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 /* This doesn't work in all compilers with void (they're buggy), so fake it
772 conn->securityData = (VOID *) 0;
773 conn->securityIndex = serviceSecurityIndex;
774 rx_SetConnDeadTime(conn, rx_connDeadTime);
775 conn->ackRate = RX_FAST_ACK_RATE;
777 conn->specific = NULL;
778 conn->challengeEvent = NULL;
779 conn->delayedAbortEvent = NULL;
780 conn->abortCount = 0;
783 RXS_NewConnection(securityObject, conn);
785 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
787 conn->refCount++; /* no lock required since only this thread knows... */
788 conn->next = rx_connHashTable[hashindex];
789 rx_connHashTable[hashindex] = conn;
790 MUTEX_ENTER(&rx_stats_mutex);
791 rx_stats.nClientConns++;
792 MUTEX_EXIT(&rx_stats_mutex);
794 MUTEX_EXIT(&rx_connHashTable_lock);
800 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
802 /* The idea is to set the dead time to a value that allows several
803 * keepalives to be dropped without timing out the connection. */
804 conn->secondsUntilDead = MAX(seconds, 6);
805 conn->secondsUntilPing = conn->secondsUntilDead / 6;
808 int rxi_lowPeerRefCount = 0;
809 int rxi_lowConnRefCount = 0;
812 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
813 * NOTE: must not be called with rx_connHashTable_lock held.
816 rxi_CleanupConnection(struct rx_connection *conn)
818 /* Notify the service exporter, if requested, that this connection
819 * is being destroyed */
820 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
821 (*conn->service->destroyConnProc) (conn);
823 /* Notify the security module that this connection is being destroyed */
824 RXS_DestroyConnection(conn->securityObject, conn);
826 /* If this is the last connection using the rx_peer struct, set its
827 * idle time to now. rxi_ReapConnections will reap it if it's still
828 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
830 MUTEX_ENTER(&rx_peerHashTable_lock);
831 if (conn->peer->refCount < 2) {
832 conn->peer->idleWhen = clock_Sec();
833 if (conn->peer->refCount < 1) {
834 conn->peer->refCount = 1;
835 MUTEX_ENTER(&rx_stats_mutex);
836 rxi_lowPeerRefCount++;
837 MUTEX_EXIT(&rx_stats_mutex);
840 conn->peer->refCount--;
841 MUTEX_EXIT(&rx_peerHashTable_lock);
843 MUTEX_ENTER(&rx_stats_mutex);
844 if (conn->type == RX_SERVER_CONNECTION)
845 rx_stats.nServerConns--;
847 rx_stats.nClientConns--;
848 MUTEX_EXIT(&rx_stats_mutex);
851 if (conn->specific) {
853 for (i = 0; i < conn->nSpecific; i++) {
854 if (conn->specific[i] && rxi_keyCreate_destructor[i])
855 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
856 conn->specific[i] = NULL;
858 free(conn->specific);
860 conn->specific = NULL;
864 MUTEX_DESTROY(&conn->conn_call_lock);
865 MUTEX_DESTROY(&conn->conn_data_lock);
866 CV_DESTROY(&conn->conn_call_cv);
868 rxi_FreeConnection(conn);
871 /* Destroy the specified connection */
873 rxi_DestroyConnection(register struct rx_connection *conn)
875 MUTEX_ENTER(&rx_connHashTable_lock);
876 rxi_DestroyConnectionNoLock(conn);
877 /* conn should be at the head of the cleanup list */
878 if (conn == rx_connCleanup_list) {
879 rx_connCleanup_list = rx_connCleanup_list->next;
880 MUTEX_EXIT(&rx_connHashTable_lock);
881 rxi_CleanupConnection(conn);
883 #ifdef RX_ENABLE_LOCKS
885 MUTEX_EXIT(&rx_connHashTable_lock);
887 #endif /* RX_ENABLE_LOCKS */
891 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
893 register struct rx_connection **conn_ptr;
894 register int havecalls = 0;
895 struct rx_packet *packet;
902 MUTEX_ENTER(&conn->conn_data_lock);
903 if (conn->refCount > 0)
906 MUTEX_ENTER(&rx_stats_mutex);
907 rxi_lowConnRefCount++;
908 MUTEX_EXIT(&rx_stats_mutex);
911 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
912 /* Busy; wait till the last guy before proceeding */
913 MUTEX_EXIT(&conn->conn_data_lock);
918 /* If the client previously called rx_NewCall, but it is still
919 * waiting, treat this as a running call, and wait to destroy the
920 * connection later when the call completes. */
921 if ((conn->type == RX_CLIENT_CONNECTION)
922 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
923 conn->flags |= RX_CONN_DESTROY_ME;
924 MUTEX_EXIT(&conn->conn_data_lock);
928 MUTEX_EXIT(&conn->conn_data_lock);
930 /* Check for extant references to this connection */
931 for (i = 0; i < RX_MAXCALLS; i++) {
932 register struct rx_call *call = conn->call[i];
935 if (conn->type == RX_CLIENT_CONNECTION) {
936 MUTEX_ENTER(&call->lock);
937 if (call->delayedAckEvent) {
938 /* Push the final acknowledgment out now--there
939 * won't be a subsequent call to acknowledge the
940 * last reply packets */
941 rxevent_Cancel(call->delayedAckEvent, call,
942 RX_CALL_REFCOUNT_DELAY);
943 if (call->state == RX_STATE_PRECALL
944 || call->state == RX_STATE_ACTIVE) {
945 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
947 rxi_AckAll(NULL, call, 0);
950 MUTEX_EXIT(&call->lock);
954 #ifdef RX_ENABLE_LOCKS
956 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
957 MUTEX_EXIT(&conn->conn_data_lock);
959 /* Someone is accessing a packet right now. */
963 #endif /* RX_ENABLE_LOCKS */
966 /* Don't destroy the connection if there are any call
967 * structures still in use */
968 MUTEX_ENTER(&conn->conn_data_lock);
969 conn->flags |= RX_CONN_DESTROY_ME;
970 MUTEX_EXIT(&conn->conn_data_lock);
975 if (conn->delayedAbortEvent) {
976 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
977 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
979 MUTEX_ENTER(&conn->conn_data_lock);
980 rxi_SendConnectionAbort(conn, packet, 0, 1);
981 MUTEX_EXIT(&conn->conn_data_lock);
982 rxi_FreePacket(packet);
986 /* Remove from connection hash table before proceeding */
988 &rx_connHashTable[CONN_HASH
989 (peer->host, peer->port, conn->cid, conn->epoch,
991 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
992 if (*conn_ptr == conn) {
993 *conn_ptr = conn->next;
997 /* if the conn that we are destroying was the last connection, then we
998 * clear rxLastConn as well */
999 if (rxLastConn == conn)
1002 /* Make sure the connection is completely reset before deleting it. */
1003 /* get rid of pending events that could zap us later */
1004 if (conn->challengeEvent)
1005 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1006 if (conn->checkReachEvent)
1007 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1009 /* Add the connection to the list of destroyed connections that
1010 * need to be cleaned up. This is necessary to avoid deadlocks
1011 * in the routines we call to inform others that this connection is
1012 * being destroyed. */
1013 conn->next = rx_connCleanup_list;
1014 rx_connCleanup_list = conn;
1017 /* Externally available version */
1019 rx_DestroyConnection(register struct rx_connection *conn)
1024 rxi_DestroyConnection(conn);
1029 rx_GetConnection(register struct rx_connection *conn)
1034 MUTEX_ENTER(&conn->conn_data_lock);
1036 MUTEX_EXIT(&conn->conn_data_lock);
1040 /* Wait for the transmit queue to no longer be busy.
1041 * requires the call->lock to be held */
1042 static void rxi_WaitforTQBusy(struct rx_call *call) {
1043 while (call->flags & RX_CALL_TQ_BUSY) {
1044 call->flags |= RX_CALL_TQ_WAIT;
1046 #ifdef RX_ENABLE_LOCKS
1047 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1048 CV_WAIT(&call->cv_tq, &call->lock);
1049 #else /* RX_ENABLE_LOCKS */
1050 osi_rxSleep(&call->tq);
1051 #endif /* RX_ENABLE_LOCKS */
1053 if (call->tqWaiters == 0) {
1054 call->flags &= ~RX_CALL_TQ_WAIT;
1058 /* Start a new rx remote procedure call, on the specified connection.
1059 * If wait is set to 1, wait for a free call channel; otherwise return
1060 * 0. Maxtime gives the maximum number of seconds this call may take,
1061 * after rx_NewCall returns. After this time interval, a call to any
1062 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1063 * For fine grain locking, we hold the conn_call_lock in order to
1064 * to ensure that we don't get signalle after we found a call in an active
1065 * state and before we go to sleep.
1068 rx_NewCall(register struct rx_connection *conn)
1071 register struct rx_call *call;
1072 struct clock queueTime;
1076 dpf(("rx_NewCall(conn %x)\n", conn));
1079 clock_GetTime(&queueTime);
1080 MUTEX_ENTER(&conn->conn_call_lock);
1083 * Check if there are others waiting for a new call.
1084 * If so, let them go first to avoid starving them.
1085 * This is a fairly simple scheme, and might not be
1086 * a complete solution for large numbers of waiters.
1088 * makeCallWaiters keeps track of the number of
1089 * threads waiting to make calls and the
1090 * RX_CONN_MAKECALL_WAITING flag bit is used to
1091 * indicate that there are indeed calls waiting.
1092 * The flag is set when the waiter is incremented.
1093 * It is only cleared in rx_EndCall when
1094 * makeCallWaiters is 0. This prevents us from
1095 * accidently destroying the connection while it
1096 * is potentially about to be used.
1098 MUTEX_ENTER(&conn->conn_data_lock);
1099 if (conn->makeCallWaiters) {
1100 conn->flags |= RX_CONN_MAKECALL_WAITING;
1101 conn->makeCallWaiters++;
1102 MUTEX_EXIT(&conn->conn_data_lock);
1104 #ifdef RX_ENABLE_LOCKS
1105 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1109 MUTEX_ENTER(&conn->conn_data_lock);
1110 conn->makeCallWaiters--;
1112 MUTEX_EXIT(&conn->conn_data_lock);
1115 for (i = 0; i < RX_MAXCALLS; i++) {
1116 call = conn->call[i];
1118 MUTEX_ENTER(&call->lock);
1119 if (call->state == RX_STATE_DALLY) {
1120 rxi_ResetCall(call, 0);
1121 (*call->callNumber)++;
1124 MUTEX_EXIT(&call->lock);
1126 call = rxi_NewCall(conn, i);
1130 if (i < RX_MAXCALLS) {
1133 MUTEX_ENTER(&conn->conn_data_lock);
1134 conn->flags |= RX_CONN_MAKECALL_WAITING;
1135 conn->makeCallWaiters++;
1136 MUTEX_EXIT(&conn->conn_data_lock);
1138 #ifdef RX_ENABLE_LOCKS
1139 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1143 MUTEX_ENTER(&conn->conn_data_lock);
1144 conn->makeCallWaiters--;
1145 MUTEX_EXIT(&conn->conn_data_lock);
1148 * Wake up anyone else who might be giving us a chance to
1149 * run (see code above that avoids resource starvation).
1151 #ifdef RX_ENABLE_LOCKS
1152 CV_BROADCAST(&conn->conn_call_cv);
1157 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1159 /* Client is initially in send mode */
1160 call->state = RX_STATE_ACTIVE;
1161 call->error = conn->error;
1163 call->mode = RX_MODE_ERROR;
1165 call->mode = RX_MODE_SENDING;
1167 /* remember start time for call in case we have hard dead time limit */
1168 call->queueTime = queueTime;
1169 clock_GetTime(&call->startTime);
1170 hzero(call->bytesSent);
1171 hzero(call->bytesRcvd);
1173 /* Turn on busy protocol. */
1174 rxi_KeepAliveOn(call);
1176 MUTEX_EXIT(&call->lock);
1177 MUTEX_EXIT(&conn->conn_call_lock);
1180 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1181 /* Now, if TQ wasn't cleared earlier, do it now. */
1182 MUTEX_ENTER(&call->lock);
1183 rxi_WaitforTQBusy(call);
1184 if (call->flags & RX_CALL_TQ_CLEARME) {
1185 rxi_ClearTransmitQueue(call, 0);
1186 queue_Init(&call->tq);
1188 MUTEX_EXIT(&call->lock);
1189 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1191 dpf(("rx_NewCall(call %x)\n", call));
1196 rxi_HasActiveCalls(register struct rx_connection *aconn)
1199 register struct rx_call *tcall;
1203 for (i = 0; i < RX_MAXCALLS; i++) {
1204 if ((tcall = aconn->call[i])) {
1205 if ((tcall->state == RX_STATE_ACTIVE)
1206 || (tcall->state == RX_STATE_PRECALL)) {
1217 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1218 register afs_int32 * aint32s)
1221 register struct rx_call *tcall;
1225 for (i = 0; i < RX_MAXCALLS; i++) {
1226 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1227 aint32s[i] = aconn->callNumber[i] + 1;
1229 aint32s[i] = aconn->callNumber[i];
1236 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1237 register afs_int32 * aint32s)
1240 register struct rx_call *tcall;
1244 for (i = 0; i < RX_MAXCALLS; i++) {
1245 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1246 aconn->callNumber[i] = aint32s[i] - 1;
1248 aconn->callNumber[i] = aint32s[i];
1254 /* Advertise a new service. A service is named locally by a UDP port
1255 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1258 char *serviceName; Name for identification purposes (e.g. the
1259 service name might be used for probing for
1262 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1263 char *serviceName, struct rx_securityClass **securityObjects,
1264 int nSecurityObjects,
1265 afs_int32(*serviceProc) (struct rx_call * acall))
1267 osi_socket socket = OSI_NULLSOCKET;
1268 register struct rx_service *tservice;
1274 if (serviceId == 0) {
1276 "rx_NewService: service id for service %s is not non-zero.\n",
1283 "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",
1291 tservice = rxi_AllocService();
1293 for (i = 0; i < RX_MAX_SERVICES; i++) {
1294 register struct rx_service *service = rx_services[i];
1296 if (port == service->servicePort && host == service->serviceHost) {
1297 if (service->serviceId == serviceId) {
1298 /* The identical service has already been
1299 * installed; if the caller was intending to
1300 * change the security classes used by this
1301 * service, he/she loses. */
1303 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1304 serviceName, serviceId, service->serviceName);
1306 rxi_FreeService(tservice);
1309 /* Different service, same port: re-use the socket
1310 * which is bound to the same port */
1311 socket = service->socket;
1314 if (socket == OSI_NULLSOCKET) {
1315 /* If we don't already have a socket (from another
1316 * service on same port) get a new one */
1317 socket = rxi_GetHostUDPSocket(host, port);
1318 if (socket == OSI_NULLSOCKET) {
1320 rxi_FreeService(tservice);
1325 service->socket = socket;
1326 service->serviceHost = host;
1327 service->servicePort = port;
1328 service->serviceId = serviceId;
1329 service->serviceName = serviceName;
1330 service->nSecurityObjects = nSecurityObjects;
1331 service->securityObjects = securityObjects;
1332 service->minProcs = 0;
1333 service->maxProcs = 1;
1334 service->idleDeadTime = 60;
1335 service->connDeadTime = rx_connDeadTime;
1336 service->executeRequestProc = serviceProc;
1337 service->checkReach = 0;
1338 rx_services[i] = service; /* not visible until now */
1344 rxi_FreeService(tservice);
1345 (osi_Msg "rx_NewService: cannot support > %d services\n",
1351 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1352 struct rx_securityClass **securityObjects, int nSecurityObjects,
1353 afs_int32(*serviceProc) (struct rx_call * acall))
1355 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1358 /* Generic request processing loop. This routine should be called
1359 * by the implementation dependent rx_ServerProc. If socketp is
1360 * non-null, it will be set to the file descriptor that this thread
1361 * is now listening on. If socketp is null, this routine will never
1364 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1366 register struct rx_call *call;
1367 register afs_int32 code;
1368 register struct rx_service *tservice = NULL;
1375 call = rx_GetCall(threadID, tservice, socketp);
1376 if (socketp && *socketp != OSI_NULLSOCKET) {
1377 /* We are now a listener thread */
1382 /* if server is restarting( typically smooth shutdown) then do not
1383 * allow any new calls.
1386 if (rx_tranquil && (call != NULL)) {
1390 MUTEX_ENTER(&call->lock);
1392 rxi_CallError(call, RX_RESTARTING);
1393 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1395 MUTEX_EXIT(&call->lock);
1399 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1400 #ifdef RX_ENABLE_LOCKS
1402 #endif /* RX_ENABLE_LOCKS */
1403 afs_termState = AFSOP_STOP_AFS;
1404 afs_osi_Wakeup(&afs_termState);
1405 #ifdef RX_ENABLE_LOCKS
1407 #endif /* RX_ENABLE_LOCKS */
1412 tservice = call->conn->service;
1414 if (tservice->beforeProc)
1415 (*tservice->beforeProc) (call);
1417 code = call->conn->service->executeRequestProc(call);
1419 if (tservice->afterProc)
1420 (*tservice->afterProc) (call, code);
1422 rx_EndCall(call, code);
1423 MUTEX_ENTER(&rx_stats_mutex);
1425 MUTEX_EXIT(&rx_stats_mutex);
1431 rx_WakeupServerProcs(void)
1433 struct rx_serverQueueEntry *np, *tqp;
1437 MUTEX_ENTER(&rx_serverPool_lock);
1439 #ifdef RX_ENABLE_LOCKS
1440 if (rx_waitForPacket)
1441 CV_BROADCAST(&rx_waitForPacket->cv);
1442 #else /* RX_ENABLE_LOCKS */
1443 if (rx_waitForPacket)
1444 osi_rxWakeup(rx_waitForPacket);
1445 #endif /* RX_ENABLE_LOCKS */
1446 MUTEX_ENTER(&freeSQEList_lock);
1447 for (np = rx_FreeSQEList; np; np = tqp) {
1448 tqp = *(struct rx_serverQueueEntry **)np;
1449 #ifdef RX_ENABLE_LOCKS
1450 CV_BROADCAST(&np->cv);
1451 #else /* RX_ENABLE_LOCKS */
1453 #endif /* RX_ENABLE_LOCKS */
1455 MUTEX_EXIT(&freeSQEList_lock);
1456 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1457 #ifdef RX_ENABLE_LOCKS
1458 CV_BROADCAST(&np->cv);
1459 #else /* RX_ENABLE_LOCKS */
1461 #endif /* RX_ENABLE_LOCKS */
1463 MUTEX_EXIT(&rx_serverPool_lock);
1468 * One thing that seems to happen is that all the server threads get
1469 * tied up on some empty or slow call, and then a whole bunch of calls
1470 * arrive at once, using up the packet pool, so now there are more
1471 * empty calls. The most critical resources here are server threads
1472 * and the free packet pool. The "doreclaim" code seems to help in
1473 * general. I think that eventually we arrive in this state: there
1474 * are lots of pending calls which do have all their packets present,
1475 * so they won't be reclaimed, are multi-packet calls, so they won't
1476 * be scheduled until later, and thus are tying up most of the free
1477 * packet pool for a very long time.
1479 * 1. schedule multi-packet calls if all the packets are present.
1480 * Probably CPU-bound operation, useful to return packets to pool.
1481 * Do what if there is a full window, but the last packet isn't here?
1482 * 3. preserve one thread which *only* runs "best" calls, otherwise
1483 * it sleeps and waits for that type of call.
1484 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1485 * the current dataquota business is badly broken. The quota isn't adjusted
1486 * to reflect how many packets are presently queued for a running call.
1487 * So, when we schedule a queued call with a full window of packets queued
1488 * up for it, that *should* free up a window full of packets for other 2d-class
1489 * calls to be able to use from the packet pool. But it doesn't.
1491 * NB. Most of the time, this code doesn't run -- since idle server threads
1492 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1493 * as a new call arrives.
1495 /* Sleep until a call arrives. Returns a pointer to the call, ready
1496 * for an rx_Read. */
1497 #ifdef RX_ENABLE_LOCKS
1499 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1501 struct rx_serverQueueEntry *sq;
1502 register struct rx_call *call = (struct rx_call *)0;
1503 struct rx_service *service = NULL;
1506 MUTEX_ENTER(&freeSQEList_lock);
1508 if ((sq = rx_FreeSQEList)) {
1509 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1510 MUTEX_EXIT(&freeSQEList_lock);
1511 } else { /* otherwise allocate a new one and return that */
1512 MUTEX_EXIT(&freeSQEList_lock);
1513 sq = (struct rx_serverQueueEntry *)
1514 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1515 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1516 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1519 MUTEX_ENTER(&rx_serverPool_lock);
1520 if (cur_service != NULL) {
1521 ReturnToServerPool(cur_service);
1524 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1525 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1527 /* Scan for eligible incoming calls. A call is not eligible
1528 * if the maximum number of calls for its service type are
1529 * already executing */
1530 /* One thread will process calls FCFS (to prevent starvation),
1531 * while the other threads may run ahead looking for calls which
1532 * have all their input data available immediately. This helps
1533 * keep threads from blocking, waiting for data from the client. */
1534 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1535 service = tcall->conn->service;
1536 if (!QuotaOK(service)) {
1539 if (tno == rxi_fcfs_thread_num
1540 || !tcall->queue_item_header.next) {
1541 /* If we're the fcfs thread , then we'll just use
1542 * this call. If we haven't been able to find an optimal
1543 * choice, and we're at the end of the list, then use a
1544 * 2d choice if one has been identified. Otherwise... */
1545 call = (choice2 ? choice2 : tcall);
1546 service = call->conn->service;
1547 } else if (!queue_IsEmpty(&tcall->rq)) {
1548 struct rx_packet *rp;
1549 rp = queue_First(&tcall->rq, rx_packet);
1550 if (rp->header.seq == 1) {
1552 || (rp->header.flags & RX_LAST_PACKET)) {
1554 } else if (rxi_2dchoice && !choice2
1555 && !(tcall->flags & RX_CALL_CLEARED)
1556 && (tcall->rprev > rxi_HardAckRate)) {
1565 ReturnToServerPool(service);
1572 MUTEX_EXIT(&rx_serverPool_lock);
1573 MUTEX_ENTER(&call->lock);
1575 if (call->flags & RX_CALL_WAIT_PROC) {
1576 call->flags &= ~RX_CALL_WAIT_PROC;
1577 MUTEX_ENTER(&rx_stats_mutex);
1579 MUTEX_EXIT(&rx_stats_mutex);
1582 if (call->state != RX_STATE_PRECALL || call->error) {
1583 MUTEX_EXIT(&call->lock);
1584 MUTEX_ENTER(&rx_serverPool_lock);
1585 ReturnToServerPool(service);
1590 if (queue_IsEmpty(&call->rq)
1591 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1592 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1594 CLEAR_CALL_QUEUE_LOCK(call);
1597 /* If there are no eligible incoming calls, add this process
1598 * to the idle server queue, to wait for one */
1602 *socketp = OSI_NULLSOCKET;
1604 sq->socketp = socketp;
1605 queue_Append(&rx_idleServerQueue, sq);
1606 #ifndef AFS_AIX41_ENV
1607 rx_waitForPacket = sq;
1609 rx_waitingForPacket = sq;
1610 #endif /* AFS_AIX41_ENV */
1612 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1614 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1615 MUTEX_EXIT(&rx_serverPool_lock);
1616 return (struct rx_call *)0;
1619 } while (!(call = sq->newcall)
1620 && !(socketp && *socketp != OSI_NULLSOCKET));
1621 MUTEX_EXIT(&rx_serverPool_lock);
1623 MUTEX_ENTER(&call->lock);
1629 MUTEX_ENTER(&freeSQEList_lock);
1630 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1631 rx_FreeSQEList = sq;
1632 MUTEX_EXIT(&freeSQEList_lock);
1635 clock_GetTime(&call->startTime);
1636 call->state = RX_STATE_ACTIVE;
1637 call->mode = RX_MODE_RECEIVING;
1638 #ifdef RX_KERNEL_TRACE
1639 if (ICL_SETACTIVE(afs_iclSetp)) {
1640 int glockOwner = ISAFS_GLOCK();
1643 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1644 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1651 rxi_calltrace(RX_CALL_START, call);
1652 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1653 call->conn->service->servicePort, call->conn->service->serviceId,
1656 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1657 MUTEX_EXIT(&call->lock);
1659 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1664 #else /* RX_ENABLE_LOCKS */
1666 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1668 struct rx_serverQueueEntry *sq;
1669 register struct rx_call *call = (struct rx_call *)0, *choice2;
1670 struct rx_service *service = NULL;
1674 MUTEX_ENTER(&freeSQEList_lock);
1676 if ((sq = rx_FreeSQEList)) {
1677 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1678 MUTEX_EXIT(&freeSQEList_lock);
1679 } else { /* otherwise allocate a new one and return that */
1680 MUTEX_EXIT(&freeSQEList_lock);
1681 sq = (struct rx_serverQueueEntry *)
1682 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1683 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1684 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1686 MUTEX_ENTER(&sq->lock);
1688 if (cur_service != NULL) {
1689 cur_service->nRequestsRunning--;
1690 if (cur_service->nRequestsRunning < cur_service->minProcs)
1694 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1695 register struct rx_call *tcall, *ncall;
1696 /* Scan for eligible incoming calls. A call is not eligible
1697 * if the maximum number of calls for its service type are
1698 * already executing */
1699 /* One thread will process calls FCFS (to prevent starvation),
1700 * while the other threads may run ahead looking for calls which
1701 * have all their input data available immediately. This helps
1702 * keep threads from blocking, waiting for data from the client. */
1703 choice2 = (struct rx_call *)0;
1704 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1705 service = tcall->conn->service;
1706 if (QuotaOK(service)) {
1707 if (tno == rxi_fcfs_thread_num
1708 || !tcall->queue_item_header.next) {
1709 /* If we're the fcfs thread, then we'll just use
1710 * this call. If we haven't been able to find an optimal
1711 * choice, and we're at the end of the list, then use a
1712 * 2d choice if one has been identified. Otherwise... */
1713 call = (choice2 ? choice2 : tcall);
1714 service = call->conn->service;
1715 } else if (!queue_IsEmpty(&tcall->rq)) {
1716 struct rx_packet *rp;
1717 rp = queue_First(&tcall->rq, rx_packet);
1718 if (rp->header.seq == 1
1720 || (rp->header.flags & RX_LAST_PACKET))) {
1722 } else if (rxi_2dchoice && !choice2
1723 && !(tcall->flags & RX_CALL_CLEARED)
1724 && (tcall->rprev > rxi_HardAckRate)) {
1737 /* we can't schedule a call if there's no data!!! */
1738 /* send an ack if there's no data, if we're missing the
1739 * first packet, or we're missing something between first
1740 * and last -- there's a "hole" in the incoming data. */
1741 if (queue_IsEmpty(&call->rq)
1742 || queue_First(&call->rq, rx_packet)->header.seq != 1
1743 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1744 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1746 call->flags &= (~RX_CALL_WAIT_PROC);
1747 service->nRequestsRunning++;
1748 /* just started call in minProcs pool, need fewer to maintain
1750 if (service->nRequestsRunning <= service->minProcs)
1754 /* MUTEX_EXIT(&call->lock); */
1756 /* If there are no eligible incoming calls, add this process
1757 * to the idle server queue, to wait for one */
1760 *socketp = OSI_NULLSOCKET;
1762 sq->socketp = socketp;
1763 queue_Append(&rx_idleServerQueue, sq);
1767 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1769 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1770 return (struct rx_call *)0;
1773 } while (!(call = sq->newcall)
1774 && !(socketp && *socketp != OSI_NULLSOCKET));
1776 MUTEX_EXIT(&sq->lock);
1778 MUTEX_ENTER(&freeSQEList_lock);
1779 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1780 rx_FreeSQEList = sq;
1781 MUTEX_EXIT(&freeSQEList_lock);
1784 clock_GetTime(&call->startTime);
1785 call->state = RX_STATE_ACTIVE;
1786 call->mode = RX_MODE_RECEIVING;
1787 #ifdef RX_KERNEL_TRACE
1788 if (ICL_SETACTIVE(afs_iclSetp)) {
1789 int glockOwner = ISAFS_GLOCK();
1792 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1793 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1800 rxi_calltrace(RX_CALL_START, call);
1801 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1802 call->conn->service->servicePort, call->conn->service->serviceId,
1805 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1812 #endif /* RX_ENABLE_LOCKS */
1816 /* Establish a procedure to be called when a packet arrives for a
1817 * call. This routine will be called at most once after each call,
1818 * and will also be called if there is an error condition on the or
1819 * the call is complete. Used by multi rx to build a selection
1820 * function which determines which of several calls is likely to be a
1821 * good one to read from.
1822 * NOTE: the way this is currently implemented it is probably only a
1823 * good idea to (1) use it immediately after a newcall (clients only)
1824 * and (2) only use it once. Other uses currently void your warranty
1827 rx_SetArrivalProc(register struct rx_call *call,
1828 register void (*proc) (register struct rx_call * call,
1830 register int index),
1831 register VOID * handle, register int arg)
1833 call->arrivalProc = proc;
1834 call->arrivalProcHandle = handle;
1835 call->arrivalProcArg = arg;
1838 /* Call is finished (possibly prematurely). Return rc to the peer, if
1839 * appropriate, and return the final error code from the conversation
1843 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1845 register struct rx_connection *conn = call->conn;
1846 register struct rx_service *service;
1852 dpf(("rx_EndCall(call %x rc %d error %d abortCode %d)\n", call, rc, call->error, call->abortCode));
1855 MUTEX_ENTER(&call->lock);
1857 if (rc == 0 && call->error == 0) {
1858 call->abortCode = 0;
1859 call->abortCount = 0;
1862 call->arrivalProc = (void (*)())0;
1863 if (rc && call->error == 0) {
1864 rxi_CallError(call, rc);
1865 /* Send an abort message to the peer if this error code has
1866 * only just been set. If it was set previously, assume the
1867 * peer has already been sent the error code or will request it
1869 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1871 if (conn->type == RX_SERVER_CONNECTION) {
1872 /* Make sure reply or at least dummy reply is sent */
1873 if (call->mode == RX_MODE_RECEIVING) {
1874 rxi_WriteProc(call, 0, 0);
1876 if (call->mode == RX_MODE_SENDING) {
1877 rxi_FlushWrite(call);
1879 service = conn->service;
1880 rxi_calltrace(RX_CALL_END, call);
1881 /* Call goes to hold state until reply packets are acknowledged */
1882 if (call->tfirst + call->nSoftAcked < call->tnext) {
1883 call->state = RX_STATE_HOLD;
1885 call->state = RX_STATE_DALLY;
1886 rxi_ClearTransmitQueue(call, 0);
1887 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1888 rxevent_Cancel(call->keepAliveEvent, call,
1889 RX_CALL_REFCOUNT_ALIVE);
1891 } else { /* Client connection */
1893 /* Make sure server receives input packets, in the case where
1894 * no reply arguments are expected */
1895 if ((call->mode == RX_MODE_SENDING)
1896 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1897 (void)rxi_ReadProc(call, &dummy, 1);
1900 /* If we had an outstanding delayed ack, be nice to the server
1901 * and force-send it now.
1903 if (call->delayedAckEvent) {
1904 rxevent_Cancel(call->delayedAckEvent, call,
1905 RX_CALL_REFCOUNT_DELAY);
1906 call->delayedAckEvent = NULL;
1907 rxi_SendDelayedAck(NULL, call, NULL);
1910 /* We need to release the call lock since it's lower than the
1911 * conn_call_lock and we don't want to hold the conn_call_lock
1912 * over the rx_ReadProc call. The conn_call_lock needs to be held
1913 * here for the case where rx_NewCall is perusing the calls on
1914 * the connection structure. We don't want to signal until
1915 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1916 * have checked this call, found it active and by the time it
1917 * goes to sleep, will have missed the signal.
1919 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
1920 * there are threads waiting to use the conn object.
1922 MUTEX_EXIT(&call->lock);
1923 MUTEX_ENTER(&conn->conn_call_lock);
1924 MUTEX_ENTER(&call->lock);
1925 MUTEX_ENTER(&conn->conn_data_lock);
1926 conn->flags |= RX_CONN_BUSY;
1927 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1928 if (conn->makeCallWaiters == 0)
1929 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1930 MUTEX_EXIT(&conn->conn_data_lock);
1931 #ifdef RX_ENABLE_LOCKS
1932 CV_BROADCAST(&conn->conn_call_cv);
1937 #ifdef RX_ENABLE_LOCKS
1939 MUTEX_EXIT(&conn->conn_data_lock);
1941 #endif /* RX_ENABLE_LOCKS */
1942 call->state = RX_STATE_DALLY;
1944 error = call->error;
1946 /* currentPacket, nLeft, and NFree must be zeroed here, because
1947 * ResetCall cannot: ResetCall may be called at splnet(), in the
1948 * kernel version, and may interrupt the macros rx_Read or
1949 * rx_Write, which run at normal priority for efficiency. */
1950 if (call->currentPacket) {
1951 queue_Prepend(&call->iovq, call->currentPacket);
1952 call->currentPacket = (struct rx_packet *)0;
1955 call->nLeft = call->nFree = call->curlen = 0;
1957 /* Free any packets from the last call to ReadvProc/WritevProc */
1958 rxi_FreePackets(0, &call->iovq);
1960 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1961 MUTEX_EXIT(&call->lock);
1962 if (conn->type == RX_CLIENT_CONNECTION) {
1963 MUTEX_EXIT(&conn->conn_call_lock);
1964 conn->flags &= ~RX_CONN_BUSY;
1968 * Map errors to the local host's errno.h format.
1970 error = ntoh_syserr_conv(error);
1974 #if !defined(KERNEL)
1976 /* Call this routine when shutting down a server or client (especially
1977 * clients). This will allow Rx to gracefully garbage collect server
1978 * connections, and reduce the number of retries that a server might
1979 * make to a dead client.
1980 * This is not quite right, since some calls may still be ongoing and
1981 * we can't lock them to destroy them. */
1985 register struct rx_connection **conn_ptr, **conn_end;
1989 if (rxinit_status == 1) {
1991 return; /* Already shutdown. */
1993 rxi_DeleteCachedConnections();
1994 if (rx_connHashTable) {
1995 MUTEX_ENTER(&rx_connHashTable_lock);
1996 for (conn_ptr = &rx_connHashTable[0], conn_end =
1997 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
1999 struct rx_connection *conn, *next;
2000 for (conn = *conn_ptr; conn; conn = next) {
2002 if (conn->type == RX_CLIENT_CONNECTION) {
2003 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2005 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2006 #ifdef RX_ENABLE_LOCKS
2007 rxi_DestroyConnectionNoLock(conn);
2008 #else /* RX_ENABLE_LOCKS */
2009 rxi_DestroyConnection(conn);
2010 #endif /* RX_ENABLE_LOCKS */
2014 #ifdef RX_ENABLE_LOCKS
2015 while (rx_connCleanup_list) {
2016 struct rx_connection *conn;
2017 conn = rx_connCleanup_list;
2018 rx_connCleanup_list = rx_connCleanup_list->next;
2019 MUTEX_EXIT(&rx_connHashTable_lock);
2020 rxi_CleanupConnection(conn);
2021 MUTEX_ENTER(&rx_connHashTable_lock);
2023 MUTEX_EXIT(&rx_connHashTable_lock);
2024 #endif /* RX_ENABLE_LOCKS */
2029 afs_winsockCleanup();
2037 /* if we wakeup packet waiter too often, can get in loop with two
2038 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2040 rxi_PacketsUnWait(void)
2042 if (!rx_waitingForPackets) {
2046 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2047 return; /* still over quota */
2050 rx_waitingForPackets = 0;
2051 #ifdef RX_ENABLE_LOCKS
2052 CV_BROADCAST(&rx_waitingForPackets_cv);
2054 osi_rxWakeup(&rx_waitingForPackets);
2060 /* ------------------Internal interfaces------------------------- */
2062 /* Return this process's service structure for the
2063 * specified socket and service */
2065 rxi_FindService(register osi_socket socket, register u_short serviceId)
2067 register struct rx_service **sp;
2068 for (sp = &rx_services[0]; *sp; sp++) {
2069 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2075 /* Allocate a call structure, for the indicated channel of the
2076 * supplied connection. The mode and state of the call must be set by
2077 * the caller. Returns the call with mutex locked. */
2079 rxi_NewCall(register struct rx_connection *conn, register int channel)
2081 register struct rx_call *call;
2082 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2083 register struct rx_call *cp; /* Call pointer temp */
2084 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2085 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2087 dpf(("rxi_NewCall(conn %x, channel %d)\n", conn, channel));
2089 /* Grab an existing call structure, or allocate a new one.
2090 * Existing call structures are assumed to have been left reset by
2092 MUTEX_ENTER(&rx_freeCallQueue_lock);
2094 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2096 * EXCEPT that the TQ might not yet be cleared out.
2097 * Skip over those with in-use TQs.
2100 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2101 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2107 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2108 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2109 call = queue_First(&rx_freeCallQueue, rx_call);
2110 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2112 MUTEX_ENTER(&rx_stats_mutex);
2113 rx_stats.nFreeCallStructs--;
2114 MUTEX_EXIT(&rx_stats_mutex);
2115 MUTEX_EXIT(&rx_freeCallQueue_lock);
2116 MUTEX_ENTER(&call->lock);
2117 CLEAR_CALL_QUEUE_LOCK(call);
2118 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2119 /* Now, if TQ wasn't cleared earlier, do it now. */
2120 if (call->flags & RX_CALL_TQ_CLEARME) {
2121 rxi_ClearTransmitQueue(call, 0);
2122 queue_Init(&call->tq);
2124 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2125 /* Bind the call to its connection structure */
2127 rxi_ResetCall(call, 1);
2129 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2131 MUTEX_EXIT(&rx_freeCallQueue_lock);
2132 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2133 MUTEX_ENTER(&call->lock);
2134 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2135 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2136 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2138 MUTEX_ENTER(&rx_stats_mutex);
2139 rx_stats.nCallStructs++;
2140 MUTEX_EXIT(&rx_stats_mutex);
2141 /* Initialize once-only items */
2142 queue_Init(&call->tq);
2143 queue_Init(&call->rq);
2144 queue_Init(&call->iovq);
2145 /* Bind the call to its connection structure (prereq for reset) */
2147 rxi_ResetCall(call, 1);
2149 call->channel = channel;
2150 call->callNumber = &conn->callNumber[channel];
2151 /* Note that the next expected call number is retained (in
2152 * conn->callNumber[i]), even if we reallocate the call structure
2154 conn->call[channel] = call;
2155 /* if the channel's never been used (== 0), we should start at 1, otherwise
2156 * the call number is valid from the last time this channel was used */
2157 if (*call->callNumber == 0)
2158 *call->callNumber = 1;
2163 /* A call has been inactive long enough that so we can throw away
2164 * state, including the call structure, which is placed on the call
2166 * Call is locked upon entry.
2167 * haveCTLock set if called from rxi_ReapConnections
2169 #ifdef RX_ENABLE_LOCKS
2171 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2172 #else /* RX_ENABLE_LOCKS */
2174 rxi_FreeCall(register struct rx_call *call)
2175 #endif /* RX_ENABLE_LOCKS */
2177 register int channel = call->channel;
2178 register struct rx_connection *conn = call->conn;
2181 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2182 (*call->callNumber)++;
2183 rxi_ResetCall(call, 0);
2184 call->conn->call[channel] = (struct rx_call *)0;
2186 MUTEX_ENTER(&rx_freeCallQueue_lock);
2187 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2188 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2189 /* A call may be free even though its transmit queue is still in use.
2190 * Since we search the call list from head to tail, put busy calls at
2191 * the head of the list, and idle calls at the tail.
2193 if (call->flags & RX_CALL_TQ_BUSY)
2194 queue_Prepend(&rx_freeCallQueue, call);
2196 queue_Append(&rx_freeCallQueue, call);
2197 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2198 queue_Append(&rx_freeCallQueue, call);
2199 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2200 MUTEX_ENTER(&rx_stats_mutex);
2201 rx_stats.nFreeCallStructs++;
2202 MUTEX_EXIT(&rx_stats_mutex);
2204 MUTEX_EXIT(&rx_freeCallQueue_lock);
2206 /* Destroy the connection if it was previously slated for
2207 * destruction, i.e. the Rx client code previously called
2208 * rx_DestroyConnection (client connections), or
2209 * rxi_ReapConnections called the same routine (server
2210 * connections). Only do this, however, if there are no
2211 * outstanding calls. Note that for fine grain locking, there appears
2212 * to be a deadlock in that rxi_FreeCall has a call locked and
2213 * DestroyConnectionNoLock locks each call in the conn. But note a
2214 * few lines up where we have removed this call from the conn.
2215 * If someone else destroys a connection, they either have no
2216 * call lock held or are going through this section of code.
2218 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2219 MUTEX_ENTER(&conn->conn_data_lock);
2221 MUTEX_EXIT(&conn->conn_data_lock);
2222 #ifdef RX_ENABLE_LOCKS
2224 rxi_DestroyConnectionNoLock(conn);
2226 rxi_DestroyConnection(conn);
2227 #else /* RX_ENABLE_LOCKS */
2228 rxi_DestroyConnection(conn);
2229 #endif /* RX_ENABLE_LOCKS */
2233 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2235 rxi_Alloc(register size_t size)
2239 MUTEX_ENTER(&rx_stats_mutex);
2241 rxi_Allocsize += (afs_int32)size;
2242 MUTEX_EXIT(&rx_stats_mutex);
2244 p = (char *)osi_Alloc(size);
2247 osi_Panic("rxi_Alloc error");
2253 rxi_Free(void *addr, register size_t size)
2255 MUTEX_ENTER(&rx_stats_mutex);
2257 rxi_Allocsize -= (afs_int32)size;
2258 MUTEX_EXIT(&rx_stats_mutex);
2260 osi_Free(addr, size);
2263 /* Find the peer process represented by the supplied (host,port)
2264 * combination. If there is no appropriate active peer structure, a
2265 * new one will be allocated and initialized
2266 * The origPeer, if set, is a pointer to a peer structure on which the
2267 * refcount will be be decremented. This is used to replace the peer
2268 * structure hanging off a connection structure */
2270 rxi_FindPeer(register afs_uint32 host, register u_short port,
2271 struct rx_peer *origPeer, int create)
2273 register struct rx_peer *pp;
2275 hashIndex = PEER_HASH(host, port);
2276 MUTEX_ENTER(&rx_peerHashTable_lock);
2277 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2278 if ((pp->host == host) && (pp->port == port))
2283 pp = rxi_AllocPeer(); /* This bzero's *pp */
2284 pp->host = host; /* set here or in InitPeerParams is zero */
2286 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2287 queue_Init(&pp->congestionQueue);
2288 queue_Init(&pp->rpcStats);
2289 pp->next = rx_peerHashTable[hashIndex];
2290 rx_peerHashTable[hashIndex] = pp;
2291 rxi_InitPeerParams(pp);
2292 MUTEX_ENTER(&rx_stats_mutex);
2293 rx_stats.nPeerStructs++;
2294 MUTEX_EXIT(&rx_stats_mutex);
2301 origPeer->refCount--;
2302 MUTEX_EXIT(&rx_peerHashTable_lock);
2307 /* Find the connection at (host, port) started at epoch, and with the
2308 * given connection id. Creates the server connection if necessary.
2309 * The type specifies whether a client connection or a server
2310 * connection is desired. In both cases, (host, port) specify the
2311 * peer's (host, pair) pair. Client connections are not made
2312 * automatically by this routine. The parameter socket gives the
2313 * socket descriptor on which the packet was received. This is used,
2314 * in the case of server connections, to check that *new* connections
2315 * come via a valid (port, serviceId). Finally, the securityIndex
2316 * parameter must match the existing index for the connection. If a
2317 * server connection is created, it will be created using the supplied
2318 * index, if the index is valid for this service */
2319 struct rx_connection *
2320 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2321 register u_short port, u_short serviceId, afs_uint32 cid,
2322 afs_uint32 epoch, int type, u_int securityIndex)
2324 int hashindex, flag;
2325 register struct rx_connection *conn;
2326 hashindex = CONN_HASH(host, port, cid, epoch, type);
2327 MUTEX_ENTER(&rx_connHashTable_lock);
2328 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2329 rx_connHashTable[hashindex],
2332 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2333 && (epoch == conn->epoch)) {
2334 register struct rx_peer *pp = conn->peer;
2335 if (securityIndex != conn->securityIndex) {
2336 /* this isn't supposed to happen, but someone could forge a packet
2337 * like this, and there seems to be some CM bug that makes this
2338 * happen from time to time -- in which case, the fileserver
2340 MUTEX_EXIT(&rx_connHashTable_lock);
2341 return (struct rx_connection *)0;
2343 if (pp->host == host && pp->port == port)
2345 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2347 /* So what happens when it's a callback connection? */
2348 if ( /*type == RX_CLIENT_CONNECTION && */
2349 (conn->epoch & 0x80000000))
2353 /* the connection rxLastConn that was used the last time is not the
2354 ** one we are looking for now. Hence, start searching in the hash */
2356 conn = rx_connHashTable[hashindex];
2361 struct rx_service *service;
2362 if (type == RX_CLIENT_CONNECTION) {
2363 MUTEX_EXIT(&rx_connHashTable_lock);
2364 return (struct rx_connection *)0;
2366 service = rxi_FindService(socket, serviceId);
2367 if (!service || (securityIndex >= service->nSecurityObjects)
2368 || (service->securityObjects[securityIndex] == 0)) {
2369 MUTEX_EXIT(&rx_connHashTable_lock);
2370 return (struct rx_connection *)0;
2372 conn = rxi_AllocConnection(); /* This bzero's the connection */
2373 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2374 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2375 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2376 conn->next = rx_connHashTable[hashindex];
2377 rx_connHashTable[hashindex] = conn;
2378 conn->peer = rxi_FindPeer(host, port, 0, 1);
2379 conn->type = RX_SERVER_CONNECTION;
2380 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2381 conn->epoch = epoch;
2382 conn->cid = cid & RX_CIDMASK;
2383 /* conn->serial = conn->lastSerial = 0; */
2384 /* conn->timeout = 0; */
2385 conn->ackRate = RX_FAST_ACK_RATE;
2386 conn->service = service;
2387 conn->serviceId = serviceId;
2388 conn->securityIndex = securityIndex;
2389 conn->securityObject = service->securityObjects[securityIndex];
2390 conn->nSpecific = 0;
2391 conn->specific = NULL;
2392 rx_SetConnDeadTime(conn, service->connDeadTime);
2393 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2394 /* Notify security object of the new connection */
2395 RXS_NewConnection(conn->securityObject, conn);
2396 /* XXXX Connection timeout? */
2397 if (service->newConnProc)
2398 (*service->newConnProc) (conn);
2399 MUTEX_ENTER(&rx_stats_mutex);
2400 rx_stats.nServerConns++;
2401 MUTEX_EXIT(&rx_stats_mutex);
2404 MUTEX_ENTER(&conn->conn_data_lock);
2406 MUTEX_EXIT(&conn->conn_data_lock);
2408 rxLastConn = conn; /* store this connection as the last conn used */
2409 MUTEX_EXIT(&rx_connHashTable_lock);
2413 /* There are two packet tracing routines available for testing and monitoring
2414 * Rx. One is called just after every packet is received and the other is
2415 * called just before every packet is sent. Received packets, have had their
2416 * headers decoded, and packets to be sent have not yet had their headers
2417 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2418 * containing the network address. Both can be modified. The return value, if
2419 * non-zero, indicates that the packet should be dropped. */
2421 int (*rx_justReceived) () = 0;
2422 int (*rx_almostSent) () = 0;
2424 /* A packet has been received off the interface. Np is the packet, socket is
2425 * the socket number it was received from (useful in determining which service
2426 * this packet corresponds to), and (host, port) reflect the host,port of the
2427 * sender. This call returns the packet to the caller if it is finished with
2428 * it, rather than de-allocating it, just as a small performance hack */
2431 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2432 afs_uint32 host, u_short port, int *tnop,
2433 struct rx_call **newcallp)
2435 register struct rx_call *call;
2436 register struct rx_connection *conn;
2438 afs_uint32 currentCallNumber;
2444 struct rx_packet *tnp;
2447 /* We don't print out the packet until now because (1) the time may not be
2448 * accurate enough until now in the lwp implementation (rx_Listener only gets
2449 * the time after the packet is read) and (2) from a protocol point of view,
2450 * this is the first time the packet has been seen */
2451 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2452 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2453 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2454 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2455 np->header.epoch, np->header.cid, np->header.callNumber,
2456 np->header.seq, np->header.flags, np));
2459 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2460 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2463 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2464 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2467 /* If an input tracer function is defined, call it with the packet and
2468 * network address. Note this function may modify its arguments. */
2469 if (rx_justReceived) {
2470 struct sockaddr_in addr;
2472 addr.sin_family = AF_INET;
2473 addr.sin_port = port;
2474 addr.sin_addr.s_addr = host;
2475 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2476 addr.sin_len = sizeof(addr);
2477 #endif /* AFS_OSF_ENV */
2478 drop = (*rx_justReceived) (np, &addr);
2479 /* drop packet if return value is non-zero */
2482 port = addr.sin_port; /* in case fcn changed addr */
2483 host = addr.sin_addr.s_addr;
2487 /* If packet was not sent by the client, then *we* must be the client */
2488 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2489 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2491 /* Find the connection (or fabricate one, if we're the server & if
2492 * necessary) associated with this packet */
2494 rxi_FindConnection(socket, host, port, np->header.serviceId,
2495 np->header.cid, np->header.epoch, type,
2496 np->header.securityIndex);
2499 /* If no connection found or fabricated, just ignore the packet.
2500 * (An argument could be made for sending an abort packet for
2505 MUTEX_ENTER(&conn->conn_data_lock);
2506 if (conn->maxSerial < np->header.serial)
2507 conn->maxSerial = np->header.serial;
2508 MUTEX_EXIT(&conn->conn_data_lock);
2510 /* If the connection is in an error state, send an abort packet and ignore
2511 * the incoming packet */
2513 /* Don't respond to an abort packet--we don't want loops! */
2514 MUTEX_ENTER(&conn->conn_data_lock);
2515 if (np->header.type != RX_PACKET_TYPE_ABORT)
2516 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2518 MUTEX_EXIT(&conn->conn_data_lock);
2522 /* Check for connection-only requests (i.e. not call specific). */
2523 if (np->header.callNumber == 0) {
2524 switch (np->header.type) {
2525 case RX_PACKET_TYPE_ABORT: {
2526 /* What if the supplied error is zero? */
2527 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2528 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2529 rxi_ConnectionError(conn, errcode);
2530 MUTEX_ENTER(&conn->conn_data_lock);
2532 MUTEX_EXIT(&conn->conn_data_lock);
2535 case RX_PACKET_TYPE_CHALLENGE:
2536 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2537 MUTEX_ENTER(&conn->conn_data_lock);
2539 MUTEX_EXIT(&conn->conn_data_lock);
2541 case RX_PACKET_TYPE_RESPONSE:
2542 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2543 MUTEX_ENTER(&conn->conn_data_lock);
2545 MUTEX_EXIT(&conn->conn_data_lock);
2547 case RX_PACKET_TYPE_PARAMS:
2548 case RX_PACKET_TYPE_PARAMS + 1:
2549 case RX_PACKET_TYPE_PARAMS + 2:
2550 /* ignore these packet types for now */
2551 MUTEX_ENTER(&conn->conn_data_lock);
2553 MUTEX_EXIT(&conn->conn_data_lock);
2558 /* Should not reach here, unless the peer is broken: send an
2560 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2561 MUTEX_ENTER(&conn->conn_data_lock);
2562 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2564 MUTEX_EXIT(&conn->conn_data_lock);
2569 channel = np->header.cid & RX_CHANNELMASK;
2570 call = conn->call[channel];
2571 #ifdef RX_ENABLE_LOCKS
2573 MUTEX_ENTER(&call->lock);
2574 /* Test to see if call struct is still attached to conn. */
2575 if (call != conn->call[channel]) {
2577 MUTEX_EXIT(&call->lock);
2578 if (type == RX_SERVER_CONNECTION) {
2579 call = conn->call[channel];
2580 /* If we started with no call attached and there is one now,
2581 * another thread is also running this routine and has gotten
2582 * the connection channel. We should drop this packet in the tests
2583 * below. If there was a call on this connection and it's now
2584 * gone, then we'll be making a new call below.
2585 * If there was previously a call and it's now different then
2586 * the old call was freed and another thread running this routine
2587 * has created a call on this channel. One of these two threads
2588 * has a packet for the old call and the code below handles those
2592 MUTEX_ENTER(&call->lock);
2594 /* This packet can't be for this call. If the new call address is
2595 * 0 then no call is running on this channel. If there is a call
2596 * then, since this is a client connection we're getting data for
2597 * it must be for the previous call.
2599 MUTEX_ENTER(&rx_stats_mutex);
2600 rx_stats.spuriousPacketsRead++;
2601 MUTEX_EXIT(&rx_stats_mutex);
2602 MUTEX_ENTER(&conn->conn_data_lock);
2604 MUTEX_EXIT(&conn->conn_data_lock);
2609 currentCallNumber = conn->callNumber[channel];
2611 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2612 if (np->header.callNumber < currentCallNumber) {
2613 MUTEX_ENTER(&rx_stats_mutex);
2614 rx_stats.spuriousPacketsRead++;
2615 MUTEX_EXIT(&rx_stats_mutex);
2616 #ifdef RX_ENABLE_LOCKS
2618 MUTEX_EXIT(&call->lock);
2620 MUTEX_ENTER(&conn->conn_data_lock);
2622 MUTEX_EXIT(&conn->conn_data_lock);
2626 MUTEX_ENTER(&conn->conn_call_lock);
2627 call = rxi_NewCall(conn, channel);
2628 MUTEX_EXIT(&conn->conn_call_lock);
2629 *call->callNumber = np->header.callNumber;
2630 if (np->header.callNumber == 0)
2631 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));
2633 call->state = RX_STATE_PRECALL;
2634 clock_GetTime(&call->queueTime);
2635 hzero(call->bytesSent);
2636 hzero(call->bytesRcvd);
2638 * If the number of queued calls exceeds the overload
2639 * threshold then abort this call.
2641 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2642 struct rx_packet *tp;
2644 rxi_CallError(call, rx_BusyError);
2645 tp = rxi_SendCallAbort(call, np, 1, 0);
2646 MUTEX_EXIT(&call->lock);
2647 MUTEX_ENTER(&conn->conn_data_lock);
2649 MUTEX_EXIT(&conn->conn_data_lock);
2650 MUTEX_ENTER(&rx_stats_mutex);
2652 MUTEX_EXIT(&rx_stats_mutex);
2655 rxi_KeepAliveOn(call);
2656 } else if (np->header.callNumber != currentCallNumber) {
2657 /* Wait until the transmit queue is idle before deciding
2658 * whether to reset the current call. Chances are that the
2659 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2662 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2663 while ((call->state == RX_STATE_ACTIVE)
2664 && (call->flags & RX_CALL_TQ_BUSY)) {
2665 call->flags |= RX_CALL_TQ_WAIT;
2667 #ifdef RX_ENABLE_LOCKS
2668 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2669 CV_WAIT(&call->cv_tq, &call->lock);
2670 #else /* RX_ENABLE_LOCKS */
2671 osi_rxSleep(&call->tq);
2672 #endif /* RX_ENABLE_LOCKS */
2674 if (call->tqWaiters == 0)
2675 call->flags &= ~RX_CALL_TQ_WAIT;
2677 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2678 /* If the new call cannot be taken right now send a busy and set
2679 * the error condition in this call, so that it terminates as
2680 * quickly as possible */
2681 if (call->state == RX_STATE_ACTIVE) {
2682 struct rx_packet *tp;
2684 rxi_CallError(call, RX_CALL_DEAD);
2685 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2687 MUTEX_EXIT(&call->lock);
2688 MUTEX_ENTER(&conn->conn_data_lock);
2690 MUTEX_EXIT(&conn->conn_data_lock);
2693 rxi_ResetCall(call, 0);
2694 *call->callNumber = np->header.callNumber;
2695 if (np->header.callNumber == 0)
2696 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));
2698 call->state = RX_STATE_PRECALL;
2699 clock_GetTime(&call->queueTime);
2700 hzero(call->bytesSent);
2701 hzero(call->bytesRcvd);
2703 * If the number of queued calls exceeds the overload
2704 * threshold then abort this call.
2706 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2707 struct rx_packet *tp;
2709 rxi_CallError(call, rx_BusyError);
2710 tp = rxi_SendCallAbort(call, np, 1, 0);
2711 MUTEX_EXIT(&call->lock);
2712 MUTEX_ENTER(&conn->conn_data_lock);
2714 MUTEX_EXIT(&conn->conn_data_lock);
2715 MUTEX_ENTER(&rx_stats_mutex);
2717 MUTEX_EXIT(&rx_stats_mutex);
2720 rxi_KeepAliveOn(call);
2722 /* Continuing call; do nothing here. */
2724 } else { /* we're the client */
2725 /* Ignore all incoming acknowledgements for calls in DALLY state */
2726 if (call && (call->state == RX_STATE_DALLY)
2727 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2728 MUTEX_ENTER(&rx_stats_mutex);
2729 rx_stats.ignorePacketDally++;
2730 MUTEX_EXIT(&rx_stats_mutex);
2731 #ifdef RX_ENABLE_LOCKS
2733 MUTEX_EXIT(&call->lock);
2736 MUTEX_ENTER(&conn->conn_data_lock);
2738 MUTEX_EXIT(&conn->conn_data_lock);
2742 /* Ignore anything that's not relevant to the current call. If there
2743 * isn't a current call, then no packet is relevant. */
2744 if (!call || (np->header.callNumber != currentCallNumber)) {
2745 MUTEX_ENTER(&rx_stats_mutex);
2746 rx_stats.spuriousPacketsRead++;
2747 MUTEX_EXIT(&rx_stats_mutex);
2748 #ifdef RX_ENABLE_LOCKS
2750 MUTEX_EXIT(&call->lock);
2753 MUTEX_ENTER(&conn->conn_data_lock);
2755 MUTEX_EXIT(&conn->conn_data_lock);
2758 /* If the service security object index stamped in the packet does not
2759 * match the connection's security index, ignore the packet */
2760 if (np->header.securityIndex != conn->securityIndex) {
2761 #ifdef RX_ENABLE_LOCKS
2762 MUTEX_EXIT(&call->lock);
2764 MUTEX_ENTER(&conn->conn_data_lock);
2766 MUTEX_EXIT(&conn->conn_data_lock);
2770 /* If we're receiving the response, then all transmit packets are
2771 * implicitly acknowledged. Get rid of them. */
2772 if (np->header.type == RX_PACKET_TYPE_DATA) {
2773 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2774 /* XXX Hack. Because we must release the global rx lock when
2775 * sending packets (osi_NetSend) we drop all acks while we're
2776 * traversing the tq in rxi_Start sending packets out because
2777 * packets may move to the freePacketQueue as result of being here!
2778 * So we drop these packets until we're safely out of the
2779 * traversing. Really ugly!
2780 * For fine grain RX locking, we set the acked field in the
2781 * packets and let rxi_Start remove them from the transmit queue.
2783 if (call->flags & RX_CALL_TQ_BUSY) {
2784 #ifdef RX_ENABLE_LOCKS
2785 rxi_SetAcksInTransmitQueue(call);
2788 return np; /* xmitting; drop packet */
2791 rxi_ClearTransmitQueue(call, 0);
2793 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2794 rxi_ClearTransmitQueue(call, 0);
2795 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2797 if (np->header.type == RX_PACKET_TYPE_ACK) {
2798 /* now check to see if this is an ack packet acknowledging that the
2799 * server actually *lost* some hard-acked data. If this happens we
2800 * ignore this packet, as it may indicate that the server restarted in
2801 * the middle of a call. It is also possible that this is an old ack
2802 * packet. We don't abort the connection in this case, because this
2803 * *might* just be an old ack packet. The right way to detect a server
2804 * restart in the midst of a call is to notice that the server epoch
2806 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2807 * XXX unacknowledged. I think that this is off-by-one, but
2808 * XXX I don't dare change it just yet, since it will
2809 * XXX interact badly with the server-restart detection
2810 * XXX code in receiveackpacket. */
2811 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2812 MUTEX_ENTER(&rx_stats_mutex);
2813 rx_stats.spuriousPacketsRead++;
2814 MUTEX_EXIT(&rx_stats_mutex);
2815 MUTEX_EXIT(&call->lock);
2816 MUTEX_ENTER(&conn->conn_data_lock);
2818 MUTEX_EXIT(&conn->conn_data_lock);
2822 } /* else not a data packet */
2825 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2826 /* Set remote user defined status from packet */
2827 call->remoteStatus = np->header.userStatus;
2829 /* Note the gap between the expected next packet and the actual
2830 * packet that arrived, when the new packet has a smaller serial number
2831 * than expected. Rioses frequently reorder packets all by themselves,
2832 * so this will be quite important with very large window sizes.
2833 * Skew is checked against 0 here to avoid any dependence on the type of
2834 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2836 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2837 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2838 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2840 MUTEX_ENTER(&conn->conn_data_lock);
2841 skew = conn->lastSerial - np->header.serial;
2842 conn->lastSerial = np->header.serial;
2843 MUTEX_EXIT(&conn->conn_data_lock);
2845 register struct rx_peer *peer;
2847 if (skew > peer->inPacketSkew) {
2848 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2850 peer->inPacketSkew = skew;
2854 /* Now do packet type-specific processing */
2855 switch (np->header.type) {
2856 case RX_PACKET_TYPE_DATA:
2857 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2860 case RX_PACKET_TYPE_ACK:
2861 /* Respond immediately to ack packets requesting acknowledgement
2863 if (np->header.flags & RX_REQUEST_ACK) {
2865 (void)rxi_SendCallAbort(call, 0, 1, 0);
2867 (void)rxi_SendAck(call, 0, np->header.serial,
2868 RX_ACK_PING_RESPONSE, 1);
2870 np = rxi_ReceiveAckPacket(call, np, 1);
2872 case RX_PACKET_TYPE_ABORT: {
2873 /* An abort packet: reset the call, passing the error up to the user. */
2874 /* What if error is zero? */
2875 /* What if the error is -1? the application will treat it as a timeout. */
2876 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
2877 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
2878 rxi_CallError(call, errdata);
2879 MUTEX_EXIT(&call->lock);
2880 MUTEX_ENTER(&conn->conn_data_lock);
2882 MUTEX_EXIT(&conn->conn_data_lock);
2883 return np; /* xmitting; drop packet */
2885 case RX_PACKET_TYPE_BUSY:
2888 case RX_PACKET_TYPE_ACKALL:
2889 /* All packets acknowledged, so we can drop all packets previously
2890 * readied for sending */
2891 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2892 /* XXX Hack. We because we can't release the global rx lock when
2893 * sending packets (osi_NetSend) we drop all ack pkts while we're
2894 * traversing the tq in rxi_Start sending packets out because
2895 * packets may move to the freePacketQueue as result of being
2896 * here! So we drop these packets until we're safely out of the
2897 * traversing. Really ugly!
2898 * For fine grain RX locking, we set the acked field in the packets
2899 * and let rxi_Start remove the packets from the transmit queue.
2901 if (call->flags & RX_CALL_TQ_BUSY) {
2902 #ifdef RX_ENABLE_LOCKS
2903 rxi_SetAcksInTransmitQueue(call);
2905 #else /* RX_ENABLE_LOCKS */
2906 MUTEX_EXIT(&call->lock);
2907 MUTEX_ENTER(&conn->conn_data_lock);
2909 MUTEX_EXIT(&conn->conn_data_lock);
2910 return np; /* xmitting; drop packet */
2911 #endif /* RX_ENABLE_LOCKS */
2913 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2914 rxi_ClearTransmitQueue(call, 0);
2917 /* Should not reach here, unless the peer is broken: send an abort
2919 rxi_CallError(call, RX_PROTOCOL_ERROR);
2920 np = rxi_SendCallAbort(call, np, 1, 0);
2923 /* Note when this last legitimate packet was received, for keep-alive
2924 * processing. Note, we delay getting the time until now in the hope that
2925 * the packet will be delivered to the user before any get time is required
2926 * (if not, then the time won't actually be re-evaluated here). */
2927 call->lastReceiveTime = clock_Sec();
2928 MUTEX_EXIT(&call->lock);
2929 MUTEX_ENTER(&conn->conn_data_lock);
2931 MUTEX_EXIT(&conn->conn_data_lock);
2935 /* return true if this is an "interesting" connection from the point of view
2936 of someone trying to debug the system */
2938 rxi_IsConnInteresting(struct rx_connection *aconn)
2941 register struct rx_call *tcall;
2943 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2945 for (i = 0; i < RX_MAXCALLS; i++) {
2946 tcall = aconn->call[i];
2948 if ((tcall->state == RX_STATE_PRECALL)
2949 || (tcall->state == RX_STATE_ACTIVE))
2951 if ((tcall->mode == RX_MODE_SENDING)
2952 || (tcall->mode == RX_MODE_RECEIVING))
2960 /* if this is one of the last few packets AND it wouldn't be used by the
2961 receiving call to immediately satisfy a read request, then drop it on
2962 the floor, since accepting it might prevent a lock-holding thread from
2963 making progress in its reading. If a call has been cleared while in
2964 the precall state then ignore all subsequent packets until the call
2965 is assigned to a thread. */
2968 TooLow(struct rx_packet *ap, struct rx_call *acall)
2971 MUTEX_ENTER(&rx_stats_mutex);
2972 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
2973 && (acall->state == RX_STATE_PRECALL))
2974 || ((rx_nFreePackets < rxi_dataQuota + 2)
2975 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
2976 && (acall->flags & RX_CALL_READER_WAIT)))) {
2979 MUTEX_EXIT(&rx_stats_mutex);
2985 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
2986 struct rx_call *acall)
2988 struct rx_call *call = acall;
2992 MUTEX_ENTER(&conn->conn_data_lock);
2993 conn->checkReachEvent = NULL;
2994 waiting = conn->flags & RX_CONN_ATTACHWAIT;
2997 MUTEX_EXIT(&conn->conn_data_lock);
3001 MUTEX_ENTER(&conn->conn_call_lock);
3002 MUTEX_ENTER(&conn->conn_data_lock);
3003 for (i = 0; i < RX_MAXCALLS; i++) {
3004 struct rx_call *tc = conn->call[i];
3005 if (tc && tc->state == RX_STATE_PRECALL) {
3011 /* Indicate that rxi_CheckReachEvent is no longer running by
3012 * clearing the flag. Must be atomic under conn_data_lock to
3013 * avoid a new call slipping by: rxi_CheckConnReach holds
3014 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3016 conn->flags &= ~RX_CONN_ATTACHWAIT;
3017 MUTEX_EXIT(&conn->conn_data_lock);
3018 MUTEX_EXIT(&conn->conn_call_lock);
3023 MUTEX_ENTER(&call->lock);
3024 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3026 MUTEX_EXIT(&call->lock);
3028 clock_GetTime(&when);
3029 when.sec += RX_CHECKREACH_TIMEOUT;
3030 MUTEX_ENTER(&conn->conn_data_lock);
3031 if (!conn->checkReachEvent) {
3033 conn->checkReachEvent =
3034 rxevent_Post(&when, rxi_CheckReachEvent, conn, NULL);
3036 MUTEX_EXIT(&conn->conn_data_lock);
3042 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3044 struct rx_service *service = conn->service;
3045 struct rx_peer *peer = conn->peer;
3046 afs_uint32 now, lastReach;
3048 if (service->checkReach == 0)
3052 MUTEX_ENTER(&peer->peer_lock);
3053 lastReach = peer->lastReachTime;
3054 MUTEX_EXIT(&peer->peer_lock);
3055 if (now - lastReach < RX_CHECKREACH_TTL)
3058 MUTEX_ENTER(&conn->conn_data_lock);
3059 if (conn->flags & RX_CONN_ATTACHWAIT) {
3060 MUTEX_EXIT(&conn->conn_data_lock);
3063 conn->flags |= RX_CONN_ATTACHWAIT;
3064 MUTEX_EXIT(&conn->conn_data_lock);
3065 if (!conn->checkReachEvent)
3066 rxi_CheckReachEvent(NULL, conn, call);
3071 /* try to attach call, if authentication is complete */
3073 TryAttach(register struct rx_call *acall, register osi_socket socket,
3074 register int *tnop, register struct rx_call **newcallp,
3077 struct rx_connection *conn = acall->conn;
3079 if (conn->type == RX_SERVER_CONNECTION
3080 && acall->state == RX_STATE_PRECALL) {
3081 /* Don't attach until we have any req'd. authentication. */
3082 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3083 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3084 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3085 /* Note: this does not necessarily succeed; there
3086 * may not any proc available
3089 rxi_ChallengeOn(acall->conn);
3094 /* A data packet has been received off the interface. This packet is
3095 * appropriate to the call (the call is in the right state, etc.). This
3096 * routine can return a packet to the caller, for re-use */
3099 rxi_ReceiveDataPacket(register struct rx_call *call,
3100 register struct rx_packet *np, int istack,
3101 osi_socket socket, afs_uint32 host, u_short port,
3102 int *tnop, struct rx_call **newcallp)
3104 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3108 afs_uint32 seq, serial, flags;
3110 struct rx_packet *tnp;
3112 MUTEX_ENTER(&rx_stats_mutex);
3113 rx_stats.dataPacketsRead++;
3114 MUTEX_EXIT(&rx_stats_mutex);
3117 /* If there are no packet buffers, drop this new packet, unless we can find
3118 * packet buffers from inactive calls */
3120 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3121 MUTEX_ENTER(&rx_freePktQ_lock);
3122 rxi_NeedMorePackets = TRUE;
3123 MUTEX_EXIT(&rx_freePktQ_lock);
3124 MUTEX_ENTER(&rx_stats_mutex);
3125 rx_stats.noPacketBuffersOnRead++;
3126 MUTEX_EXIT(&rx_stats_mutex);
3127 call->rprev = np->header.serial;
3128 rxi_calltrace(RX_TRACE_DROP, call);
3129 dpf(("packet %x dropped on receipt - quota problems", np));
3131 rxi_ClearReceiveQueue(call);
3132 clock_GetTime(&when);
3133 clock_Add(&when, &rx_softAckDelay);
3134 if (!call->delayedAckEvent
3135 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3136 rxevent_Cancel(call->delayedAckEvent, call,
3137 RX_CALL_REFCOUNT_DELAY);
3138 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3139 call->delayedAckEvent =
3140 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3142 /* we've damaged this call already, might as well do it in. */
3148 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3149 * packet is one of several packets transmitted as a single
3150 * datagram. Do not send any soft or hard acks until all packets
3151 * in a jumbogram have been processed. Send negative acks right away.
3153 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3154 /* tnp is non-null when there are more packets in the
3155 * current jumbo gram */
3162 seq = np->header.seq;
3163 serial = np->header.serial;
3164 flags = np->header.flags;
3166 /* If the call is in an error state, send an abort message */
3168 return rxi_SendCallAbort(call, np, istack, 0);
3170 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3171 * AFS 3.5 jumbogram. */
3172 if (flags & RX_JUMBO_PACKET) {
3173 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3178 if (np->header.spare != 0) {
3179 MUTEX_ENTER(&call->conn->conn_data_lock);
3180 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3181 MUTEX_EXIT(&call->conn->conn_data_lock);
3184 /* The usual case is that this is the expected next packet */
3185 if (seq == call->rnext) {
3187 /* Check to make sure it is not a duplicate of one already queued */
3188 if (queue_IsNotEmpty(&call->rq)
3189 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3190 MUTEX_ENTER(&rx_stats_mutex);
3191 rx_stats.dupPacketsRead++;
3192 MUTEX_EXIT(&rx_stats_mutex);
3193 dpf(("packet %x dropped on receipt - duplicate", np));
3194 rxevent_Cancel(call->delayedAckEvent, call,
3195 RX_CALL_REFCOUNT_DELAY);
3196 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3202 /* It's the next packet. Stick it on the receive queue
3203 * for this call. Set newPackets to make sure we wake
3204 * the reader once all packets have been processed */
3205 queue_Prepend(&call->rq, np);
3207 np = NULL; /* We can't use this anymore */
3210 /* If an ack is requested then set a flag to make sure we
3211 * send an acknowledgement for this packet */
3212 if (flags & RX_REQUEST_ACK) {
3213 ackNeeded = RX_ACK_REQUESTED;
3216 /* Keep track of whether we have received the last packet */
3217 if (flags & RX_LAST_PACKET) {
3218 call->flags |= RX_CALL_HAVE_LAST;
3222 /* Check whether we have all of the packets for this call */
3223 if (call->flags & RX_CALL_HAVE_LAST) {
3224 afs_uint32 tseq; /* temporary sequence number */
3225 struct rx_packet *tp; /* Temporary packet pointer */
3226 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3228 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3229 if (tseq != tp->header.seq)
3231 if (tp->header.flags & RX_LAST_PACKET) {
3232 call->flags |= RX_CALL_RECEIVE_DONE;
3239 /* Provide asynchronous notification for those who want it
3240 * (e.g. multi rx) */
3241 if (call->arrivalProc) {
3242 (*call->arrivalProc) (call, call->arrivalProcHandle,
3243 call->arrivalProcArg);
3244 call->arrivalProc = (void (*)())0;
3247 /* Update last packet received */
3250 /* If there is no server process serving this call, grab
3251 * one, if available. We only need to do this once. If a
3252 * server thread is available, this thread becomes a server
3253 * thread and the server thread becomes a listener thread. */
3255 TryAttach(call, socket, tnop, newcallp, 0);
3258 /* This is not the expected next packet. */
3260 /* Determine whether this is a new or old packet, and if it's
3261 * a new one, whether it fits into the current receive window.
3262 * Also figure out whether the packet was delivered in sequence.
3263 * We use the prev variable to determine whether the new packet
3264 * is the successor of its immediate predecessor in the
3265 * receive queue, and the missing flag to determine whether
3266 * any of this packets predecessors are missing. */
3268 afs_uint32 prev; /* "Previous packet" sequence number */
3269 struct rx_packet *tp; /* Temporary packet pointer */
3270 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3271 int missing; /* Are any predecessors missing? */
3273 /* If the new packet's sequence number has been sent to the
3274 * application already, then this is a duplicate */
3275 if (seq < call->rnext) {
3276 MUTEX_ENTER(&rx_stats_mutex);
3277 rx_stats.dupPacketsRead++;
3278 MUTEX_EXIT(&rx_stats_mutex);
3279 rxevent_Cancel(call->delayedAckEvent, call,
3280 RX_CALL_REFCOUNT_DELAY);
3281 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3287 /* If the sequence number is greater than what can be
3288 * accomodated by the current window, then send a negative
3289 * acknowledge and drop the packet */
3290 if ((call->rnext + call->rwind) <= seq) {
3291 rxevent_Cancel(call->delayedAckEvent, call,
3292 RX_CALL_REFCOUNT_DELAY);
3293 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3300 /* Look for the packet in the queue of old received packets */
3301 for (prev = call->rnext - 1, missing =
3302 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3303 /*Check for duplicate packet */
3304 if (seq == tp->header.seq) {
3305 MUTEX_ENTER(&rx_stats_mutex);
3306 rx_stats.dupPacketsRead++;
3307 MUTEX_EXIT(&rx_stats_mutex);
3308 rxevent_Cancel(call->delayedAckEvent, call,
3309 RX_CALL_REFCOUNT_DELAY);
3310 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3316 /* If we find a higher sequence packet, break out and
3317 * insert the new packet here. */
3318 if (seq < tp->header.seq)
3320 /* Check for missing packet */
3321 if (tp->header.seq != prev + 1) {
3325 prev = tp->header.seq;
3328 /* Keep track of whether we have received the last packet. */
3329 if (flags & RX_LAST_PACKET) {
3330 call->flags |= RX_CALL_HAVE_LAST;
3333 /* It's within the window: add it to the the receive queue.
3334 * tp is left by the previous loop either pointing at the
3335 * packet before which to insert the new packet, or at the
3336 * queue head if the queue is empty or the packet should be
3338 queue_InsertBefore(tp, np);
3342 /* Check whether we have all of the packets for this call */
3343 if ((call->flags & RX_CALL_HAVE_LAST)
3344 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3345 afs_uint32 tseq; /* temporary sequence number */
3348 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3349 if (tseq != tp->header.seq)
3351 if (tp->header.flags & RX_LAST_PACKET) {
3352 call->flags |= RX_CALL_RECEIVE_DONE;
3359 /* We need to send an ack of the packet is out of sequence,
3360 * or if an ack was requested by the peer. */
3361 if (seq != prev + 1 || missing) {
3362 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3363 } else if (flags & RX_REQUEST_ACK) {
3364 ackNeeded = RX_ACK_REQUESTED;
3367 /* Acknowledge the last packet for each call */
3368 if (flags & RX_LAST_PACKET) {
3379 * If the receiver is waiting for an iovec, fill the iovec
3380 * using the data from the receive queue */
3381 if (call->flags & RX_CALL_IOVEC_WAIT) {
3382 didHardAck = rxi_FillReadVec(call, serial);
3383 /* the call may have been aborted */
3392 /* Wakeup the reader if any */
3393 if ((call->flags & RX_CALL_READER_WAIT)
3394 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3395 || (call->iovNext >= call->iovMax)
3396 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3397 call->flags &= ~RX_CALL_READER_WAIT;
3398 #ifdef RX_ENABLE_LOCKS
3399 CV_BROADCAST(&call->cv_rq);
3401 osi_rxWakeup(&call->rq);
3407 * Send an ack when requested by the peer, or once every
3408 * rxi_SoftAckRate packets until the last packet has been
3409 * received. Always send a soft ack for the last packet in
3410 * the server's reply. */
3412 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3413 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3414 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3415 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3416 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3417 } else if (call->nSoftAcks) {
3418 clock_GetTime(&when);
3419 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3420 clock_Add(&when, &rx_lastAckDelay);
3422 clock_Add(&when, &rx_softAckDelay);
3424 if (!call->delayedAckEvent
3425 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3426 rxevent_Cancel(call->delayedAckEvent, call,
3427 RX_CALL_REFCOUNT_DELAY);
3428 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3429 call->delayedAckEvent =
3430 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3432 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3433 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3440 static void rxi_ComputeRate();
3444 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3446 struct rx_peer *peer = conn->peer;
3448 MUTEX_ENTER(&peer->peer_lock);
3449 peer->lastReachTime = clock_Sec();
3450 MUTEX_EXIT(&peer->peer_lock);
3452 MUTEX_ENTER(&conn->conn_data_lock);
3453 if (conn->flags & RX_CONN_ATTACHWAIT) {
3456 conn->flags &= ~RX_CONN_ATTACHWAIT;
3457 MUTEX_EXIT(&conn->conn_data_lock);
3459 for (i = 0; i < RX_MAXCALLS; i++) {
3460 struct rx_call *call = conn->call[i];
3463 MUTEX_ENTER(&call->lock);
3464 /* tnop can be null if newcallp is null */
3465 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3467 MUTEX_EXIT(&call->lock);
3471 MUTEX_EXIT(&conn->conn_data_lock);
3475 rx_ack_reason(int reason)
3478 case RX_ACK_REQUESTED:
3480 case RX_ACK_DUPLICATE:
3482 case RX_ACK_OUT_OF_SEQUENCE:
3484 case RX_ACK_EXCEEDS_WINDOW:
3486 case RX_ACK_NOSPACE:
3490 case RX_ACK_PING_RESPONSE:
3502 /* rxi_ComputePeerNetStats
3504 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3505 * estimates (like RTT and throughput) based on ack packets. Caller
3506 * must ensure that the packet in question is the right one (i.e.
3507 * serial number matches).
3510 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3511 struct rx_ackPacket *ap, struct rx_packet *np)
3513 struct rx_peer *peer = call->conn->peer;
3515 /* Use RTT if not delayed by client. */
3516 if (ap->reason != RX_ACK_DELAY)
3517 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3519 rxi_ComputeRate(peer, call, p, np, ap->reason);
3523 /* The real smarts of the whole thing. */
3525 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3528 struct rx_ackPacket *ap;
3530 register struct rx_packet *tp;
3531 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3532 register struct rx_connection *conn = call->conn;
3533 struct rx_peer *peer = conn->peer;
3536 /* because there are CM's that are bogus, sending weird values for this. */
3537 afs_uint32 skew = 0;
3542 int newAckCount = 0;
3543 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3544 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3546 MUTEX_ENTER(&rx_stats_mutex);
3547 rx_stats.ackPacketsRead++;
3548 MUTEX_EXIT(&rx_stats_mutex);
3549 ap = (struct rx_ackPacket *)rx_DataOf(np);
3550 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3552 return np; /* truncated ack packet */
3554 /* depends on ack packet struct */
3555 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3556 first = ntohl(ap->firstPacket);
3557 serial = ntohl(ap->serial);
3558 /* temporarily disabled -- needs to degrade over time
3559 * skew = ntohs(ap->maxSkew); */
3561 /* Ignore ack packets received out of order */
3562 if (first < call->tfirst) {
3566 if (np->header.flags & RX_SLOW_START_OK) {
3567 call->flags |= RX_CALL_SLOW_START_OK;
3570 if (ap->reason == RX_ACK_PING_RESPONSE)
3571 rxi_UpdatePeerReach(conn, call);
3575 if (rxdebug_active) {
3579 len = _snprintf(msg, sizeof(msg),
3580 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3581 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3582 ntohl(ap->serial), ntohl(ap->previousPacket),
3583 (unsigned int)np->header.seq, (unsigned int)skew,
3584 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3588 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3589 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3593 OutputDebugString(msg);
3595 #else /* AFS_NT40_ENV */
3598 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3599 ap->reason, ntohl(ap->previousPacket),
3600 (unsigned int)np->header.seq, (unsigned int)serial,
3601 (unsigned int)skew, ntohl(ap->firstPacket));
3604 for (offset = 0; offset < nAcks; offset++)
3605 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3610 #endif /* AFS_NT40_ENV */
3613 /* Update the outgoing packet skew value to the latest value of
3614 * the peer's incoming packet skew value. The ack packet, of
3615 * course, could arrive out of order, but that won't affect things
3617 MUTEX_ENTER(&peer->peer_lock);
3618 peer->outPacketSkew = skew;
3620 /* Check for packets that no longer need to be transmitted, and
3621 * discard them. This only applies to packets positively
3622 * acknowledged as having been sent to the peer's upper level.
3623 * All other packets must be retained. So only packets with
3624 * sequence numbers < ap->firstPacket are candidates. */
3625 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3626 if (tp->header.seq >= first)
3628 call->tfirst = tp->header.seq + 1;
3630 && (tp->header.serial == serial || tp->firstSerial == serial))
3631 rxi_ComputePeerNetStats(call, tp, ap, np);
3632 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3635 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3636 /* XXX Hack. Because we have to release the global rx lock when sending
3637 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3638 * in rxi_Start sending packets out because packets may move to the
3639 * freePacketQueue as result of being here! So we drop these packets until
3640 * we're safely out of the traversing. Really ugly!
3641 * To make it even uglier, if we're using fine grain locking, we can
3642 * set the ack bits in the packets and have rxi_Start remove the packets
3643 * when it's done transmitting.
3645 if (call->flags & RX_CALL_TQ_BUSY) {
3646 #ifdef RX_ENABLE_LOCKS
3647 tp->flags |= RX_PKTFLAG_ACKED;
3648 call->flags |= RX_CALL_TQ_SOME_ACKED;
3649 #else /* RX_ENABLE_LOCKS */
3651 #endif /* RX_ENABLE_LOCKS */
3653 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3656 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3661 /* Give rate detector a chance to respond to ping requests */
3662 if (ap->reason == RX_ACK_PING_RESPONSE) {
3663 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3667 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3669 /* Now go through explicit acks/nacks and record the results in
3670 * the waiting packets. These are packets that can't be released
3671 * yet, even with a positive acknowledge. This positive
3672 * acknowledge only means the packet has been received by the
3673 * peer, not that it will be retained long enough to be sent to
3674 * the peer's upper level. In addition, reset the transmit timers
3675 * of any missing packets (those packets that must be missing
3676 * because this packet was out of sequence) */
3678 call->nSoftAcked = 0;
3679 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3680 /* Update round trip time if the ack was stimulated on receipt
3682 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3683 #ifdef RX_ENABLE_LOCKS
3684 if (tp->header.seq >= first)
3685 #endif /* RX_ENABLE_LOCKS */
3686 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3688 && (tp->header.serial == serial || tp->firstSerial == serial))
3689 rxi_ComputePeerNetStats(call, tp, ap, np);
3691 /* Set the acknowledge flag per packet based on the
3692 * information in the ack packet. An acknowlegded packet can
3693 * be downgraded when the server has discarded a packet it
3694 * soacked previously, or when an ack packet is received
3695 * out of sequence. */
3696 if (tp->header.seq < first) {
3697 /* Implicit ack information */
3698 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3701 tp->flags |= RX_PKTFLAG_ACKED;
3702 } else if (tp->header.seq < first + nAcks) {
3703 /* Explicit ack information: set it in the packet appropriately */
3704 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3705 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3707 tp->flags |= RX_PKTFLAG_ACKED;
3714 } else /* RX_ACK_TYPE_NACK */ {
3715 tp->flags &= ~RX_PKTFLAG_ACKED;
3719 tp->flags &= ~RX_PKTFLAG_ACKED;
3723 /* If packet isn't yet acked, and it has been transmitted at least
3724 * once, reset retransmit time using latest timeout
3725 * ie, this should readjust the retransmit timer for all outstanding
3726 * packets... So we don't just retransmit when we should know better*/
3728 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3729 tp->retryTime = tp->timeSent;
3730 clock_Add(&tp->retryTime, &peer->timeout);
3731 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3732 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3736 /* If the window has been extended by this acknowledge packet,
3737 * then wakeup a sender waiting in alloc for window space, or try
3738 * sending packets now, if he's been sitting on packets due to
3739 * lack of window space */
3740 if (call->tnext < (call->tfirst + call->twind)) {
3741 #ifdef RX_ENABLE_LOCKS
3742 CV_SIGNAL(&call->cv_twind);
3744 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3745 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3746 osi_rxWakeup(&call->twind);
3749 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3750 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3754 /* if the ack packet has a receivelen field hanging off it,
3755 * update our state */
3756 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3759 /* If the ack packet has a "recommended" size that is less than
3760 * what I am using now, reduce my size to match */
3761 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3762 (int)sizeof(afs_int32), &tSize);
3763 tSize = (afs_uint32) ntohl(tSize);
3764 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3766 /* Get the maximum packet size to send to this peer */
3767 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
3769 tSize = (afs_uint32) ntohl(tSize);
3770 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3771 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3773 /* sanity check - peer might have restarted with different params.
3774 * If peer says "send less", dammit, send less... Peer should never
3775 * be unable to accept packets of the size that prior AFS versions would
3776 * send without asking. */
3777 if (peer->maxMTU != tSize) {
3778 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
3780 peer->maxMTU = tSize;
3781 peer->MTU = MIN(tSize, peer->MTU);
3782 call->MTU = MIN(call->MTU, tSize);
3785 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3788 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3789 (int)sizeof(afs_int32), &tSize);
3790 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3791 if (tSize < call->twind) { /* smaller than our send */
3792 call->twind = tSize; /* window, we must send less... */
3793 call->ssthresh = MIN(call->twind, call->ssthresh);
3796 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3797 * network MTU confused with the loopback MTU. Calculate the
3798 * maximum MTU here for use in the slow start code below.
3800 maxMTU = peer->maxMTU;
3801 /* Did peer restart with older RX version? */
3802 if (peer->maxDgramPackets > 1) {
3803 peer->maxDgramPackets = 1;
3805 } else if (np->length >=
3806 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3809 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3810 sizeof(afs_int32), &tSize);
3811 tSize = (afs_uint32) ntohl(tSize);
3813 * As of AFS 3.5 we set the send window to match the receive window.
3815 if (tSize < call->twind) {
3816 call->twind = tSize;
3817 call->ssthresh = MIN(call->twind, call->ssthresh);
3818 } else if (tSize > call->twind) {
3819 call->twind = tSize;
3823 * As of AFS 3.5, a jumbogram is more than one fixed size
3824 * packet transmitted in a single UDP datagram. If the remote
3825 * MTU is smaller than our local MTU then never send a datagram
3826 * larger than the natural MTU.
3829 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3830 sizeof(afs_int32), &tSize);
3831 maxDgramPackets = (afs_uint32) ntohl(tSize);
3832 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3834 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
3835 maxDgramPackets = MIN(maxDgramPackets, tSize);
3836 if (maxDgramPackets > 1) {
3837 peer->maxDgramPackets = maxDgramPackets;
3838 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3840 peer->maxDgramPackets = 1;
3841 call->MTU = peer->natMTU;
3843 } else if (peer->maxDgramPackets > 1) {
3844 /* Restarted with lower version of RX */
3845 peer->maxDgramPackets = 1;
3847 } else if (peer->maxDgramPackets > 1
3848 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3849 /* Restarted with lower version of RX */
3850 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3851 peer->natMTU = OLD_MAX_PACKET_SIZE;
3852 peer->MTU = OLD_MAX_PACKET_SIZE;
3853 peer->maxDgramPackets = 1;
3854 peer->nDgramPackets = 1;
3856 call->MTU = OLD_MAX_PACKET_SIZE;
3861 * Calculate how many datagrams were successfully received after
3862 * the first missing packet and adjust the negative ack counter
3867 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3868 if (call->nNacks < nNacked) {
3869 call->nNacks = nNacked;
3878 if (call->flags & RX_CALL_FAST_RECOVER) {
3880 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3882 call->flags &= ~RX_CALL_FAST_RECOVER;
3883 call->cwind = call->nextCwind;
3884 call->nextCwind = 0;
3887 call->nCwindAcks = 0;
3888 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3889 /* Three negative acks in a row trigger congestion recovery */
3890 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3891 MUTEX_EXIT(&peer->peer_lock);
3892 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3893 /* someone else is waiting to start recovery */
3896 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3897 rxi_WaitforTQBusy(call);
3898 MUTEX_ENTER(&peer->peer_lock);
3899 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3900 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3901 call->flags |= RX_CALL_FAST_RECOVER;
3902 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3904 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3905 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3906 call->nextCwind = call->ssthresh;
3909 peer->MTU = call->MTU;
3910 peer->cwind = call->nextCwind;
3911 peer->nDgramPackets = call->nDgramPackets;
3913 call->congestSeq = peer->congestSeq;
3914 /* Reset the resend times on the packets that were nacked
3915 * so we will retransmit as soon as the window permits*/
3916 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3918 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3919 clock_Zero(&tp->retryTime);
3921 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3926 /* If cwind is smaller than ssthresh, then increase
3927 * the window one packet for each ack we receive (exponential
3929 * If cwind is greater than or equal to ssthresh then increase
3930 * the congestion window by one packet for each cwind acks we
3931 * receive (linear growth). */
3932 if (call->cwind < call->ssthresh) {
3934 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3935 call->nCwindAcks = 0;
3937 call->nCwindAcks += newAckCount;
3938 if (call->nCwindAcks >= call->cwind) {
3939 call->nCwindAcks = 0;
3940 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3944 * If we have received several acknowledgements in a row then
3945 * it is time to increase the size of our datagrams
3947 if ((int)call->nAcks > rx_nDgramThreshold) {
3948 if (peer->maxDgramPackets > 1) {
3949 if (call->nDgramPackets < peer->maxDgramPackets) {
3950 call->nDgramPackets++;
3952 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3953 } else if (call->MTU < peer->maxMTU) {
3954 call->MTU += peer->natMTU;
3955 call->MTU = MIN(call->MTU, peer->maxMTU);
3961 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3963 /* Servers need to hold the call until all response packets have
3964 * been acknowledged. Soft acks are good enough since clients
3965 * are not allowed to clear their receive queues. */
3966 if (call->state == RX_STATE_HOLD
3967 && call->tfirst + call->nSoftAcked >= call->tnext) {
3968 call->state = RX_STATE_DALLY;
3969 rxi_ClearTransmitQueue(call, 0);
3970 } else if (!queue_IsEmpty(&call->tq)) {
3971 rxi_Start(0, call, 0, istack);
3976 /* Received a response to a challenge packet */
3978 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
3979 register struct rx_packet *np, int istack)
3983 /* Ignore the packet if we're the client */
3984 if (conn->type == RX_CLIENT_CONNECTION)
3987 /* If already authenticated, ignore the packet (it's probably a retry) */
3988 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
3991 /* Otherwise, have the security object evaluate the response packet */
3992 error = RXS_CheckResponse(conn->securityObject, conn, np);
3994 /* If the response is invalid, reset the connection, sending
3995 * an abort to the peer */
3999 rxi_ConnectionError(conn, error);
4000 MUTEX_ENTER(&conn->conn_data_lock);
4001 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4002 MUTEX_EXIT(&conn->conn_data_lock);
4005 /* If the response is valid, any calls waiting to attach
4006 * servers can now do so */
4009 for (i = 0; i < RX_MAXCALLS; i++) {
4010 struct rx_call *call = conn->call[i];
4012 MUTEX_ENTER(&call->lock);
4013 if (call->state == RX_STATE_PRECALL)
4014 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4015 /* tnop can be null if newcallp is null */
4016 MUTEX_EXIT(&call->lock);
4020 /* Update the peer reachability information, just in case
4021 * some calls went into attach-wait while we were waiting
4022 * for authentication..
4024 rxi_UpdatePeerReach(conn, NULL);
4029 /* A client has received an authentication challenge: the security
4030 * object is asked to cough up a respectable response packet to send
4031 * back to the server. The server is responsible for retrying the
4032 * challenge if it fails to get a response. */
4035 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
4036 register struct rx_packet *np, int istack)
4040 /* Ignore the challenge if we're the server */
4041 if (conn->type == RX_SERVER_CONNECTION)
4044 /* Ignore the challenge if the connection is otherwise idle; someone's
4045 * trying to use us as an oracle. */
4046 if (!rxi_HasActiveCalls(conn))
4049 /* Send the security object the challenge packet. It is expected to fill
4050 * in the response. */
4051 error = RXS_GetResponse(conn->securityObject, conn, np);
4053 /* If the security object is unable to return a valid response, reset the
4054 * connection and send an abort to the peer. Otherwise send the response
4055 * packet to the peer connection. */
4057 rxi_ConnectionError(conn, error);
4058 MUTEX_ENTER(&conn->conn_data_lock);
4059 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4060 MUTEX_EXIT(&conn->conn_data_lock);
4062 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4063 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4069 /* Find an available server process to service the current request in
4070 * the given call structure. If one isn't available, queue up this
4071 * call so it eventually gets one */
4073 rxi_AttachServerProc(register struct rx_call *call,
4074 register osi_socket socket, register int *tnop,
4075 register struct rx_call **newcallp)
4077 register struct rx_serverQueueEntry *sq;
4078 register struct rx_service *service = call->conn->service;
4079 register int haveQuota = 0;
4081 /* May already be attached */
4082 if (call->state == RX_STATE_ACTIVE)
4085 MUTEX_ENTER(&rx_serverPool_lock);
4087 haveQuota = QuotaOK(service);
4088 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4089 /* If there are no processes available to service this call,
4090 * put the call on the incoming call queue (unless it's
4091 * already on the queue).
4093 #ifdef RX_ENABLE_LOCKS
4095 ReturnToServerPool(service);
4096 #endif /* RX_ENABLE_LOCKS */
4098 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4099 call->flags |= RX_CALL_WAIT_PROC;
4100 MUTEX_ENTER(&rx_stats_mutex);
4103 MUTEX_EXIT(&rx_stats_mutex);
4104 rxi_calltrace(RX_CALL_ARRIVAL, call);
4105 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4106 queue_Append(&rx_incomingCallQueue, call);
4109 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4111 /* If hot threads are enabled, and both newcallp and sq->socketp
4112 * are non-null, then this thread will process the call, and the
4113 * idle server thread will start listening on this threads socket.
4116 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4119 *sq->socketp = socket;
4120 clock_GetTime(&call->startTime);
4121 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4125 if (call->flags & RX_CALL_WAIT_PROC) {
4126 /* Conservative: I don't think this should happen */
4127 call->flags &= ~RX_CALL_WAIT_PROC;
4128 if (queue_IsOnQueue(call)) {
4130 MUTEX_ENTER(&rx_stats_mutex);
4132 MUTEX_EXIT(&rx_stats_mutex);
4135 call->state = RX_STATE_ACTIVE;
4136 call->mode = RX_MODE_RECEIVING;
4137 #ifdef RX_KERNEL_TRACE
4139 int glockOwner = ISAFS_GLOCK();
4142 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4143 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4149 if (call->flags & RX_CALL_CLEARED) {
4150 /* send an ack now to start the packet flow up again */
4151 call->flags &= ~RX_CALL_CLEARED;
4152 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4154 #ifdef RX_ENABLE_LOCKS
4157 service->nRequestsRunning++;
4158 if (service->nRequestsRunning <= service->minProcs)
4164 MUTEX_EXIT(&rx_serverPool_lock);
4167 /* Delay the sending of an acknowledge event for a short while, while
4168 * a new call is being prepared (in the case of a client) or a reply
4169 * is being prepared (in the case of a server). Rather than sending
4170 * an ack packet, an ACKALL packet is sent. */
4172 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4174 #ifdef RX_ENABLE_LOCKS
4176 MUTEX_ENTER(&call->lock);
4177 call->delayedAckEvent = NULL;
4178 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4180 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4181 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4183 MUTEX_EXIT(&call->lock);
4184 #else /* RX_ENABLE_LOCKS */
4186 call->delayedAckEvent = NULL;
4187 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4188 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4189 #endif /* RX_ENABLE_LOCKS */
4193 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4196 #ifdef RX_ENABLE_LOCKS
4198 MUTEX_ENTER(&call->lock);
4199 if (event == call->delayedAckEvent)
4200 call->delayedAckEvent = NULL;
4201 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4203 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4205 MUTEX_EXIT(&call->lock);
4206 #else /* RX_ENABLE_LOCKS */
4208 call->delayedAckEvent = NULL;
4209 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4210 #endif /* RX_ENABLE_LOCKS */
4214 #ifdef RX_ENABLE_LOCKS
4215 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4216 * clearing them out.
4219 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4221 register struct rx_packet *p, *tp;
4224 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4225 p->flags |= RX_PKTFLAG_ACKED;
4229 call->flags |= RX_CALL_TQ_CLEARME;
4230 call->flags |= RX_CALL_TQ_SOME_ACKED;
4233 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4234 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4235 call->tfirst = call->tnext;
4236 call->nSoftAcked = 0;
4238 if (call->flags & RX_CALL_FAST_RECOVER) {
4239 call->flags &= ~RX_CALL_FAST_RECOVER;
4240 call->cwind = call->nextCwind;
4241 call->nextCwind = 0;
4244 CV_SIGNAL(&call->cv_twind);
4246 #endif /* RX_ENABLE_LOCKS */
4248 /* Clear out the transmit queue for the current call (all packets have
4249 * been received by peer) */
4251 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4253 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4254 register struct rx_packet *p, *tp;
4256 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4258 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4259 p->flags |= RX_PKTFLAG_ACKED;
4263 call->flags |= RX_CALL_TQ_CLEARME;
4264 call->flags |= RX_CALL_TQ_SOME_ACKED;
4267 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4268 rxi_FreePackets(0, &call->tq);
4269 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4270 call->flags &= ~RX_CALL_TQ_CLEARME;
4272 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4274 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4275 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4276 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4277 call->nSoftAcked = 0;
4279 if (call->flags & RX_CALL_FAST_RECOVER) {
4280 call->flags &= ~RX_CALL_FAST_RECOVER;
4281 call->cwind = call->nextCwind;
4283 #ifdef RX_ENABLE_LOCKS
4284 CV_SIGNAL(&call->cv_twind);
4286 osi_rxWakeup(&call->twind);
4291 rxi_ClearReceiveQueue(register struct rx_call *call)
4293 if (queue_IsNotEmpty(&call->rq)) {
4294 rx_packetReclaims += rxi_FreePackets(0, &call->rq);
4295 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4297 if (call->state == RX_STATE_PRECALL) {
4298 call->flags |= RX_CALL_CLEARED;
4302 /* Send an abort packet for the specified call */
4304 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4305 int istack, int force)
4313 /* Clients should never delay abort messages */
4314 if (rx_IsClientConn(call->conn))
4317 if (call->abortCode != call->error) {
4318 call->abortCode = call->error;
4319 call->abortCount = 0;
4322 if (force || rxi_callAbortThreshhold == 0
4323 || call->abortCount < rxi_callAbortThreshhold) {
4324 if (call->delayedAbortEvent) {
4325 rxevent_Cancel(call->delayedAbortEvent, call,
4326 RX_CALL_REFCOUNT_ABORT);
4328 error = htonl(call->error);
4331 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4332 (char *)&error, sizeof(error), istack);
4333 } else if (!call->delayedAbortEvent) {
4334 clock_GetTime(&when);
4335 clock_Addmsec(&when, rxi_callAbortDelay);
4336 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4337 call->delayedAbortEvent =
4338 rxevent_Post(&when, rxi_SendDelayedCallAbort, call, 0);
4343 /* Send an abort packet for the specified connection. Packet is an
4344 * optional pointer to a packet that can be used to send the abort.
4345 * Once the number of abort messages reaches the threshhold, an
4346 * event is scheduled to send the abort. Setting the force flag
4347 * overrides sending delayed abort messages.
4349 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4350 * to send the abort packet.
4353 rxi_SendConnectionAbort(register struct rx_connection *conn,
4354 struct rx_packet *packet, int istack, int force)
4362 /* Clients should never delay abort messages */
4363 if (rx_IsClientConn(conn))
4366 if (force || rxi_connAbortThreshhold == 0
4367 || conn->abortCount < rxi_connAbortThreshhold) {
4368 if (conn->delayedAbortEvent) {
4369 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4371 error = htonl(conn->error);
4373 MUTEX_EXIT(&conn->conn_data_lock);
4375 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4376 RX_PACKET_TYPE_ABORT, (char *)&error,
4377 sizeof(error), istack);
4378 MUTEX_ENTER(&conn->conn_data_lock);
4379 } else if (!conn->delayedAbortEvent) {
4380 clock_GetTime(&when);
4381 clock_Addmsec(&when, rxi_connAbortDelay);
4382 conn->delayedAbortEvent =
4383 rxevent_Post(&when, rxi_SendDelayedConnAbort, conn, 0);
4388 /* Associate an error all of the calls owned by a connection. Called
4389 * with error non-zero. This is only for really fatal things, like
4390 * bad authentication responses. The connection itself is set in
4391 * error at this point, so that future packets received will be
4394 rxi_ConnectionError(register struct rx_connection *conn,
4395 register afs_int32 error)
4400 dpf(("rxi_ConnectionError conn %x error %d", conn, error));
4402 MUTEX_ENTER(&conn->conn_data_lock);
4403 if (conn->challengeEvent)
4404 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4405 if (conn->checkReachEvent) {
4406 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4407 conn->checkReachEvent = 0;
4408 conn->flags &= ~RX_CONN_ATTACHWAIT;
4411 MUTEX_EXIT(&conn->conn_data_lock);
4412 for (i = 0; i < RX_MAXCALLS; i++) {
4413 struct rx_call *call = conn->call[i];
4415 MUTEX_ENTER(&call->lock);
4416 rxi_CallError(call, error);
4417 MUTEX_EXIT(&call->lock);
4420 conn->error = error;
4421 MUTEX_ENTER(&rx_stats_mutex);
4422 rx_stats.fatalErrors++;
4423 MUTEX_EXIT(&rx_stats_mutex);
4428 rxi_CallError(register struct rx_call *call, afs_int32 error)
4430 dpf(("rxi_CallError call %x error %d call->error %d", call, error, call->error));
4432 error = call->error;
4434 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4435 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4436 rxi_ResetCall(call, 0);
4439 rxi_ResetCall(call, 0);
4441 call->error = error;
4442 call->mode = RX_MODE_ERROR;
4445 /* Reset various fields in a call structure, and wakeup waiting
4446 * processes. Some fields aren't changed: state & mode are not
4447 * touched (these must be set by the caller), and bufptr, nLeft, and
4448 * nFree are not reset, since these fields are manipulated by
4449 * unprotected macros, and may only be reset by non-interrupting code.
4452 /* this code requires that call->conn be set properly as a pre-condition. */
4453 #endif /* ADAPT_WINDOW */
4456 rxi_ResetCall(register struct rx_call *call, register int newcall)
4459 register struct rx_peer *peer;
4460 struct rx_packet *packet;
4462 dpf(("rxi_ResetCall(call %x, newcall %d)\n", call, newcall));
4464 /* Notify anyone who is waiting for asynchronous packet arrival */
4465 if (call->arrivalProc) {
4466 (*call->arrivalProc) (call, call->arrivalProcHandle,
4467 call->arrivalProcArg);
4468 call->arrivalProc = (void (*)())0;
4471 if (call->delayedAbortEvent) {
4472 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4473 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4475 rxi_SendCallAbort(call, packet, 0, 1);
4476 rxi_FreePacket(packet);
4481 * Update the peer with the congestion information in this call
4482 * so other calls on this connection can pick up where this call
4483 * left off. If the congestion sequence numbers don't match then
4484 * another call experienced a retransmission.
4486 peer = call->conn->peer;
4487 MUTEX_ENTER(&peer->peer_lock);
4489 if (call->congestSeq == peer->congestSeq) {
4490 peer->cwind = MAX(peer->cwind, call->cwind);
4491 peer->MTU = MAX(peer->MTU, call->MTU);
4492 peer->nDgramPackets =
4493 MAX(peer->nDgramPackets, call->nDgramPackets);
4496 call->abortCode = 0;
4497 call->abortCount = 0;
4499 if (peer->maxDgramPackets > 1) {
4500 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4502 call->MTU = peer->MTU;
4504 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4505 call->ssthresh = rx_maxSendWindow;
4506 call->nDgramPackets = peer->nDgramPackets;
4507 call->congestSeq = peer->congestSeq;
4508 MUTEX_EXIT(&peer->peer_lock);
4510 flags = call->flags;
4511 rxi_ClearReceiveQueue(call);
4512 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4513 if (flags & RX_CALL_TQ_BUSY) {
4514 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4515 call->flags |= (flags & RX_CALL_TQ_WAIT);
4517 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4519 rxi_ClearTransmitQueue(call, 0);
4520 queue_Init(&call->tq);
4521 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4522 dpf(("rcall %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4525 while (call->tqWaiters) {
4526 #ifdef RX_ENABLE_LOCKS
4527 CV_BROADCAST(&call->cv_tq);
4528 #else /* RX_ENABLE_LOCKS */
4529 osi_rxWakeup(&call->tq);
4530 #endif /* RX_ENABLE_LOCKS */
4534 queue_Init(&call->rq);
4536 call->rwind = rx_initReceiveWindow;
4537 call->twind = rx_initSendWindow;
4538 call->nSoftAcked = 0;
4539 call->nextCwind = 0;
4542 call->nCwindAcks = 0;
4543 call->nSoftAcks = 0;
4544 call->nHardAcks = 0;
4546 call->tfirst = call->rnext = call->tnext = 1;
4548 call->lastAcked = 0;
4549 call->localStatus = call->remoteStatus = 0;
4551 if (flags & RX_CALL_READER_WAIT) {
4552 #ifdef RX_ENABLE_LOCKS
4553 CV_BROADCAST(&call->cv_rq);
4555 osi_rxWakeup(&call->rq);
4558 if (flags & RX_CALL_WAIT_PACKETS) {
4559 MUTEX_ENTER(&rx_freePktQ_lock);
4560 rxi_PacketsUnWait(); /* XXX */
4561 MUTEX_EXIT(&rx_freePktQ_lock);
4563 #ifdef RX_ENABLE_LOCKS
4564 CV_SIGNAL(&call->cv_twind);
4566 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4567 osi_rxWakeup(&call->twind);
4570 #ifdef RX_ENABLE_LOCKS
4571 /* The following ensures that we don't mess with any queue while some
4572 * other thread might also be doing so. The call_queue_lock field is
4573 * is only modified under the call lock. If the call is in the process
4574 * of being removed from a queue, the call is not locked until the
4575 * the queue lock is dropped and only then is the call_queue_lock field
4576 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4577 * Note that any other routine which removes a call from a queue has to
4578 * obtain the queue lock before examing the queue and removing the call.
4580 if (call->call_queue_lock) {
4581 MUTEX_ENTER(call->call_queue_lock);
4582 if (queue_IsOnQueue(call)) {
4584 if (flags & RX_CALL_WAIT_PROC) {
4585 MUTEX_ENTER(&rx_stats_mutex);
4587 MUTEX_EXIT(&rx_stats_mutex);
4590 MUTEX_EXIT(call->call_queue_lock);
4591 CLEAR_CALL_QUEUE_LOCK(call);
4593 #else /* RX_ENABLE_LOCKS */
4594 if (queue_IsOnQueue(call)) {
4596 if (flags & RX_CALL_WAIT_PROC)
4599 #endif /* RX_ENABLE_LOCKS */
4601 rxi_KeepAliveOff(call);
4602 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4605 /* Send an acknowledge for the indicated packet (seq,serial) of the
4606 * indicated call, for the indicated reason (reason). This
4607 * acknowledge will specifically acknowledge receiving the packet, and
4608 * will also specify which other packets for this call have been
4609 * received. This routine returns the packet that was used to the
4610 * caller. The caller is responsible for freeing it or re-using it.
4611 * This acknowledgement also returns the highest sequence number
4612 * actually read out by the higher level to the sender; the sender
4613 * promises to keep around packets that have not been read by the
4614 * higher level yet (unless, of course, the sender decides to abort
4615 * the call altogether). Any of p, seq, serial, pflags, or reason may
4616 * be set to zero without ill effect. That is, if they are zero, they
4617 * will not convey any information.
4618 * NOW there is a trailer field, after the ack where it will safely be
4619 * ignored by mundanes, which indicates the maximum size packet this
4620 * host can swallow. */
4622 register struct rx_packet *optionalPacket; use to send ack (or null)
4623 int seq; Sequence number of the packet we are acking
4624 int serial; Serial number of the packet
4625 int pflags; Flags field from packet header
4626 int reason; Reason an acknowledge was prompted
4630 rxi_SendAck(register struct rx_call *call,
4631 register struct rx_packet *optionalPacket, int serial, int reason,
4634 struct rx_ackPacket *ap;
4635 register struct rx_packet *rqp;
4636 register struct rx_packet *nxp; /* For queue_Scan */
4637 register struct rx_packet *p;
4640 #ifdef RX_ENABLE_TSFPQ
4641 struct rx_ts_info_t * rx_ts_info;
4645 * Open the receive window once a thread starts reading packets
4647 if (call->rnext > 1) {
4648 call->rwind = rx_maxReceiveWindow;
4651 call->nHardAcks = 0;
4652 call->nSoftAcks = 0;
4653 if (call->rnext > call->lastAcked)
4654 call->lastAcked = call->rnext;
4658 rx_computelen(p, p->length); /* reset length, you never know */
4659 } /* where that's been... */
4660 #ifdef RX_ENABLE_TSFPQ
4662 RX_TS_INFO_GET(rx_ts_info);
4663 if ((p = rx_ts_info->local_special_packet)) {
4664 rx_computelen(p, p->length);
4665 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4666 rx_ts_info->local_special_packet = p;
4667 } else { /* We won't send the ack, but don't panic. */
4668 return optionalPacket;
4672 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4673 /* We won't send the ack, but don't panic. */
4674 return optionalPacket;
4679 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4682 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4683 #ifndef RX_ENABLE_TSFPQ
4684 if (!optionalPacket)
4687 return optionalPacket;
4689 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4690 if (rx_Contiguous(p) < templ) {
4691 #ifndef RX_ENABLE_TSFPQ
4692 if (!optionalPacket)
4695 return optionalPacket;
4700 /* MTUXXX failing to send an ack is very serious. We should */
4701 /* try as hard as possible to send even a partial ack; it's */
4702 /* better than nothing. */
4703 ap = (struct rx_ackPacket *)rx_DataOf(p);
4704 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4705 ap->reason = reason;
4707 /* The skew computation used to be bogus, I think it's better now. */
4708 /* We should start paying attention to skew. XXX */
4709 ap->serial = htonl(serial);
4710 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4712 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4713 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4715 /* No fear of running out of ack packet here because there can only be at most
4716 * one window full of unacknowledged packets. The window size must be constrained
4717 * to be less than the maximum ack size, of course. Also, an ack should always
4718 * fit into a single packet -- it should not ever be fragmented. */
4719 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4720 if (!rqp || !call->rq.next
4721 || (rqp->header.seq > (call->rnext + call->rwind))) {
4722 #ifndef RX_ENABLE_TSFPQ
4723 if (!optionalPacket)
4726 rxi_CallError(call, RX_CALL_DEAD);
4727 return optionalPacket;
4730 while (rqp->header.seq > call->rnext + offset)
4731 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4732 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4734 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4735 #ifndef RX_ENABLE_TSFPQ
4736 if (!optionalPacket)
4739 rxi_CallError(call, RX_CALL_DEAD);
4740 return optionalPacket;
4745 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4747 /* these are new for AFS 3.3 */
4748 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4749 templ = htonl(templ);
4750 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4751 templ = htonl(call->conn->peer->ifMTU);
4752 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4753 sizeof(afs_int32), &templ);
4755 /* new for AFS 3.4 */
4756 templ = htonl(call->rwind);
4757 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4758 sizeof(afs_int32), &templ);
4760 /* new for AFS 3.5 */
4761 templ = htonl(call->conn->peer->ifDgramPackets);
4762 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4763 sizeof(afs_int32), &templ);
4765 p->header.serviceId = call->conn->serviceId;
4766 p->header.cid = (call->conn->cid | call->channel);
4767 p->header.callNumber = *call->callNumber;
4769 p->header.securityIndex = call->conn->securityIndex;
4770 p->header.epoch = call->conn->epoch;
4771 p->header.type = RX_PACKET_TYPE_ACK;
4772 p->header.flags = RX_SLOW_START_OK;
4773 if (reason == RX_ACK_PING) {
4774 p->header.flags |= RX_REQUEST_ACK;
4776 clock_GetTime(&call->pingRequestTime);
4779 if (call->conn->type == RX_CLIENT_CONNECTION)
4780 p->header.flags |= RX_CLIENT_INITIATED;
4784 if (rxdebug_active) {
4788 len = _snprintf(msg, sizeof(msg),
4789 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4790 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4791 ntohl(ap->serial), ntohl(ap->previousPacket),
4792 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
4793 ap->nAcks, ntohs(ap->bufferSpace) );
4797 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
4798 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4802 OutputDebugString(msg);
4804 #else /* AFS_NT40_ENV */
4806 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
4807 ap->reason, ntohl(ap->previousPacket),
4808 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4810 for (offset = 0; offset < ap->nAcks; offset++)
4811 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4816 #endif /* AFS_NT40_ENV */
4819 register int i, nbytes = p->length;
4821 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4822 if (nbytes <= p->wirevec[i].iov_len) {
4823 register int savelen, saven;
4825 savelen = p->wirevec[i].iov_len;
4827 p->wirevec[i].iov_len = nbytes;
4829 rxi_Send(call, p, istack);
4830 p->wirevec[i].iov_len = savelen;
4834 nbytes -= p->wirevec[i].iov_len;
4837 MUTEX_ENTER(&rx_stats_mutex);
4838 rx_stats.ackPacketsSent++;
4839 MUTEX_EXIT(&rx_stats_mutex);
4840 #ifndef RX_ENABLE_TSFPQ
4841 if (!optionalPacket)
4844 return optionalPacket; /* Return packet for re-use by caller */
4847 /* Send all of the packets in the list in single datagram */
4849 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4850 int istack, int moreFlag, struct clock *now,
4851 struct clock *retryTime, int resending)
4856 struct rx_connection *conn = call->conn;
4857 struct rx_peer *peer = conn->peer;
4859 MUTEX_ENTER(&peer->peer_lock);
4862 peer->reSends += len;
4863 MUTEX_ENTER(&rx_stats_mutex);
4864 rx_stats.dataPacketsSent += len;
4865 MUTEX_EXIT(&rx_stats_mutex);
4866 MUTEX_EXIT(&peer->peer_lock);
4868 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4872 /* Set the packet flags and schedule the resend events */
4873 /* Only request an ack for the last packet in the list */
4874 for (i = 0; i < len; i++) {
4875 list[i]->retryTime = *retryTime;
4876 if (list[i]->header.serial) {
4877 /* Exponentially backoff retry times */
4878 if (list[i]->backoff < MAXBACKOFF) {
4879 /* so it can't stay == 0 */
4880 list[i]->backoff = (list[i]->backoff << 1) + 1;
4883 clock_Addmsec(&(list[i]->retryTime),
4884 ((afs_uint32) list[i]->backoff) << 8);
4887 /* Wait a little extra for the ack on the last packet */
4888 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4889 clock_Addmsec(&(list[i]->retryTime), 400);
4892 /* Record the time sent */
4893 list[i]->timeSent = *now;
4895 /* Ask for an ack on retransmitted packets, on every other packet
4896 * if the peer doesn't support slow start. Ask for an ack on every
4897 * packet until the congestion window reaches the ack rate. */
4898 if (list[i]->header.serial) {
4900 MUTEX_ENTER(&rx_stats_mutex);
4901 rx_stats.dataPacketsReSent++;
4902 MUTEX_EXIT(&rx_stats_mutex);
4904 /* improved RTO calculation- not Karn */
4905 list[i]->firstSent = *now;
4906 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4907 || (!(call->flags & RX_CALL_SLOW_START_OK)
4908 && (list[i]->header.seq & 1)))) {
4913 MUTEX_ENTER(&peer->peer_lock);
4917 MUTEX_ENTER(&rx_stats_mutex);
4918 rx_stats.dataPacketsSent++;
4919 MUTEX_EXIT(&rx_stats_mutex);
4920 MUTEX_EXIT(&peer->peer_lock);
4922 /* Tag this packet as not being the last in this group,
4923 * for the receiver's benefit */
4924 if (i < len - 1 || moreFlag) {
4925 list[i]->header.flags |= RX_MORE_PACKETS;
4928 /* Install the new retransmit time for the packet, and
4929 * record the time sent */
4930 list[i]->timeSent = *now;
4934 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4937 /* Since we're about to send a data packet to the peer, it's
4938 * safe to nuke any scheduled end-of-packets ack */
4939 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4941 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4942 MUTEX_EXIT(&call->lock);
4944 rxi_SendPacketList(call, conn, list, len, istack);
4946 rxi_SendPacket(call, conn, list[0], istack);
4948 MUTEX_ENTER(&call->lock);
4949 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4951 /* Update last send time for this call (for keep-alive
4952 * processing), and for the connection (so that we can discover
4953 * idle connections) */
4954 conn->lastSendTime = call->lastSendTime = clock_Sec();
4957 /* When sending packets we need to follow these rules:
4958 * 1. Never send more than maxDgramPackets in a jumbogram.
4959 * 2. Never send a packet with more than two iovecs in a jumbogram.
4960 * 3. Never send a retransmitted packet in a jumbogram.
4961 * 4. Never send more than cwind/4 packets in a jumbogram
4962 * We always keep the last list we should have sent so we
4963 * can set the RX_MORE_PACKETS flags correctly.
4966 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4967 int istack, struct clock *now, struct clock *retryTime,
4970 int i, cnt, lastCnt = 0;
4971 struct rx_packet **listP, **lastP = 0;
4972 struct rx_peer *peer = call->conn->peer;
4973 int morePackets = 0;
4975 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4976 /* Does the current packet force us to flush the current list? */
4978 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4979 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4981 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
4983 /* If the call enters an error state stop sending, or if
4984 * we entered congestion recovery mode, stop sending */
4985 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4993 /* Add the current packet to the list if it hasn't been acked.
4994 * Otherwise adjust the list pointer to skip the current packet. */
4995 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
4997 /* Do we need to flush the list? */
4998 if (cnt >= (int)peer->maxDgramPackets
4999 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5000 || list[i]->header.serial
5001 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5003 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5004 retryTime, resending);
5005 /* If the call enters an error state stop sending, or if
5006 * we entered congestion recovery mode, stop sending */
5008 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5013 listP = &list[i + 1];
5018 osi_Panic("rxi_SendList error");
5020 listP = &list[i + 1];
5024 /* Send the whole list when the call is in receive mode, when
5025 * the call is in eof mode, when we are in fast recovery mode,
5026 * and when we have the last packet */
5027 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5028 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5029 || (call->flags & RX_CALL_FAST_RECOVER)) {
5030 /* Check for the case where the current list contains
5031 * an acked packet. Since we always send retransmissions
5032 * in a separate packet, we only need to check the first
5033 * packet in the list */
5034 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5038 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5039 retryTime, resending);
5040 /* If the call enters an error state stop sending, or if
5041 * we entered congestion recovery mode, stop sending */
5042 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5046 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5049 } else if (lastCnt > 0) {
5050 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5055 #ifdef RX_ENABLE_LOCKS
5056 /* Call rxi_Start, below, but with the call lock held. */
5058 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
5059 void *arg1, int istack)
5061 MUTEX_ENTER(&call->lock);
5062 rxi_Start(event, call, arg1, istack);
5063 MUTEX_EXIT(&call->lock);
5065 #endif /* RX_ENABLE_LOCKS */
5067 /* This routine is called when new packets are readied for
5068 * transmission and when retransmission may be necessary, or when the
5069 * transmission window or burst count are favourable. This should be
5070 * better optimized for new packets, the usual case, now that we've
5071 * got rid of queues of send packets. XXXXXXXXXXX */
5073 rxi_Start(struct rxevent *event, register struct rx_call *call,
5074 void *arg1, int istack)
5076 struct rx_packet *p;
5077 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
5078 struct rx_peer *peer = call->conn->peer;
5079 struct clock now, retryTime;
5083 struct rx_packet **xmitList;
5086 /* If rxi_Start is being called as a result of a resend event,
5087 * then make sure that the event pointer is removed from the call
5088 * structure, since there is no longer a per-call retransmission
5090 if (event && event == call->resendEvent) {
5091 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5092 call->resendEvent = NULL;
5094 if (queue_IsEmpty(&call->tq)) {
5098 /* Timeouts trigger congestion recovery */
5099 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5100 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5101 /* someone else is waiting to start recovery */
5104 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5105 rxi_WaitforTQBusy(call);
5106 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5107 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5108 call->flags |= RX_CALL_FAST_RECOVER;
5109 if (peer->maxDgramPackets > 1) {
5110 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5112 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5114 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5115 call->nDgramPackets = 1;
5117 call->nextCwind = 1;
5120 MUTEX_ENTER(&peer->peer_lock);
5121 peer->MTU = call->MTU;
5122 peer->cwind = call->cwind;
5123 peer->nDgramPackets = 1;
5125 call->congestSeq = peer->congestSeq;
5126 MUTEX_EXIT(&peer->peer_lock);
5127 /* Clear retry times on packets. Otherwise, it's possible for
5128 * some packets in the queue to force resends at rates faster
5129 * than recovery rates.
5131 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5132 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5133 clock_Zero(&p->retryTime);
5138 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5139 MUTEX_ENTER(&rx_stats_mutex);
5140 rx_tq_debug.rxi_start_in_error++;
5141 MUTEX_EXIT(&rx_stats_mutex);
5146 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5147 /* Get clock to compute the re-transmit time for any packets
5148 * in this burst. Note, if we back off, it's reasonable to
5149 * back off all of the packets in the same manner, even if
5150 * some of them have been retransmitted more times than more
5151 * recent additions */
5152 clock_GetTime(&now);
5153 retryTime = now; /* initialize before use */
5154 MUTEX_ENTER(&peer->peer_lock);
5155 clock_Add(&retryTime, &peer->timeout);
5156 MUTEX_EXIT(&peer->peer_lock);
5158 /* Send (or resend) any packets that need it, subject to
5159 * window restrictions and congestion burst control
5160 * restrictions. Ask for an ack on the last packet sent in
5161 * this burst. For now, we're relying upon the window being
5162 * considerably bigger than the largest number of packets that
5163 * are typically sent at once by one initial call to
5164 * rxi_Start. This is probably bogus (perhaps we should ask
5165 * for an ack when we're half way through the current
5166 * window?). Also, for non file transfer applications, this
5167 * may end up asking for an ack for every packet. Bogus. XXXX
5170 * But check whether we're here recursively, and let the other guy
5173 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5174 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5175 call->flags |= RX_CALL_TQ_BUSY;
5177 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5179 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5180 call->flags &= ~RX_CALL_NEED_START;
5181 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5183 maxXmitPackets = MIN(call->twind, call->cwind);
5184 xmitList = (struct rx_packet **)
5185 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5186 if (xmitList == NULL)
5187 osi_Panic("rxi_Start, failed to allocate xmit list");
5188 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5189 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5190 /* We shouldn't be sending packets if a thread is waiting
5191 * to initiate congestion recovery */
5195 && (call->flags & RX_CALL_FAST_RECOVER)) {
5196 /* Only send one packet during fast recovery */
5199 if ((p->flags & RX_PKTFLAG_FREE)
5200 || (!queue_IsEnd(&call->tq, nxp)
5201 && (nxp->flags & RX_PKTFLAG_FREE))
5202 || (p == (struct rx_packet *)&rx_freePacketQueue)
5203 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5204 osi_Panic("rxi_Start: xmit queue clobbered");
5206 if (p->flags & RX_PKTFLAG_ACKED) {
5207 MUTEX_ENTER(&rx_stats_mutex);
5208 rx_stats.ignoreAckedPacket++;
5209 MUTEX_EXIT(&rx_stats_mutex);
5210 continue; /* Ignore this packet if it has been acknowledged */
5213 /* Turn off all flags except these ones, which are the same
5214 * on each transmission */
5215 p->header.flags &= RX_PRESET_FLAGS;
5217 if (p->header.seq >=
5218 call->tfirst + MIN((int)call->twind,
5219 (int)(call->nSoftAcked +
5221 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5222 /* Note: if we're waiting for more window space, we can
5223 * still send retransmits; hence we don't return here, but
5224 * break out to schedule a retransmit event */
5225 dpf(("call %d waiting for window",
5226 *(call->callNumber)));
5230 /* Transmit the packet if it needs to be sent. */
5231 if (!clock_Lt(&now, &p->retryTime)) {
5232 if (nXmitPackets == maxXmitPackets) {
5233 rxi_SendXmitList(call, xmitList, nXmitPackets,
5234 istack, &now, &retryTime,
5236 osi_Free(xmitList, maxXmitPackets *
5237 sizeof(struct rx_packet *));
5240 xmitList[nXmitPackets++] = p;
5244 /* xmitList now hold pointers to all of the packets that are
5245 * ready to send. Now we loop to send the packets */
5246 if (nXmitPackets > 0) {
5247 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5248 &now, &retryTime, resending);
5251 maxXmitPackets * sizeof(struct rx_packet *));
5253 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5255 * TQ references no longer protected by this flag; they must remain
5256 * protected by the global lock.
5258 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5259 call->flags &= ~RX_CALL_TQ_BUSY;
5260 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5261 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5262 #ifdef RX_ENABLE_LOCKS
5263 osirx_AssertMine(&call->lock, "rxi_Start start");
5264 CV_BROADCAST(&call->cv_tq);
5265 #else /* RX_ENABLE_LOCKS */
5266 osi_rxWakeup(&call->tq);
5267 #endif /* RX_ENABLE_LOCKS */
5272 /* We went into the error state while sending packets. Now is
5273 * the time to reset the call. This will also inform the using
5274 * process that the call is in an error state.
5276 MUTEX_ENTER(&rx_stats_mutex);
5277 rx_tq_debug.rxi_start_aborted++;
5278 MUTEX_EXIT(&rx_stats_mutex);
5279 call->flags &= ~RX_CALL_TQ_BUSY;
5280 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5281 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5282 #ifdef RX_ENABLE_LOCKS
5283 osirx_AssertMine(&call->lock, "rxi_Start middle");
5284 CV_BROADCAST(&call->cv_tq);
5285 #else /* RX_ENABLE_LOCKS */
5286 osi_rxWakeup(&call->tq);
5287 #endif /* RX_ENABLE_LOCKS */
5289 rxi_CallError(call, call->error);
5292 #ifdef RX_ENABLE_LOCKS
5293 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5294 register int missing;
5295 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5296 /* Some packets have received acks. If they all have, we can clear
5297 * the transmit queue.
5300 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5301 if (p->header.seq < call->tfirst
5302 && (p->flags & RX_PKTFLAG_ACKED)) {
5309 call->flags |= RX_CALL_TQ_CLEARME;
5311 #endif /* RX_ENABLE_LOCKS */
5312 /* Don't bother doing retransmits if the TQ is cleared. */
5313 if (call->flags & RX_CALL_TQ_CLEARME) {
5314 rxi_ClearTransmitQueue(call, 1);
5316 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5319 /* Always post a resend event, if there is anything in the
5320 * queue, and resend is possible. There should be at least
5321 * one unacknowledged packet in the queue ... otherwise none
5322 * of these packets should be on the queue in the first place.
5324 if (call->resendEvent) {
5325 /* Cancel the existing event and post a new one */
5326 rxevent_Cancel(call->resendEvent, call,
5327 RX_CALL_REFCOUNT_RESEND);
5330 /* The retry time is the retry time on the first unacknowledged
5331 * packet inside the current window */
5333 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5334 /* Don't set timers for packets outside the window */
5335 if (p->header.seq >= call->tfirst + call->twind) {
5339 if (!(p->flags & RX_PKTFLAG_ACKED)
5340 && !clock_IsZero(&p->retryTime)) {
5342 retryTime = p->retryTime;
5347 /* Post a new event to re-run rxi_Start when retries may be needed */
5348 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5349 #ifdef RX_ENABLE_LOCKS
5350 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5352 rxevent_Post2(&retryTime, rxi_StartUnlocked,
5353 (void *)call, 0, istack);
5354 #else /* RX_ENABLE_LOCKS */
5356 rxevent_Post2(&retryTime, rxi_Start, (void *)call,
5358 #endif /* RX_ENABLE_LOCKS */
5361 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5362 } while (call->flags & RX_CALL_NEED_START);
5364 * TQ references no longer protected by this flag; they must remain
5365 * protected by the global lock.
5367 call->flags &= ~RX_CALL_TQ_BUSY;
5368 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5369 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5370 #ifdef RX_ENABLE_LOCKS
5371 osirx_AssertMine(&call->lock, "rxi_Start end");
5372 CV_BROADCAST(&call->cv_tq);
5373 #else /* RX_ENABLE_LOCKS */
5374 osi_rxWakeup(&call->tq);
5375 #endif /* RX_ENABLE_LOCKS */
5378 call->flags |= RX_CALL_NEED_START;
5380 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5382 if (call->resendEvent) {
5383 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5388 /* Also adjusts the keep alive parameters for the call, to reflect
5389 * that we have just sent a packet (so keep alives aren't sent
5392 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5395 register struct rx_connection *conn = call->conn;
5397 /* Stamp each packet with the user supplied status */
5398 p->header.userStatus = call->localStatus;
5400 /* Allow the security object controlling this call's security to
5401 * make any last-minute changes to the packet */
5402 RXS_SendPacket(conn->securityObject, call, p);
5404 /* Since we're about to send SOME sort of packet to the peer, it's
5405 * safe to nuke any scheduled end-of-packets ack */
5406 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5408 /* Actually send the packet, filling in more connection-specific fields */
5409 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5410 MUTEX_EXIT(&call->lock);
5411 rxi_SendPacket(call, conn, p, istack);
5412 MUTEX_ENTER(&call->lock);
5413 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5415 /* Update last send time for this call (for keep-alive
5416 * processing), and for the connection (so that we can discover
5417 * idle connections) */
5418 conn->lastSendTime = call->lastSendTime = clock_Sec();
5422 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5423 * that things are fine. Also called periodically to guarantee that nothing
5424 * falls through the cracks (e.g. (error + dally) connections have keepalive
5425 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5427 * haveCTLock Set if calling from rxi_ReapConnections
5429 #ifdef RX_ENABLE_LOCKS
5431 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5432 #else /* RX_ENABLE_LOCKS */
5434 rxi_CheckCall(register struct rx_call *call)
5435 #endif /* RX_ENABLE_LOCKS */
5437 register struct rx_connection *conn = call->conn;
5439 afs_uint32 deadTime;
5441 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5442 if (call->flags & RX_CALL_TQ_BUSY) {
5443 /* Call is active and will be reset by rxi_Start if it's
5444 * in an error state.
5449 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5451 (((afs_uint32) conn->secondsUntilDead << 10) +
5452 ((afs_uint32) conn->peer->rtt >> 3) +
5453 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5455 /* These are computed to the second (+- 1 second). But that's
5456 * good enough for these values, which should be a significant
5457 * number of seconds. */
5458 if (now > (call->lastReceiveTime + deadTime)) {
5459 if (call->state == RX_STATE_ACTIVE) {
5460 rxi_CallError(call, RX_CALL_DEAD);
5463 #ifdef RX_ENABLE_LOCKS
5464 /* Cancel pending events */
5465 rxevent_Cancel(call->delayedAckEvent, call,
5466 RX_CALL_REFCOUNT_DELAY);
5467 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5468 rxevent_Cancel(call->keepAliveEvent, call,
5469 RX_CALL_REFCOUNT_ALIVE);
5470 if (call->refCount == 0) {
5471 rxi_FreeCall(call, haveCTLock);
5475 #else /* RX_ENABLE_LOCKS */
5478 #endif /* RX_ENABLE_LOCKS */
5480 /* Non-active calls are destroyed if they are not responding
5481 * to pings; active calls are simply flagged in error, so the
5482 * attached process can die reasonably gracefully. */
5484 /* see if we have a non-activity timeout */
5485 if (call->startWait && conn->idleDeadTime
5486 && ((call->startWait + conn->idleDeadTime) < now)) {
5487 if (call->state == RX_STATE_ACTIVE) {
5488 rxi_CallError(call, RX_CALL_TIMEOUT);
5492 /* see if we have a hard timeout */
5493 if (conn->hardDeadTime
5494 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5495 if (call->state == RX_STATE_ACTIVE)
5496 rxi_CallError(call, RX_CALL_TIMEOUT);
5503 /* When a call is in progress, this routine is called occasionally to
5504 * make sure that some traffic has arrived (or been sent to) the peer.
5505 * If nothing has arrived in a reasonable amount of time, the call is
5506 * declared dead; if nothing has been sent for a while, we send a
5507 * keep-alive packet (if we're actually trying to keep the call alive)
5510 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5513 struct rx_connection *conn;
5516 MUTEX_ENTER(&call->lock);
5517 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5518 if (event == call->keepAliveEvent)
5519 call->keepAliveEvent = NULL;
5522 #ifdef RX_ENABLE_LOCKS
5523 if (rxi_CheckCall(call, 0)) {
5524 MUTEX_EXIT(&call->lock);
5527 #else /* RX_ENABLE_LOCKS */
5528 if (rxi_CheckCall(call))
5530 #endif /* RX_ENABLE_LOCKS */
5532 /* Don't try to keep alive dallying calls */
5533 if (call->state == RX_STATE_DALLY) {
5534 MUTEX_EXIT(&call->lock);
5539 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5540 /* Don't try to send keepalives if there is unacknowledged data */
5541 /* the rexmit code should be good enough, this little hack
5542 * doesn't quite work XXX */
5543 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5545 rxi_ScheduleKeepAliveEvent(call);
5546 MUTEX_EXIT(&call->lock);
5551 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5553 if (!call->keepAliveEvent) {
5555 clock_GetTime(&when);
5556 when.sec += call->conn->secondsUntilPing;
5557 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5558 call->keepAliveEvent =
5559 rxevent_Post(&when, rxi_KeepAliveEvent, call, 0);
5563 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5565 rxi_KeepAliveOn(register struct rx_call *call)
5567 /* Pretend last packet received was received now--i.e. if another
5568 * packet isn't received within the keep alive time, then the call
5569 * will die; Initialize last send time to the current time--even
5570 * if a packet hasn't been sent yet. This will guarantee that a
5571 * keep-alive is sent within the ping time */
5572 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5573 rxi_ScheduleKeepAliveEvent(call);
5576 /* This routine is called to send connection abort messages
5577 * that have been delayed to throttle looping clients. */
5579 rxi_SendDelayedConnAbort(struct rxevent *event,
5580 register struct rx_connection *conn, char *dummy)
5583 struct rx_packet *packet;
5585 MUTEX_ENTER(&conn->conn_data_lock);
5586 conn->delayedAbortEvent = NULL;
5587 error = htonl(conn->error);
5589 MUTEX_EXIT(&conn->conn_data_lock);
5590 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5593 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5594 RX_PACKET_TYPE_ABORT, (char *)&error,
5596 rxi_FreePacket(packet);
5600 /* This routine is called to send call abort messages
5601 * that have been delayed to throttle looping clients. */
5603 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5607 struct rx_packet *packet;
5609 MUTEX_ENTER(&call->lock);
5610 call->delayedAbortEvent = NULL;
5611 error = htonl(call->error);
5613 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5616 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5617 (char *)&error, sizeof(error), 0);
5618 rxi_FreePacket(packet);
5620 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5621 MUTEX_EXIT(&call->lock);
5624 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5625 * seconds) to ask the client to authenticate itself. The routine
5626 * issues a challenge to the client, which is obtained from the
5627 * security object associated with the connection */
5629 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5630 void *arg1, int tries)
5632 conn->challengeEvent = NULL;
5633 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5634 register struct rx_packet *packet;
5638 /* We've failed to authenticate for too long.
5639 * Reset any calls waiting for authentication;
5640 * they are all in RX_STATE_PRECALL.
5644 MUTEX_ENTER(&conn->conn_call_lock);
5645 for (i = 0; i < RX_MAXCALLS; i++) {
5646 struct rx_call *call = conn->call[i];
5648 MUTEX_ENTER(&call->lock);
5649 if (call->state == RX_STATE_PRECALL) {
5650 rxi_CallError(call, RX_CALL_DEAD);
5651 rxi_SendCallAbort(call, NULL, 0, 0);
5653 MUTEX_EXIT(&call->lock);
5656 MUTEX_EXIT(&conn->conn_call_lock);
5660 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5662 /* If there's no packet available, do this later. */
5663 RXS_GetChallenge(conn->securityObject, conn, packet);
5664 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5665 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5666 rxi_FreePacket(packet);
5668 clock_GetTime(&when);
5669 when.sec += RX_CHALLENGE_TIMEOUT;
5670 conn->challengeEvent =
5671 rxevent_Post2(&when, rxi_ChallengeEvent, conn, 0,
5676 /* Call this routine to start requesting the client to authenticate
5677 * itself. This will continue until authentication is established,
5678 * the call times out, or an invalid response is returned. The
5679 * security object associated with the connection is asked to create
5680 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5681 * defined earlier. */
5683 rxi_ChallengeOn(register struct rx_connection *conn)
5685 if (!conn->challengeEvent) {
5686 RXS_CreateChallenge(conn->securityObject, conn);
5687 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5692 /* Compute round trip time of the packet provided, in *rttp.
5695 /* rxi_ComputeRoundTripTime is called with peer locked. */
5696 /* sentp and/or peer may be null */
5698 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5699 register struct clock *sentp,
5700 register struct rx_peer *peer)
5702 struct clock thisRtt, *rttp = &thisRtt;
5704 register int rtt_timeout;
5706 clock_GetTime(rttp);
5708 if (clock_Lt(rttp, sentp)) {
5710 return; /* somebody set the clock back, don't count this time. */
5712 clock_Sub(rttp, sentp);
5713 MUTEX_ENTER(&rx_stats_mutex);
5714 if (clock_Lt(rttp, &rx_stats.minRtt))
5715 rx_stats.minRtt = *rttp;
5716 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5717 if (rttp->sec > 60) {
5718 MUTEX_EXIT(&rx_stats_mutex);
5719 return; /* somebody set the clock ahead */
5721 rx_stats.maxRtt = *rttp;
5723 clock_Add(&rx_stats.totalRtt, rttp);
5724 rx_stats.nRttSamples++;
5725 MUTEX_EXIT(&rx_stats_mutex);
5727 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5729 /* Apply VanJacobson round-trip estimations */
5734 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5735 * srtt is stored as fixed point with 3 bits after the binary
5736 * point (i.e., scaled by 8). The following magic is
5737 * equivalent to the smoothing algorithm in rfc793 with an
5738 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5739 * srtt*8 = srtt*8 + rtt - srtt
5740 * srtt = srtt + rtt/8 - srtt/8
5743 delta = MSEC(rttp) - (peer->rtt >> 3);
5747 * We accumulate a smoothed rtt variance (actually, a smoothed
5748 * mean difference), then set the retransmit timer to smoothed
5749 * rtt + 4 times the smoothed variance (was 2x in van's original
5750 * paper, but 4x works better for me, and apparently for him as
5752 * rttvar is stored as
5753 * fixed point with 2 bits after the binary point (scaled by
5754 * 4). The following is equivalent to rfc793 smoothing with
5755 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5756 * replaces rfc793's wired-in beta.
5757 * dev*4 = dev*4 + (|actual - expected| - dev)
5763 delta -= (peer->rtt_dev >> 2);
5764 peer->rtt_dev += delta;
5766 /* I don't have a stored RTT so I start with this value. Since I'm
5767 * probably just starting a call, and will be pushing more data down
5768 * this, I expect congestion to increase rapidly. So I fudge a
5769 * little, and I set deviance to half the rtt. In practice,
5770 * deviance tends to approach something a little less than
5771 * half the smoothed rtt. */
5772 peer->rtt = (MSEC(rttp) << 3) + 8;
5773 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5775 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5776 * the other of these connections is usually in a user process, and can
5777 * be switched and/or swapped out. So on fast, reliable networks, the
5778 * timeout would otherwise be too short.
5780 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5781 clock_Zero(&(peer->timeout));
5782 clock_Addmsec(&(peer->timeout), rtt_timeout);
5784 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)));
5788 /* Find all server connections that have not been active for a long time, and
5791 rxi_ReapConnections(void)
5794 clock_GetTime(&now);
5796 /* Find server connection structures that haven't been used for
5797 * greater than rx_idleConnectionTime */
5799 struct rx_connection **conn_ptr, **conn_end;
5800 int i, havecalls = 0;
5801 MUTEX_ENTER(&rx_connHashTable_lock);
5802 for (conn_ptr = &rx_connHashTable[0], conn_end =
5803 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5805 struct rx_connection *conn, *next;
5806 struct rx_call *call;
5810 for (conn = *conn_ptr; conn; conn = next) {
5811 /* XXX -- Shouldn't the connection be locked? */
5814 for (i = 0; i < RX_MAXCALLS; i++) {
5815 call = conn->call[i];
5818 MUTEX_ENTER(&call->lock);
5819 #ifdef RX_ENABLE_LOCKS
5820 result = rxi_CheckCall(call, 1);
5821 #else /* RX_ENABLE_LOCKS */
5822 result = rxi_CheckCall(call);
5823 #endif /* RX_ENABLE_LOCKS */
5824 MUTEX_EXIT(&call->lock);
5826 /* If CheckCall freed the call, it might
5827 * have destroyed the connection as well,
5828 * which screws up the linked lists.
5834 if (conn->type == RX_SERVER_CONNECTION) {
5835 /* This only actually destroys the connection if
5836 * there are no outstanding calls */
5837 MUTEX_ENTER(&conn->conn_data_lock);
5838 if (!havecalls && !conn->refCount
5839 && ((conn->lastSendTime + rx_idleConnectionTime) <
5841 conn->refCount++; /* it will be decr in rx_DestroyConn */
5842 MUTEX_EXIT(&conn->conn_data_lock);
5843 #ifdef RX_ENABLE_LOCKS
5844 rxi_DestroyConnectionNoLock(conn);
5845 #else /* RX_ENABLE_LOCKS */
5846 rxi_DestroyConnection(conn);
5847 #endif /* RX_ENABLE_LOCKS */
5849 #ifdef RX_ENABLE_LOCKS
5851 MUTEX_EXIT(&conn->conn_data_lock);
5853 #endif /* RX_ENABLE_LOCKS */
5857 #ifdef RX_ENABLE_LOCKS
5858 while (rx_connCleanup_list) {
5859 struct rx_connection *conn;
5860 conn = rx_connCleanup_list;
5861 rx_connCleanup_list = rx_connCleanup_list->next;
5862 MUTEX_EXIT(&rx_connHashTable_lock);
5863 rxi_CleanupConnection(conn);
5864 MUTEX_ENTER(&rx_connHashTable_lock);
5866 MUTEX_EXIT(&rx_connHashTable_lock);
5867 #endif /* RX_ENABLE_LOCKS */
5870 /* Find any peer structures that haven't been used (haven't had an
5871 * associated connection) for greater than rx_idlePeerTime */
5873 struct rx_peer **peer_ptr, **peer_end;
5875 MUTEX_ENTER(&rx_rpc_stats);
5876 MUTEX_ENTER(&rx_peerHashTable_lock);
5877 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5878 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5880 struct rx_peer *peer, *next, *prev;
5881 for (prev = peer = *peer_ptr; peer; peer = next) {
5883 code = MUTEX_TRYENTER(&peer->peer_lock);
5884 if ((code) && (peer->refCount == 0)
5885 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5886 rx_interface_stat_p rpc_stat, nrpc_stat;
5888 MUTEX_EXIT(&peer->peer_lock);
5889 MUTEX_DESTROY(&peer->peer_lock);
5891 (&peer->rpcStats, rpc_stat, nrpc_stat,
5892 rx_interface_stat)) {
5893 unsigned int num_funcs;
5896 queue_Remove(&rpc_stat->queue_header);
5897 queue_Remove(&rpc_stat->all_peers);
5898 num_funcs = rpc_stat->stats[0].func_total;
5900 sizeof(rx_interface_stat_t) +
5901 rpc_stat->stats[0].func_total *
5902 sizeof(rx_function_entry_v1_t);
5904 rxi_Free(rpc_stat, space);
5905 rxi_rpc_peer_stat_cnt -= num_funcs;
5908 MUTEX_ENTER(&rx_stats_mutex);
5909 rx_stats.nPeerStructs--;
5910 MUTEX_EXIT(&rx_stats_mutex);
5911 if (peer == *peer_ptr) {
5918 MUTEX_EXIT(&peer->peer_lock);
5924 MUTEX_EXIT(&rx_peerHashTable_lock);
5925 MUTEX_EXIT(&rx_rpc_stats);
5928 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5929 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5930 * GC, just below. Really, we shouldn't have to keep moving packets from
5931 * one place to another, but instead ought to always know if we can
5932 * afford to hold onto a packet in its particular use. */
5933 MUTEX_ENTER(&rx_freePktQ_lock);
5934 if (rx_waitingForPackets) {
5935 rx_waitingForPackets = 0;
5936 #ifdef RX_ENABLE_LOCKS
5937 CV_BROADCAST(&rx_waitingForPackets_cv);
5939 osi_rxWakeup(&rx_waitingForPackets);
5942 MUTEX_EXIT(&rx_freePktQ_lock);
5944 now.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5945 rxevent_Post(&now, rxi_ReapConnections, 0, 0);
5949 /* rxs_Release - This isn't strictly necessary but, since the macro name from
5950 * rx.h is sort of strange this is better. This is called with a security
5951 * object before it is discarded. Each connection using a security object has
5952 * its own refcount to the object so it won't actually be freed until the last
5953 * connection is destroyed.
5955 * This is the only rxs module call. A hold could also be written but no one
5959 rxs_Release(struct rx_securityClass *aobj)
5961 return RXS_Close(aobj);
5965 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
5966 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
5967 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
5968 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
5970 /* Adjust our estimate of the transmission rate to this peer, given
5971 * that the packet p was just acked. We can adjust peer->timeout and
5972 * call->twind. Pragmatically, this is called
5973 * only with packets of maximal length.
5974 * Called with peer and call locked.
5978 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
5979 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
5981 afs_int32 xferSize, xferMs;
5982 register afs_int32 minTime;
5985 /* Count down packets */
5986 if (peer->rateFlag > 0)
5988 /* Do nothing until we're enabled */
5989 if (peer->rateFlag != 0)
5994 /* Count only when the ack seems legitimate */
5995 switch (ackReason) {
5996 case RX_ACK_REQUESTED:
5998 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6002 case RX_ACK_PING_RESPONSE:
6003 if (p) /* want the response to ping-request, not data send */
6005 clock_GetTime(&newTO);
6006 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6007 clock_Sub(&newTO, &call->pingRequestTime);
6008 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6012 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6019 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));
6021 /* Track only packets that are big enough. */
6022 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6026 /* absorb RTT data (in milliseconds) for these big packets */
6027 if (peer->smRtt == 0) {
6028 peer->smRtt = xferMs;
6030 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6035 if (peer->countDown) {
6039 peer->countDown = 10; /* recalculate only every so often */
6041 /* In practice, we can measure only the RTT for full packets,
6042 * because of the way Rx acks the data that it receives. (If it's
6043 * smaller than a full packet, it often gets implicitly acked
6044 * either by the call response (from a server) or by the next call
6045 * (from a client), and either case confuses transmission times
6046 * with processing times.) Therefore, replace the above
6047 * more-sophisticated processing with a simpler version, where the
6048 * smoothed RTT is kept for full-size packets, and the time to
6049 * transmit a windowful of full-size packets is simply RTT *
6050 * windowSize. Again, we take two steps:
6051 - ensure the timeout is large enough for a single packet's RTT;
6052 - ensure that the window is small enough to fit in the desired timeout.*/
6054 /* First, the timeout check. */
6055 minTime = peer->smRtt;
6056 /* Get a reasonable estimate for a timeout period */
6058 newTO.sec = minTime / 1000;
6059 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6061 /* Increase the timeout period so that we can always do at least
6062 * one packet exchange */
6063 if (clock_Gt(&newTO, &peer->timeout)) {
6065 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));
6067 peer->timeout = newTO;
6070 /* Now, get an estimate for the transmit window size. */
6071 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6072 /* Now, convert to the number of full packets that could fit in a
6073 * reasonable fraction of that interval */
6074 minTime /= (peer->smRtt << 1);
6075 xferSize = minTime; /* (make a copy) */
6077 /* Now clamp the size to reasonable bounds. */
6080 else if (minTime > rx_Window)
6081 minTime = rx_Window;
6082 /* if (minTime != peer->maxWindow) {
6083 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6084 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6085 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6087 peer->maxWindow = minTime;
6088 elide... call->twind = minTime;
6092 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6093 * Discern this by calculating the timeout necessary for rx_Window
6095 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6096 /* calculate estimate for transmission interval in milliseconds */
6097 minTime = rx_Window * peer->smRtt;
6098 if (minTime < 1000) {
6099 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
6100 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6101 peer->timeout.usec, peer->smRtt, peer->packetSize));
6103 newTO.sec = 0; /* cut back on timeout by half a second */
6104 newTO.usec = 500000;
6105 clock_Sub(&peer->timeout, &newTO);
6110 } /* end of rxi_ComputeRate */
6111 #endif /* ADAPT_WINDOW */
6119 #define TRACE_OPTION_DEBUGLOG 4
6127 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6128 0, KEY_QUERY_VALUE, &parmKey);
6129 if (code != ERROR_SUCCESS)
6132 dummyLen = sizeof(TraceOption);
6133 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6134 (BYTE *) &TraceOption, &dummyLen);
6135 if (code == ERROR_SUCCESS) {
6136 rxdebug_active = (TraceOption & TRACE_OPTION_DEBUGLOG) ? 1 : 0;
6138 RegCloseKey (parmKey);
6139 #endif /* AFS_NT40_ENV */
6144 rx_DebugOnOff(int on)
6146 rxdebug_active = on;
6148 #endif /* AFS_NT40_ENV */
6151 /* Don't call this debugging routine directly; use dpf */
6153 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
6154 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
6162 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6165 len = _snprintf(msg, sizeof(msg)-2,
6166 tformat, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10,
6167 a11, a12, a13, a14, a15);
6169 if (msg[len-1] != '\n') {
6173 OutputDebugString(msg);
6178 clock_GetTime(&now);
6179 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
6180 (unsigned int)now.usec / 1000);
6181 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
6188 * This function is used to process the rx_stats structure that is local
6189 * to a process as well as an rx_stats structure received from a remote
6190 * process (via rxdebug). Therefore, it needs to do minimal version
6194 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
6195 afs_int32 freePackets, char version)
6199 if (size != sizeof(struct rx_stats)) {
6201 "Unexpected size of stats structure: was %d, expected %d\n",
6202 size, sizeof(struct rx_stats));
6205 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6208 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6209 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6210 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6211 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6212 s->specialPktAllocFailures);
6214 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6215 s->receivePktAllocFailures, s->sendPktAllocFailures,
6216 s->specialPktAllocFailures);
6220 " greedy %d, " "bogusReads %d (last from host %x), "
6221 "noPackets %d, " "noBuffers %d, " "selects %d, "
6222 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6223 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6224 s->selects, s->sendSelects);
6226 fprintf(file, " packets read: ");
6227 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6228 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6230 fprintf(file, "\n");
6233 " other read counters: data %d, " "ack %d, " "dup %d "
6234 "spurious %d " "dally %d\n", s->dataPacketsRead,
6235 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6236 s->ignorePacketDally);
6238 fprintf(file, " packets sent: ");
6239 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6240 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6242 fprintf(file, "\n");
6245 " other send counters: ack %d, " "data %d (not resends), "
6246 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6247 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6248 s->dataPacketsPushed, s->ignoreAckedPacket);
6251 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6252 s->netSendFailures, (int)s->fatalErrors);
6254 if (s->nRttSamples) {
6255 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6256 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6258 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6259 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6263 " %d server connections, " "%d client connections, "
6264 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6265 s->nServerConns, s->nClientConns, s->nPeerStructs,
6266 s->nCallStructs, s->nFreeCallStructs);
6268 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6269 fprintf(file, " %d clock updates\n", clock_nUpdates);
6274 /* for backward compatibility */
6276 rx_PrintStats(FILE * file)
6278 MUTEX_ENTER(&rx_stats_mutex);
6279 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6281 MUTEX_EXIT(&rx_stats_mutex);
6285 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6287 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6288 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6289 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6292 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6293 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6294 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6297 " Packet size %d, " "max in packet skew %d, "
6298 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6299 (int)peer->outPacketSkew);
6302 #ifdef AFS_PTHREAD_ENV
6304 * This mutex protects the following static variables:
6308 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6309 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6311 #define LOCK_RX_DEBUG
6312 #define UNLOCK_RX_DEBUG
6313 #endif /* AFS_PTHREAD_ENV */
6316 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6317 u_char type, void *inputData, size_t inputLength,
6318 void *outputData, size_t outputLength)
6320 static afs_int32 counter = 100;
6321 time_t waitTime, waitCount, startTime, endTime;
6322 struct rx_header theader;
6324 register afs_int32 code;
6325 struct timeval tv_now, tv_wake, tv_delta;
6326 struct sockaddr_in taddr, faddr;
6331 startTime = time(0);
6337 tp = &tbuffer[sizeof(struct rx_header)];
6338 taddr.sin_family = AF_INET;
6339 taddr.sin_port = remotePort;
6340 taddr.sin_addr.s_addr = remoteAddr;
6341 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6342 taddr.sin_len = sizeof(struct sockaddr_in);
6345 memset(&theader, 0, sizeof(theader));
6346 theader.epoch = htonl(999);
6348 theader.callNumber = htonl(counter);
6351 theader.type = type;
6352 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6353 theader.serviceId = 0;
6355 memcpy(tbuffer, &theader, sizeof(theader));
6356 memcpy(tp, inputData, inputLength);
6358 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6359 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6361 /* see if there's a packet available */
6362 gettimeofday(&tv_wake,0);
6363 tv_wake.tv_sec += waitTime;
6366 FD_SET(socket, &imask);
6367 tv_delta.tv_sec = tv_wake.tv_sec;
6368 tv_delta.tv_usec = tv_wake.tv_usec;
6369 gettimeofday(&tv_now, 0);
6371 if (tv_delta.tv_usec < tv_now.tv_usec) {
6373 tv_delta.tv_usec += 1000000;
6376 tv_delta.tv_usec -= tv_now.tv_usec;
6378 if (tv_delta.tv_sec < tv_now.tv_sec) {
6382 tv_delta.tv_sec -= tv_now.tv_sec;
6384 code = select(socket + 1, &imask, 0, 0, &tv_delta);
6385 if (code == 1 && FD_ISSET(socket, &imask)) {
6386 /* now receive a packet */
6387 faddrLen = sizeof(struct sockaddr_in);
6389 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6390 (struct sockaddr *)&faddr, &faddrLen);
6393 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6394 if (counter == ntohl(theader.callNumber))
6402 /* see if we've timed out */
6410 code -= sizeof(struct rx_header);
6411 if (code > outputLength)
6412 code = outputLength;
6413 memcpy(outputData, tp, code);
6418 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6419 afs_uint16 remotePort, struct rx_debugStats * stat,
6420 afs_uint32 * supportedValues)
6422 struct rx_debugIn in;
6425 *supportedValues = 0;
6426 in.type = htonl(RX_DEBUGI_GETSTATS);
6429 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6430 &in, sizeof(in), stat, sizeof(*stat));
6433 * If the call was successful, fixup the version and indicate
6434 * what contents of the stat structure are valid.
6435 * Also do net to host conversion of fields here.
6439 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6440 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6442 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6443 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6445 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6446 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6448 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6449 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6451 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6452 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6454 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6455 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6457 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6458 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6460 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6461 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6464 stat->nFreePackets = ntohl(stat->nFreePackets);
6465 stat->packetReclaims = ntohl(stat->packetReclaims);
6466 stat->callsExecuted = ntohl(stat->callsExecuted);
6467 stat->nWaiting = ntohl(stat->nWaiting);
6468 stat->idleThreads = ntohl(stat->idleThreads);
6475 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6476 afs_uint16 remotePort, struct rx_stats * stat,
6477 afs_uint32 * supportedValues)
6479 struct rx_debugIn in;
6480 afs_int32 *lp = (afs_int32 *) stat;
6485 * supportedValues is currently unused, but added to allow future
6486 * versioning of this function.
6489 *supportedValues = 0;
6490 in.type = htonl(RX_DEBUGI_RXSTATS);
6492 memset(stat, 0, sizeof(*stat));
6494 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6495 &in, sizeof(in), stat, sizeof(*stat));
6500 * Do net to host conversion here
6503 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6512 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6513 afs_uint16 remotePort, size_t version_length,
6517 return MakeDebugCall(socket, remoteAddr, remotePort,
6518 RX_PACKET_TYPE_VERSION, a, 1, version,
6523 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6524 afs_uint16 remotePort, afs_int32 * nextConnection,
6525 int allConnections, afs_uint32 debugSupportedValues,
6526 struct rx_debugConn * conn,
6527 afs_uint32 * supportedValues)
6529 struct rx_debugIn in;
6534 * supportedValues is currently unused, but added to allow future
6535 * versioning of this function.
6538 *supportedValues = 0;
6539 if (allConnections) {
6540 in.type = htonl(RX_DEBUGI_GETALLCONN);
6542 in.type = htonl(RX_DEBUGI_GETCONN);
6544 in.index = htonl(*nextConnection);
6545 memset(conn, 0, sizeof(*conn));
6547 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6548 &in, sizeof(in), conn, sizeof(*conn));
6551 *nextConnection += 1;
6554 * Convert old connection format to new structure.
6557 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6558 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6559 #define MOVEvL(a) (conn->a = vL->a)
6561 /* any old or unrecognized version... */
6562 for (i = 0; i < RX_MAXCALLS; i++) {
6563 MOVEvL(callState[i]);
6564 MOVEvL(callMode[i]);
6565 MOVEvL(callFlags[i]);
6566 MOVEvL(callOther[i]);
6568 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6569 MOVEvL(secStats.type);
6570 MOVEvL(secStats.level);
6571 MOVEvL(secStats.flags);
6572 MOVEvL(secStats.expires);
6573 MOVEvL(secStats.packetsReceived);
6574 MOVEvL(secStats.packetsSent);
6575 MOVEvL(secStats.bytesReceived);
6576 MOVEvL(secStats.bytesSent);
6581 * Do net to host conversion here
6583 * I don't convert host or port since we are most likely
6584 * going to want these in NBO.
6586 conn->cid = ntohl(conn->cid);
6587 conn->serial = ntohl(conn->serial);
6588 for (i = 0; i < RX_MAXCALLS; i++) {
6589 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6591 conn->error = ntohl(conn->error);
6592 conn->secStats.flags = ntohl(conn->secStats.flags);
6593 conn->secStats.expires = ntohl(conn->secStats.expires);
6594 conn->secStats.packetsReceived =
6595 ntohl(conn->secStats.packetsReceived);
6596 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6597 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6598 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6599 conn->epoch = ntohl(conn->epoch);
6600 conn->natMTU = ntohl(conn->natMTU);
6607 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6608 afs_uint16 remotePort, afs_int32 * nextPeer,
6609 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6610 afs_uint32 * supportedValues)
6612 struct rx_debugIn in;
6616 * supportedValues is currently unused, but added to allow future
6617 * versioning of this function.
6620 *supportedValues = 0;
6621 in.type = htonl(RX_DEBUGI_GETPEER);
6622 in.index = htonl(*nextPeer);
6623 memset(peer, 0, sizeof(*peer));
6625 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6626 &in, sizeof(in), peer, sizeof(*peer));
6632 * Do net to host conversion here
6634 * I don't convert host or port since we are most likely
6635 * going to want these in NBO.
6637 peer->ifMTU = ntohs(peer->ifMTU);
6638 peer->idleWhen = ntohl(peer->idleWhen);
6639 peer->refCount = ntohs(peer->refCount);
6640 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6641 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6642 peer->rtt = ntohl(peer->rtt);
6643 peer->rtt_dev = ntohl(peer->rtt_dev);
6644 peer->timeout.sec = ntohl(peer->timeout.sec);
6645 peer->timeout.usec = ntohl(peer->timeout.usec);
6646 peer->nSent = ntohl(peer->nSent);
6647 peer->reSends = ntohl(peer->reSends);
6648 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6649 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6650 peer->rateFlag = ntohl(peer->rateFlag);
6651 peer->natMTU = ntohs(peer->natMTU);
6652 peer->maxMTU = ntohs(peer->maxMTU);
6653 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6654 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6655 peer->MTU = ntohs(peer->MTU);
6656 peer->cwind = ntohs(peer->cwind);
6657 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6658 peer->congestSeq = ntohs(peer->congestSeq);
6659 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6660 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6661 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6662 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6667 #endif /* RXDEBUG */
6672 struct rx_serverQueueEntry *np;
6675 register struct rx_call *call;
6676 register struct rx_serverQueueEntry *sq;
6680 if (rxinit_status == 1) {
6682 return; /* Already shutdown. */
6686 #ifndef AFS_PTHREAD_ENV
6687 FD_ZERO(&rx_selectMask);
6688 #endif /* AFS_PTHREAD_ENV */
6689 rxi_dataQuota = RX_MAX_QUOTA;
6690 #ifndef AFS_PTHREAD_ENV
6692 #endif /* AFS_PTHREAD_ENV */
6695 #ifndef AFS_PTHREAD_ENV
6696 #ifndef AFS_USE_GETTIMEOFDAY
6698 #endif /* AFS_USE_GETTIMEOFDAY */
6699 #endif /* AFS_PTHREAD_ENV */
6701 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6702 call = queue_First(&rx_freeCallQueue, rx_call);
6704 rxi_Free(call, sizeof(struct rx_call));
6707 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6708 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6714 struct rx_peer **peer_ptr, **peer_end;
6715 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6716 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6718 struct rx_peer *peer, *next;
6719 for (peer = *peer_ptr; peer; peer = next) {
6720 rx_interface_stat_p rpc_stat, nrpc_stat;
6723 (&peer->rpcStats, rpc_stat, nrpc_stat,
6724 rx_interface_stat)) {
6725 unsigned int num_funcs;
6728 queue_Remove(&rpc_stat->queue_header);
6729 queue_Remove(&rpc_stat->all_peers);
6730 num_funcs = rpc_stat->stats[0].func_total;
6732 sizeof(rx_interface_stat_t) +
6733 rpc_stat->stats[0].func_total *
6734 sizeof(rx_function_entry_v1_t);
6736 rxi_Free(rpc_stat, space);
6737 MUTEX_ENTER(&rx_rpc_stats);
6738 rxi_rpc_peer_stat_cnt -= num_funcs;
6739 MUTEX_EXIT(&rx_rpc_stats);
6743 MUTEX_ENTER(&rx_stats_mutex);
6744 rx_stats.nPeerStructs--;
6745 MUTEX_EXIT(&rx_stats_mutex);
6749 for (i = 0; i < RX_MAX_SERVICES; i++) {
6751 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6753 for (i = 0; i < rx_hashTableSize; i++) {
6754 register struct rx_connection *tc, *ntc;
6755 MUTEX_ENTER(&rx_connHashTable_lock);
6756 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6758 for (j = 0; j < RX_MAXCALLS; j++) {
6760 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6763 rxi_Free(tc, sizeof(*tc));
6765 MUTEX_EXIT(&rx_connHashTable_lock);
6768 MUTEX_ENTER(&freeSQEList_lock);
6770 while ((np = rx_FreeSQEList)) {
6771 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6772 MUTEX_DESTROY(&np->lock);
6773 rxi_Free(np, sizeof(*np));
6776 MUTEX_EXIT(&freeSQEList_lock);
6777 MUTEX_DESTROY(&freeSQEList_lock);
6778 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6779 MUTEX_DESTROY(&rx_connHashTable_lock);
6780 MUTEX_DESTROY(&rx_peerHashTable_lock);
6781 MUTEX_DESTROY(&rx_serverPool_lock);
6783 osi_Free(rx_connHashTable,
6784 rx_hashTableSize * sizeof(struct rx_connection *));
6785 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6787 UNPIN(rx_connHashTable,
6788 rx_hashTableSize * sizeof(struct rx_connection *));
6789 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6791 rxi_FreeAllPackets();
6793 MUTEX_ENTER(&rx_stats_mutex);
6794 rxi_dataQuota = RX_MAX_QUOTA;
6795 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6796 MUTEX_EXIT(&rx_stats_mutex);
6802 #ifdef RX_ENABLE_LOCKS
6804 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6806 if (!MUTEX_ISMINE(lockaddr))
6807 osi_Panic("Lock not held: %s", msg);
6809 #endif /* RX_ENABLE_LOCKS */
6814 * Routines to implement connection specific data.
6818 rx_KeyCreate(rx_destructor_t rtn)
6821 MUTEX_ENTER(&rxi_keyCreate_lock);
6822 key = rxi_keyCreate_counter++;
6823 rxi_keyCreate_destructor = (rx_destructor_t *)
6824 realloc((void *)rxi_keyCreate_destructor,
6825 (key + 1) * sizeof(rx_destructor_t));
6826 rxi_keyCreate_destructor[key] = rtn;
6827 MUTEX_EXIT(&rxi_keyCreate_lock);
6832 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6835 MUTEX_ENTER(&conn->conn_data_lock);
6836 if (!conn->specific) {
6837 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6838 for (i = 0; i < key; i++)
6839 conn->specific[i] = NULL;
6840 conn->nSpecific = key + 1;
6841 conn->specific[key] = ptr;
6842 } else if (key >= conn->nSpecific) {
6843 conn->specific = (void **)
6844 realloc(conn->specific, (key + 1) * sizeof(void *));
6845 for (i = conn->nSpecific; i < key; i++)
6846 conn->specific[i] = NULL;
6847 conn->nSpecific = key + 1;
6848 conn->specific[key] = ptr;
6850 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6851 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6852 conn->specific[key] = ptr;
6854 MUTEX_EXIT(&conn->conn_data_lock);
6858 rx_GetSpecific(struct rx_connection *conn, int key)
6861 MUTEX_ENTER(&conn->conn_data_lock);
6862 if (key >= conn->nSpecific)
6865 ptr = conn->specific[key];
6866 MUTEX_EXIT(&conn->conn_data_lock);
6870 #endif /* !KERNEL */
6873 * processStats is a queue used to store the statistics for the local
6874 * process. Its contents are similar to the contents of the rpcStats
6875 * queue on a rx_peer structure, but the actual data stored within
6876 * this queue contains totals across the lifetime of the process (assuming
6877 * the stats have not been reset) - unlike the per peer structures
6878 * which can come and go based upon the peer lifetime.
6881 static struct rx_queue processStats = { &processStats, &processStats };
6884 * peerStats is a queue used to store the statistics for all peer structs.
6885 * Its contents are the union of all the peer rpcStats queues.
6888 static struct rx_queue peerStats = { &peerStats, &peerStats };
6891 * rxi_monitor_processStats is used to turn process wide stat collection
6895 static int rxi_monitor_processStats = 0;
6898 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6901 static int rxi_monitor_peerStats = 0;
6904 * rxi_AddRpcStat - given all of the information for a particular rpc
6905 * call, create (if needed) and update the stat totals for the rpc.
6909 * IN stats - the queue of stats that will be updated with the new value
6911 * IN rxInterface - a unique number that identifies the rpc interface
6913 * IN currentFunc - the index of the function being invoked
6915 * IN totalFunc - the total number of functions in this interface
6917 * IN queueTime - the amount of time this function waited for a thread
6919 * IN execTime - the amount of time this function invocation took to execute
6921 * IN bytesSent - the number bytes sent by this invocation
6923 * IN bytesRcvd - the number bytes received by this invocation
6925 * IN isServer - if true, this invocation was made to a server
6927 * IN remoteHost - the ip address of the remote host
6929 * IN remotePort - the port of the remote host
6931 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6933 * INOUT counter - if a new stats structure is allocated, the counter will
6934 * be updated with the new number of allocated stat structures
6942 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6943 afs_uint32 currentFunc, afs_uint32 totalFunc,
6944 struct clock *queueTime, struct clock *execTime,
6945 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6946 afs_uint32 remoteHost, afs_uint32 remotePort,
6947 int addToPeerList, unsigned int *counter)
6950 rx_interface_stat_p rpc_stat, nrpc_stat;
6953 * See if there's already a structure for this interface
6956 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6957 if ((rpc_stat->stats[0].interfaceId == rxInterface)
6958 && (rpc_stat->stats[0].remote_is_server == isServer))
6963 * Didn't find a match so allocate a new structure and add it to the
6967 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
6968 || (rpc_stat->stats[0].interfaceId != rxInterface)
6969 || (rpc_stat->stats[0].remote_is_server != isServer)) {
6974 sizeof(rx_interface_stat_t) +
6975 totalFunc * sizeof(rx_function_entry_v1_t);
6977 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
6978 if (rpc_stat == NULL) {
6982 *counter += totalFunc;
6983 for (i = 0; i < totalFunc; i++) {
6984 rpc_stat->stats[i].remote_peer = remoteHost;
6985 rpc_stat->stats[i].remote_port = remotePort;
6986 rpc_stat->stats[i].remote_is_server = isServer;
6987 rpc_stat->stats[i].interfaceId = rxInterface;
6988 rpc_stat->stats[i].func_total = totalFunc;
6989 rpc_stat->stats[i].func_index = i;
6990 hzero(rpc_stat->stats[i].invocations);
6991 hzero(rpc_stat->stats[i].bytes_sent);
6992 hzero(rpc_stat->stats[i].bytes_rcvd);
6993 rpc_stat->stats[i].queue_time_sum.sec = 0;
6994 rpc_stat->stats[i].queue_time_sum.usec = 0;
6995 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
6996 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
6997 rpc_stat->stats[i].queue_time_min.sec = 9999999;
6998 rpc_stat->stats[i].queue_time_min.usec = 9999999;
6999 rpc_stat->stats[i].queue_time_max.sec = 0;
7000 rpc_stat->stats[i].queue_time_max.usec = 0;
7001 rpc_stat->stats[i].execution_time_sum.sec = 0;
7002 rpc_stat->stats[i].execution_time_sum.usec = 0;
7003 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7004 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7005 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7006 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7007 rpc_stat->stats[i].execution_time_max.sec = 0;
7008 rpc_stat->stats[i].execution_time_max.usec = 0;
7010 queue_Prepend(stats, rpc_stat);
7011 if (addToPeerList) {
7012 queue_Prepend(&peerStats, &rpc_stat->all_peers);
7017 * Increment the stats for this function
7020 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
7021 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
7022 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
7023 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
7024 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7025 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7026 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7028 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7029 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7031 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7032 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7034 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7035 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7037 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7038 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7046 * rx_IncrementTimeAndCount - increment the times and count for a particular
7051 * IN peer - the peer who invoked the rpc
7053 * IN rxInterface - a unique number that identifies the rpc interface
7055 * IN currentFunc - the index of the function being invoked
7057 * IN totalFunc - the total number of functions in this interface
7059 * IN queueTime - the amount of time this function waited for a thread
7061 * IN execTime - the amount of time this function invocation took to execute
7063 * IN bytesSent - the number bytes sent by this invocation
7065 * IN bytesRcvd - the number bytes received by this invocation
7067 * IN isServer - if true, this invocation was made to a server
7075 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7076 afs_uint32 currentFunc, afs_uint32 totalFunc,
7077 struct clock *queueTime, struct clock *execTime,
7078 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7082 MUTEX_ENTER(&rx_rpc_stats);
7083 MUTEX_ENTER(&peer->peer_lock);
7085 if (rxi_monitor_peerStats) {
7086 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7087 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7088 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7091 if (rxi_monitor_processStats) {
7092 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7093 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7094 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7097 MUTEX_EXIT(&peer->peer_lock);
7098 MUTEX_EXIT(&rx_rpc_stats);
7103 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7107 * IN callerVersion - the rpc stat version of the caller.
7109 * IN count - the number of entries to marshall.
7111 * IN stats - pointer to stats to be marshalled.
7113 * OUT ptr - Where to store the marshalled data.
7120 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7121 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7127 * We only support the first version
7129 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7130 *(ptr++) = stats->remote_peer;
7131 *(ptr++) = stats->remote_port;
7132 *(ptr++) = stats->remote_is_server;
7133 *(ptr++) = stats->interfaceId;
7134 *(ptr++) = stats->func_total;
7135 *(ptr++) = stats->func_index;
7136 *(ptr++) = hgethi(stats->invocations);
7137 *(ptr++) = hgetlo(stats->invocations);
7138 *(ptr++) = hgethi(stats->bytes_sent);
7139 *(ptr++) = hgetlo(stats->bytes_sent);
7140 *(ptr++) = hgethi(stats->bytes_rcvd);
7141 *(ptr++) = hgetlo(stats->bytes_rcvd);
7142 *(ptr++) = stats->queue_time_sum.sec;
7143 *(ptr++) = stats->queue_time_sum.usec;
7144 *(ptr++) = stats->queue_time_sum_sqr.sec;
7145 *(ptr++) = stats->queue_time_sum_sqr.usec;
7146 *(ptr++) = stats->queue_time_min.sec;
7147 *(ptr++) = stats->queue_time_min.usec;
7148 *(ptr++) = stats->queue_time_max.sec;
7149 *(ptr++) = stats->queue_time_max.usec;
7150 *(ptr++) = stats->execution_time_sum.sec;
7151 *(ptr++) = stats->execution_time_sum.usec;
7152 *(ptr++) = stats->execution_time_sum_sqr.sec;
7153 *(ptr++) = stats->execution_time_sum_sqr.usec;
7154 *(ptr++) = stats->execution_time_min.sec;
7155 *(ptr++) = stats->execution_time_min.usec;
7156 *(ptr++) = stats->execution_time_max.sec;
7157 *(ptr++) = stats->execution_time_max.usec;
7163 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7168 * IN callerVersion - the rpc stat version of the caller
7170 * OUT myVersion - the rpc stat version of this function
7172 * OUT clock_sec - local time seconds
7174 * OUT clock_usec - local time microseconds
7176 * OUT allocSize - the number of bytes allocated to contain stats
7178 * OUT statCount - the number stats retrieved from this process.
7180 * OUT stats - the actual stats retrieved from this process.
7184 * Returns void. If successful, stats will != NULL.
7188 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7189 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7190 size_t * allocSize, afs_uint32 * statCount,
7191 afs_uint32 ** stats)
7201 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7204 * Check to see if stats are enabled
7207 MUTEX_ENTER(&rx_rpc_stats);
7208 if (!rxi_monitor_processStats) {
7209 MUTEX_EXIT(&rx_rpc_stats);
7213 clock_GetTime(&now);
7214 *clock_sec = now.sec;
7215 *clock_usec = now.usec;
7218 * Allocate the space based upon the caller version
7220 * If the client is at an older version than we are,
7221 * we return the statistic data in the older data format, but
7222 * we still return our version number so the client knows we
7223 * are maintaining more data than it can retrieve.
7226 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7227 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7228 *statCount = rxi_rpc_process_stat_cnt;
7231 * This can't happen yet, but in the future version changes
7232 * can be handled by adding additional code here
7236 if (space > (size_t) 0) {
7238 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7241 rx_interface_stat_p rpc_stat, nrpc_stat;
7245 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7247 * Copy the data based upon the caller version
7249 rx_MarshallProcessRPCStats(callerVersion,
7250 rpc_stat->stats[0].func_total,
7251 rpc_stat->stats, &ptr);
7257 MUTEX_EXIT(&rx_rpc_stats);
7262 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7266 * IN callerVersion - the rpc stat version of the caller
7268 * OUT myVersion - the rpc stat version of this function
7270 * OUT clock_sec - local time seconds
7272 * OUT clock_usec - local time microseconds
7274 * OUT allocSize - the number of bytes allocated to contain stats
7276 * OUT statCount - the number of stats retrieved from the individual
7279 * OUT stats - the actual stats retrieved from the individual peer structures.
7283 * Returns void. If successful, stats will != NULL.
7287 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7288 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7289 size_t * allocSize, afs_uint32 * statCount,
7290 afs_uint32 ** stats)
7300 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7303 * Check to see if stats are enabled
7306 MUTEX_ENTER(&rx_rpc_stats);
7307 if (!rxi_monitor_peerStats) {
7308 MUTEX_EXIT(&rx_rpc_stats);
7312 clock_GetTime(&now);
7313 *clock_sec = now.sec;
7314 *clock_usec = now.usec;
7317 * Allocate the space based upon the caller version
7319 * If the client is at an older version than we are,
7320 * we return the statistic data in the older data format, but
7321 * we still return our version number so the client knows we
7322 * are maintaining more data than it can retrieve.
7325 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7326 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7327 *statCount = rxi_rpc_peer_stat_cnt;
7330 * This can't happen yet, but in the future version changes
7331 * can be handled by adding additional code here
7335 if (space > (size_t) 0) {
7337 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7340 rx_interface_stat_p rpc_stat, nrpc_stat;
7344 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7346 * We have to fix the offset of rpc_stat since we are
7347 * keeping this structure on two rx_queues. The rx_queue
7348 * package assumes that the rx_queue member is the first
7349 * member of the structure. That is, rx_queue assumes that
7350 * any one item is only on one queue at a time. We are
7351 * breaking that assumption and so we have to do a little
7352 * math to fix our pointers.
7355 fix_offset = (char *)rpc_stat;
7356 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7357 rpc_stat = (rx_interface_stat_p) fix_offset;
7360 * Copy the data based upon the caller version
7362 rx_MarshallProcessRPCStats(callerVersion,
7363 rpc_stat->stats[0].func_total,
7364 rpc_stat->stats, &ptr);
7370 MUTEX_EXIT(&rx_rpc_stats);
7375 * rx_FreeRPCStats - free memory allocated by
7376 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7380 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7381 * rx_RetrievePeerRPCStats
7383 * IN allocSize - the number of bytes in stats.
7391 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7393 rxi_Free(stats, allocSize);
7397 * rx_queryProcessRPCStats - see if process rpc stat collection is
7398 * currently enabled.
7404 * Returns 0 if stats are not enabled != 0 otherwise
7408 rx_queryProcessRPCStats(void)
7411 MUTEX_ENTER(&rx_rpc_stats);
7412 rc = rxi_monitor_processStats;
7413 MUTEX_EXIT(&rx_rpc_stats);
7418 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7424 * Returns 0 if stats are not enabled != 0 otherwise
7428 rx_queryPeerRPCStats(void)
7431 MUTEX_ENTER(&rx_rpc_stats);
7432 rc = rxi_monitor_peerStats;
7433 MUTEX_EXIT(&rx_rpc_stats);
7438 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7448 rx_enableProcessRPCStats(void)
7450 MUTEX_ENTER(&rx_rpc_stats);
7451 rx_enable_stats = 1;
7452 rxi_monitor_processStats = 1;
7453 MUTEX_EXIT(&rx_rpc_stats);
7457 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7467 rx_enablePeerRPCStats(void)
7469 MUTEX_ENTER(&rx_rpc_stats);
7470 rx_enable_stats = 1;
7471 rxi_monitor_peerStats = 1;
7472 MUTEX_EXIT(&rx_rpc_stats);
7476 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7486 rx_disableProcessRPCStats(void)
7488 rx_interface_stat_p rpc_stat, nrpc_stat;
7491 MUTEX_ENTER(&rx_rpc_stats);
7494 * Turn off process statistics and if peer stats is also off, turn
7498 rxi_monitor_processStats = 0;
7499 if (rxi_monitor_peerStats == 0) {
7500 rx_enable_stats = 0;
7503 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7504 unsigned int num_funcs = 0;
7507 queue_Remove(rpc_stat);
7508 num_funcs = rpc_stat->stats[0].func_total;
7510 sizeof(rx_interface_stat_t) +
7511 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7513 rxi_Free(rpc_stat, space);
7514 rxi_rpc_process_stat_cnt -= num_funcs;
7516 MUTEX_EXIT(&rx_rpc_stats);
7520 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7530 rx_disablePeerRPCStats(void)
7532 struct rx_peer **peer_ptr, **peer_end;
7535 MUTEX_ENTER(&rx_rpc_stats);
7538 * Turn off peer statistics and if process stats is also off, turn
7542 rxi_monitor_peerStats = 0;
7543 if (rxi_monitor_processStats == 0) {
7544 rx_enable_stats = 0;
7547 MUTEX_ENTER(&rx_peerHashTable_lock);
7548 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7549 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7551 struct rx_peer *peer, *next, *prev;
7552 for (prev = peer = *peer_ptr; peer; peer = next) {
7554 code = MUTEX_TRYENTER(&peer->peer_lock);
7556 rx_interface_stat_p rpc_stat, nrpc_stat;
7559 (&peer->rpcStats, rpc_stat, nrpc_stat,
7560 rx_interface_stat)) {
7561 unsigned int num_funcs = 0;
7564 queue_Remove(&rpc_stat->queue_header);
7565 queue_Remove(&rpc_stat->all_peers);
7566 num_funcs = rpc_stat->stats[0].func_total;
7568 sizeof(rx_interface_stat_t) +
7569 rpc_stat->stats[0].func_total *
7570 sizeof(rx_function_entry_v1_t);
7572 rxi_Free(rpc_stat, space);
7573 rxi_rpc_peer_stat_cnt -= num_funcs;
7575 MUTEX_EXIT(&peer->peer_lock);
7576 if (prev == *peer_ptr) {
7586 MUTEX_EXIT(&rx_peerHashTable_lock);
7587 MUTEX_EXIT(&rx_rpc_stats);
7591 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7596 * IN clearFlag - flag indicating which stats to clear
7604 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7606 rx_interface_stat_p rpc_stat, nrpc_stat;
7608 MUTEX_ENTER(&rx_rpc_stats);
7610 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7611 unsigned int num_funcs = 0, i;
7612 num_funcs = rpc_stat->stats[0].func_total;
7613 for (i = 0; i < num_funcs; i++) {
7614 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7615 hzero(rpc_stat->stats[i].invocations);
7617 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7618 hzero(rpc_stat->stats[i].bytes_sent);
7620 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7621 hzero(rpc_stat->stats[i].bytes_rcvd);
7623 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7624 rpc_stat->stats[i].queue_time_sum.sec = 0;
7625 rpc_stat->stats[i].queue_time_sum.usec = 0;
7627 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7628 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7629 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7631 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7632 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7633 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7635 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7636 rpc_stat->stats[i].queue_time_max.sec = 0;
7637 rpc_stat->stats[i].queue_time_max.usec = 0;
7639 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7640 rpc_stat->stats[i].execution_time_sum.sec = 0;
7641 rpc_stat->stats[i].execution_time_sum.usec = 0;
7643 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7644 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7645 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7647 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7648 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7649 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7651 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7652 rpc_stat->stats[i].execution_time_max.sec = 0;
7653 rpc_stat->stats[i].execution_time_max.usec = 0;
7658 MUTEX_EXIT(&rx_rpc_stats);
7662 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7667 * IN clearFlag - flag indicating which stats to clear
7675 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7677 rx_interface_stat_p rpc_stat, nrpc_stat;
7679 MUTEX_ENTER(&rx_rpc_stats);
7681 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7682 unsigned int num_funcs = 0, i;
7685 * We have to fix the offset of rpc_stat since we are
7686 * keeping this structure on two rx_queues. The rx_queue
7687 * package assumes that the rx_queue member is the first
7688 * member of the structure. That is, rx_queue assumes that
7689 * any one item is only on one queue at a time. We are
7690 * breaking that assumption and so we have to do a little
7691 * math to fix our pointers.
7694 fix_offset = (char *)rpc_stat;
7695 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7696 rpc_stat = (rx_interface_stat_p) fix_offset;
7698 num_funcs = rpc_stat->stats[0].func_total;
7699 for (i = 0; i < num_funcs; i++) {
7700 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7701 hzero(rpc_stat->stats[i].invocations);
7703 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7704 hzero(rpc_stat->stats[i].bytes_sent);
7706 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7707 hzero(rpc_stat->stats[i].bytes_rcvd);
7709 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7710 rpc_stat->stats[i].queue_time_sum.sec = 0;
7711 rpc_stat->stats[i].queue_time_sum.usec = 0;
7713 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7714 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7715 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7717 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7718 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7719 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7721 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7722 rpc_stat->stats[i].queue_time_max.sec = 0;
7723 rpc_stat->stats[i].queue_time_max.usec = 0;
7725 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7726 rpc_stat->stats[i].execution_time_sum.sec = 0;
7727 rpc_stat->stats[i].execution_time_sum.usec = 0;
7729 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7730 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7731 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7733 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7734 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7735 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7737 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7738 rpc_stat->stats[i].execution_time_max.sec = 0;
7739 rpc_stat->stats[i].execution_time_max.usec = 0;
7744 MUTEX_EXIT(&rx_rpc_stats);
7748 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7749 * is authorized to enable/disable/clear RX statistics.
7751 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7754 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7756 rxi_rxstat_userok = proc;
7760 rx_RxStatUserOk(struct rx_call *call)
7762 if (!rxi_rxstat_userok)
7764 return rxi_rxstat_userok(call);
7769 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
7770 * function in the MSVC runtime DLL (msvcrt.dll).
7772 * Note: the system serializes calls to this function.
7775 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
7776 DWORD reason, /* reason function is being called */
7777 LPVOID reserved) /* reserved for future use */
7780 case DLL_PROCESS_ATTACH:
7781 /* library is being attached to a process */
7785 case DLL_PROCESS_DETACH: