta;
5781 /* I don't have a stored RTT so I start with this value. Since I'm
5782 * probably just starting a call, and will be pushing more data down
5783 * this, I expect congestion to increase rapidly. So I fudge a
5784 * little, and I set deviance to half the rtt. In practice,
5785 * deviance tends to approach something a little less than
5786 * half the smoothed rtt. */
5787 peer->rtt = (MSEC(rttp) << 3) + 8;
5788 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5790 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5791 * the other of these connections is usually in a user process, and can
5792 * be switched and/or swapped out. So on fast, reliable networks, the
5793 * timeout would otherwise be too short.
5795 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5796 clock_Zero(&(peer->timeout));
5797 clock_Addmsec(&(peer->timeout), rtt_timeout);
5799 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)));
5803 /* Find all server connections that have not been active for a long time, and
5806 rxi_ReapConnections(void)
5809 clock_GetTime(&now);
5811 /* Find server connection structures that haven't been used for
5812 * greater than rx_idleConnectionTime */
5814 struct rx_connection **conn_ptr, **conn_end;
5815 int i, havecalls = 0;
5816 MUTEX_ENTER(&rx_connHashTable_lock);
5817 for (conn_ptr = &rx_connHashTable[0], conn_end =
5818 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5820 struct rx_connection *conn, *next;
5821 struct rx_call *call;
5825 for (conn = *conn_ptr; conn; conn = next) {
5826 /* XXX -- Shouldn't the connection be locked? */
5829 for (i = 0; i < RX_MAXCALLS; i++) {
5830 call = conn->call[i];
5833 MUTEX_ENTER(&call->lock);
5834 #ifdef RX_ENABLE_LOCKS
5835 result = rxi_CheckCall(call, 1);
5836 #else /* RX_ENABLE_LOCKS */
5837 result = rxi_CheckCall(call);
5838 #endif /* RX_ENABLE_LOCKS */
5839 MUTEX_EXIT(&call->lock);
5841 /* If CheckCall freed the call, it might
5842 * have destroyed the connection as well,
5843 * which screws up the linked lists.
5849 if (conn->type == RX_SERVER_CONNECTION) {
5850 /* This only actually destroys the connection if
5851 * there are no outstanding calls */
5852 MUTEX_ENTER(&conn->conn_data_lock);
5853 if (!havecalls && !conn->refCount
5854 && ((conn->lastSendTime + rx_idleConnectionTime) <
5856 conn->refCount++; /* it will be decr in rx_DestroyConn */
5857 MUTEX_EXIT(&conn->conn_data_lock);
5858 #ifdef RX_ENABLE_LOCKS
5859 rxi_DestroyConnectionNoLock(conn);
5860 #else /* RX_ENABLE_LOCKS */
5861 rxi_DestroyConnection(conn);
5862 #endif /* RX_ENABLE_LOCKS */
5864 #ifdef RX_ENABLE_LOCKS
5866 MUTEX_EXIT(&conn->conn_data_lock);
5868 #endif /* RX_ENABLE_LOCKS */
5872 #ifdef RX_ENABLE_LOCKS
5873 while (rx_connCleanup_list) {
5874 struct rx_connection *conn;
5875 conn = rx_connCleanup_list;
5876 rx_connCleanup_list = rx_connCleanup_list->next;
5877 MUTEX_EXIT(&rx_connHashTable_lock);
5878 rxi_CleanupConnection(conn);
5879 MUTEX_ENTER(&rx_connHashTable_lock);
5881 MUTEX_EXIT(&rx_connHashTable_lock);
5882 #endif /* RX_ENABLE_LOCKS */
5885 /* Find any peer structures that haven't been used (haven't had an
5886 * associated connection) for greater than rx_idlePeerTime */
5888 struct rx_peer **peer_ptr, **peer_end;
5890 MUTEX_ENTER(&rx_rpc_stats);
5891 MUTEX_ENTER(&rx_peerHashTable_lock);
5892 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5893 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5895 struct rx_peer *peer, *next, *prev;
5896 for (prev = peer = *peer_ptr; peer; peer = next) {
5898 code = MUTEX_TRYENTER(&peer->peer_lock);
5899 if ((code) && (peer->refCount == 0)
5900 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5901 rx_interface_stat_p rpc_stat, nrpc_stat;
5903 MUTEX_EXIT(&peer->peer_lock);
5904 MUTEX_DESTROY(&peer->peer_lock);
5906 (&peer->rpcStats, rpc_stat, nrpc_stat,
5907 rx_interface_stat)) {
5908 unsigned int num_funcs;
5911 queue_Remove(&rpc_stat->queue_header);
5912 queue_Remove(&rpc_stat->all_peers);
5913 num_funcs = rpc_stat->stats[0].func_total;
5915 sizeof(rx_interface_stat_t) +
5916 rpc_stat->stats[0].func_total *
5917 sizeof(rx_function_entry_v1_t);
5919 rxi_Free(rpc_stat, space);
5920 rxi_rpc_peer_stat_cnt -= num_funcs;
5923 MUTEX_ENTER(&rx_stats_mutex);
5924 rx_stats.nPeerStructs--;
5925 MUTEX_EXIT(&rx_stats_mutex);
5926 if (peer == *peer_ptr) {
5933 MUTEX_EXIT(&peer->peer_lock);
5939 MUTEX_EXIT(&rx_peerHashTable_lock);
5940 MUTEX_EXIT(&rx_rpc_stats);
5943 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5944 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5945 * GC, just below. Really, we shouldn't have to keep moving packets from
5946 * one place to another, but instead ought to always know if we can
5947 * afford to hold onto a packet in its particular use. */
5948 MUTEX_ENTER(&rx_freePktQ_lock);
5949 if (rx_waitingForPackets) {
5950 rx_waitingForPackets = 0;
5951 #ifdef RX_ENABLE_LOCKS
5952 CV_BROADCAST(&rx_waitingForPackets_cv);
5954 osi_rxWakeup(&rx_waitingForPackets);
5957 MUTEX_EXIT(&rx_freePktQ_lock);
5959 now.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5960 rxevent_Post(&now, rxi_ReapConnections, 0, 0);
5964 /* rxs_Release - This isn't strictly necessary but, since the macro name from
5965 * rx.h is sort of strange this is better. This is called with a security
5966 * object before it is discarded. Each connection using a security object has
5967 * its own refcount to the object so it won't actually be freed until the last
5968 * connection is destroyed.
5970 * This is the only rxs module call. A hold could also be written but no one
5974 rxs_Release(struct rx_securityClass *aobj)
5976 return RXS_Close(aobj);
5980 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
5981 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
5982 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
5983 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
5985 /* Adjust our estimate of the transmission rate to this peer, given
5986 * that the packet p was just acked. We can adjust peer->timeout and
5987 * call->twind. Pragmatically, this is called
5988 * only with packets of maximal length.
5989 * Called with peer and call locked.
5993 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
5994 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
5996 afs_int32 xferSize, xferMs;
5997 register afs_int32 minTime;
6000 /* Count down packets */
6001 if (peer->rateFlag > 0)
6003 /* Do nothing until we're enabled */
6004 if (peer->rateFlag != 0)
6009 /* Count only when the ack seems legitimate */
6010 switch (ackReason) {
6011 case RX_ACK_REQUESTED:
6013 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6017 case RX_ACK_PING_RESPONSE:
6018 if (p) /* want the response to ping-request, not data send */
6020 clock_GetTime(&newTO);
6021 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6022 clock_Sub(&newTO, &call->pingRequestTime);
6023 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6027 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6034 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));
6036 /* Track only packets that are big enough. */
6037 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6041 /* absorb RTT data (in milliseconds) for these big packets */
6042 if (peer->smRtt == 0) {
6043 peer->smRtt = xferMs;
6045 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6050 if (peer->countDown) {
6054 peer->countDown = 10; /* recalculate only every so often */
6056 /* In practice, we can measure only the RTT for full packets,
6057 * because of the way Rx acks the data that it receives. (If it's
6058 * smaller than a full packet, it often gets implicitly acked
6059 * either by the call response (from a server) or by the next call
6060 * (from a client), and either case confuses transmission times
6061 * with processing times.) Therefore, replace the above
6062 * more-sophisticated processing with a simpler version, where the
6063 * smoothed RTT is kept for full-size packets, and the time to
6064 * transmit a windowful of full-size packets is simply RTT *
6065 * windowSize. Again, we take two steps:
6066 - ensure the timeout is large enough for a single packet's RTT;
6067 - ensure that the window is small enough to fit in the desired timeout.*/
6069 /* First, the timeout check. */
6070 minTime = peer->smRtt;
6071 /* Get a reasonable estimate for a timeout period */
6073 newTO.sec = minTime / 1000;
6074 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6076 /* Increase the timeout period so that we can always do at least
6077 * one packet exchange */
6078 if (clock_Gt(&newTO, &peer->timeout)) {
6080 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));
6082 peer->timeout = newTO;
6085 /* Now, get an estimate for the transmit window size. */
6086 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6087 /* Now, convert to the number of full packets that could fit in a
6088 * reasonable fraction of that interval */
6089 minTime /= (peer->smRtt << 1);
6090 xferSize = minTime; /* (make a copy) */
6092 /* Now clamp the size to reasonable bounds. */
6095 else if (minTime > rx_Window)
6096 minTime = rx_Window;
6097 /* if (minTime != peer->maxWindow) {
6098 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6099 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6100 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6102 peer->maxWindow = minTime;
6103 elide... call->twind = minTime;
6107 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6108 * Discern this by calculating the timeout necessary for rx_Window
6110 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6111 /* calculate estimate for transmission interval in milliseconds */
6112 minTime = rx_Window * peer->smRtt;
6113 if (minTime < 1000) {
6114 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
6115 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6116 peer->timeout.usec, peer->smRtt, peer->packetSize));
6118 newTO.sec = 0; /* cut back on timeout by half a second */
6119 newTO.usec = 500000;
6120 clock_Sub(&peer->timeout, &newTO);
6125 } /* end of rxi_ComputeRate */
6126 #endif /* ADAPT_WINDOW */
6134 #define TRACE_OPTION_DEBUGLOG 4
6142 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6143 0, KEY_QUERY_VALUE, &parmKey);
6144 if (code != ERROR_SUCCESS)
6147 dummyLen = sizeof(TraceOption);
6148 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6149 (BYTE *) &TraceOption, &dummyLen);
6150 if (code == ERROR_SUCCESS) {
6151 rxdebug_active = (TraceOption & TRACE_OPTION_DEBUGLOG) ? 1 : 0;
6153 RegCloseKey (parmKey);
6154 #endif /* AFS_NT40_ENV */
6159 rx_DebugOnOff(int on)
6161 rxdebug_active = on;
6163 #endif /* AFS_NT40_ENV */
6166 /* Don't call this debugging routine directly; use dpf */
6168 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
6169 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
6177 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6180 len = _snprintf(msg, sizeof(msg)-2,
6181 tformat, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10,
6182 a11, a12, a13, a14, a15);
6184 if (msg[len-1] != '\n') {
6188 OutputDebugString(msg);
6193 clock_GetTime(&now);
6194 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
6195 (unsigned int)now.usec / 1000);
6196 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
6203 * This function is used to process the rx_stats structure that is local
6204 * to a process as well as an rx_stats structure received from a remote
6205 * process (via rxdebug). Therefore, it needs to do minimal version
6209 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
6210 afs_int32 freePackets, char version)
6214 if (size != sizeof(struct rx_stats)) {
6216 "Unexpected size of stats structure: was %d, expected %d\n",
6217 size, sizeof(struct rx_stats));
6220 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6223 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6224 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6225 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6226 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6227 s->specialPktAllocFailures);
6229 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6230 s->receivePktAllocFailures, s->sendPktAllocFailures,
6231 s->specialPktAllocFailures);
6235 " greedy %d, " "bogusReads %d (last from host %x), "
6236 "noPackets %d, " "noBuffers %d, " "selects %d, "
6237 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6238 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6239 s->selects, s->sendSelects);
6241 fprintf(file, " packets read: ");
6242 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6243 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6245 fprintf(file, "\n");
6248 " other read counters: data %d, " "ack %d, " "dup %d "
6249 "spurious %d " "dally %d\n", s->dataPacketsRead,
6250 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6251 s->ignorePacketDally);
6253 fprintf(file, " packets sent: ");
6254 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6255 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6257 fprintf(file, "\n");
6260 " other send counters: ack %d, " "data %d (not resends), "
6261 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6262 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6263 s->dataPacketsPushed, s->ignoreAckedPacket);
6266 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6267 s->netSendFailures, (int)s->fatalErrors);
6269 if (s->nRttSamples) {
6270 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6271 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6273 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6274 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6278 " %d server connections, " "%d client connections, "
6279 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6280 s->nServerConns, s->nClientConns, s->nPeerStructs,
6281 s->nCallStructs, s->nFreeCallStructs);
6283 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6284 fprintf(file, " %d clock updates\n", clock_nUpdates);
6289 /* for backward compatibility */
6291 rx_PrintStats(FILE * file)
6293 MUTEX_ENTER(&rx_stats_mutex);
6294 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6296 MUTEX_EXIT(&rx_stats_mutex);
6300 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6302 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6303 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6304 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6307 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6308 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6309 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6312 " Packet size %d, " "max in packet skew %d, "
6313 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6314 (int)peer->outPacketSkew);
6317 #ifdef AFS_PTHREAD_ENV
6319 * This mutex protects the following static variables:
6323 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6324 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6326 #define LOCK_RX_DEBUG
6327 #define UNLOCK_RX_DEBUG
6328 #endif /* AFS_PTHREAD_ENV */
6331 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6332 u_char type, void *inputData, size_t inputLength,
6333 void *outputData, size_t outputLength)
6335 static afs_int32 counter = 100;
6336 time_t waitTime, waitCount, startTime, endTime;
6337 struct rx_header theader;
6339 register afs_int32 code;
6340 struct timeval tv_now, tv_wake, tv_delta;
6341 struct sockaddr_in taddr, faddr;
6346 startTime = time(0);
6352 tp = &tbuffer[sizeof(struct rx_header)];
6353 taddr.sin_family = AF_INET;
6354 taddr.sin_port = remotePort;
6355 taddr.sin_addr.s_addr = remoteAddr;
6356 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6357 taddr.sin_len = sizeof(struct sockaddr_in);
6360 memset(&theader, 0, sizeof(theader));
6361 theader.epoch = htonl(999);
6363 theader.callNumber = htonl(counter);
6366 theader.type = type;
6367 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6368 theader.serviceId = 0;
6370 memcpy(tbuffer, &theader, sizeof(theader));
6371 memcpy(tp, inputData, inputLength);
6373 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6374 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6376 /* see if there's a packet available */
6377 gettimeofday(&tv_wake,0);
6378 tv_wake.tv_sec += waitTime;
6381 FD_SET(socket, &imask);
6382 tv_delta.tv_sec = tv_wake.tv_sec;
6383 tv_delta.tv_usec = tv_wake.tv_usec;
6384 gettimeofday(&tv_now, 0);
6386 if (tv_delta.tv_usec < tv_now.tv_usec) {
6388 tv_delta.tv_usec += 1000000;
6391 tv_delta.tv_usec -= tv_now.tv_usec;
6393 if (tv_delta.tv_sec < tv_now.tv_sec) {
6397 tv_delta.tv_sec -= tv_now.tv_sec;
6399 code = select(socket + 1, &imask, 0, 0, &tv_delta);
6400 if (code == 1 && FD_ISSET(socket, &imask)) {
6401 /* now receive a packet */
6402 faddrLen = sizeof(struct sockaddr_in);
6404 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6405 (struct sockaddr *)&faddr, &faddrLen);
6408 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6409 if (counter == ntohl(theader.callNumber))
6417 /* see if we've timed out */
6425 code -= sizeof(struct rx_header);
6426 if (code > outputLength)
6427 code = outputLength;
6428 memcpy(outputData, tp, code);
6433 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6434 afs_uint16 remotePort, struct rx_debugStats * stat,
6435 afs_uint32 * supportedValues)
6437 struct rx_debugIn in;
6440 *supportedValues = 0;
6441 in.type = htonl(RX_DEBUGI_GETSTATS);
6444 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6445 &in, sizeof(in), stat, sizeof(*stat));
6448 * If the call was successful, fixup the version and indicate
6449 * what contents of the stat structure are valid.
6450 * Also do net to host conversion of fields here.
6454 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6455 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6457 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6458 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6460 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6461 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6463 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6464 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6466 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6467 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6469 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6470 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6472 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6473 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6475 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6476 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6479 stat->nFreePackets = ntohl(stat->nFreePackets);
6480 stat->packetReclaims = ntohl(stat->packetReclaims);
6481 stat->callsExecuted = ntohl(stat->callsExecuted);
6482 stat->nWaiting = ntohl(stat->nWaiting);
6483 stat->idleThreads = ntohl(stat->idleThreads);
6490 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6491 afs_uint16 remotePort, struct rx_stats * stat,
6492 afs_uint32 * supportedValues)
6494 struct rx_debugIn in;
6495 afs_int32 *lp = (afs_int32 *) stat;
6500 * supportedValues is currently unused, but added to allow future
6501 * versioning of this function.
6504 *supportedValues = 0;
6505 in.type = htonl(RX_DEBUGI_RXSTATS);
6507 memset(stat, 0, sizeof(*stat));
6509 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6510 &in, sizeof(in), stat, sizeof(*stat));
6515 * Do net to host conversion here
6518 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6527 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6528 afs_uint16 remotePort, size_t version_length,
6532 return MakeDebugCall(socket, remoteAddr, remotePort,
6533 RX_PACKET_TYPE_VERSION, a, 1, version,
6538 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6539 afs_uint16 remotePort, afs_int32 * nextConnection,
6540 int allConnections, afs_uint32 debugSupportedValues,
6541 struct rx_debugConn * conn,
6542 afs_uint32 * supportedValues)
6544 struct rx_debugIn in;
6549 * supportedValues is currently unused, but added to allow future
6550 * versioning of this function.
6553 *supportedValues = 0;
6554 if (allConnections) {
6555 in.type = htonl(RX_DEBUGI_GETALLCONN);
6557 in.type = htonl(RX_DEBUGI_GETCONN);
6559 in.index = htonl(*nextConnection);
6560 memset(conn, 0, sizeof(*conn));
6562 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6563 &in, sizeof(in), conn, sizeof(*conn));
6566 *nextConnection += 1;
6569 * Convert old connection format to new structure.
6572 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6573 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6574 #define MOVEvL(a) (conn->a = vL->a)
6576 /* any old or unrecognized version... */
6577 for (i = 0; i < RX_MAXCALLS; i++) {
6578 MOVEvL(callState[i]);
6579 MOVEvL(callMode[i]);
6580 MOVEvL(callFlags[i]);
6581 MOVEvL(callOther[i]);
6583 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6584 MOVEvL(secStats.type);
6585 MOVEvL(secStats.level);
6586 MOVEvL(secStats.flags);
6587 MOVEvL(secStats.expires);
6588 MOVEvL(secStats.packetsReceived);
6589 MOVEvL(secStats.packetsSent);
6590 MOVEvL(secStats.bytesReceived);
6591 MOVEvL(secStats.bytesSent);
6596 * Do net to host conversion here
6598 * I don't convert host or port since we are most likely
6599 * going to want these in NBO.
6601 conn->cid = ntohl(conn->cid);
6602 conn->serial = ntohl(conn->serial);
6603 for (i = 0; i < RX_MAXCALLS; i++) {
6604 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6606 conn->error = ntohl(conn->error);
6607 conn->secStats.flags = ntohl(conn->secStats.flags);
6608 conn->secStats.expires = ntohl(conn->secStats.expires);
6609 conn->secStats.packetsReceived =
6610 ntohl(conn->secStats.packetsReceived);
6611 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6612 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6613 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6614 conn->epoch = ntohl(conn->epoch);
6615 conn->natMTU = ntohl(conn->natMTU);
6622 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6623 afs_uint16 remotePort, afs_int32 * nextPeer,
6624 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6625 afs_uint32 * supportedValues)
6627 struct rx_debugIn in;
6631 * supportedValues is currently unused, but added to allow future
6632 * versioning of this function.
6635 *supportedValues = 0;
6636 in.type = htonl(RX_DEBUGI_GETPEER);
6637 in.index = htonl(*nextPeer);
6638 memset(peer, 0, sizeof(*peer));
6640 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6641 &in, sizeof(in), peer, sizeof(*peer));
6647 * Do net to host conversion here
6649 * I don't convert host or port since we are most likely
6650 * going to want these in NBO.
6652 peer->ifMTU = ntohs(peer->ifMTU);
6653 peer->idleWhen = ntohl(peer->idleWhen);
6654 peer->refCount = ntohs(peer->refCount);
6655 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6656 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6657 peer->rtt = ntohl(peer->rtt);
6658 peer->rtt_dev = ntohl(peer->rtt_dev);
6659 peer->timeout.sec = ntohl(peer->timeout.sec);
6660 peer->timeout.usec = ntohl(peer->timeout.usec);
6661 peer->nSent = ntohl(peer->nSent);
6662 peer->reSends = ntohl(peer->reSends);
6663 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6664 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6665 peer->rateFlag = ntohl(peer->rateFlag);
6666 peer->natMTU = ntohs(peer->natMTU);
6667 peer->maxMTU = ntohs(peer->maxMTU);
6668 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6669 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6670 peer->MTU = ntohs(peer->MTU);
6671 peer->cwind = ntohs(peer->cwind);
6672 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6673 peer->congestSeq = ntohs(peer->congestSeq);
6674 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6675 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6676 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6677 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6682 #endif /* RXDEBUG */
6687 struct rx_serverQueueEntry *np;
6690 register struct rx_call *call;
6691 register struct rx_serverQueueEntry *sq;
6695 if (rxinit_status == 1) {
6697 return; /* Already shutdown. */
6701 #ifndef AFS_PTHREAD_ENV
6702 FD_ZERO(&rx_selectMask);
6703 #endif /* AFS_PTHREAD_ENV */
6704 rxi_dataQuota = RX_MAX_QUOTA;
6705 #ifndef AFS_PTHREAD_ENV
6707 #endif /* AFS_PTHREAD_ENV */
6710 #ifndef AFS_PTHREAD_ENV
6711 #ifndef AFS_USE_GETTIMEOFDAY
6713 #endif /* AFS_USE_GETTIMEOFDAY */
6714 #endif /* AFS_PTHREAD_ENV */
6716 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6717 call = queue_First(&rx_freeCallQueue, rx_call);
6719 rxi_Free(call, sizeof(struct rx_call));
6722 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6723 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6729 struct rx_peer **peer_ptr, **peer_end;
6730 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6731 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6733 struct rx_peer *peer, *next;
6734 for (peer = *peer_ptr; peer; peer = next) {
6735 rx_interface_stat_p rpc_stat, nrpc_stat;
6738 (&peer->rpcStats, rpc_stat, nrpc_stat,
6739 rx_interface_stat)) {
6740 unsigned int num_funcs;
6743 queue_Remove(&rpc_stat->queue_header);
6744 queue_Remove(&rpc_stat->all_peers);
6745 num_funcs = rpc_stat->stats[0].func_total;
6747 sizeof(rx_interface_stat_t) +
6748 rpc_stat->stats[0].func_total *
6749 sizeof(rx_function_entry_v1_t);
6751 rxi_Free(rpc_stat, space);
6752 MUTEX_ENTER(&rx_rpc_stats);
6753 rxi_rpc_peer_stat_cnt -= num_funcs;
6754 MUTEX_EXIT(&rx_rpc_stats);
6758 MUTEX_ENTER(&rx_stats_mutex);
6759 rx_stats.nPeerStructs--;
6760 MUTEX_EXIT(&rx_stats_mutex);
6764 for (i = 0; i < RX_MAX_SERVICES; i++) {
6766 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6768 for (i = 0; i < rx_hashTableSize; i++) {
6769 register struct rx_connection *tc, *ntc;
6770 MUTEX_ENTER(&rx_connHashTable_lock);
6771 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6773 for (j = 0; j < RX_MAXCALLS; j++) {
6775 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6778 rxi_Free(tc, sizeof(*tc));
6780 MUTEX_EXIT(&rx_connHashTable_lock);
6783 MUTEX_ENTER(&freeSQEList_lock);
6785 while ((np = rx_FreeSQEList)) {
6786 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6787 MUTEX_DESTROY(&np->lock);
6788 rxi_Free(np, sizeof(*np));
6791 MUTEX_EXIT(&freeSQEList_lock);
6792 MUTEX_DESTROY(&freeSQEList_lock);
6793 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6794 MUTEX_DESTROY(&rx_connHashTable_lock);
6795 MUTEX_DESTROY(&rx_peerHashTable_lock);
6796 MUTEX_DESTROY(&rx_serverPool_lock);
6798 osi_Free(rx_connHashTable,
6799 rx_hashTableSize * sizeof(struct rx_connection *));
6800 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6802 UNPIN(rx_connHashTable,
6803 rx_hashTableSize * sizeof(struct rx_connection *));
6804 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6806 rxi_FreeAllPackets();
6808 MUTEX_ENTER(&rx_stats_mutex);
6809 rxi_dataQuota = RX_MAX_QUOTA;
6810 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6811 MUTEX_EXIT(&rx_stats_mutex);
6817 #ifdef RX_ENABLE_LOCKS
6819 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6821 if (!MUTEX_ISMINE(lockaddr))
6822 osi_Panic("Lock not held: %s", msg);
6824 #endif /* RX_ENABLE_LOCKS */
6829 * Routines to implement connection specific data.
6833 rx_KeyCreate(rx_destructor_t rtn)
6836 MUTEX_ENTER(&rxi_keyCreate_lock);
6837 key = rxi_keyCreate_counter++;
6838 rxi_keyCreate_destructor = (rx_destructor_t *)
6839 realloc((void *)rxi_keyCreate_destructor,
6840 (key + 1) * sizeof(rx_destructor_t));
6841 rxi_keyCreate_destructor[key] = rtn;
6842 MUTEX_EXIT(&rxi_keyCreate_lock);
6847 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6850 MUTEX_ENTER(&conn->conn_data_lock);
6851 if (!conn->specific) {
6852 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6853 for (i = 0; i < key; i++)
6854 conn->specific[i] = NULL;
6855 conn->nSpecific = key + 1;
6856 conn->specific[key] = ptr;
6857 } else if (key >= conn->nSpecific) {
6858 conn->specific = (void **)
6859 realloc(conn->specific, (key + 1) * sizeof(void *));
6860 for (i = conn->nSpecific; i < key; i++)
6861 conn->specific[i] = NULL;
6862 conn->nSpecific = key + 1;
6863 conn->specific[key] = ptr;
6865 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6866 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6867 conn->specific[key] = ptr;
6869 MUTEX_EXIT(&conn->conn_data_lock);
6873 rx_GetSpecific(struct rx_connection *conn, int key)
6876 MUTEX_ENTER(&conn->conn_data_lock);
6877 if (key >= conn->nSpecific)
6880 ptr = conn->specific[key];
6881 MUTEX_EXIT(&conn->conn_data_lock);
6885 #endif /* !KERNEL */
6888 * processStats is a queue used to store the statistics for the local
6889 * process. Its contents are similar to the contents of the rpcStats
6890 * queue on a rx_peer structure, but the actual data stored within
6891 * this queue contains totals across the lifetime of the process (assuming
6892 * the stats have not been reset) - unlike the per peer structures
6893 * which can come and go based upon the peer lifetime.
6896 static struct rx_queue processStats = { &processStats, &processStats };
6899 * peerStats is a queue used to store the statistics for all peer structs.
6900 * Its contents are the union of all the peer rpcStats queues.
6903 static struct rx_queue peerStats = { &peerStats, &peerStats };
6906 * rxi_monitor_processStats is used to turn process wide stat collection
6910 static int rxi_monitor_processStats = 0;
6913 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6916 static int rxi_monitor_peerStats = 0;
6919 * rxi_AddRpcStat - given all of the information for a particular rpc
6920 * call, create (if needed) and update the stat totals for the rpc.
6924 * IN stats - the queue of stats that will be updated with the new value
6926 * IN rxInterface - a unique number that identifies the rpc interface
6928 * IN currentFunc - the index of the function being invoked
6930 * IN totalFunc - the total number of functions in this interface
6932 * IN queueTime - the amount of time this function waited for a thread
6934 * IN execTime - the amount of time this function invocation took to execute
6936 * IN bytesSent - the number bytes sent by this invocation
6938 * IN bytesRcvd - the number bytes received by this invocation
6940 * IN isServer - if true, this invocation was made to a server
6942 * IN remoteHost - the ip address of the remote host
6944 * IN remotePort - the port of the remote host
6946 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6948 * INOUT counter - if a new stats structure is allocated, the counter will
6949 * be updated with the new number of allocated stat structures
6957 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6958 afs_uint32 currentFunc, afs_uint32 totalFunc,
6959 struct clock *queueTime, struct clock *execTime,
6960 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6961 afs_uint32 remoteHost, afs_uint32 remotePort,
6962 int addToPeerList, unsigned int *counter)
6965 rx_interface_stat_p rpc_stat, nrpc_stat;
6968 * See if there's already a structure for this interface
6971 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6972 if ((rpc_stat->stats[0].interfaceId == rxInterface)
6973 && (rpc_stat->stats[0].remote_is_server == isServer))
6978 * Didn't find a match so allocate a new structure and add it to the
6982 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
6983 || (rpc_stat->stats[0].interfaceId != rxInterface)
6984 || (rpc_stat->stats[0].remote_is_server != isServer)) {
6989 sizeof(rx_interface_stat_t) +
6990 totalFunc * sizeof(rx_function_entry_v1_t);
6992 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
6993 if (rpc_stat == NULL) {
6997 *counter += totalFunc;
6998 for (i = 0; i < totalFunc; i++) {
6999 rpc_stat->stats[i].remote_peer = remoteHost;
7000 rpc_stat->stats[i].remote_port = remotePort;
7001 rpc_stat->stats[i].remote_is_server = isServer;
7002 rpc_stat->stats[i].interfaceId = rxInterface;
7003 rpc_stat->stats[i].func_total = totalFunc;
7004 rpc_stat->stats[i].func_index = i;
7005 hzero(rpc_stat->stats[i].invocations);
7006 hzero(rpc_stat->stats[i].bytes_sent);
7007 hzero(rpc_stat->stats[i].bytes_rcvd);
7008 rpc_stat->stats[i].queue_time_sum.sec = 0;
7009 rpc_stat->stats[i].queue_time_sum.usec = 0;
7010 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7011 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7012 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7013 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7014 rpc_stat->stats[i].queue_time_max.sec = 0;
7015 rpc_stat->stats[i].queue_time_max.usec = 0;
7016 rpc_stat->stats[i].execution_time_sum.sec = 0;
7017 rpc_stat->stats[i].execution_time_sum.usec = 0;
7018 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7019 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7020 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7021 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7022 rpc_stat->stats[i].execution_time_max.sec = 0;
7023 rpc_stat->stats[i].execution_time_max.usec = 0;
7025 queue_Prepend(stats, rpc_stat);
7026 if (addToPeerList) {
7027 queue_Prepend(&peerStats, &rpc_stat->all_peers);
7032 * Increment the stats for this function
7035 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
7036 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
7037 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
7038 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
7039 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7040 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7041 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7043 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7044 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7046 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7047 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7049 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7050 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7052 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7053 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7061 * rx_IncrementTimeAndCount - increment the times and count for a particular
7066 * IN peer - the peer who invoked the rpc
7068 * IN rxInterface - a unique number that identifies the rpc interface
7070 * IN currentFunc - the index of the function being invoked
7072 * IN totalFunc - the total number of functions in this interface
7074 * IN queueTime - the amount of time this function waited for a thread
7076 * IN execTime - the amount of time this function invocation took to execute
7078 * IN bytesSent - the number bytes sent by this invocation
7080 * IN bytesRcvd - the number bytes received by this invocation
7082 * IN isServer - if true, this invocation was made to a server
7090 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7091 afs_uint32 currentFunc, afs_uint32 totalFunc,
7092 struct clock *queueTime, struct clock *execTime,
7093 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7097 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
7100 MUTEX_ENTER(&rx_rpc_stats);
7101 MUTEX_ENTER(&peer->peer_lock);
7103 if (rxi_monitor_peerStats) {
7104 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7105 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7106 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7109 if (rxi_monitor_processStats) {
7110 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7111 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7112 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7115 MUTEX_EXIT(&peer->peer_lock);
7116 MUTEX_EXIT(&rx_rpc_stats);
7121 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7125 * IN callerVersion - the rpc stat version of the caller.
7127 * IN count - the number of entries to marshall.
7129 * IN stats - pointer to stats to be marshalled.
7131 * OUT ptr - Where to store the marshalled data.
7138 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7139 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7145 * We only support the first version
7147 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7148 *(ptr++) = stats->remote_peer;
7149 *(ptr++) = stats->remote_port;
7150 *(ptr++) = stats->remote_is_server;
7151 *(ptr++) = stats->interfaceId;
7152 *(ptr++) = stats->func_total;
7153 *(ptr++) = stats->func_index;
7154 *(ptr++) = hgethi(stats->invocations);
7155 *(ptr++) = hgetlo(stats->invocations);
7156 *(ptr++) = hgethi(stats->bytes_sent);
7157 *(ptr++) = hgetlo(stats->bytes_sent);
7158 *(ptr++) = hgethi(stats->bytes_rcvd);
7159 *(ptr++) = hgetlo(stats->bytes_rcvd);
7160 *(ptr++) = stats->queue_time_sum.sec;
7161 *(ptr++) = stats->queue_time_sum.usec;
7162 *(ptr++) = stats->queue_time_sum_sqr.sec;
7163 *(ptr++) = stats->queue_time_sum_sqr.usec;
7164 *(ptr++) = stats->queue_time_min.sec;
7165 *(ptr++) = stats->queue_time_min.usec;
7166 *(ptr++) = stats->queue_time_max.sec;
7167 *(ptr++) = stats->queue_time_max.usec;
7168 *(ptr++) = stats->execution_time_sum.sec;
7169 *(ptr++) = stats->execution_time_sum.usec;
7170 *(ptr++) = stats->execution_time_sum_sqr.sec;
7171 *(ptr++) = stats->execution_time_sum_sqr.usec;
7172 *(ptr++) = stats->execution_time_min.sec;
7173 *(ptr++) = stats->execution_time_min.usec;
7174 *(ptr++) = stats->execution_time_max.sec;
7175 *(ptr++) = stats->execution_time_max.usec;
7181 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7186 * IN callerVersion - the rpc stat version of the caller
7188 * OUT myVersion - the rpc stat version of this function
7190 * OUT clock_sec - local time seconds
7192 * OUT clock_usec - local time microseconds
7194 * OUT allocSize - the number of bytes allocated to contain stats
7196 * OUT statCount - the number stats retrieved from this process.
7198 * OUT stats - the actual stats retrieved from this process.
7202 * Returns void. If successful, stats will != NULL.
7206 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7207 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7208 size_t * allocSize, afs_uint32 * statCount,
7209 afs_uint32 ** stats)
7219 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7222 * Check to see if stats are enabled
7225 MUTEX_ENTER(&rx_rpc_stats);
7226 if (!rxi_monitor_processStats) {
7227 MUTEX_EXIT(&rx_rpc_stats);
7231 clock_GetTime(&now);
7232 *clock_sec = now.sec;
7233 *clock_usec = now.usec;
7236 * Allocate the space based upon the caller version
7238 * If the client is at an older version than we are,
7239 * we return the statistic data in the older data format, but
7240 * we still return our version number so the client knows we
7241 * are maintaining more data than it can retrieve.
7244 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7245 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7246 *statCount = rxi_rpc_process_stat_cnt;
7249 * This can't happen yet, but in the future version changes
7250 * can be handled by adding additional code here
7254 if (space > (size_t) 0) {
7256 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7259 rx_interface_stat_p rpc_stat, nrpc_stat;
7263 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7265 * Copy the data based upon the caller version
7267 rx_MarshallProcessRPCStats(callerVersion,
7268 rpc_stat->stats[0].func_total,
7269 rpc_stat->stats, &ptr);
7275 MUTEX_EXIT(&rx_rpc_stats);
7280 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7284 * IN callerVersion - the rpc stat version of the caller
7286 * OUT myVersion - the rpc stat version of this function
7288 * OUT clock_sec - local time seconds
7290 * OUT clock_usec - local time microseconds
7292 * OUT allocSize - the number of bytes allocated to contain stats
7294 * OUT statCount - the number of stats retrieved from the individual
7297 * OUT stats - the actual stats retrieved from the individual peer structures.
7301 * Returns void. If successful, stats will != NULL.
7305 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7306 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7307 size_t * allocSize, afs_uint32 * statCount,
7308 afs_uint32 ** stats)
7318 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7321 * Check to see if stats are enabled
7324 MUTEX_ENTER(&rx_rpc_stats);
7325 if (!rxi_monitor_peerStats) {
7326 MUTEX_EXIT(&rx_rpc_stats);
7330 clock_GetTime(&now);
7331 *clock_sec = now.sec;
7332 *clock_usec = now.usec;
7335 * Allocate the space based upon the caller version
7337 * If the client is at an older version than we are,
7338 * we return the statistic data in the older data format, but
7339 * we still return our version number so the client knows we
7340 * are maintaining more data than it can retrieve.
7343 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7344 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7345 *statCount = rxi_rpc_peer_stat_cnt;
7348 * This can't happen yet, but in the future version changes
7349 * can be handled by adding additional code here
7353 if (space > (size_t) 0) {
7355 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7358 rx_interface_stat_p rpc_stat, nrpc_stat;
7362 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7364 * We have to fix the offset of rpc_stat since we are
7365 * keeping this structure on two rx_queues. The rx_queue
7366 * package assumes that the rx_queue member is the first
7367 * member of the structure. That is, rx_queue assumes that
7368 * any one item is only on one queue at a time. We are
7369 * breaking that assumption and so we have to do a little
7370 * math to fix our pointers.
7373 fix_offset = (char *)rpc_stat;
7374 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7375 rpc_stat = (rx_interface_stat_p) fix_offset;
7378 * Copy the data based upon the caller version
7380 rx_MarshallProcessRPCStats(callerVersion,
7381 rpc_stat->stats[0].func_total,
7382 rpc_stat->stats, &ptr);
7388 MUTEX_EXIT(&rx_rpc_stats);
7393 * rx_FreeRPCStats - free memory allocated by
7394 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7398 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7399 * rx_RetrievePeerRPCStats
7401 * IN allocSize - the number of bytes in stats.
7409 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7411 rxi_Free(stats, allocSize);
7415 * rx_queryProcessRPCStats - see if process rpc stat collection is
7416 * currently enabled.
7422 * Returns 0 if stats are not enabled != 0 otherwise
7426 rx_queryProcessRPCStats(void)
7429 MUTEX_ENTER(&rx_rpc_stats);
7430 rc = rxi_monitor_processStats;
7431 MUTEX_EXIT(&rx_rpc_stats);
7436 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7442 * Returns 0 if stats are not enabled != 0 otherwise
7446 rx_queryPeerRPCStats(void)
7449 MUTEX_ENTER(&rx_rpc_stats);
7450 rc = rxi_monitor_peerStats;
7451 MUTEX_EXIT(&rx_rpc_stats);
7456 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7466 rx_enableProcessRPCStats(void)
7468 MUTEX_ENTER(&rx_rpc_stats);
7469 rx_enable_stats = 1;
7470 rxi_monitor_processStats = 1;
7471 MUTEX_EXIT(&rx_rpc_stats);
7475 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7485 rx_enablePeerRPCStats(void)
7487 MUTEX_ENTER(&rx_rpc_stats);
7488 rx_enable_stats = 1;
7489 rxi_monitor_peerStats = 1;
7490 MUTEX_EXIT(&rx_rpc_stats);
7494 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7504 rx_disableProcessRPCStats(void)
7506 rx_interface_stat_p rpc_stat, nrpc_stat;
7509 MUTEX_ENTER(&rx_rpc_stats);
7512 * Turn off process statistics and if peer stats is also off, turn
7516 rxi_monitor_processStats = 0;
7517 if (rxi_monitor_peerStats == 0) {
7518 rx_enable_stats = 0;
7521 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7522 unsigned int num_funcs = 0;
7525 queue_Remove(rpc_stat);
7526 num_funcs = rpc_stat->stats[0].func_total;
7528 sizeof(rx_interface_stat_t) +
7529 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7531 rxi_Free(rpc_stat, space);
7532 rxi_rpc_process_stat_cnt -= num_funcs;
7534 MUTEX_EXIT(&rx_rpc_stats);
7538 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7548 rx_disablePeerRPCStats(void)
7550 struct rx_peer **peer_ptr, **peer_end;
7553 MUTEX_ENTER(&rx_rpc_stats);
7556 * Turn off peer statistics and if process stats is also off, turn
7560 rxi_monitor_peerStats = 0;
7561 if (rxi_monitor_processStats == 0) {
7562 rx_enable_stats = 0;
7565 MUTEX_ENTER(&rx_peerHashTable_lock);
7566 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7567 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7569 struct rx_peer *peer, *next, *prev;
7570 for (prev = peer = *peer_ptr; peer; peer = next) {
7572 code = MUTEX_TRYENTER(&peer->peer_lock);
7574 rx_interface_stat_p rpc_stat, nrpc_stat;
7577 (&peer->rpcStats, rpc_stat, nrpc_stat,
7578 rx_interface_stat)) {
7579 unsigned int num_funcs = 0;
7582 queue_Remove(&rpc_stat->queue_header);
7583 queue_Remove(&rpc_stat->all_peers);
7584 num_funcs = rpc_stat->stats[0].func_total;
7586 sizeof(rx_interface_stat_t) +
7587 rpc_stat->stats[0].func_total *
7588 sizeof(rx_function_entry_v1_t);
7590 rxi_Free(rpc_stat, space);
7591 rxi_rpc_peer_stat_cnt -= num_funcs;
7593 MUTEX_EXIT(&peer->peer_lock);
7594 if (prev == *peer_ptr) {
7604 MUTEX_EXIT(&rx_peerHashTable_lock);
7605 MUTEX_EXIT(&rx_rpc_stats);
7609 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7614 * IN clearFlag - flag indicating which stats to clear
7622 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7624 rx_interface_stat_p rpc_stat, nrpc_stat;
7626 MUTEX_ENTER(&rx_rpc_stats);
7628 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7629 unsigned int num_funcs = 0, i;
7630 num_funcs = rpc_stat->stats[0].func_total;
7631 for (i = 0; i < num_funcs; i++) {
7632 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7633 hzero(rpc_stat->stats[i].invocations);
7635 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7636 hzero(rpc_stat->stats[i].bytes_sent);
7638 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7639 hzero(rpc_stat->stats[i].bytes_rcvd);
7641 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7642 rpc_stat->stats[i].queue_time_sum.sec = 0;
7643 rpc_stat->stats[i].queue_time_sum.usec = 0;
7645 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7646 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7647 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7649 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7650 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7651 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7653 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7654 rpc_stat->stats[i].queue_time_max.sec = 0;
7655 rpc_stat->stats[i].queue_time_max.usec = 0;
7657 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7658 rpc_stat->stats[i].execution_time_sum.sec = 0;
7659 rpc_stat->stats[i].execution_time_sum.usec = 0;
7661 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7662 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7663 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7665 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7666 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7667 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7669 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7670 rpc_stat->stats[i].execution_time_max.sec = 0;
7671 rpc_stat->stats[i].execution_time_max.usec = 0;
7676 MUTEX_EXIT(&rx_rpc_stats);
7680 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7685 * IN clearFlag - flag indicating which stats to clear
7693 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7695 rx_interface_stat_p rpc_stat, nrpc_stat;
7697 MUTEX_ENTER(&rx_rpc_stats);
7699 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7700 unsigned int num_funcs = 0, i;
7703 * We have to fix the offset of rpc_stat since we are
7704 * keeping this structure on two rx_queues. The rx_queue
7705 * package assumes that the rx_queue member is the first
7706 * member of the structure. That is, rx_queue assumes that
7707 * any one item is only on one queue at a time. We are
7708 * breaking that assumption and so we have to do a little
7709 * math to fix our pointers.
7712 fix_offset = (char *)rpc_stat;
7713 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7714 rpc_stat = (rx_interface_stat_p) fix_offset;
7716 num_funcs = rpc_stat->stats[0].func_total;
7717 for (i = 0; i < num_funcs; i++) {
7718 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7719 hzero(rpc_stat->stats[i].invocations);
7721 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7722 hzero(rpc_stat->stats[i].bytes_sent);
7724 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7725 hzero(rpc_stat->stats[i].bytes_rcvd);
7727 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7728 rpc_stat->stats[i].queue_time_sum.sec = 0;
7729 rpc_stat->stats[i].queue_time_sum.usec = 0;
7731 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7732 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7733 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7735 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7736 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7737 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7739 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7740 rpc_stat->stats[i].queue_time_max.sec = 0;
7741 rpc_stat->stats[i].queue_time_max.usec = 0;
7743 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7744 rpc_stat->stats[i].execution_time_sum.sec = 0;
7745 rpc_stat->stats[i].execution_time_sum.usec = 0;
7747 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7748 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7749 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7751 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7752 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7753 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7755 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7756 rpc_stat->stats[i].execution_time_max.sec = 0;
7757 rpc_stat->stats[i].execution_time_max.usec = 0;
7762 MUTEX_EXIT(&rx_rpc_stats);
7766 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7767 * is authorized to enable/disable/clear RX statistics.
7769 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7772 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7774 rxi_rxstat_userok = proc;
7778 rx_RxStatUserOk(struct rx_call *call)
7780 if (!rxi_rxstat_userok)
7782 return rxi_rxstat_userok(call);
7787 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
7788 * function in the MSVC runtime DLL (msvcrt.dll).
7790 * Note: the system serializes calls to this function.
7793 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
7794 DWORD reason, /* reason function is being called */
7795 LPVOID reserved) /* reserved for future use */
7798 case DLL_PROCESS_ATTACH:
7799 /* library is being attached to a process */
7803 case DLL_PROCESS_DETACH: