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 = 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(NULL); /* Never returns */
727 #ifdef RX_ENABLE_TSFPQ
728 /* no use leaving packets around in this thread's local queue if
729 * it isn't getting donated to the server thread pool.
731 rxi_FlushLocalPacketsTSFPQ();
732 #endif /* RX_ENABLE_TSFPQ */
736 /* Create a new client connection to the specified service, using the
737 * specified security object to implement the security model for this
739 struct rx_connection *
740 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
741 register struct rx_securityClass *securityObject,
742 int serviceSecurityIndex)
746 register struct rx_connection *conn;
751 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", ntohl(shost), ntohs(sport), sservice, securityObject, serviceSecurityIndex));
753 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
754 * the case of kmem_alloc? */
755 conn = rxi_AllocConnection();
756 #ifdef RX_ENABLE_LOCKS
757 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
758 MUTEX_INIT(&conn->conn_data_lock, "conn call lock", MUTEX_DEFAULT, 0);
759 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
762 MUTEX_ENTER(&rx_connHashTable_lock);
763 cid = (rx_nextCid += RX_MAXCALLS);
764 conn->type = RX_CLIENT_CONNECTION;
766 conn->epoch = rx_epoch;
767 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
768 conn->serviceId = sservice;
769 conn->securityObject = securityObject;
770 conn->securityData = (void *) 0;
771 conn->securityIndex = serviceSecurityIndex;
772 rx_SetConnDeadTime(conn, rx_connDeadTime);
773 conn->ackRate = RX_FAST_ACK_RATE;
775 conn->specific = NULL;
776 conn->challengeEvent = NULL;
777 conn->delayedAbortEvent = NULL;
778 conn->abortCount = 0;
781 RXS_NewConnection(securityObject, conn);
783 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
785 conn->refCount++; /* no lock required since only this thread knows... */
786 conn->next = rx_connHashTable[hashindex];
787 rx_connHashTable[hashindex] = conn;
788 rx_MutexIncrement(rx_stats.nClientConns, rx_stats_mutex);
789 MUTEX_EXIT(&rx_connHashTable_lock);
795 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
797 /* The idea is to set the dead time to a value that allows several
798 * keepalives to be dropped without timing out the connection. */
799 conn->secondsUntilDead = MAX(seconds, 6);
800 conn->secondsUntilPing = conn->secondsUntilDead / 6;
803 int rxi_lowPeerRefCount = 0;
804 int rxi_lowConnRefCount = 0;
807 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
808 * NOTE: must not be called with rx_connHashTable_lock held.
811 rxi_CleanupConnection(struct rx_connection *conn)
813 /* Notify the service exporter, if requested, that this connection
814 * is being destroyed */
815 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
816 (*conn->service->destroyConnProc) (conn);
818 /* Notify the security module that this connection is being destroyed */
819 RXS_DestroyConnection(conn->securityObject, conn);
821 /* If this is the last connection using the rx_peer struct, set its
822 * idle time to now. rxi_ReapConnections will reap it if it's still
823 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
825 MUTEX_ENTER(&rx_peerHashTable_lock);
826 if (conn->peer->refCount < 2) {
827 conn->peer->idleWhen = clock_Sec();
828 if (conn->peer->refCount < 1) {
829 conn->peer->refCount = 1;
830 MUTEX_ENTER(&rx_stats_mutex);
831 rxi_lowPeerRefCount++;
832 MUTEX_EXIT(&rx_stats_mutex);
835 conn->peer->refCount--;
836 MUTEX_EXIT(&rx_peerHashTable_lock);
838 if (conn->type == RX_SERVER_CONNECTION)
839 rx_MutexDecrement(rx_stats.nServerConns, rx_stats_mutex);
841 rx_MutexDecrement(rx_stats.nClientConns, rx_stats_mutex);
843 if (conn->specific) {
845 for (i = 0; i < conn->nSpecific; i++) {
846 if (conn->specific[i] && rxi_keyCreate_destructor[i])
847 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
848 conn->specific[i] = NULL;
850 free(conn->specific);
852 conn->specific = NULL;
856 MUTEX_DESTROY(&conn->conn_call_lock);
857 MUTEX_DESTROY(&conn->conn_data_lock);
858 CV_DESTROY(&conn->conn_call_cv);
860 rxi_FreeConnection(conn);
863 /* Destroy the specified connection */
865 rxi_DestroyConnection(register struct rx_connection *conn)
867 MUTEX_ENTER(&rx_connHashTable_lock);
868 rxi_DestroyConnectionNoLock(conn);
869 /* conn should be at the head of the cleanup list */
870 if (conn == rx_connCleanup_list) {
871 rx_connCleanup_list = rx_connCleanup_list->next;
872 MUTEX_EXIT(&rx_connHashTable_lock);
873 rxi_CleanupConnection(conn);
875 #ifdef RX_ENABLE_LOCKS
877 MUTEX_EXIT(&rx_connHashTable_lock);
879 #endif /* RX_ENABLE_LOCKS */
883 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
885 register struct rx_connection **conn_ptr;
886 register int havecalls = 0;
887 struct rx_packet *packet;
894 MUTEX_ENTER(&conn->conn_data_lock);
895 if (conn->refCount > 0)
898 MUTEX_ENTER(&rx_stats_mutex);
899 rxi_lowConnRefCount++;
900 MUTEX_EXIT(&rx_stats_mutex);
903 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
904 /* Busy; wait till the last guy before proceeding */
905 MUTEX_EXIT(&conn->conn_data_lock);
910 /* If the client previously called rx_NewCall, but it is still
911 * waiting, treat this as a running call, and wait to destroy the
912 * connection later when the call completes. */
913 if ((conn->type == RX_CLIENT_CONNECTION)
914 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
915 conn->flags |= RX_CONN_DESTROY_ME;
916 MUTEX_EXIT(&conn->conn_data_lock);
920 MUTEX_EXIT(&conn->conn_data_lock);
922 /* Check for extant references to this connection */
923 for (i = 0; i < RX_MAXCALLS; i++) {
924 register struct rx_call *call = conn->call[i];
927 if (conn->type == RX_CLIENT_CONNECTION) {
928 MUTEX_ENTER(&call->lock);
929 if (call->delayedAckEvent) {
930 /* Push the final acknowledgment out now--there
931 * won't be a subsequent call to acknowledge the
932 * last reply packets */
933 rxevent_Cancel(call->delayedAckEvent, call,
934 RX_CALL_REFCOUNT_DELAY);
935 if (call->state == RX_STATE_PRECALL
936 || call->state == RX_STATE_ACTIVE) {
937 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
939 rxi_AckAll(NULL, call, 0);
942 MUTEX_EXIT(&call->lock);
946 #ifdef RX_ENABLE_LOCKS
948 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
949 MUTEX_EXIT(&conn->conn_data_lock);
951 /* Someone is accessing a packet right now. */
955 #endif /* RX_ENABLE_LOCKS */
958 /* Don't destroy the connection if there are any call
959 * structures still in use */
960 MUTEX_ENTER(&conn->conn_data_lock);
961 conn->flags |= RX_CONN_DESTROY_ME;
962 MUTEX_EXIT(&conn->conn_data_lock);
967 if (conn->delayedAbortEvent) {
968 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
969 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
971 MUTEX_ENTER(&conn->conn_data_lock);
972 rxi_SendConnectionAbort(conn, packet, 0, 1);
973 MUTEX_EXIT(&conn->conn_data_lock);
974 rxi_FreePacket(packet);
978 /* Remove from connection hash table before proceeding */
980 &rx_connHashTable[CONN_HASH
981 (peer->host, peer->port, conn->cid, conn->epoch,
983 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
984 if (*conn_ptr == conn) {
985 *conn_ptr = conn->next;
989 /* if the conn that we are destroying was the last connection, then we
990 * clear rxLastConn as well */
991 if (rxLastConn == conn)
994 /* Make sure the connection is completely reset before deleting it. */
995 /* get rid of pending events that could zap us later */
996 if (conn->challengeEvent)
997 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
998 if (conn->checkReachEvent)
999 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1001 /* Add the connection to the list of destroyed connections that
1002 * need to be cleaned up. This is necessary to avoid deadlocks
1003 * in the routines we call to inform others that this connection is
1004 * being destroyed. */
1005 conn->next = rx_connCleanup_list;
1006 rx_connCleanup_list = conn;
1009 /* Externally available version */
1011 rx_DestroyConnection(register struct rx_connection *conn)
1016 rxi_DestroyConnection(conn);
1021 rx_GetConnection(register struct rx_connection *conn)
1026 MUTEX_ENTER(&conn->conn_data_lock);
1028 MUTEX_EXIT(&conn->conn_data_lock);
1032 /* Wait for the transmit queue to no longer be busy.
1033 * requires the call->lock to be held */
1034 static void rxi_WaitforTQBusy(struct rx_call *call) {
1035 while (call->flags & RX_CALL_TQ_BUSY) {
1036 call->flags |= RX_CALL_TQ_WAIT;
1038 #ifdef RX_ENABLE_LOCKS
1039 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1040 CV_WAIT(&call->cv_tq, &call->lock);
1041 #else /* RX_ENABLE_LOCKS */
1042 osi_rxSleep(&call->tq);
1043 #endif /* RX_ENABLE_LOCKS */
1045 if (call->tqWaiters == 0) {
1046 call->flags &= ~RX_CALL_TQ_WAIT;
1050 /* Start a new rx remote procedure call, on the specified connection.
1051 * If wait is set to 1, wait for a free call channel; otherwise return
1052 * 0. Maxtime gives the maximum number of seconds this call may take,
1053 * after rx_NewCall returns. After this time interval, a call to any
1054 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1055 * For fine grain locking, we hold the conn_call_lock in order to
1056 * to ensure that we don't get signalle after we found a call in an active
1057 * state and before we go to sleep.
1060 rx_NewCall(register struct rx_connection *conn)
1063 register struct rx_call *call;
1064 struct clock queueTime;
1068 dpf(("rx_NewCall(conn %x)\n", conn));
1071 clock_GetTime(&queueTime);
1072 MUTEX_ENTER(&conn->conn_call_lock);
1075 * Check if there are others waiting for a new call.
1076 * If so, let them go first to avoid starving them.
1077 * This is a fairly simple scheme, and might not be
1078 * a complete solution for large numbers of waiters.
1080 * makeCallWaiters keeps track of the number of
1081 * threads waiting to make calls and the
1082 * RX_CONN_MAKECALL_WAITING flag bit is used to
1083 * indicate that there are indeed calls waiting.
1084 * The flag is set when the waiter is incremented.
1085 * It is only cleared in rx_EndCall when
1086 * makeCallWaiters is 0. This prevents us from
1087 * accidently destroying the connection while it
1088 * is potentially about to be used.
1090 MUTEX_ENTER(&conn->conn_data_lock);
1091 if (conn->makeCallWaiters) {
1092 conn->flags |= RX_CONN_MAKECALL_WAITING;
1093 conn->makeCallWaiters++;
1094 MUTEX_EXIT(&conn->conn_data_lock);
1096 #ifdef RX_ENABLE_LOCKS
1097 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1101 MUTEX_ENTER(&conn->conn_data_lock);
1102 conn->makeCallWaiters--;
1104 MUTEX_EXIT(&conn->conn_data_lock);
1107 for (i = 0; i < RX_MAXCALLS; i++) {
1108 call = conn->call[i];
1110 MUTEX_ENTER(&call->lock);
1111 if (call->state == RX_STATE_DALLY) {
1112 rxi_ResetCall(call, 0);
1113 (*call->callNumber)++;
1116 MUTEX_EXIT(&call->lock);
1118 call = rxi_NewCall(conn, i);
1122 if (i < RX_MAXCALLS) {
1125 MUTEX_ENTER(&conn->conn_data_lock);
1126 conn->flags |= RX_CONN_MAKECALL_WAITING;
1127 conn->makeCallWaiters++;
1128 MUTEX_EXIT(&conn->conn_data_lock);
1130 #ifdef RX_ENABLE_LOCKS
1131 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1135 MUTEX_ENTER(&conn->conn_data_lock);
1136 conn->makeCallWaiters--;
1137 MUTEX_EXIT(&conn->conn_data_lock);
1140 * Wake up anyone else who might be giving us a chance to
1141 * run (see code above that avoids resource starvation).
1143 #ifdef RX_ENABLE_LOCKS
1144 CV_BROADCAST(&conn->conn_call_cv);
1149 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1151 /* Client is initially in send mode */
1152 call->state = RX_STATE_ACTIVE;
1153 call->error = conn->error;
1155 call->mode = RX_MODE_ERROR;
1157 call->mode = RX_MODE_SENDING;
1159 /* remember start time for call in case we have hard dead time limit */
1160 call->queueTime = queueTime;
1161 clock_GetTime(&call->startTime);
1162 hzero(call->bytesSent);
1163 hzero(call->bytesRcvd);
1165 /* Turn on busy protocol. */
1166 rxi_KeepAliveOn(call);
1168 MUTEX_EXIT(&call->lock);
1169 MUTEX_EXIT(&conn->conn_call_lock);
1172 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1173 /* Now, if TQ wasn't cleared earlier, do it now. */
1174 MUTEX_ENTER(&call->lock);
1175 rxi_WaitforTQBusy(call);
1176 if (call->flags & RX_CALL_TQ_CLEARME) {
1177 rxi_ClearTransmitQueue(call, 0);
1178 queue_Init(&call->tq);
1180 MUTEX_EXIT(&call->lock);
1181 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1183 dpf(("rx_NewCall(call %x)\n", call));
1188 rxi_HasActiveCalls(register struct rx_connection *aconn)
1191 register struct rx_call *tcall;
1195 for (i = 0; i < RX_MAXCALLS; i++) {
1196 if ((tcall = aconn->call[i])) {
1197 if ((tcall->state == RX_STATE_ACTIVE)
1198 || (tcall->state == RX_STATE_PRECALL)) {
1209 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1210 register afs_int32 * aint32s)
1213 register struct rx_call *tcall;
1217 for (i = 0; i < RX_MAXCALLS; i++) {
1218 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1219 aint32s[i] = aconn->callNumber[i] + 1;
1221 aint32s[i] = aconn->callNumber[i];
1228 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1229 register afs_int32 * aint32s)
1232 register struct rx_call *tcall;
1236 for (i = 0; i < RX_MAXCALLS; i++) {
1237 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1238 aconn->callNumber[i] = aint32s[i] - 1;
1240 aconn->callNumber[i] = aint32s[i];
1246 /* Advertise a new service. A service is named locally by a UDP port
1247 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1250 char *serviceName; Name for identification purposes (e.g. the
1251 service name might be used for probing for
1254 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1255 char *serviceName, struct rx_securityClass **securityObjects,
1256 int nSecurityObjects,
1257 afs_int32(*serviceProc) (struct rx_call * acall))
1259 osi_socket socket = OSI_NULLSOCKET;
1260 register struct rx_service *tservice;
1266 if (serviceId == 0) {
1268 "rx_NewService: service id for service %s is not non-zero.\n",
1275 "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",
1283 tservice = rxi_AllocService();
1285 for (i = 0; i < RX_MAX_SERVICES; i++) {
1286 register struct rx_service *service = rx_services[i];
1288 if (port == service->servicePort && host == service->serviceHost) {
1289 if (service->serviceId == serviceId) {
1290 /* The identical service has already been
1291 * installed; if the caller was intending to
1292 * change the security classes used by this
1293 * service, he/she loses. */
1295 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1296 serviceName, serviceId, service->serviceName);
1298 rxi_FreeService(tservice);
1301 /* Different service, same port: re-use the socket
1302 * which is bound to the same port */
1303 socket = service->socket;
1306 if (socket == OSI_NULLSOCKET) {
1307 /* If we don't already have a socket (from another
1308 * service on same port) get a new one */
1309 socket = rxi_GetHostUDPSocket(host, port);
1310 if (socket == OSI_NULLSOCKET) {
1312 rxi_FreeService(tservice);
1317 service->socket = socket;
1318 service->serviceHost = host;
1319 service->servicePort = port;
1320 service->serviceId = serviceId;
1321 service->serviceName = serviceName;
1322 service->nSecurityObjects = nSecurityObjects;
1323 service->securityObjects = securityObjects;
1324 service->minProcs = 0;
1325 service->maxProcs = 1;
1326 service->idleDeadTime = 60;
1327 service->connDeadTime = rx_connDeadTime;
1328 service->executeRequestProc = serviceProc;
1329 service->checkReach = 0;
1330 rx_services[i] = service; /* not visible until now */
1336 rxi_FreeService(tservice);
1337 (osi_Msg "rx_NewService: cannot support > %d services\n",
1342 /* Set configuration options for all of a service's security objects */
1345 rx_SetSecurityConfiguration(struct rx_service *service,
1346 rx_securityConfigVariables type,
1350 for (i = 0; i<service->nSecurityObjects; i++) {
1351 if (service->securityObjects[i]) {
1352 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1360 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1361 struct rx_securityClass **securityObjects, int nSecurityObjects,
1362 afs_int32(*serviceProc) (struct rx_call * acall))
1364 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1367 /* Generic request processing loop. This routine should be called
1368 * by the implementation dependent rx_ServerProc. If socketp is
1369 * non-null, it will be set to the file descriptor that this thread
1370 * is now listening on. If socketp is null, this routine will never
1373 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1375 register struct rx_call *call;
1376 register afs_int32 code;
1377 register struct rx_service *tservice = NULL;
1384 call = rx_GetCall(threadID, tservice, socketp);
1385 if (socketp && *socketp != OSI_NULLSOCKET) {
1386 /* We are now a listener thread */
1391 /* if server is restarting( typically smooth shutdown) then do not
1392 * allow any new calls.
1395 if (rx_tranquil && (call != NULL)) {
1399 MUTEX_ENTER(&call->lock);
1401 rxi_CallError(call, RX_RESTARTING);
1402 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1404 MUTEX_EXIT(&call->lock);
1408 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1409 #ifdef RX_ENABLE_LOCKS
1411 #endif /* RX_ENABLE_LOCKS */
1412 afs_termState = AFSOP_STOP_AFS;
1413 afs_osi_Wakeup(&afs_termState);
1414 #ifdef RX_ENABLE_LOCKS
1416 #endif /* RX_ENABLE_LOCKS */
1421 tservice = call->conn->service;
1423 if (tservice->beforeProc)
1424 (*tservice->beforeProc) (call);
1426 code = call->conn->service->executeRequestProc(call);
1428 if (tservice->afterProc)
1429 (*tservice->afterProc) (call, code);
1431 rx_EndCall(call, code);
1432 MUTEX_ENTER(&rx_stats_mutex);
1434 MUTEX_EXIT(&rx_stats_mutex);
1440 rx_WakeupServerProcs(void)
1442 struct rx_serverQueueEntry *np, *tqp;
1446 MUTEX_ENTER(&rx_serverPool_lock);
1448 #ifdef RX_ENABLE_LOCKS
1449 if (rx_waitForPacket)
1450 CV_BROADCAST(&rx_waitForPacket->cv);
1451 #else /* RX_ENABLE_LOCKS */
1452 if (rx_waitForPacket)
1453 osi_rxWakeup(rx_waitForPacket);
1454 #endif /* RX_ENABLE_LOCKS */
1455 MUTEX_ENTER(&freeSQEList_lock);
1456 for (np = rx_FreeSQEList; np; np = tqp) {
1457 tqp = *(struct rx_serverQueueEntry **)np;
1458 #ifdef RX_ENABLE_LOCKS
1459 CV_BROADCAST(&np->cv);
1460 #else /* RX_ENABLE_LOCKS */
1462 #endif /* RX_ENABLE_LOCKS */
1464 MUTEX_EXIT(&freeSQEList_lock);
1465 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1466 #ifdef RX_ENABLE_LOCKS
1467 CV_BROADCAST(&np->cv);
1468 #else /* RX_ENABLE_LOCKS */
1470 #endif /* RX_ENABLE_LOCKS */
1472 MUTEX_EXIT(&rx_serverPool_lock);
1477 * One thing that seems to happen is that all the server threads get
1478 * tied up on some empty or slow call, and then a whole bunch of calls
1479 * arrive at once, using up the packet pool, so now there are more
1480 * empty calls. The most critical resources here are server threads
1481 * and the free packet pool. The "doreclaim" code seems to help in
1482 * general. I think that eventually we arrive in this state: there
1483 * are lots of pending calls which do have all their packets present,
1484 * so they won't be reclaimed, are multi-packet calls, so they won't
1485 * be scheduled until later, and thus are tying up most of the free
1486 * packet pool for a very long time.
1488 * 1. schedule multi-packet calls if all the packets are present.
1489 * Probably CPU-bound operation, useful to return packets to pool.
1490 * Do what if there is a full window, but the last packet isn't here?
1491 * 3. preserve one thread which *only* runs "best" calls, otherwise
1492 * it sleeps and waits for that type of call.
1493 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1494 * the current dataquota business is badly broken. The quota isn't adjusted
1495 * to reflect how many packets are presently queued for a running call.
1496 * So, when we schedule a queued call with a full window of packets queued
1497 * up for it, that *should* free up a window full of packets for other 2d-class
1498 * calls to be able to use from the packet pool. But it doesn't.
1500 * NB. Most of the time, this code doesn't run -- since idle server threads
1501 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1502 * as a new call arrives.
1504 /* Sleep until a call arrives. Returns a pointer to the call, ready
1505 * for an rx_Read. */
1506 #ifdef RX_ENABLE_LOCKS
1508 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1510 struct rx_serverQueueEntry *sq;
1511 register struct rx_call *call = (struct rx_call *)0;
1512 struct rx_service *service = NULL;
1515 MUTEX_ENTER(&freeSQEList_lock);
1517 if ((sq = rx_FreeSQEList)) {
1518 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1519 MUTEX_EXIT(&freeSQEList_lock);
1520 } else { /* otherwise allocate a new one and return that */
1521 MUTEX_EXIT(&freeSQEList_lock);
1522 sq = (struct rx_serverQueueEntry *)
1523 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1524 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1525 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1528 MUTEX_ENTER(&rx_serverPool_lock);
1529 if (cur_service != NULL) {
1530 ReturnToServerPool(cur_service);
1533 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1534 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1536 /* Scan for eligible incoming calls. A call is not eligible
1537 * if the maximum number of calls for its service type are
1538 * already executing */
1539 /* One thread will process calls FCFS (to prevent starvation),
1540 * while the other threads may run ahead looking for calls which
1541 * have all their input data available immediately. This helps
1542 * keep threads from blocking, waiting for data from the client. */
1543 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1544 service = tcall->conn->service;
1545 if (!QuotaOK(service)) {
1548 if (tno == rxi_fcfs_thread_num
1549 || !tcall->queue_item_header.next) {
1550 /* If we're the fcfs thread , then we'll just use
1551 * this call. If we haven't been able to find an optimal
1552 * choice, and we're at the end of the list, then use a
1553 * 2d choice if one has been identified. Otherwise... */
1554 call = (choice2 ? choice2 : tcall);
1555 service = call->conn->service;
1556 } else if (!queue_IsEmpty(&tcall->rq)) {
1557 struct rx_packet *rp;
1558 rp = queue_First(&tcall->rq, rx_packet);
1559 if (rp->header.seq == 1) {
1561 || (rp->header.flags & RX_LAST_PACKET)) {
1563 } else if (rxi_2dchoice && !choice2
1564 && !(tcall->flags & RX_CALL_CLEARED)
1565 && (tcall->rprev > rxi_HardAckRate)) {
1574 ReturnToServerPool(service);
1581 MUTEX_EXIT(&rx_serverPool_lock);
1582 MUTEX_ENTER(&call->lock);
1584 if (call->flags & RX_CALL_WAIT_PROC) {
1585 call->flags &= ~RX_CALL_WAIT_PROC;
1586 MUTEX_ENTER(&rx_stats_mutex);
1588 MUTEX_EXIT(&rx_stats_mutex);
1591 if (call->state != RX_STATE_PRECALL || call->error) {
1592 MUTEX_EXIT(&call->lock);
1593 MUTEX_ENTER(&rx_serverPool_lock);
1594 ReturnToServerPool(service);
1599 if (queue_IsEmpty(&call->rq)
1600 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1601 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1603 CLEAR_CALL_QUEUE_LOCK(call);
1606 /* If there are no eligible incoming calls, add this process
1607 * to the idle server queue, to wait for one */
1611 *socketp = OSI_NULLSOCKET;
1613 sq->socketp = socketp;
1614 queue_Append(&rx_idleServerQueue, sq);
1615 #ifndef AFS_AIX41_ENV
1616 rx_waitForPacket = sq;
1618 rx_waitingForPacket = sq;
1619 #endif /* AFS_AIX41_ENV */
1621 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1623 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1624 MUTEX_EXIT(&rx_serverPool_lock);
1625 return (struct rx_call *)0;
1628 } while (!(call = sq->newcall)
1629 && !(socketp && *socketp != OSI_NULLSOCKET));
1630 MUTEX_EXIT(&rx_serverPool_lock);
1632 MUTEX_ENTER(&call->lock);
1638 MUTEX_ENTER(&freeSQEList_lock);
1639 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1640 rx_FreeSQEList = sq;
1641 MUTEX_EXIT(&freeSQEList_lock);
1644 clock_GetTime(&call->startTime);
1645 call->state = RX_STATE_ACTIVE;
1646 call->mode = RX_MODE_RECEIVING;
1647 #ifdef RX_KERNEL_TRACE
1648 if (ICL_SETACTIVE(afs_iclSetp)) {
1649 int glockOwner = ISAFS_GLOCK();
1652 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1653 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1660 rxi_calltrace(RX_CALL_START, call);
1661 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1662 call->conn->service->servicePort, call->conn->service->serviceId,
1665 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1666 MUTEX_EXIT(&call->lock);
1668 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1673 #else /* RX_ENABLE_LOCKS */
1675 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1677 struct rx_serverQueueEntry *sq;
1678 register struct rx_call *call = (struct rx_call *)0, *choice2;
1679 struct rx_service *service = NULL;
1683 MUTEX_ENTER(&freeSQEList_lock);
1685 if ((sq = rx_FreeSQEList)) {
1686 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1687 MUTEX_EXIT(&freeSQEList_lock);
1688 } else { /* otherwise allocate a new one and return that */
1689 MUTEX_EXIT(&freeSQEList_lock);
1690 sq = (struct rx_serverQueueEntry *)
1691 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1692 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1693 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1695 MUTEX_ENTER(&sq->lock);
1697 if (cur_service != NULL) {
1698 cur_service->nRequestsRunning--;
1699 if (cur_service->nRequestsRunning < cur_service->minProcs)
1703 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1704 register struct rx_call *tcall, *ncall;
1705 /* Scan for eligible incoming calls. A call is not eligible
1706 * if the maximum number of calls for its service type are
1707 * already executing */
1708 /* One thread will process calls FCFS (to prevent starvation),
1709 * while the other threads may run ahead looking for calls which
1710 * have all their input data available immediately. This helps
1711 * keep threads from blocking, waiting for data from the client. */
1712 choice2 = (struct rx_call *)0;
1713 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1714 service = tcall->conn->service;
1715 if (QuotaOK(service)) {
1716 if (tno == rxi_fcfs_thread_num
1717 || !tcall->queue_item_header.next) {
1718 /* If we're the fcfs thread, then we'll just use
1719 * this call. If we haven't been able to find an optimal
1720 * choice, and we're at the end of the list, then use a
1721 * 2d choice if one has been identified. Otherwise... */
1722 call = (choice2 ? choice2 : tcall);
1723 service = call->conn->service;
1724 } else if (!queue_IsEmpty(&tcall->rq)) {
1725 struct rx_packet *rp;
1726 rp = queue_First(&tcall->rq, rx_packet);
1727 if (rp->header.seq == 1
1729 || (rp->header.flags & RX_LAST_PACKET))) {
1731 } else if (rxi_2dchoice && !choice2
1732 && !(tcall->flags & RX_CALL_CLEARED)
1733 && (tcall->rprev > rxi_HardAckRate)) {
1746 /* we can't schedule a call if there's no data!!! */
1747 /* send an ack if there's no data, if we're missing the
1748 * first packet, or we're missing something between first
1749 * and last -- there's a "hole" in the incoming data. */
1750 if (queue_IsEmpty(&call->rq)
1751 || queue_First(&call->rq, rx_packet)->header.seq != 1
1752 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1753 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1755 call->flags &= (~RX_CALL_WAIT_PROC);
1756 service->nRequestsRunning++;
1757 /* just started call in minProcs pool, need fewer to maintain
1759 if (service->nRequestsRunning <= service->minProcs)
1763 /* MUTEX_EXIT(&call->lock); */
1765 /* If there are no eligible incoming calls, add this process
1766 * to the idle server queue, to wait for one */
1769 *socketp = OSI_NULLSOCKET;
1771 sq->socketp = socketp;
1772 queue_Append(&rx_idleServerQueue, sq);
1776 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1778 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1779 return (struct rx_call *)0;
1782 } while (!(call = sq->newcall)
1783 && !(socketp && *socketp != OSI_NULLSOCKET));
1785 MUTEX_EXIT(&sq->lock);
1787 MUTEX_ENTER(&freeSQEList_lock);
1788 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1789 rx_FreeSQEList = sq;
1790 MUTEX_EXIT(&freeSQEList_lock);
1793 clock_GetTime(&call->startTime);
1794 call->state = RX_STATE_ACTIVE;
1795 call->mode = RX_MODE_RECEIVING;
1796 #ifdef RX_KERNEL_TRACE
1797 if (ICL_SETACTIVE(afs_iclSetp)) {
1798 int glockOwner = ISAFS_GLOCK();
1801 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1802 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1809 rxi_calltrace(RX_CALL_START, call);
1810 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1811 call->conn->service->servicePort, call->conn->service->serviceId,
1814 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1821 #endif /* RX_ENABLE_LOCKS */
1825 /* Establish a procedure to be called when a packet arrives for a
1826 * call. This routine will be called at most once after each call,
1827 * and will also be called if there is an error condition on the or
1828 * the call is complete. Used by multi rx to build a selection
1829 * function which determines which of several calls is likely to be a
1830 * good one to read from.
1831 * NOTE: the way this is currently implemented it is probably only a
1832 * good idea to (1) use it immediately after a newcall (clients only)
1833 * and (2) only use it once. Other uses currently void your warranty
1836 rx_SetArrivalProc(register struct rx_call *call,
1837 register void (*proc) (register struct rx_call * call,
1839 register int index),
1840 register void * handle, register int arg)
1842 call->arrivalProc = proc;
1843 call->arrivalProcHandle = handle;
1844 call->arrivalProcArg = arg;
1847 /* Call is finished (possibly prematurely). Return rc to the peer, if
1848 * appropriate, and return the final error code from the conversation
1852 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1854 register struct rx_connection *conn = call->conn;
1855 register struct rx_service *service;
1861 dpf(("rx_EndCall(call %x rc %d error %d abortCode %d)\n", call, rc, call->error, call->abortCode));
1864 MUTEX_ENTER(&call->lock);
1866 if (rc == 0 && call->error == 0) {
1867 call->abortCode = 0;
1868 call->abortCount = 0;
1871 call->arrivalProc = (void (*)())0;
1872 if (rc && call->error == 0) {
1873 rxi_CallError(call, rc);
1874 /* Send an abort message to the peer if this error code has
1875 * only just been set. If it was set previously, assume the
1876 * peer has already been sent the error code or will request it
1878 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1880 if (conn->type == RX_SERVER_CONNECTION) {
1881 /* Make sure reply or at least dummy reply is sent */
1882 if (call->mode == RX_MODE_RECEIVING) {
1883 rxi_WriteProc(call, 0, 0);
1885 if (call->mode == RX_MODE_SENDING) {
1886 rxi_FlushWrite(call);
1888 service = conn->service;
1889 rxi_calltrace(RX_CALL_END, call);
1890 /* Call goes to hold state until reply packets are acknowledged */
1891 if (call->tfirst + call->nSoftAcked < call->tnext) {
1892 call->state = RX_STATE_HOLD;
1894 call->state = RX_STATE_DALLY;
1895 rxi_ClearTransmitQueue(call, 0);
1896 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1897 rxevent_Cancel(call->keepAliveEvent, call,
1898 RX_CALL_REFCOUNT_ALIVE);
1900 } else { /* Client connection */
1902 /* Make sure server receives input packets, in the case where
1903 * no reply arguments are expected */
1904 if ((call->mode == RX_MODE_SENDING)
1905 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1906 (void)rxi_ReadProc(call, &dummy, 1);
1909 /* If we had an outstanding delayed ack, be nice to the server
1910 * and force-send it now.
1912 if (call->delayedAckEvent) {
1913 rxevent_Cancel(call->delayedAckEvent, call,
1914 RX_CALL_REFCOUNT_DELAY);
1915 call->delayedAckEvent = NULL;
1916 rxi_SendDelayedAck(NULL, call, NULL);
1919 /* We need to release the call lock since it's lower than the
1920 * conn_call_lock and we don't want to hold the conn_call_lock
1921 * over the rx_ReadProc call. The conn_call_lock needs to be held
1922 * here for the case where rx_NewCall is perusing the calls on
1923 * the connection structure. We don't want to signal until
1924 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1925 * have checked this call, found it active and by the time it
1926 * goes to sleep, will have missed the signal.
1928 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
1929 * there are threads waiting to use the conn object.
1931 MUTEX_EXIT(&call->lock);
1932 MUTEX_ENTER(&conn->conn_call_lock);
1933 MUTEX_ENTER(&call->lock);
1934 MUTEX_ENTER(&conn->conn_data_lock);
1935 conn->flags |= RX_CONN_BUSY;
1936 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1937 if (conn->makeCallWaiters == 0)
1938 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1939 MUTEX_EXIT(&conn->conn_data_lock);
1940 #ifdef RX_ENABLE_LOCKS
1941 CV_BROADCAST(&conn->conn_call_cv);
1946 #ifdef RX_ENABLE_LOCKS
1948 MUTEX_EXIT(&conn->conn_data_lock);
1950 #endif /* RX_ENABLE_LOCKS */
1951 call->state = RX_STATE_DALLY;
1953 error = call->error;
1955 /* currentPacket, nLeft, and NFree must be zeroed here, because
1956 * ResetCall cannot: ResetCall may be called at splnet(), in the
1957 * kernel version, and may interrupt the macros rx_Read or
1958 * rx_Write, which run at normal priority for efficiency. */
1959 if (call->currentPacket) {
1960 queue_Prepend(&call->iovq, call->currentPacket);
1961 call->currentPacket = (struct rx_packet *)0;
1964 call->nLeft = call->nFree = call->curlen = 0;
1966 /* Free any packets from the last call to ReadvProc/WritevProc */
1967 rxi_FreePackets(0, &call->iovq);
1969 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1970 MUTEX_EXIT(&call->lock);
1971 if (conn->type == RX_CLIENT_CONNECTION) {
1972 MUTEX_EXIT(&conn->conn_call_lock);
1973 conn->flags &= ~RX_CONN_BUSY;
1977 * Map errors to the local host's errno.h format.
1979 error = ntoh_syserr_conv(error);
1983 #if !defined(KERNEL)
1985 /* Call this routine when shutting down a server or client (especially
1986 * clients). This will allow Rx to gracefully garbage collect server
1987 * connections, and reduce the number of retries that a server might
1988 * make to a dead client.
1989 * This is not quite right, since some calls may still be ongoing and
1990 * we can't lock them to destroy them. */
1994 register struct rx_connection **conn_ptr, **conn_end;
1998 if (rxinit_status == 1) {
2000 return; /* Already shutdown. */
2002 rxi_DeleteCachedConnections();
2003 if (rx_connHashTable) {
2004 MUTEX_ENTER(&rx_connHashTable_lock);
2005 for (conn_ptr = &rx_connHashTable[0], conn_end =
2006 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2008 struct rx_connection *conn, *next;
2009 for (conn = *conn_ptr; conn; conn = next) {
2011 if (conn->type == RX_CLIENT_CONNECTION) {
2012 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2014 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2015 #ifdef RX_ENABLE_LOCKS
2016 rxi_DestroyConnectionNoLock(conn);
2017 #else /* RX_ENABLE_LOCKS */
2018 rxi_DestroyConnection(conn);
2019 #endif /* RX_ENABLE_LOCKS */
2023 #ifdef RX_ENABLE_LOCKS
2024 while (rx_connCleanup_list) {
2025 struct rx_connection *conn;
2026 conn = rx_connCleanup_list;
2027 rx_connCleanup_list = rx_connCleanup_list->next;
2028 MUTEX_EXIT(&rx_connHashTable_lock);
2029 rxi_CleanupConnection(conn);
2030 MUTEX_ENTER(&rx_connHashTable_lock);
2032 MUTEX_EXIT(&rx_connHashTable_lock);
2033 #endif /* RX_ENABLE_LOCKS */
2038 afs_winsockCleanup();
2046 /* if we wakeup packet waiter too often, can get in loop with two
2047 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2049 rxi_PacketsUnWait(void)
2051 if (!rx_waitingForPackets) {
2055 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2056 return; /* still over quota */
2059 rx_waitingForPackets = 0;
2060 #ifdef RX_ENABLE_LOCKS
2061 CV_BROADCAST(&rx_waitingForPackets_cv);
2063 osi_rxWakeup(&rx_waitingForPackets);
2069 /* ------------------Internal interfaces------------------------- */
2071 /* Return this process's service structure for the
2072 * specified socket and service */
2074 rxi_FindService(register osi_socket socket, register u_short serviceId)
2076 register struct rx_service **sp;
2077 for (sp = &rx_services[0]; *sp; sp++) {
2078 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2084 /* Allocate a call structure, for the indicated channel of the
2085 * supplied connection. The mode and state of the call must be set by
2086 * the caller. Returns the call with mutex locked. */
2088 rxi_NewCall(register struct rx_connection *conn, register int channel)
2090 register struct rx_call *call;
2091 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2092 register struct rx_call *cp; /* Call pointer temp */
2093 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2094 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2096 dpf(("rxi_NewCall(conn %x, channel %d)\n", conn, channel));
2098 /* Grab an existing call structure, or allocate a new one.
2099 * Existing call structures are assumed to have been left reset by
2101 MUTEX_ENTER(&rx_freeCallQueue_lock);
2103 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2105 * EXCEPT that the TQ might not yet be cleared out.
2106 * Skip over those with in-use TQs.
2109 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2110 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2116 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2117 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2118 call = queue_First(&rx_freeCallQueue, rx_call);
2119 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2121 rx_MutexDecrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2122 MUTEX_EXIT(&rx_freeCallQueue_lock);
2123 MUTEX_ENTER(&call->lock);
2124 CLEAR_CALL_QUEUE_LOCK(call);
2125 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2126 /* Now, if TQ wasn't cleared earlier, do it now. */
2127 if (call->flags & RX_CALL_TQ_CLEARME) {
2128 rxi_ClearTransmitQueue(call, 0);
2129 queue_Init(&call->tq);
2131 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2132 /* Bind the call to its connection structure */
2134 rxi_ResetCall(call, 1);
2136 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2138 MUTEX_EXIT(&rx_freeCallQueue_lock);
2139 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2140 MUTEX_ENTER(&call->lock);
2141 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2142 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2143 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2145 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2146 /* Initialize once-only items */
2147 queue_Init(&call->tq);
2148 queue_Init(&call->rq);
2149 queue_Init(&call->iovq);
2150 /* Bind the call to its connection structure (prereq for reset) */
2152 rxi_ResetCall(call, 1);
2154 call->channel = channel;
2155 call->callNumber = &conn->callNumber[channel];
2156 /* Note that the next expected call number is retained (in
2157 * conn->callNumber[i]), even if we reallocate the call structure
2159 conn->call[channel] = call;
2160 /* if the channel's never been used (== 0), we should start at 1, otherwise
2161 * the call number is valid from the last time this channel was used */
2162 if (*call->callNumber == 0)
2163 *call->callNumber = 1;
2168 /* A call has been inactive long enough that so we can throw away
2169 * state, including the call structure, which is placed on the call
2171 * Call is locked upon entry.
2172 * haveCTLock set if called from rxi_ReapConnections
2174 #ifdef RX_ENABLE_LOCKS
2176 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2177 #else /* RX_ENABLE_LOCKS */
2179 rxi_FreeCall(register struct rx_call *call)
2180 #endif /* RX_ENABLE_LOCKS */
2182 register int channel = call->channel;
2183 register struct rx_connection *conn = call->conn;
2186 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2187 (*call->callNumber)++;
2188 rxi_ResetCall(call, 0);
2189 call->conn->call[channel] = (struct rx_call *)0;
2191 MUTEX_ENTER(&rx_freeCallQueue_lock);
2192 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2193 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2194 /* A call may be free even though its transmit queue is still in use.
2195 * Since we search the call list from head to tail, put busy calls at
2196 * the head of the list, and idle calls at the tail.
2198 if (call->flags & RX_CALL_TQ_BUSY)
2199 queue_Prepend(&rx_freeCallQueue, call);
2201 queue_Append(&rx_freeCallQueue, call);
2202 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2203 queue_Append(&rx_freeCallQueue, call);
2204 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2205 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2206 MUTEX_EXIT(&rx_freeCallQueue_lock);
2208 /* Destroy the connection if it was previously slated for
2209 * destruction, i.e. the Rx client code previously called
2210 * rx_DestroyConnection (client connections), or
2211 * rxi_ReapConnections called the same routine (server
2212 * connections). Only do this, however, if there are no
2213 * outstanding calls. Note that for fine grain locking, there appears
2214 * to be a deadlock in that rxi_FreeCall has a call locked and
2215 * DestroyConnectionNoLock locks each call in the conn. But note a
2216 * few lines up where we have removed this call from the conn.
2217 * If someone else destroys a connection, they either have no
2218 * call lock held or are going through this section of code.
2220 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2221 MUTEX_ENTER(&conn->conn_data_lock);
2223 MUTEX_EXIT(&conn->conn_data_lock);
2224 #ifdef RX_ENABLE_LOCKS
2226 rxi_DestroyConnectionNoLock(conn);
2228 rxi_DestroyConnection(conn);
2229 #else /* RX_ENABLE_LOCKS */
2230 rxi_DestroyConnection(conn);
2231 #endif /* RX_ENABLE_LOCKS */
2235 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2237 rxi_Alloc(register size_t size)
2241 rx_MutexAdd1Increment2(rxi_Allocsize, (afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2242 p = (char *)osi_Alloc(size);
2245 osi_Panic("rxi_Alloc error");
2251 rxi_Free(void *addr, register size_t size)
2253 rx_MutexAdd1Decrement2(rxi_Allocsize, -(afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2254 osi_Free(addr, size);
2257 /* Find the peer process represented by the supplied (host,port)
2258 * combination. If there is no appropriate active peer structure, a
2259 * new one will be allocated and initialized
2260 * The origPeer, if set, is a pointer to a peer structure on which the
2261 * refcount will be be decremented. This is used to replace the peer
2262 * structure hanging off a connection structure */
2264 rxi_FindPeer(register afs_uint32 host, register u_short port,
2265 struct rx_peer *origPeer, int create)
2267 register struct rx_peer *pp;
2269 hashIndex = PEER_HASH(host, port);
2270 MUTEX_ENTER(&rx_peerHashTable_lock);
2271 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2272 if ((pp->host == host) && (pp->port == port))
2277 pp = rxi_AllocPeer(); /* This bzero's *pp */
2278 pp->host = host; /* set here or in InitPeerParams is zero */
2280 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2281 queue_Init(&pp->congestionQueue);
2282 queue_Init(&pp->rpcStats);
2283 pp->next = rx_peerHashTable[hashIndex];
2284 rx_peerHashTable[hashIndex] = pp;
2285 rxi_InitPeerParams(pp);
2286 rx_MutexIncrement(rx_stats.nPeerStructs, rx_stats_mutex);
2293 origPeer->refCount--;
2294 MUTEX_EXIT(&rx_peerHashTable_lock);
2299 /* Find the connection at (host, port) started at epoch, and with the
2300 * given connection id. Creates the server connection if necessary.
2301 * The type specifies whether a client connection or a server
2302 * connection is desired. In both cases, (host, port) specify the
2303 * peer's (host, pair) pair. Client connections are not made
2304 * automatically by this routine. The parameter socket gives the
2305 * socket descriptor on which the packet was received. This is used,
2306 * in the case of server connections, to check that *new* connections
2307 * come via a valid (port, serviceId). Finally, the securityIndex
2308 * parameter must match the existing index for the connection. If a
2309 * server connection is created, it will be created using the supplied
2310 * index, if the index is valid for this service */
2311 struct rx_connection *
2312 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2313 register u_short port, u_short serviceId, afs_uint32 cid,
2314 afs_uint32 epoch, int type, u_int securityIndex)
2316 int hashindex, flag;
2317 register struct rx_connection *conn;
2318 hashindex = CONN_HASH(host, port, cid, epoch, type);
2319 MUTEX_ENTER(&rx_connHashTable_lock);
2320 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2321 rx_connHashTable[hashindex],
2324 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2325 && (epoch == conn->epoch)) {
2326 register struct rx_peer *pp = conn->peer;
2327 if (securityIndex != conn->securityIndex) {
2328 /* this isn't supposed to happen, but someone could forge a packet
2329 * like this, and there seems to be some CM bug that makes this
2330 * happen from time to time -- in which case, the fileserver
2332 MUTEX_EXIT(&rx_connHashTable_lock);
2333 return (struct rx_connection *)0;
2335 if (pp->host == host && pp->port == port)
2337 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2339 /* So what happens when it's a callback connection? */
2340 if ( /*type == RX_CLIENT_CONNECTION && */
2341 (conn->epoch & 0x80000000))
2345 /* the connection rxLastConn that was used the last time is not the
2346 ** one we are looking for now. Hence, start searching in the hash */
2348 conn = rx_connHashTable[hashindex];
2353 struct rx_service *service;
2354 if (type == RX_CLIENT_CONNECTION) {
2355 MUTEX_EXIT(&rx_connHashTable_lock);
2356 return (struct rx_connection *)0;
2358 service = rxi_FindService(socket, serviceId);
2359 if (!service || (securityIndex >= service->nSecurityObjects)
2360 || (service->securityObjects[securityIndex] == 0)) {
2361 MUTEX_EXIT(&rx_connHashTable_lock);
2362 return (struct rx_connection *)0;
2364 conn = rxi_AllocConnection(); /* This bzero's the connection */
2365 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2366 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2367 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2368 conn->next = rx_connHashTable[hashindex];
2369 rx_connHashTable[hashindex] = conn;
2370 conn->peer = rxi_FindPeer(host, port, 0, 1);
2371 conn->type = RX_SERVER_CONNECTION;
2372 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2373 conn->epoch = epoch;
2374 conn->cid = cid & RX_CIDMASK;
2375 /* conn->serial = conn->lastSerial = 0; */
2376 /* conn->timeout = 0; */
2377 conn->ackRate = RX_FAST_ACK_RATE;
2378 conn->service = service;
2379 conn->serviceId = serviceId;
2380 conn->securityIndex = securityIndex;
2381 conn->securityObject = service->securityObjects[securityIndex];
2382 conn->nSpecific = 0;
2383 conn->specific = NULL;
2384 rx_SetConnDeadTime(conn, service->connDeadTime);
2385 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2386 /* Notify security object of the new connection */
2387 RXS_NewConnection(conn->securityObject, conn);
2388 /* XXXX Connection timeout? */
2389 if (service->newConnProc)
2390 (*service->newConnProc) (conn);
2391 rx_MutexIncrement(rx_stats.nServerConns, rx_stats_mutex);
2394 MUTEX_ENTER(&conn->conn_data_lock);
2396 MUTEX_EXIT(&conn->conn_data_lock);
2398 rxLastConn = conn; /* store this connection as the last conn used */
2399 MUTEX_EXIT(&rx_connHashTable_lock);
2403 /* There are two packet tracing routines available for testing and monitoring
2404 * Rx. One is called just after every packet is received and the other is
2405 * called just before every packet is sent. Received packets, have had their
2406 * headers decoded, and packets to be sent have not yet had their headers
2407 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2408 * containing the network address. Both can be modified. The return value, if
2409 * non-zero, indicates that the packet should be dropped. */
2411 int (*rx_justReceived) () = 0;
2412 int (*rx_almostSent) () = 0;
2414 /* A packet has been received off the interface. Np is the packet, socket is
2415 * the socket number it was received from (useful in determining which service
2416 * this packet corresponds to), and (host, port) reflect the host,port of the
2417 * sender. This call returns the packet to the caller if it is finished with
2418 * it, rather than de-allocating it, just as a small performance hack */
2421 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2422 afs_uint32 host, u_short port, int *tnop,
2423 struct rx_call **newcallp)
2425 register struct rx_call *call;
2426 register struct rx_connection *conn;
2428 afs_uint32 currentCallNumber;
2434 struct rx_packet *tnp;
2437 /* We don't print out the packet until now because (1) the time may not be
2438 * accurate enough until now in the lwp implementation (rx_Listener only gets
2439 * the time after the packet is read) and (2) from a protocol point of view,
2440 * this is the first time the packet has been seen */
2441 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2442 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2443 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2444 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2445 np->header.epoch, np->header.cid, np->header.callNumber,
2446 np->header.seq, np->header.flags, np));
2449 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2450 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2453 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2454 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2457 /* If an input tracer function is defined, call it with the packet and
2458 * network address. Note this function may modify its arguments. */
2459 if (rx_justReceived) {
2460 struct sockaddr_in addr;
2462 addr.sin_family = AF_INET;
2463 addr.sin_port = port;
2464 addr.sin_addr.s_addr = host;
2465 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2466 addr.sin_len = sizeof(addr);
2467 #endif /* AFS_OSF_ENV */
2468 drop = (*rx_justReceived) (np, &addr);
2469 /* drop packet if return value is non-zero */
2472 port = addr.sin_port; /* in case fcn changed addr */
2473 host = addr.sin_addr.s_addr;
2477 /* If packet was not sent by the client, then *we* must be the client */
2478 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2479 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2481 /* Find the connection (or fabricate one, if we're the server & if
2482 * necessary) associated with this packet */
2484 rxi_FindConnection(socket, host, port, np->header.serviceId,
2485 np->header.cid, np->header.epoch, type,
2486 np->header.securityIndex);
2489 /* If no connection found or fabricated, just ignore the packet.
2490 * (An argument could be made for sending an abort packet for
2495 MUTEX_ENTER(&conn->conn_data_lock);
2496 if (conn->maxSerial < np->header.serial)
2497 conn->maxSerial = np->header.serial;
2498 MUTEX_EXIT(&conn->conn_data_lock);
2500 /* If the connection is in an error state, send an abort packet and ignore
2501 * the incoming packet */
2503 /* Don't respond to an abort packet--we don't want loops! */
2504 MUTEX_ENTER(&conn->conn_data_lock);
2505 if (np->header.type != RX_PACKET_TYPE_ABORT)
2506 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2508 MUTEX_EXIT(&conn->conn_data_lock);
2512 /* Check for connection-only requests (i.e. not call specific). */
2513 if (np->header.callNumber == 0) {
2514 switch (np->header.type) {
2515 case RX_PACKET_TYPE_ABORT: {
2516 /* What if the supplied error is zero? */
2517 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2518 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2519 rxi_ConnectionError(conn, errcode);
2520 MUTEX_ENTER(&conn->conn_data_lock);
2522 MUTEX_EXIT(&conn->conn_data_lock);
2525 case RX_PACKET_TYPE_CHALLENGE:
2526 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2527 MUTEX_ENTER(&conn->conn_data_lock);
2529 MUTEX_EXIT(&conn->conn_data_lock);
2531 case RX_PACKET_TYPE_RESPONSE:
2532 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2533 MUTEX_ENTER(&conn->conn_data_lock);
2535 MUTEX_EXIT(&conn->conn_data_lock);
2537 case RX_PACKET_TYPE_PARAMS:
2538 case RX_PACKET_TYPE_PARAMS + 1:
2539 case RX_PACKET_TYPE_PARAMS + 2:
2540 /* ignore these packet types for now */
2541 MUTEX_ENTER(&conn->conn_data_lock);
2543 MUTEX_EXIT(&conn->conn_data_lock);
2548 /* Should not reach here, unless the peer is broken: send an
2550 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2551 MUTEX_ENTER(&conn->conn_data_lock);
2552 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2554 MUTEX_EXIT(&conn->conn_data_lock);
2559 channel = np->header.cid & RX_CHANNELMASK;
2560 call = conn->call[channel];
2561 #ifdef RX_ENABLE_LOCKS
2563 MUTEX_ENTER(&call->lock);
2564 /* Test to see if call struct is still attached to conn. */
2565 if (call != conn->call[channel]) {
2567 MUTEX_EXIT(&call->lock);
2568 if (type == RX_SERVER_CONNECTION) {
2569 call = conn->call[channel];
2570 /* If we started with no call attached and there is one now,
2571 * another thread is also running this routine and has gotten
2572 * the connection channel. We should drop this packet in the tests
2573 * below. If there was a call on this connection and it's now
2574 * gone, then we'll be making a new call below.
2575 * If there was previously a call and it's now different then
2576 * the old call was freed and another thread running this routine
2577 * has created a call on this channel. One of these two threads
2578 * has a packet for the old call and the code below handles those
2582 MUTEX_ENTER(&call->lock);
2584 /* This packet can't be for this call. If the new call address is
2585 * 0 then no call is running on this channel. If there is a call
2586 * then, since this is a client connection we're getting data for
2587 * it must be for the previous call.
2589 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2590 MUTEX_ENTER(&conn->conn_data_lock);
2592 MUTEX_EXIT(&conn->conn_data_lock);
2597 currentCallNumber = conn->callNumber[channel];
2599 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2600 if (np->header.callNumber < currentCallNumber) {
2601 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2602 #ifdef RX_ENABLE_LOCKS
2604 MUTEX_EXIT(&call->lock);
2606 MUTEX_ENTER(&conn->conn_data_lock);
2608 MUTEX_EXIT(&conn->conn_data_lock);
2612 MUTEX_ENTER(&conn->conn_call_lock);
2613 call = rxi_NewCall(conn, channel);
2614 MUTEX_EXIT(&conn->conn_call_lock);
2615 *call->callNumber = np->header.callNumber;
2616 if (np->header.callNumber == 0)
2617 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));
2619 call->state = RX_STATE_PRECALL;
2620 clock_GetTime(&call->queueTime);
2621 hzero(call->bytesSent);
2622 hzero(call->bytesRcvd);
2624 * If the number of queued calls exceeds the overload
2625 * threshold then abort this call.
2627 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2628 struct rx_packet *tp;
2630 rxi_CallError(call, rx_BusyError);
2631 tp = rxi_SendCallAbort(call, np, 1, 0);
2632 MUTEX_EXIT(&call->lock);
2633 MUTEX_ENTER(&conn->conn_data_lock);
2635 MUTEX_EXIT(&conn->conn_data_lock);
2636 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2639 rxi_KeepAliveOn(call);
2640 } else if (np->header.callNumber != currentCallNumber) {
2641 /* Wait until the transmit queue is idle before deciding
2642 * whether to reset the current call. Chances are that the
2643 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2646 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2647 while ((call->state == RX_STATE_ACTIVE)
2648 && (call->flags & RX_CALL_TQ_BUSY)) {
2649 call->flags |= RX_CALL_TQ_WAIT;
2651 #ifdef RX_ENABLE_LOCKS
2652 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2653 CV_WAIT(&call->cv_tq, &call->lock);
2654 #else /* RX_ENABLE_LOCKS */
2655 osi_rxSleep(&call->tq);
2656 #endif /* RX_ENABLE_LOCKS */
2658 if (call->tqWaiters == 0)
2659 call->flags &= ~RX_CALL_TQ_WAIT;
2661 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2662 /* If the new call cannot be taken right now send a busy and set
2663 * the error condition in this call, so that it terminates as
2664 * quickly as possible */
2665 if (call->state == RX_STATE_ACTIVE) {
2666 struct rx_packet *tp;
2668 rxi_CallError(call, RX_CALL_DEAD);
2669 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2671 MUTEX_EXIT(&call->lock);
2672 MUTEX_ENTER(&conn->conn_data_lock);
2674 MUTEX_EXIT(&conn->conn_data_lock);
2677 rxi_ResetCall(call, 0);
2678 *call->callNumber = np->header.callNumber;
2679 if (np->header.callNumber == 0)
2680 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %lx resend %d.%0.3d len %d", np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port), np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq, np->header.flags, (unsigned long)np, np->retryTime.sec, np->retryTime.usec / 1000, np->length));
2682 call->state = RX_STATE_PRECALL;
2683 clock_GetTime(&call->queueTime);
2684 hzero(call->bytesSent);
2685 hzero(call->bytesRcvd);
2687 * If the number of queued calls exceeds the overload
2688 * threshold then abort this call.
2690 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2691 struct rx_packet *tp;
2693 rxi_CallError(call, rx_BusyError);
2694 tp = rxi_SendCallAbort(call, np, 1, 0);
2695 MUTEX_EXIT(&call->lock);
2696 MUTEX_ENTER(&conn->conn_data_lock);
2698 MUTEX_EXIT(&conn->conn_data_lock);
2699 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2702 rxi_KeepAliveOn(call);
2704 /* Continuing call; do nothing here. */
2706 } else { /* we're the client */
2707 /* Ignore all incoming acknowledgements for calls in DALLY state */
2708 if (call && (call->state == RX_STATE_DALLY)
2709 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2710 rx_MutexIncrement(rx_stats.ignorePacketDally, rx_stats_mutex);
2711 #ifdef RX_ENABLE_LOCKS
2713 MUTEX_EXIT(&call->lock);
2716 MUTEX_ENTER(&conn->conn_data_lock);
2718 MUTEX_EXIT(&conn->conn_data_lock);
2722 /* Ignore anything that's not relevant to the current call. If there
2723 * isn't a current call, then no packet is relevant. */
2724 if (!call || (np->header.callNumber != currentCallNumber)) {
2725 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2726 #ifdef RX_ENABLE_LOCKS
2728 MUTEX_EXIT(&call->lock);
2731 MUTEX_ENTER(&conn->conn_data_lock);
2733 MUTEX_EXIT(&conn->conn_data_lock);
2736 /* If the service security object index stamped in the packet does not
2737 * match the connection's security index, ignore the packet */
2738 if (np->header.securityIndex != conn->securityIndex) {
2739 #ifdef RX_ENABLE_LOCKS
2740 MUTEX_EXIT(&call->lock);
2742 MUTEX_ENTER(&conn->conn_data_lock);
2744 MUTEX_EXIT(&conn->conn_data_lock);
2748 /* If we're receiving the response, then all transmit packets are
2749 * implicitly acknowledged. Get rid of them. */
2750 if (np->header.type == RX_PACKET_TYPE_DATA) {
2751 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2752 /* XXX Hack. Because we must release the global rx lock when
2753 * sending packets (osi_NetSend) we drop all acks while we're
2754 * traversing the tq in rxi_Start sending packets out because
2755 * packets may move to the freePacketQueue as result of being here!
2756 * So we drop these packets until we're safely out of the
2757 * traversing. Really ugly!
2758 * For fine grain RX locking, we set the acked field in the
2759 * packets and let rxi_Start remove them from the transmit queue.
2761 if (call->flags & RX_CALL_TQ_BUSY) {
2762 #ifdef RX_ENABLE_LOCKS
2763 rxi_SetAcksInTransmitQueue(call);
2766 return np; /* xmitting; drop packet */
2769 rxi_ClearTransmitQueue(call, 0);
2771 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2772 rxi_ClearTransmitQueue(call, 0);
2773 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2775 if (np->header.type == RX_PACKET_TYPE_ACK) {
2776 /* now check to see if this is an ack packet acknowledging that the
2777 * server actually *lost* some hard-acked data. If this happens we
2778 * ignore this packet, as it may indicate that the server restarted in
2779 * the middle of a call. It is also possible that this is an old ack
2780 * packet. We don't abort the connection in this case, because this
2781 * *might* just be an old ack packet. The right way to detect a server
2782 * restart in the midst of a call is to notice that the server epoch
2784 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2785 * XXX unacknowledged. I think that this is off-by-one, but
2786 * XXX I don't dare change it just yet, since it will
2787 * XXX interact badly with the server-restart detection
2788 * XXX code in receiveackpacket. */
2789 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2790 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2791 MUTEX_EXIT(&call->lock);
2792 MUTEX_ENTER(&conn->conn_data_lock);
2794 MUTEX_EXIT(&conn->conn_data_lock);
2798 } /* else not a data packet */
2801 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2802 /* Set remote user defined status from packet */
2803 call->remoteStatus = np->header.userStatus;
2805 /* Note the gap between the expected next packet and the actual
2806 * packet that arrived, when the new packet has a smaller serial number
2807 * than expected. Rioses frequently reorder packets all by themselves,
2808 * so this will be quite important with very large window sizes.
2809 * Skew is checked against 0 here to avoid any dependence on the type of
2810 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2812 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2813 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2814 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2816 MUTEX_ENTER(&conn->conn_data_lock);
2817 skew = conn->lastSerial - np->header.serial;
2818 conn->lastSerial = np->header.serial;
2819 MUTEX_EXIT(&conn->conn_data_lock);
2821 register struct rx_peer *peer;
2823 if (skew > peer->inPacketSkew) {
2824 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2826 peer->inPacketSkew = skew;
2830 /* Now do packet type-specific processing */
2831 switch (np->header.type) {
2832 case RX_PACKET_TYPE_DATA:
2833 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2836 case RX_PACKET_TYPE_ACK:
2837 /* Respond immediately to ack packets requesting acknowledgement
2839 if (np->header.flags & RX_REQUEST_ACK) {
2841 (void)rxi_SendCallAbort(call, 0, 1, 0);
2843 (void)rxi_SendAck(call, 0, np->header.serial,
2844 RX_ACK_PING_RESPONSE, 1);
2846 np = rxi_ReceiveAckPacket(call, np, 1);
2848 case RX_PACKET_TYPE_ABORT: {
2849 /* An abort packet: reset the call, passing the error up to the user. */
2850 /* What if error is zero? */
2851 /* What if the error is -1? the application will treat it as a timeout. */
2852 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
2853 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
2854 rxi_CallError(call, errdata);
2855 MUTEX_EXIT(&call->lock);
2856 MUTEX_ENTER(&conn->conn_data_lock);
2858 MUTEX_EXIT(&conn->conn_data_lock);
2859 return np; /* xmitting; drop packet */
2861 case RX_PACKET_TYPE_BUSY:
2864 case RX_PACKET_TYPE_ACKALL:
2865 /* All packets acknowledged, so we can drop all packets previously
2866 * readied for sending */
2867 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2868 /* XXX Hack. We because we can't release the global rx lock when
2869 * sending packets (osi_NetSend) we drop all ack pkts while we're
2870 * traversing the tq in rxi_Start sending packets out because
2871 * packets may move to the freePacketQueue as result of being
2872 * here! So we drop these packets until we're safely out of the
2873 * traversing. Really ugly!
2874 * For fine grain RX locking, we set the acked field in the packets
2875 * and let rxi_Start remove the packets from the transmit queue.
2877 if (call->flags & RX_CALL_TQ_BUSY) {
2878 #ifdef RX_ENABLE_LOCKS
2879 rxi_SetAcksInTransmitQueue(call);
2881 #else /* RX_ENABLE_LOCKS */
2882 MUTEX_EXIT(&call->lock);
2883 MUTEX_ENTER(&conn->conn_data_lock);
2885 MUTEX_EXIT(&conn->conn_data_lock);
2886 return np; /* xmitting; drop packet */
2887 #endif /* RX_ENABLE_LOCKS */
2889 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2890 rxi_ClearTransmitQueue(call, 0);
2893 /* Should not reach here, unless the peer is broken: send an abort
2895 rxi_CallError(call, RX_PROTOCOL_ERROR);
2896 np = rxi_SendCallAbort(call, np, 1, 0);
2899 /* Note when this last legitimate packet was received, for keep-alive
2900 * processing. Note, we delay getting the time until now in the hope that
2901 * the packet will be delivered to the user before any get time is required
2902 * (if not, then the time won't actually be re-evaluated here). */
2903 call->lastReceiveTime = clock_Sec();
2904 MUTEX_EXIT(&call->lock);
2905 MUTEX_ENTER(&conn->conn_data_lock);
2907 MUTEX_EXIT(&conn->conn_data_lock);
2911 /* return true if this is an "interesting" connection from the point of view
2912 of someone trying to debug the system */
2914 rxi_IsConnInteresting(struct rx_connection *aconn)
2917 register struct rx_call *tcall;
2919 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2921 for (i = 0; i < RX_MAXCALLS; i++) {
2922 tcall = aconn->call[i];
2924 if ((tcall->state == RX_STATE_PRECALL)
2925 || (tcall->state == RX_STATE_ACTIVE))
2927 if ((tcall->mode == RX_MODE_SENDING)
2928 || (tcall->mode == RX_MODE_RECEIVING))
2936 /* if this is one of the last few packets AND it wouldn't be used by the
2937 receiving call to immediately satisfy a read request, then drop it on
2938 the floor, since accepting it might prevent a lock-holding thread from
2939 making progress in its reading. If a call has been cleared while in
2940 the precall state then ignore all subsequent packets until the call
2941 is assigned to a thread. */
2944 TooLow(struct rx_packet *ap, struct rx_call *acall)
2947 MUTEX_ENTER(&rx_stats_mutex);
2948 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
2949 && (acall->state == RX_STATE_PRECALL))
2950 || ((rx_nFreePackets < rxi_dataQuota + 2)
2951 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
2952 && (acall->flags & RX_CALL_READER_WAIT)))) {
2955 MUTEX_EXIT(&rx_stats_mutex);
2961 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
2962 struct rx_call *acall)
2964 struct rx_call *call = acall;
2965 struct clock when, now;
2968 MUTEX_ENTER(&conn->conn_data_lock);
2969 conn->checkReachEvent = NULL;
2970 waiting = conn->flags & RX_CONN_ATTACHWAIT;
2973 MUTEX_EXIT(&conn->conn_data_lock);
2977 MUTEX_ENTER(&conn->conn_call_lock);
2978 MUTEX_ENTER(&conn->conn_data_lock);
2979 for (i = 0; i < RX_MAXCALLS; i++) {
2980 struct rx_call *tc = conn->call[i];
2981 if (tc && tc->state == RX_STATE_PRECALL) {
2987 /* Indicate that rxi_CheckReachEvent is no longer running by
2988 * clearing the flag. Must be atomic under conn_data_lock to
2989 * avoid a new call slipping by: rxi_CheckConnReach holds
2990 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
2992 conn->flags &= ~RX_CONN_ATTACHWAIT;
2993 MUTEX_EXIT(&conn->conn_data_lock);
2994 MUTEX_EXIT(&conn->conn_call_lock);
2999 MUTEX_ENTER(&call->lock);
3000 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3002 MUTEX_EXIT(&call->lock);
3004 clock_GetTime(&now);
3006 when.sec += RX_CHECKREACH_TIMEOUT;
3007 MUTEX_ENTER(&conn->conn_data_lock);
3008 if (!conn->checkReachEvent) {
3010 conn->checkReachEvent =
3011 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3014 MUTEX_EXIT(&conn->conn_data_lock);
3020 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3022 struct rx_service *service = conn->service;
3023 struct rx_peer *peer = conn->peer;
3024 afs_uint32 now, lastReach;
3026 if (service->checkReach == 0)
3030 MUTEX_ENTER(&peer->peer_lock);
3031 lastReach = peer->lastReachTime;
3032 MUTEX_EXIT(&peer->peer_lock);
3033 if (now - lastReach < RX_CHECKREACH_TTL)
3036 MUTEX_ENTER(&conn->conn_data_lock);
3037 if (conn->flags & RX_CONN_ATTACHWAIT) {
3038 MUTEX_EXIT(&conn->conn_data_lock);
3041 conn->flags |= RX_CONN_ATTACHWAIT;
3042 MUTEX_EXIT(&conn->conn_data_lock);
3043 if (!conn->checkReachEvent)
3044 rxi_CheckReachEvent(NULL, conn, call);
3049 /* try to attach call, if authentication is complete */
3051 TryAttach(register struct rx_call *acall, register osi_socket socket,
3052 register int *tnop, register struct rx_call **newcallp,
3055 struct rx_connection *conn = acall->conn;
3057 if (conn->type == RX_SERVER_CONNECTION
3058 && acall->state == RX_STATE_PRECALL) {
3059 /* Don't attach until we have any req'd. authentication. */
3060 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3061 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3062 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3063 /* Note: this does not necessarily succeed; there
3064 * may not any proc available
3067 rxi_ChallengeOn(acall->conn);
3072 /* A data packet has been received off the interface. This packet is
3073 * appropriate to the call (the call is in the right state, etc.). This
3074 * routine can return a packet to the caller, for re-use */
3077 rxi_ReceiveDataPacket(register struct rx_call *call,
3078 register struct rx_packet *np, int istack,
3079 osi_socket socket, afs_uint32 host, u_short port,
3080 int *tnop, struct rx_call **newcallp)
3082 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3086 afs_uint32 seq, serial, flags;
3088 struct rx_packet *tnp;
3089 struct clock when, now;
3090 rx_MutexIncrement(rx_stats.dataPacketsRead, rx_stats_mutex);
3093 /* If there are no packet buffers, drop this new packet, unless we can find
3094 * packet buffers from inactive calls */
3096 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3097 MUTEX_ENTER(&rx_freePktQ_lock);
3098 rxi_NeedMorePackets = TRUE;
3099 MUTEX_EXIT(&rx_freePktQ_lock);
3100 rx_MutexIncrement(rx_stats.noPacketBuffersOnRead, rx_stats_mutex);
3101 call->rprev = np->header.serial;
3102 rxi_calltrace(RX_TRACE_DROP, call);
3103 dpf(("packet %x dropped on receipt - quota problems", np));
3105 rxi_ClearReceiveQueue(call);
3106 clock_GetTime(&now);
3108 clock_Add(&when, &rx_softAckDelay);
3109 if (!call->delayedAckEvent
3110 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3111 rxevent_Cancel(call->delayedAckEvent, call,
3112 RX_CALL_REFCOUNT_DELAY);
3113 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3114 call->delayedAckEvent =
3115 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3117 /* we've damaged this call already, might as well do it in. */
3123 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3124 * packet is one of several packets transmitted as a single
3125 * datagram. Do not send any soft or hard acks until all packets
3126 * in a jumbogram have been processed. Send negative acks right away.
3128 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3129 /* tnp is non-null when there are more packets in the
3130 * current jumbo gram */
3137 seq = np->header.seq;
3138 serial = np->header.serial;
3139 flags = np->header.flags;
3141 /* If the call is in an error state, send an abort message */
3143 return rxi_SendCallAbort(call, np, istack, 0);
3145 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3146 * AFS 3.5 jumbogram. */
3147 if (flags & RX_JUMBO_PACKET) {
3148 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3153 if (np->header.spare != 0) {
3154 MUTEX_ENTER(&call->conn->conn_data_lock);
3155 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3156 MUTEX_EXIT(&call->conn->conn_data_lock);
3159 /* The usual case is that this is the expected next packet */
3160 if (seq == call->rnext) {
3162 /* Check to make sure it is not a duplicate of one already queued */
3163 if (queue_IsNotEmpty(&call->rq)
3164 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3165 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3166 dpf(("packet %x dropped on receipt - duplicate", np));
3167 rxevent_Cancel(call->delayedAckEvent, call,
3168 RX_CALL_REFCOUNT_DELAY);
3169 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3175 /* It's the next packet. Stick it on the receive queue
3176 * for this call. Set newPackets to make sure we wake
3177 * the reader once all packets have been processed */
3178 queue_Prepend(&call->rq, np);
3180 np = NULL; /* We can't use this anymore */
3183 /* If an ack is requested then set a flag to make sure we
3184 * send an acknowledgement for this packet */
3185 if (flags & RX_REQUEST_ACK) {
3186 ackNeeded = RX_ACK_REQUESTED;
3189 /* Keep track of whether we have received the last packet */
3190 if (flags & RX_LAST_PACKET) {
3191 call->flags |= RX_CALL_HAVE_LAST;
3195 /* Check whether we have all of the packets for this call */
3196 if (call->flags & RX_CALL_HAVE_LAST) {
3197 afs_uint32 tseq; /* temporary sequence number */
3198 struct rx_packet *tp; /* Temporary packet pointer */
3199 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3201 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3202 if (tseq != tp->header.seq)
3204 if (tp->header.flags & RX_LAST_PACKET) {
3205 call->flags |= RX_CALL_RECEIVE_DONE;
3212 /* Provide asynchronous notification for those who want it
3213 * (e.g. multi rx) */
3214 if (call->arrivalProc) {
3215 (*call->arrivalProc) (call, call->arrivalProcHandle,
3216 call->arrivalProcArg);
3217 call->arrivalProc = (void (*)())0;
3220 /* Update last packet received */
3223 /* If there is no server process serving this call, grab
3224 * one, if available. We only need to do this once. If a
3225 * server thread is available, this thread becomes a server
3226 * thread and the server thread becomes a listener thread. */
3228 TryAttach(call, socket, tnop, newcallp, 0);
3231 /* This is not the expected next packet. */
3233 /* Determine whether this is a new or old packet, and if it's
3234 * a new one, whether it fits into the current receive window.
3235 * Also figure out whether the packet was delivered in sequence.
3236 * We use the prev variable to determine whether the new packet
3237 * is the successor of its immediate predecessor in the
3238 * receive queue, and the missing flag to determine whether
3239 * any of this packets predecessors are missing. */
3241 afs_uint32 prev; /* "Previous packet" sequence number */
3242 struct rx_packet *tp; /* Temporary packet pointer */
3243 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3244 int missing; /* Are any predecessors missing? */
3246 /* If the new packet's sequence number has been sent to the
3247 * application already, then this is a duplicate */
3248 if (seq < call->rnext) {
3249 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3250 rxevent_Cancel(call->delayedAckEvent, call,
3251 RX_CALL_REFCOUNT_DELAY);
3252 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3258 /* If the sequence number is greater than what can be
3259 * accomodated by the current window, then send a negative
3260 * acknowledge and drop the packet */
3261 if ((call->rnext + call->rwind) <= seq) {
3262 rxevent_Cancel(call->delayedAckEvent, call,
3263 RX_CALL_REFCOUNT_DELAY);
3264 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3271 /* Look for the packet in the queue of old received packets */
3272 for (prev = call->rnext - 1, missing =
3273 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3274 /*Check for duplicate packet */
3275 if (seq == tp->header.seq) {
3276 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3277 rxevent_Cancel(call->delayedAckEvent, call,
3278 RX_CALL_REFCOUNT_DELAY);
3279 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3285 /* If we find a higher sequence packet, break out and
3286 * insert the new packet here. */
3287 if (seq < tp->header.seq)
3289 /* Check for missing packet */
3290 if (tp->header.seq != prev + 1) {
3294 prev = tp->header.seq;
3297 /* Keep track of whether we have received the last packet. */
3298 if (flags & RX_LAST_PACKET) {
3299 call->flags |= RX_CALL_HAVE_LAST;
3302 /* It's within the window: add it to the the receive queue.
3303 * tp is left by the previous loop either pointing at the
3304 * packet before which to insert the new packet, or at the
3305 * queue head if the queue is empty or the packet should be
3307 queue_InsertBefore(tp, np);
3311 /* Check whether we have all of the packets for this call */
3312 if ((call->flags & RX_CALL_HAVE_LAST)
3313 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3314 afs_uint32 tseq; /* temporary sequence number */
3317 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3318 if (tseq != tp->header.seq)
3320 if (tp->header.flags & RX_LAST_PACKET) {
3321 call->flags |= RX_CALL_RECEIVE_DONE;
3328 /* We need to send an ack of the packet is out of sequence,
3329 * or if an ack was requested by the peer. */
3330 if (seq != prev + 1 || missing) {
3331 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3332 } else if (flags & RX_REQUEST_ACK) {
3333 ackNeeded = RX_ACK_REQUESTED;
3336 /* Acknowledge the last packet for each call */
3337 if (flags & RX_LAST_PACKET) {
3348 * If the receiver is waiting for an iovec, fill the iovec
3349 * using the data from the receive queue */
3350 if (call->flags & RX_CALL_IOVEC_WAIT) {
3351 didHardAck = rxi_FillReadVec(call, serial);
3352 /* the call may have been aborted */
3361 /* Wakeup the reader if any */
3362 if ((call->flags & RX_CALL_READER_WAIT)
3363 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3364 || (call->iovNext >= call->iovMax)
3365 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3366 call->flags &= ~RX_CALL_READER_WAIT;
3367 #ifdef RX_ENABLE_LOCKS
3368 CV_BROADCAST(&call->cv_rq);
3370 osi_rxWakeup(&call->rq);
3376 * Send an ack when requested by the peer, or once every
3377 * rxi_SoftAckRate packets until the last packet has been
3378 * received. Always send a soft ack for the last packet in
3379 * the server's reply. */
3381 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3382 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3383 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3384 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3385 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3386 } else if (call->nSoftAcks) {
3387 clock_GetTime(&now);
3389 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3390 clock_Add(&when, &rx_lastAckDelay);
3392 clock_Add(&when, &rx_softAckDelay);
3394 if (!call->delayedAckEvent
3395 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3396 rxevent_Cancel(call->delayedAckEvent, call,
3397 RX_CALL_REFCOUNT_DELAY);
3398 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3399 call->delayedAckEvent =
3400 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3402 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3403 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3410 static void rxi_ComputeRate();
3414 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3416 struct rx_peer *peer = conn->peer;
3418 MUTEX_ENTER(&peer->peer_lock);
3419 peer->lastReachTime = clock_Sec();
3420 MUTEX_EXIT(&peer->peer_lock);
3422 MUTEX_ENTER(&conn->conn_data_lock);
3423 if (conn->flags & RX_CONN_ATTACHWAIT) {
3426 conn->flags &= ~RX_CONN_ATTACHWAIT;
3427 MUTEX_EXIT(&conn->conn_data_lock);
3429 for (i = 0; i < RX_MAXCALLS; i++) {
3430 struct rx_call *call = conn->call[i];
3433 MUTEX_ENTER(&call->lock);
3434 /* tnop can be null if newcallp is null */
3435 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3437 MUTEX_EXIT(&call->lock);
3441 MUTEX_EXIT(&conn->conn_data_lock);
3445 rx_ack_reason(int reason)
3448 case RX_ACK_REQUESTED:
3450 case RX_ACK_DUPLICATE:
3452 case RX_ACK_OUT_OF_SEQUENCE:
3454 case RX_ACK_EXCEEDS_WINDOW:
3456 case RX_ACK_NOSPACE:
3460 case RX_ACK_PING_RESPONSE:
3472 /* rxi_ComputePeerNetStats
3474 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3475 * estimates (like RTT and throughput) based on ack packets. Caller
3476 * must ensure that the packet in question is the right one (i.e.
3477 * serial number matches).
3480 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3481 struct rx_ackPacket *ap, struct rx_packet *np)
3483 struct rx_peer *peer = call->conn->peer;
3485 /* Use RTT if not delayed by client. */
3486 if (ap->reason != RX_ACK_DELAY)
3487 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3489 rxi_ComputeRate(peer, call, p, np, ap->reason);
3493 /* The real smarts of the whole thing. */
3495 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3498 struct rx_ackPacket *ap;
3500 register struct rx_packet *tp;
3501 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3502 register struct rx_connection *conn = call->conn;
3503 struct rx_peer *peer = conn->peer;
3506 /* because there are CM's that are bogus, sending weird values for this. */
3507 afs_uint32 skew = 0;
3512 int newAckCount = 0;
3513 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3514 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3516 rx_MutexIncrement(rx_stats.ackPacketsRead, rx_stats_mutex);
3517 ap = (struct rx_ackPacket *)rx_DataOf(np);
3518 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3520 return np; /* truncated ack packet */
3522 /* depends on ack packet struct */
3523 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3524 first = ntohl(ap->firstPacket);
3525 serial = ntohl(ap->serial);
3526 /* temporarily disabled -- needs to degrade over time
3527 * skew = ntohs(ap->maxSkew); */
3529 /* Ignore ack packets received out of order */
3530 if (first < call->tfirst) {
3534 if (np->header.flags & RX_SLOW_START_OK) {
3535 call->flags |= RX_CALL_SLOW_START_OK;
3538 if (ap->reason == RX_ACK_PING_RESPONSE)
3539 rxi_UpdatePeerReach(conn, call);
3543 if (rxdebug_active) {
3547 len = _snprintf(msg, sizeof(msg),
3548 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3549 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3550 ntohl(ap->serial), ntohl(ap->previousPacket),
3551 (unsigned int)np->header.seq, (unsigned int)skew,
3552 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3556 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3557 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3561 OutputDebugString(msg);
3563 #else /* AFS_NT40_ENV */
3566 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3567 ap->reason, ntohl(ap->previousPacket),
3568 (unsigned int)np->header.seq, (unsigned int)serial,
3569 (unsigned int)skew, ntohl(ap->firstPacket));
3572 for (offset = 0; offset < nAcks; offset++)
3573 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3578 #endif /* AFS_NT40_ENV */
3581 /* Update the outgoing packet skew value to the latest value of
3582 * the peer's incoming packet skew value. The ack packet, of
3583 * course, could arrive out of order, but that won't affect things
3585 MUTEX_ENTER(&peer->peer_lock);
3586 peer->outPacketSkew = skew;
3588 /* Check for packets that no longer need to be transmitted, and
3589 * discard them. This only applies to packets positively
3590 * acknowledged as having been sent to the peer's upper level.
3591 * All other packets must be retained. So only packets with
3592 * sequence numbers < ap->firstPacket are candidates. */
3593 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3594 if (tp->header.seq >= first)
3596 call->tfirst = tp->header.seq + 1;
3598 && (tp->header.serial == serial || tp->firstSerial == serial))
3599 rxi_ComputePeerNetStats(call, tp, ap, np);
3600 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3603 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3604 /* XXX Hack. Because we have to release the global rx lock when sending
3605 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3606 * in rxi_Start sending packets out because packets may move to the
3607 * freePacketQueue as result of being here! So we drop these packets until
3608 * we're safely out of the traversing. Really ugly!
3609 * To make it even uglier, if we're using fine grain locking, we can
3610 * set the ack bits in the packets and have rxi_Start remove the packets
3611 * when it's done transmitting.
3613 if (call->flags & RX_CALL_TQ_BUSY) {
3614 #ifdef RX_ENABLE_LOCKS
3615 tp->flags |= RX_PKTFLAG_ACKED;
3616 call->flags |= RX_CALL_TQ_SOME_ACKED;
3617 #else /* RX_ENABLE_LOCKS */
3619 #endif /* RX_ENABLE_LOCKS */
3621 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3624 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3629 /* Give rate detector a chance to respond to ping requests */
3630 if (ap->reason == RX_ACK_PING_RESPONSE) {
3631 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3635 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3637 /* Now go through explicit acks/nacks and record the results in
3638 * the waiting packets. These are packets that can't be released
3639 * yet, even with a positive acknowledge. This positive
3640 * acknowledge only means the packet has been received by the
3641 * peer, not that it will be retained long enough to be sent to
3642 * the peer's upper level. In addition, reset the transmit timers
3643 * of any missing packets (those packets that must be missing
3644 * because this packet was out of sequence) */
3646 call->nSoftAcked = 0;
3647 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3648 /* Update round trip time if the ack was stimulated on receipt
3650 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3651 #ifdef RX_ENABLE_LOCKS
3652 if (tp->header.seq >= first)
3653 #endif /* RX_ENABLE_LOCKS */
3654 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3656 && (tp->header.serial == serial || tp->firstSerial == serial))
3657 rxi_ComputePeerNetStats(call, tp, ap, np);
3659 /* Set the acknowledge flag per packet based on the
3660 * information in the ack packet. An acknowlegded packet can
3661 * be downgraded when the server has discarded a packet it
3662 * soacked previously, or when an ack packet is received
3663 * out of sequence. */
3664 if (tp->header.seq < first) {
3665 /* Implicit ack information */
3666 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3669 tp->flags |= RX_PKTFLAG_ACKED;
3670 } else if (tp->header.seq < first + nAcks) {
3671 /* Explicit ack information: set it in the packet appropriately */
3672 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3673 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3675 tp->flags |= RX_PKTFLAG_ACKED;
3682 } else /* RX_ACK_TYPE_NACK */ {
3683 tp->flags &= ~RX_PKTFLAG_ACKED;
3687 tp->flags &= ~RX_PKTFLAG_ACKED;
3691 /* If packet isn't yet acked, and it has been transmitted at least
3692 * once, reset retransmit time using latest timeout
3693 * ie, this should readjust the retransmit timer for all outstanding
3694 * packets... So we don't just retransmit when we should know better*/
3696 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3697 tp->retryTime = tp->timeSent;
3698 clock_Add(&tp->retryTime, &peer->timeout);
3699 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3700 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3704 /* If the window has been extended by this acknowledge packet,
3705 * then wakeup a sender waiting in alloc for window space, or try
3706 * sending packets now, if he's been sitting on packets due to
3707 * lack of window space */
3708 if (call->tnext < (call->tfirst + call->twind)) {
3709 #ifdef RX_ENABLE_LOCKS
3710 CV_SIGNAL(&call->cv_twind);
3712 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3713 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3714 osi_rxWakeup(&call->twind);
3717 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3718 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3722 /* if the ack packet has a receivelen field hanging off it,
3723 * update our state */
3724 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3727 /* If the ack packet has a "recommended" size that is less than
3728 * what I am using now, reduce my size to match */
3729 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3730 (int)sizeof(afs_int32), &tSize);
3731 tSize = (afs_uint32) ntohl(tSize);
3732 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3734 /* Get the maximum packet size to send to this peer */
3735 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
3737 tSize = (afs_uint32) ntohl(tSize);
3738 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3739 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3741 /* sanity check - peer might have restarted with different params.
3742 * If peer says "send less", dammit, send less... Peer should never
3743 * be unable to accept packets of the size that prior AFS versions would
3744 * send without asking. */
3745 if (peer->maxMTU != tSize) {
3746 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
3748 peer->maxMTU = tSize;
3749 peer->MTU = MIN(tSize, peer->MTU);
3750 call->MTU = MIN(call->MTU, tSize);
3753 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3756 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3757 (int)sizeof(afs_int32), &tSize);
3758 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3759 if (tSize < call->twind) { /* smaller than our send */
3760 call->twind = tSize; /* window, we must send less... */
3761 call->ssthresh = MIN(call->twind, call->ssthresh);
3764 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3765 * network MTU confused with the loopback MTU. Calculate the
3766 * maximum MTU here for use in the slow start code below.
3768 maxMTU = peer->maxMTU;
3769 /* Did peer restart with older RX version? */
3770 if (peer->maxDgramPackets > 1) {
3771 peer->maxDgramPackets = 1;
3773 } else if (np->length >=
3774 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3777 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3778 sizeof(afs_int32), &tSize);
3779 tSize = (afs_uint32) ntohl(tSize);
3781 * As of AFS 3.5 we set the send window to match the receive window.
3783 if (tSize < call->twind) {
3784 call->twind = tSize;
3785 call->ssthresh = MIN(call->twind, call->ssthresh);
3786 } else if (tSize > call->twind) {
3787 call->twind = tSize;
3791 * As of AFS 3.5, a jumbogram is more than one fixed size
3792 * packet transmitted in a single UDP datagram. If the remote
3793 * MTU is smaller than our local MTU then never send a datagram
3794 * larger than the natural MTU.
3797 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3798 sizeof(afs_int32), &tSize);
3799 maxDgramPackets = (afs_uint32) ntohl(tSize);
3800 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3802 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
3803 maxDgramPackets = MIN(maxDgramPackets, tSize);
3804 if (maxDgramPackets > 1) {
3805 peer->maxDgramPackets = maxDgramPackets;
3806 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3808 peer->maxDgramPackets = 1;
3809 call->MTU = peer->natMTU;
3811 } else if (peer->maxDgramPackets > 1) {
3812 /* Restarted with lower version of RX */
3813 peer->maxDgramPackets = 1;
3815 } else if (peer->maxDgramPackets > 1
3816 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3817 /* Restarted with lower version of RX */
3818 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3819 peer->natMTU = OLD_MAX_PACKET_SIZE;
3820 peer->MTU = OLD_MAX_PACKET_SIZE;
3821 peer->maxDgramPackets = 1;
3822 peer->nDgramPackets = 1;
3824 call->MTU = OLD_MAX_PACKET_SIZE;
3829 * Calculate how many datagrams were successfully received after
3830 * the first missing packet and adjust the negative ack counter
3835 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3836 if (call->nNacks < nNacked) {
3837 call->nNacks = nNacked;
3846 if (call->flags & RX_CALL_FAST_RECOVER) {
3848 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3850 call->flags &= ~RX_CALL_FAST_RECOVER;
3851 call->cwind = call->nextCwind;
3852 call->nextCwind = 0;
3855 call->nCwindAcks = 0;
3856 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3857 /* Three negative acks in a row trigger congestion recovery */
3858 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3859 MUTEX_EXIT(&peer->peer_lock);
3860 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3861 /* someone else is waiting to start recovery */
3864 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3865 rxi_WaitforTQBusy(call);
3866 MUTEX_ENTER(&peer->peer_lock);
3867 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3868 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3869 call->flags |= RX_CALL_FAST_RECOVER;
3870 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3872 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3873 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3874 call->nextCwind = call->ssthresh;
3877 peer->MTU = call->MTU;
3878 peer->cwind = call->nextCwind;
3879 peer->nDgramPackets = call->nDgramPackets;
3881 call->congestSeq = peer->congestSeq;
3882 /* Reset the resend times on the packets that were nacked
3883 * so we will retransmit as soon as the window permits*/
3884 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3886 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3887 clock_Zero(&tp->retryTime);
3889 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3894 /* If cwind is smaller than ssthresh, then increase
3895 * the window one packet for each ack we receive (exponential
3897 * If cwind is greater than or equal to ssthresh then increase
3898 * the congestion window by one packet for each cwind acks we
3899 * receive (linear growth). */
3900 if (call->cwind < call->ssthresh) {
3902 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3903 call->nCwindAcks = 0;
3905 call->nCwindAcks += newAckCount;
3906 if (call->nCwindAcks >= call->cwind) {
3907 call->nCwindAcks = 0;
3908 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3912 * If we have received several acknowledgements in a row then
3913 * it is time to increase the size of our datagrams
3915 if ((int)call->nAcks > rx_nDgramThreshold) {
3916 if (peer->maxDgramPackets > 1) {
3917 if (call->nDgramPackets < peer->maxDgramPackets) {
3918 call->nDgramPackets++;
3920 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3921 } else if (call->MTU < peer->maxMTU) {
3922 call->MTU += peer->natMTU;
3923 call->MTU = MIN(call->MTU, peer->maxMTU);
3929 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3931 /* Servers need to hold the call until all response packets have
3932 * been acknowledged. Soft acks are good enough since clients
3933 * are not allowed to clear their receive queues. */
3934 if (call->state == RX_STATE_HOLD
3935 && call->tfirst + call->nSoftAcked >= call->tnext) {
3936 call->state = RX_STATE_DALLY;
3937 rxi_ClearTransmitQueue(call, 0);
3938 } else if (!queue_IsEmpty(&call->tq)) {
3939 rxi_Start(0, call, 0, istack);
3944 /* Received a response to a challenge packet */
3946 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
3947 register struct rx_packet *np, int istack)
3951 /* Ignore the packet if we're the client */
3952 if (conn->type == RX_CLIENT_CONNECTION)
3955 /* If already authenticated, ignore the packet (it's probably a retry) */
3956 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
3959 /* Otherwise, have the security object evaluate the response packet */
3960 error = RXS_CheckResponse(conn->securityObject, conn, np);
3962 /* If the response is invalid, reset the connection, sending
3963 * an abort to the peer */
3967 rxi_ConnectionError(conn, error);
3968 MUTEX_ENTER(&conn->conn_data_lock);
3969 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3970 MUTEX_EXIT(&conn->conn_data_lock);
3973 /* If the response is valid, any calls waiting to attach
3974 * servers can now do so */
3977 for (i = 0; i < RX_MAXCALLS; i++) {
3978 struct rx_call *call = conn->call[i];
3980 MUTEX_ENTER(&call->lock);
3981 if (call->state == RX_STATE_PRECALL)
3982 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
3983 /* tnop can be null if newcallp is null */
3984 MUTEX_EXIT(&call->lock);
3988 /* Update the peer reachability information, just in case
3989 * some calls went into attach-wait while we were waiting
3990 * for authentication..
3992 rxi_UpdatePeerReach(conn, NULL);
3997 /* A client has received an authentication challenge: the security
3998 * object is asked to cough up a respectable response packet to send
3999 * back to the server. The server is responsible for retrying the
4000 * challenge if it fails to get a response. */
4003 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
4004 register struct rx_packet *np, int istack)
4008 /* Ignore the challenge if we're the server */
4009 if (conn->type == RX_SERVER_CONNECTION)
4012 /* Ignore the challenge if the connection is otherwise idle; someone's
4013 * trying to use us as an oracle. */
4014 if (!rxi_HasActiveCalls(conn))
4017 /* Send the security object the challenge packet. It is expected to fill
4018 * in the response. */
4019 error = RXS_GetResponse(conn->securityObject, conn, np);
4021 /* If the security object is unable to return a valid response, reset the
4022 * connection and send an abort to the peer. Otherwise send the response
4023 * packet to the peer connection. */
4025 rxi_ConnectionError(conn, error);
4026 MUTEX_ENTER(&conn->conn_data_lock);
4027 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4028 MUTEX_EXIT(&conn->conn_data_lock);
4030 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4031 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4037 /* Find an available server process to service the current request in
4038 * the given call structure. If one isn't available, queue up this
4039 * call so it eventually gets one */
4041 rxi_AttachServerProc(register struct rx_call *call,
4042 register osi_socket socket, register int *tnop,
4043 register struct rx_call **newcallp)
4045 register struct rx_serverQueueEntry *sq;
4046 register struct rx_service *service = call->conn->service;
4047 register int haveQuota = 0;
4049 /* May already be attached */
4050 if (call->state == RX_STATE_ACTIVE)
4053 MUTEX_ENTER(&rx_serverPool_lock);
4055 haveQuota = QuotaOK(service);
4056 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4057 /* If there are no processes available to service this call,
4058 * put the call on the incoming call queue (unless it's
4059 * already on the queue).
4061 #ifdef RX_ENABLE_LOCKS
4063 ReturnToServerPool(service);
4064 #endif /* RX_ENABLE_LOCKS */
4066 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4067 call->flags |= RX_CALL_WAIT_PROC;
4068 MUTEX_ENTER(&rx_stats_mutex);
4071 MUTEX_EXIT(&rx_stats_mutex);
4072 rxi_calltrace(RX_CALL_ARRIVAL, call);
4073 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4074 queue_Append(&rx_incomingCallQueue, call);
4077 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4079 /* If hot threads are enabled, and both newcallp and sq->socketp
4080 * are non-null, then this thread will process the call, and the
4081 * idle server thread will start listening on this threads socket.
4084 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4087 *sq->socketp = socket;
4088 clock_GetTime(&call->startTime);
4089 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4093 if (call->flags & RX_CALL_WAIT_PROC) {
4094 /* Conservative: I don't think this should happen */
4095 call->flags &= ~RX_CALL_WAIT_PROC;
4096 if (queue_IsOnQueue(call)) {
4098 MUTEX_ENTER(&rx_stats_mutex);
4100 MUTEX_EXIT(&rx_stats_mutex);
4103 call->state = RX_STATE_ACTIVE;
4104 call->mode = RX_MODE_RECEIVING;
4105 #ifdef RX_KERNEL_TRACE
4107 int glockOwner = ISAFS_GLOCK();
4110 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4111 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4117 if (call->flags & RX_CALL_CLEARED) {
4118 /* send an ack now to start the packet flow up again */
4119 call->flags &= ~RX_CALL_CLEARED;
4120 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4122 #ifdef RX_ENABLE_LOCKS
4125 service->nRequestsRunning++;
4126 if (service->nRequestsRunning <= service->minProcs)
4132 MUTEX_EXIT(&rx_serverPool_lock);
4135 /* Delay the sending of an acknowledge event for a short while, while
4136 * a new call is being prepared (in the case of a client) or a reply
4137 * is being prepared (in the case of a server). Rather than sending
4138 * an ack packet, an ACKALL packet is sent. */
4140 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4142 #ifdef RX_ENABLE_LOCKS
4144 MUTEX_ENTER(&call->lock);
4145 call->delayedAckEvent = NULL;
4146 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4148 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4149 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4151 MUTEX_EXIT(&call->lock);
4152 #else /* RX_ENABLE_LOCKS */
4154 call->delayedAckEvent = NULL;
4155 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4156 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4157 #endif /* RX_ENABLE_LOCKS */
4161 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4164 #ifdef RX_ENABLE_LOCKS
4166 MUTEX_ENTER(&call->lock);
4167 if (event == call->delayedAckEvent)
4168 call->delayedAckEvent = NULL;
4169 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4171 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4173 MUTEX_EXIT(&call->lock);
4174 #else /* RX_ENABLE_LOCKS */
4176 call->delayedAckEvent = NULL;
4177 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4178 #endif /* RX_ENABLE_LOCKS */
4182 #ifdef RX_ENABLE_LOCKS
4183 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4184 * clearing them out.
4187 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4189 register struct rx_packet *p, *tp;
4192 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4193 p->flags |= RX_PKTFLAG_ACKED;
4197 call->flags |= RX_CALL_TQ_CLEARME;
4198 call->flags |= RX_CALL_TQ_SOME_ACKED;
4201 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4202 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4203 call->tfirst = call->tnext;
4204 call->nSoftAcked = 0;
4206 if (call->flags & RX_CALL_FAST_RECOVER) {
4207 call->flags &= ~RX_CALL_FAST_RECOVER;
4208 call->cwind = call->nextCwind;
4209 call->nextCwind = 0;
4212 CV_SIGNAL(&call->cv_twind);
4214 #endif /* RX_ENABLE_LOCKS */
4216 /* Clear out the transmit queue for the current call (all packets have
4217 * been received by peer) */
4219 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4221 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4222 register struct rx_packet *p, *tp;
4224 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4226 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4227 p->flags |= RX_PKTFLAG_ACKED;
4231 call->flags |= RX_CALL_TQ_CLEARME;
4232 call->flags |= RX_CALL_TQ_SOME_ACKED;
4235 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4236 rxi_FreePackets(0, &call->tq);
4237 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4238 call->flags &= ~RX_CALL_TQ_CLEARME;
4240 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4242 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4243 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4244 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4245 call->nSoftAcked = 0;
4247 if (call->flags & RX_CALL_FAST_RECOVER) {
4248 call->flags &= ~RX_CALL_FAST_RECOVER;
4249 call->cwind = call->nextCwind;
4251 #ifdef RX_ENABLE_LOCKS
4252 CV_SIGNAL(&call->cv_twind);
4254 osi_rxWakeup(&call->twind);
4259 rxi_ClearReceiveQueue(register struct rx_call *call)
4261 if (queue_IsNotEmpty(&call->rq)) {
4262 rx_packetReclaims += rxi_FreePackets(0, &call->rq);
4263 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4265 if (call->state == RX_STATE_PRECALL) {
4266 call->flags |= RX_CALL_CLEARED;
4270 /* Send an abort packet for the specified call */
4272 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4273 int istack, int force)
4276 struct clock when, now;
4281 /* Clients should never delay abort messages */
4282 if (rx_IsClientConn(call->conn))
4285 if (call->abortCode != call->error) {
4286 call->abortCode = call->error;
4287 call->abortCount = 0;
4290 if (force || rxi_callAbortThreshhold == 0
4291 || call->abortCount < rxi_callAbortThreshhold) {
4292 if (call->delayedAbortEvent) {
4293 rxevent_Cancel(call->delayedAbortEvent, call,
4294 RX_CALL_REFCOUNT_ABORT);
4296 error = htonl(call->error);
4299 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4300 (char *)&error, sizeof(error), istack);
4301 } else if (!call->delayedAbortEvent) {
4302 clock_GetTime(&now);
4304 clock_Addmsec(&when, rxi_callAbortDelay);
4305 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4306 call->delayedAbortEvent =
4307 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4312 /* Send an abort packet for the specified connection. Packet is an
4313 * optional pointer to a packet that can be used to send the abort.
4314 * Once the number of abort messages reaches the threshhold, an
4315 * event is scheduled to send the abort. Setting the force flag
4316 * overrides sending delayed abort messages.
4318 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4319 * to send the abort packet.
4322 rxi_SendConnectionAbort(register struct rx_connection *conn,
4323 struct rx_packet *packet, int istack, int force)
4326 struct clock when, now;
4331 /* Clients should never delay abort messages */
4332 if (rx_IsClientConn(conn))
4335 if (force || rxi_connAbortThreshhold == 0
4336 || conn->abortCount < rxi_connAbortThreshhold) {
4337 if (conn->delayedAbortEvent) {
4338 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4340 error = htonl(conn->error);
4342 MUTEX_EXIT(&conn->conn_data_lock);
4344 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4345 RX_PACKET_TYPE_ABORT, (char *)&error,
4346 sizeof(error), istack);
4347 MUTEX_ENTER(&conn->conn_data_lock);
4348 } else if (!conn->delayedAbortEvent) {
4349 clock_GetTime(&now);
4351 clock_Addmsec(&when, rxi_connAbortDelay);
4352 conn->delayedAbortEvent =
4353 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4358 /* Associate an error all of the calls owned by a connection. Called
4359 * with error non-zero. This is only for really fatal things, like
4360 * bad authentication responses. The connection itself is set in
4361 * error at this point, so that future packets received will be
4364 rxi_ConnectionError(register struct rx_connection *conn,
4365 register afs_int32 error)
4370 dpf(("rxi_ConnectionError conn %x error %d", conn, error));
4372 MUTEX_ENTER(&conn->conn_data_lock);
4373 if (conn->challengeEvent)
4374 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4375 if (conn->checkReachEvent) {
4376 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4377 conn->checkReachEvent = 0;
4378 conn->flags &= ~RX_CONN_ATTACHWAIT;
4381 MUTEX_EXIT(&conn->conn_data_lock);
4382 for (i = 0; i < RX_MAXCALLS; i++) {
4383 struct rx_call *call = conn->call[i];
4385 MUTEX_ENTER(&call->lock);
4386 rxi_CallError(call, error);
4387 MUTEX_EXIT(&call->lock);
4390 conn->error = error;
4391 rx_MutexIncrement(rx_stats.fatalErrors, rx_stats_mutex);
4396 rxi_CallError(register struct rx_call *call, afs_int32 error)
4398 dpf(("rxi_CallError call %x error %d call->error %d", call, error, call->error));
4400 error = call->error;
4402 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4403 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4404 rxi_ResetCall(call, 0);
4407 rxi_ResetCall(call, 0);
4409 call->error = error;
4410 call->mode = RX_MODE_ERROR;
4413 /* Reset various fields in a call structure, and wakeup waiting
4414 * processes. Some fields aren't changed: state & mode are not
4415 * touched (these must be set by the caller), and bufptr, nLeft, and
4416 * nFree are not reset, since these fields are manipulated by
4417 * unprotected macros, and may only be reset by non-interrupting code.
4420 /* this code requires that call->conn be set properly as a pre-condition. */
4421 #endif /* ADAPT_WINDOW */
4424 rxi_ResetCall(register struct rx_call *call, register int newcall)
4427 register struct rx_peer *peer;
4428 struct rx_packet *packet;
4430 dpf(("rxi_ResetCall(call %x, newcall %d)\n", call, newcall));
4432 /* Notify anyone who is waiting for asynchronous packet arrival */
4433 if (call->arrivalProc) {
4434 (*call->arrivalProc) (call, call->arrivalProcHandle,
4435 call->arrivalProcArg);
4436 call->arrivalProc = (void (*)())0;
4439 if (call->delayedAbortEvent) {
4440 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4441 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4443 rxi_SendCallAbort(call, packet, 0, 1);
4444 rxi_FreePacket(packet);
4449 * Update the peer with the congestion information in this call
4450 * so other calls on this connection can pick up where this call
4451 * left off. If the congestion sequence numbers don't match then
4452 * another call experienced a retransmission.
4454 peer = call->conn->peer;
4455 MUTEX_ENTER(&peer->peer_lock);
4457 if (call->congestSeq == peer->congestSeq) {
4458 peer->cwind = MAX(peer->cwind, call->cwind);
4459 peer->MTU = MAX(peer->MTU, call->MTU);
4460 peer->nDgramPackets =
4461 MAX(peer->nDgramPackets, call->nDgramPackets);
4464 call->abortCode = 0;
4465 call->abortCount = 0;
4467 if (peer->maxDgramPackets > 1) {
4468 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4470 call->MTU = peer->MTU;
4472 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4473 call->ssthresh = rx_maxSendWindow;
4474 call->nDgramPackets = peer->nDgramPackets;
4475 call->congestSeq = peer->congestSeq;
4476 MUTEX_EXIT(&peer->peer_lock);
4478 flags = call->flags;
4479 rxi_ClearReceiveQueue(call);
4480 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4481 if (flags & RX_CALL_TQ_BUSY) {
4482 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4483 call->flags |= (flags & RX_CALL_TQ_WAIT);
4485 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4487 rxi_ClearTransmitQueue(call, 0);
4488 queue_Init(&call->tq);
4489 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4490 dpf(("rcall %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4493 while (call->tqWaiters) {
4494 #ifdef RX_ENABLE_LOCKS
4495 CV_BROADCAST(&call->cv_tq);
4496 #else /* RX_ENABLE_LOCKS */
4497 osi_rxWakeup(&call->tq);
4498 #endif /* RX_ENABLE_LOCKS */
4502 queue_Init(&call->rq);
4504 call->rwind = rx_initReceiveWindow;
4505 call->twind = rx_initSendWindow;
4506 call->nSoftAcked = 0;
4507 call->nextCwind = 0;
4510 call->nCwindAcks = 0;
4511 call->nSoftAcks = 0;
4512 call->nHardAcks = 0;
4514 call->tfirst = call->rnext = call->tnext = 1;
4516 call->lastAcked = 0;
4517 call->localStatus = call->remoteStatus = 0;
4519 if (flags & RX_CALL_READER_WAIT) {
4520 #ifdef RX_ENABLE_LOCKS
4521 CV_BROADCAST(&call->cv_rq);
4523 osi_rxWakeup(&call->rq);
4526 if (flags & RX_CALL_WAIT_PACKETS) {
4527 MUTEX_ENTER(&rx_freePktQ_lock);
4528 rxi_PacketsUnWait(); /* XXX */
4529 MUTEX_EXIT(&rx_freePktQ_lock);
4531 #ifdef RX_ENABLE_LOCKS
4532 CV_SIGNAL(&call->cv_twind);
4534 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4535 osi_rxWakeup(&call->twind);
4538 #ifdef RX_ENABLE_LOCKS
4539 /* The following ensures that we don't mess with any queue while some
4540 * other thread might also be doing so. The call_queue_lock field is
4541 * is only modified under the call lock. If the call is in the process
4542 * of being removed from a queue, the call is not locked until the
4543 * the queue lock is dropped and only then is the call_queue_lock field
4544 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4545 * Note that any other routine which removes a call from a queue has to
4546 * obtain the queue lock before examing the queue and removing the call.
4548 if (call->call_queue_lock) {
4549 MUTEX_ENTER(call->call_queue_lock);
4550 if (queue_IsOnQueue(call)) {
4552 if (flags & RX_CALL_WAIT_PROC) {
4553 MUTEX_ENTER(&rx_stats_mutex);
4555 MUTEX_EXIT(&rx_stats_mutex);
4558 MUTEX_EXIT(call->call_queue_lock);
4559 CLEAR_CALL_QUEUE_LOCK(call);
4561 #else /* RX_ENABLE_LOCKS */
4562 if (queue_IsOnQueue(call)) {
4564 if (flags & RX_CALL_WAIT_PROC)
4567 #endif /* RX_ENABLE_LOCKS */
4569 rxi_KeepAliveOff(call);
4570 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4573 /* Send an acknowledge for the indicated packet (seq,serial) of the
4574 * indicated call, for the indicated reason (reason). This
4575 * acknowledge will specifically acknowledge receiving the packet, and
4576 * will also specify which other packets for this call have been
4577 * received. This routine returns the packet that was used to the
4578 * caller. The caller is responsible for freeing it or re-using it.
4579 * This acknowledgement also returns the highest sequence number
4580 * actually read out by the higher level to the sender; the sender
4581 * promises to keep around packets that have not been read by the
4582 * higher level yet (unless, of course, the sender decides to abort
4583 * the call altogether). Any of p, seq, serial, pflags, or reason may
4584 * be set to zero without ill effect. That is, if they are zero, they
4585 * will not convey any information.
4586 * NOW there is a trailer field, after the ack where it will safely be
4587 * ignored by mundanes, which indicates the maximum size packet this
4588 * host can swallow. */
4590 register struct rx_packet *optionalPacket; use to send ack (or null)
4591 int seq; Sequence number of the packet we are acking
4592 int serial; Serial number of the packet
4593 int pflags; Flags field from packet header
4594 int reason; Reason an acknowledge was prompted
4598 rxi_SendAck(register struct rx_call *call,
4599 register struct rx_packet *optionalPacket, int serial, int reason,
4602 struct rx_ackPacket *ap;
4603 register struct rx_packet *rqp;
4604 register struct rx_packet *nxp; /* For queue_Scan */
4605 register struct rx_packet *p;
4608 #ifdef RX_ENABLE_TSFPQ
4609 struct rx_ts_info_t * rx_ts_info;
4613 * Open the receive window once a thread starts reading packets
4615 if (call->rnext > 1) {
4616 call->rwind = rx_maxReceiveWindow;
4619 call->nHardAcks = 0;
4620 call->nSoftAcks = 0;
4621 if (call->rnext > call->lastAcked)
4622 call->lastAcked = call->rnext;
4626 rx_computelen(p, p->length); /* reset length, you never know */
4627 } /* where that's been... */
4628 #ifdef RX_ENABLE_TSFPQ
4630 RX_TS_INFO_GET(rx_ts_info);
4631 if ((p = rx_ts_info->local_special_packet)) {
4632 rx_computelen(p, p->length);
4633 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4634 rx_ts_info->local_special_packet = p;
4635 } else { /* We won't send the ack, but don't panic. */
4636 return optionalPacket;
4640 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4641 /* We won't send the ack, but don't panic. */
4642 return optionalPacket;
4647 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4650 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4651 #ifndef RX_ENABLE_TSFPQ
4652 if (!optionalPacket)
4655 return optionalPacket;
4657 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4658 if (rx_Contiguous(p) < templ) {
4659 #ifndef RX_ENABLE_TSFPQ
4660 if (!optionalPacket)
4663 return optionalPacket;
4668 /* MTUXXX failing to send an ack is very serious. We should */
4669 /* try as hard as possible to send even a partial ack; it's */
4670 /* better than nothing. */
4671 ap = (struct rx_ackPacket *)rx_DataOf(p);
4672 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4673 ap->reason = reason;
4675 /* The skew computation used to be bogus, I think it's better now. */
4676 /* We should start paying attention to skew. XXX */
4677 ap->serial = htonl(serial);
4678 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4680 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4681 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4683 /* No fear of running out of ack packet here because there can only be at most
4684 * one window full of unacknowledged packets. The window size must be constrained
4685 * to be less than the maximum ack size, of course. Also, an ack should always
4686 * fit into a single packet -- it should not ever be fragmented. */
4687 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4688 if (!rqp || !call->rq.next
4689 || (rqp->header.seq > (call->rnext + call->rwind))) {
4690 #ifndef RX_ENABLE_TSFPQ
4691 if (!optionalPacket)
4694 rxi_CallError(call, RX_CALL_DEAD);
4695 return optionalPacket;
4698 while (rqp->header.seq > call->rnext + offset)
4699 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4700 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4702 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4703 #ifndef RX_ENABLE_TSFPQ
4704 if (!optionalPacket)
4707 rxi_CallError(call, RX_CALL_DEAD);
4708 return optionalPacket;
4713 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4715 /* these are new for AFS 3.3 */
4716 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4717 templ = htonl(templ);
4718 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4719 templ = htonl(call->conn->peer->ifMTU);
4720 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4721 sizeof(afs_int32), &templ);
4723 /* new for AFS 3.4 */
4724 templ = htonl(call->rwind);
4725 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4726 sizeof(afs_int32), &templ);
4728 /* new for AFS 3.5 */
4729 templ = htonl(call->conn->peer->ifDgramPackets);
4730 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4731 sizeof(afs_int32), &templ);
4733 p->header.serviceId = call->conn->serviceId;
4734 p->header.cid = (call->conn->cid | call->channel);
4735 p->header.callNumber = *call->callNumber;
4737 p->header.securityIndex = call->conn->securityIndex;
4738 p->header.epoch = call->conn->epoch;
4739 p->header.type = RX_PACKET_TYPE_ACK;
4740 p->header.flags = RX_SLOW_START_OK;
4741 if (reason == RX_ACK_PING) {
4742 p->header.flags |= RX_REQUEST_ACK;
4744 clock_GetTime(&call->pingRequestTime);
4747 if (call->conn->type == RX_CLIENT_CONNECTION)
4748 p->header.flags |= RX_CLIENT_INITIATED;
4752 if (rxdebug_active) {
4756 len = _snprintf(msg, sizeof(msg),
4757 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4758 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4759 ntohl(ap->serial), ntohl(ap->previousPacket),
4760 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
4761 ap->nAcks, ntohs(ap->bufferSpace) );
4765 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
4766 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4770 OutputDebugString(msg);
4772 #else /* AFS_NT40_ENV */
4774 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
4775 ap->reason, ntohl(ap->previousPacket),
4776 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4778 for (offset = 0; offset < ap->nAcks; offset++)
4779 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4784 #endif /* AFS_NT40_ENV */
4787 register int i, nbytes = p->length;
4789 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4790 if (nbytes <= p->wirevec[i].iov_len) {
4791 register int savelen, saven;
4793 savelen = p->wirevec[i].iov_len;
4795 p->wirevec[i].iov_len = nbytes;
4797 rxi_Send(call, p, istack);
4798 p->wirevec[i].iov_len = savelen;
4802 nbytes -= p->wirevec[i].iov_len;
4805 rx_MutexIncrement(rx_stats.ackPacketsSent, rx_stats_mutex);
4806 #ifndef RX_ENABLE_TSFPQ
4807 if (!optionalPacket)
4810 return optionalPacket; /* Return packet for re-use by caller */
4813 /* Send all of the packets in the list in single datagram */
4815 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4816 int istack, int moreFlag, struct clock *now,
4817 struct clock *retryTime, int resending)
4822 struct rx_connection *conn = call->conn;
4823 struct rx_peer *peer = conn->peer;
4825 MUTEX_ENTER(&peer->peer_lock);
4828 peer->reSends += len;
4829 rx_MutexIncrement(rx_stats.dataPacketsSent, rx_stats_mutex);
4830 MUTEX_EXIT(&peer->peer_lock);
4832 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4836 /* Set the packet flags and schedule the resend events */
4837 /* Only request an ack for the last packet in the list */
4838 for (i = 0; i < len; i++) {
4839 list[i]->retryTime = *retryTime;
4840 if (list[i]->header.serial) {
4841 /* Exponentially backoff retry times */
4842 if (list[i]->backoff < MAXBACKOFF) {
4843 /* so it can't stay == 0 */
4844 list[i]->backoff = (list[i]->backoff << 1) + 1;
4847 clock_Addmsec(&(list[i]->retryTime),
4848 ((afs_uint32) list[i]->backoff) << 8);
4851 /* Wait a little extra for the ack on the last packet */
4852 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4853 clock_Addmsec(&(list[i]->retryTime), 400);
4856 /* Record the time sent */
4857 list[i]->timeSent = *now;
4859 /* Ask for an ack on retransmitted packets, on every other packet
4860 * if the peer doesn't support slow start. Ask for an ack on every
4861 * packet until the congestion window reaches the ack rate. */
4862 if (list[i]->header.serial) {
4864 rx_MutexIncrement(rx_stats.dataPacketsReSent, rx_stats_mutex);
4866 /* improved RTO calculation- not Karn */
4867 list[i]->firstSent = *now;
4868 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4869 || (!(call->flags & RX_CALL_SLOW_START_OK)
4870 && (list[i]->header.seq & 1)))) {
4875 MUTEX_ENTER(&peer->peer_lock);
4879 rx_MutexIncrement(rx_stats.dataPacketsSent, rx_stats_mutex);
4880 MUTEX_EXIT(&peer->peer_lock);
4882 /* Tag this packet as not being the last in this group,
4883 * for the receiver's benefit */
4884 if (i < len - 1 || moreFlag) {
4885 list[i]->header.flags |= RX_MORE_PACKETS;
4888 /* Install the new retransmit time for the packet, and
4889 * record the time sent */
4890 list[i]->timeSent = *now;
4894 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4897 /* Since we're about to send a data packet to the peer, it's
4898 * safe to nuke any scheduled end-of-packets ack */
4899 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4901 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4902 MUTEX_EXIT(&call->lock);
4904 rxi_SendPacketList(call, conn, list, len, istack);
4906 rxi_SendPacket(call, conn, list[0], istack);
4908 MUTEX_ENTER(&call->lock);
4909 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4911 /* Update last send time for this call (for keep-alive
4912 * processing), and for the connection (so that we can discover
4913 * idle connections) */
4914 conn->lastSendTime = call->lastSendTime = clock_Sec();
4917 /* When sending packets we need to follow these rules:
4918 * 1. Never send more than maxDgramPackets in a jumbogram.
4919 * 2. Never send a packet with more than two iovecs in a jumbogram.
4920 * 3. Never send a retransmitted packet in a jumbogram.
4921 * 4. Never send more than cwind/4 packets in a jumbogram
4922 * We always keep the last list we should have sent so we
4923 * can set the RX_MORE_PACKETS flags correctly.
4926 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4927 int istack, struct clock *now, struct clock *retryTime,
4930 int i, cnt, lastCnt = 0;
4931 struct rx_packet **listP, **lastP = 0;
4932 struct rx_peer *peer = call->conn->peer;
4933 int morePackets = 0;
4935 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4936 /* Does the current packet force us to flush the current list? */
4938 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4939 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4941 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
4943 /* If the call enters an error state stop sending, or if
4944 * we entered congestion recovery mode, stop sending */
4945 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4953 /* Add the current packet to the list if it hasn't been acked.
4954 * Otherwise adjust the list pointer to skip the current packet. */
4955 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
4957 /* Do we need to flush the list? */
4958 if (cnt >= (int)peer->maxDgramPackets
4959 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
4960 || list[i]->header.serial
4961 || list[i]->length != RX_JUMBOBUFFERSIZE) {
4963 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
4964 retryTime, resending);
4965 /* If the call enters an error state stop sending, or if
4966 * we entered congestion recovery mode, stop sending */
4968 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4973 listP = &list[i + 1];
4978 osi_Panic("rxi_SendList error");
4980 listP = &list[i + 1];
4984 /* Send the whole list when the call is in receive mode, when
4985 * the call is in eof mode, when we are in fast recovery mode,
4986 * and when we have the last packet */
4987 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
4988 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
4989 || (call->flags & RX_CALL_FAST_RECOVER)) {
4990 /* Check for the case where the current list contains
4991 * an acked packet. Since we always send retransmissions
4992 * in a separate packet, we only need to check the first
4993 * packet in the list */
4994 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
4998 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
4999 retryTime, resending);
5000 /* If the call enters an error state stop sending, or if
5001 * we entered congestion recovery mode, stop sending */
5002 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5006 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5009 } else if (lastCnt > 0) {
5010 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5015 #ifdef RX_ENABLE_LOCKS
5016 /* Call rxi_Start, below, but with the call lock held. */
5018 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
5019 void *arg1, int istack)
5021 MUTEX_ENTER(&call->lock);
5022 rxi_Start(event, call, arg1, istack);
5023 MUTEX_EXIT(&call->lock);
5025 #endif /* RX_ENABLE_LOCKS */
5027 /* This routine is called when new packets are readied for
5028 * transmission and when retransmission may be necessary, or when the
5029 * transmission window or burst count are favourable. This should be
5030 * better optimized for new packets, the usual case, now that we've
5031 * got rid of queues of send packets. XXXXXXXXXXX */
5033 rxi_Start(struct rxevent *event, register struct rx_call *call,
5034 void *arg1, int istack)
5036 struct rx_packet *p;
5037 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
5038 struct rx_peer *peer = call->conn->peer;
5039 struct clock now, usenow, retryTime;
5043 struct rx_packet **xmitList;
5046 /* If rxi_Start is being called as a result of a resend event,
5047 * then make sure that the event pointer is removed from the call
5048 * structure, since there is no longer a per-call retransmission
5050 if (event && event == call->resendEvent) {
5051 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5052 call->resendEvent = NULL;
5054 if (queue_IsEmpty(&call->tq)) {
5058 /* Timeouts trigger congestion recovery */
5059 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5060 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5061 /* someone else is waiting to start recovery */
5064 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5065 rxi_WaitforTQBusy(call);
5066 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5067 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5068 call->flags |= RX_CALL_FAST_RECOVER;
5069 if (peer->maxDgramPackets > 1) {
5070 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5072 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5074 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5075 call->nDgramPackets = 1;
5077 call->nextCwind = 1;
5080 MUTEX_ENTER(&peer->peer_lock);
5081 peer->MTU = call->MTU;
5082 peer->cwind = call->cwind;
5083 peer->nDgramPackets = 1;
5085 call->congestSeq = peer->congestSeq;
5086 MUTEX_EXIT(&peer->peer_lock);
5087 /* Clear retry times on packets. Otherwise, it's possible for
5088 * some packets in the queue to force resends at rates faster
5089 * than recovery rates.
5091 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5092 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5093 clock_Zero(&p->retryTime);
5098 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5099 rx_MutexIncrement(rx_tq_debug.rxi_start_in_error, rx_stats_mutex);
5104 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5105 /* Get clock to compute the re-transmit time for any packets
5106 * in this burst. Note, if we back off, it's reasonable to
5107 * back off all of the packets in the same manner, even if
5108 * some of them have been retransmitted more times than more
5110 * Do a dance to avoid blocking after setting now. */
5111 clock_Zero(&retryTime);
5112 MUTEX_ENTER(&peer->peer_lock);
5113 clock_Add(&retryTime, &peer->timeout);
5114 MUTEX_EXIT(&peer->peer_lock);
5115 clock_GetTime(&now);
5116 clock_Add(&retryTime, &now);
5118 /* Send (or resend) any packets that need it, subject to
5119 * window restrictions and congestion burst control
5120 * restrictions. Ask for an ack on the last packet sent in
5121 * this burst. For now, we're relying upon the window being
5122 * considerably bigger than the largest number of packets that
5123 * are typically sent at once by one initial call to
5124 * rxi_Start. This is probably bogus (perhaps we should ask
5125 * for an ack when we're half way through the current
5126 * window?). Also, for non file transfer applications, this
5127 * may end up asking for an ack for every packet. Bogus. XXXX
5130 * But check whether we're here recursively, and let the other guy
5133 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5134 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5135 call->flags |= RX_CALL_TQ_BUSY;
5137 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5139 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5140 call->flags &= ~RX_CALL_NEED_START;
5141 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5143 maxXmitPackets = MIN(call->twind, call->cwind);
5144 xmitList = (struct rx_packet **)
5145 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5146 if (xmitList == NULL)
5147 osi_Panic("rxi_Start, failed to allocate xmit list");
5148 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5149 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5150 /* We shouldn't be sending packets if a thread is waiting
5151 * to initiate congestion recovery */
5155 && (call->flags & RX_CALL_FAST_RECOVER)) {
5156 /* Only send one packet during fast recovery */
5159 if ((p->flags & RX_PKTFLAG_FREE)
5160 || (!queue_IsEnd(&call->tq, nxp)
5161 && (nxp->flags & RX_PKTFLAG_FREE))
5162 || (p == (struct rx_packet *)&rx_freePacketQueue)
5163 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5164 osi_Panic("rxi_Start: xmit queue clobbered");
5166 if (p->flags & RX_PKTFLAG_ACKED) {
5167 /* Since we may block, don't trust this */
5168 usenow.sec = usenow.usec = 0;
5169 rx_MutexIncrement(rx_stats.ignoreAckedPacket, rx_stats_mutex);
5170 continue; /* Ignore this packet if it has been acknowledged */
5173 /* Turn off all flags except these ones, which are the same
5174 * on each transmission */
5175 p->header.flags &= RX_PRESET_FLAGS;
5177 if (p->header.seq >=
5178 call->tfirst + MIN((int)call->twind,
5179 (int)(call->nSoftAcked +
5181 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5182 /* Note: if we're waiting for more window space, we can
5183 * still send retransmits; hence we don't return here, but
5184 * break out to schedule a retransmit event */
5185 dpf(("call %d waiting for window",
5186 *(call->callNumber)));
5190 /* Transmit the packet if it needs to be sent. */
5191 if (!clock_Lt(&now, &p->retryTime)) {
5192 if (nXmitPackets == maxXmitPackets) {
5193 rxi_SendXmitList(call, xmitList, nXmitPackets,
5194 istack, &now, &retryTime,
5196 osi_Free(xmitList, maxXmitPackets *
5197 sizeof(struct rx_packet *));
5200 xmitList[nXmitPackets++] = p;
5204 /* xmitList now hold pointers to all of the packets that are
5205 * ready to send. Now we loop to send the packets */
5206 if (nXmitPackets > 0) {
5207 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5208 &now, &retryTime, resending);
5211 maxXmitPackets * sizeof(struct rx_packet *));
5213 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5215 * TQ references no longer protected by this flag; they must remain
5216 * protected by the global lock.
5218 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5219 call->flags &= ~RX_CALL_TQ_BUSY;
5220 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5221 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5222 #ifdef RX_ENABLE_LOCKS
5223 osirx_AssertMine(&call->lock, "rxi_Start start");
5224 CV_BROADCAST(&call->cv_tq);
5225 #else /* RX_ENABLE_LOCKS */
5226 osi_rxWakeup(&call->tq);
5227 #endif /* RX_ENABLE_LOCKS */
5232 /* We went into the error state while sending packets. Now is
5233 * the time to reset the call. This will also inform the using
5234 * process that the call is in an error state.
5236 rx_MutexIncrement(rx_tq_debug.rxi_start_aborted, rx_stats_mutex);
5237 call->flags &= ~RX_CALL_TQ_BUSY;
5238 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5239 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5240 #ifdef RX_ENABLE_LOCKS
5241 osirx_AssertMine(&call->lock, "rxi_Start middle");
5242 CV_BROADCAST(&call->cv_tq);
5243 #else /* RX_ENABLE_LOCKS */
5244 osi_rxWakeup(&call->tq);
5245 #endif /* RX_ENABLE_LOCKS */
5247 rxi_CallError(call, call->error);
5250 #ifdef RX_ENABLE_LOCKS
5251 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5252 register int missing;
5253 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5254 /* Some packets have received acks. If they all have, we can clear
5255 * the transmit queue.
5258 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5259 if (p->header.seq < call->tfirst
5260 && (p->flags & RX_PKTFLAG_ACKED)) {
5267 call->flags |= RX_CALL_TQ_CLEARME;
5269 #endif /* RX_ENABLE_LOCKS */
5270 /* Don't bother doing retransmits if the TQ is cleared. */
5271 if (call->flags & RX_CALL_TQ_CLEARME) {
5272 rxi_ClearTransmitQueue(call, 1);
5274 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5277 /* Always post a resend event, if there is anything in the
5278 * queue, and resend is possible. There should be at least
5279 * one unacknowledged packet in the queue ... otherwise none
5280 * of these packets should be on the queue in the first place.
5282 if (call->resendEvent) {
5283 /* Cancel the existing event and post a new one */
5284 rxevent_Cancel(call->resendEvent, call,
5285 RX_CALL_REFCOUNT_RESEND);
5288 /* The retry time is the retry time on the first unacknowledged
5289 * packet inside the current window */
5291 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5292 /* Don't set timers for packets outside the window */
5293 if (p->header.seq >= call->tfirst + call->twind) {
5297 if (!(p->flags & RX_PKTFLAG_ACKED)
5298 && !clock_IsZero(&p->retryTime)) {
5300 retryTime = p->retryTime;
5305 /* Post a new event to re-run rxi_Start when retries may be needed */
5306 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5307 #ifdef RX_ENABLE_LOCKS
5308 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5310 rxevent_PostNow2(&retryTime, &usenow,
5312 (void *)call, 0, istack);
5313 #else /* RX_ENABLE_LOCKS */
5315 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5316 (void *)call, 0, istack);
5317 #endif /* RX_ENABLE_LOCKS */
5320 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5321 } while (call->flags & RX_CALL_NEED_START);
5323 * TQ references no longer protected by this flag; they must remain
5324 * protected by the global lock.
5326 call->flags &= ~RX_CALL_TQ_BUSY;
5327 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5328 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5329 #ifdef RX_ENABLE_LOCKS
5330 osirx_AssertMine(&call->lock, "rxi_Start end");
5331 CV_BROADCAST(&call->cv_tq);
5332 #else /* RX_ENABLE_LOCKS */
5333 osi_rxWakeup(&call->tq);
5334 #endif /* RX_ENABLE_LOCKS */
5337 call->flags |= RX_CALL_NEED_START;
5339 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5341 if (call->resendEvent) {
5342 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5347 /* Also adjusts the keep alive parameters for the call, to reflect
5348 * that we have just sent a packet (so keep alives aren't sent
5351 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5354 register struct rx_connection *conn = call->conn;
5356 /* Stamp each packet with the user supplied status */
5357 p->header.userStatus = call->localStatus;
5359 /* Allow the security object controlling this call's security to
5360 * make any last-minute changes to the packet */
5361 RXS_SendPacket(conn->securityObject, call, p);
5363 /* Since we're about to send SOME sort of packet to the peer, it's
5364 * safe to nuke any scheduled end-of-packets ack */
5365 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5367 /* Actually send the packet, filling in more connection-specific fields */
5368 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5369 MUTEX_EXIT(&call->lock);
5370 rxi_SendPacket(call, conn, p, istack);
5371 MUTEX_ENTER(&call->lock);
5372 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5374 /* Update last send time for this call (for keep-alive
5375 * processing), and for the connection (so that we can discover
5376 * idle connections) */
5377 conn->lastSendTime = call->lastSendTime = clock_Sec();
5381 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5382 * that things are fine. Also called periodically to guarantee that nothing
5383 * falls through the cracks (e.g. (error + dally) connections have keepalive
5384 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5386 * haveCTLock Set if calling from rxi_ReapConnections
5388 #ifdef RX_ENABLE_LOCKS
5390 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5391 #else /* RX_ENABLE_LOCKS */
5393 rxi_CheckCall(register struct rx_call *call)
5394 #endif /* RX_ENABLE_LOCKS */
5396 register struct rx_connection *conn = call->conn;
5398 afs_uint32 deadTime;
5400 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5401 if (call->flags & RX_CALL_TQ_BUSY) {
5402 /* Call is active and will be reset by rxi_Start if it's
5403 * in an error state.
5408 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5410 (((afs_uint32) conn->secondsUntilDead << 10) +
5411 ((afs_uint32) conn->peer->rtt >> 3) +
5412 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5414 /* These are computed to the second (+- 1 second). But that's
5415 * good enough for these values, which should be a significant
5416 * number of seconds. */
5417 if (now > (call->lastReceiveTime + deadTime)) {
5418 if (call->state == RX_STATE_ACTIVE) {
5419 rxi_CallError(call, RX_CALL_DEAD);
5422 #ifdef RX_ENABLE_LOCKS
5423 /* Cancel pending events */
5424 rxevent_Cancel(call->delayedAckEvent, call,
5425 RX_CALL_REFCOUNT_DELAY);
5426 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5427 rxevent_Cancel(call->keepAliveEvent, call,
5428 RX_CALL_REFCOUNT_ALIVE);
5429 if (call->refCount == 0) {
5430 rxi_FreeCall(call, haveCTLock);
5434 #else /* RX_ENABLE_LOCKS */
5437 #endif /* RX_ENABLE_LOCKS */
5439 /* Non-active calls are destroyed if they are not responding
5440 * to pings; active calls are simply flagged in error, so the
5441 * attached process can die reasonably gracefully. */
5443 /* see if we have a non-activity timeout */
5444 if (call->startWait && conn->idleDeadTime
5445 && ((call->startWait + conn->idleDeadTime) < now)) {
5446 if (call->state == RX_STATE_ACTIVE) {
5447 rxi_CallError(call, RX_CALL_TIMEOUT);
5451 /* see if we have a hard timeout */
5452 if (conn->hardDeadTime
5453 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5454 if (call->state == RX_STATE_ACTIVE)
5455 rxi_CallError(call, RX_CALL_TIMEOUT);
5462 /* When a call is in progress, this routine is called occasionally to
5463 * make sure that some traffic has arrived (or been sent to) the peer.
5464 * If nothing has arrived in a reasonable amount of time, the call is
5465 * declared dead; if nothing has been sent for a while, we send a
5466 * keep-alive packet (if we're actually trying to keep the call alive)
5469 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5472 struct rx_connection *conn;
5475 MUTEX_ENTER(&call->lock);
5476 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5477 if (event == call->keepAliveEvent)
5478 call->keepAliveEvent = NULL;
5481 #ifdef RX_ENABLE_LOCKS
5482 if (rxi_CheckCall(call, 0)) {
5483 MUTEX_EXIT(&call->lock);
5486 #else /* RX_ENABLE_LOCKS */
5487 if (rxi_CheckCall(call))
5489 #endif /* RX_ENABLE_LOCKS */
5491 /* Don't try to keep alive dallying calls */
5492 if (call->state == RX_STATE_DALLY) {
5493 MUTEX_EXIT(&call->lock);
5498 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5499 /* Don't try to send keepalives if there is unacknowledged data */
5500 /* the rexmit code should be good enough, this little hack
5501 * doesn't quite work XXX */
5502 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5504 rxi_ScheduleKeepAliveEvent(call);
5505 MUTEX_EXIT(&call->lock);
5510 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5512 if (!call->keepAliveEvent) {
5513 struct clock when, now;
5514 clock_GetTime(&now);
5516 when.sec += call->conn->secondsUntilPing;
5517 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5518 call->keepAliveEvent =
5519 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
5523 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5525 rxi_KeepAliveOn(register struct rx_call *call)
5527 /* Pretend last packet received was received now--i.e. if another
5528 * packet isn't received within the keep alive time, then the call
5529 * will die; Initialize last send time to the current time--even
5530 * if a packet hasn't been sent yet. This will guarantee that a
5531 * keep-alive is sent within the ping time */
5532 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5533 rxi_ScheduleKeepAliveEvent(call);
5536 /* This routine is called to send connection abort messages
5537 * that have been delayed to throttle looping clients. */
5539 rxi_SendDelayedConnAbort(struct rxevent *event,
5540 register struct rx_connection *conn, char *dummy)
5543 struct rx_packet *packet;
5545 MUTEX_ENTER(&conn->conn_data_lock);
5546 conn->delayedAbortEvent = NULL;
5547 error = htonl(conn->error);
5549 MUTEX_EXIT(&conn->conn_data_lock);
5550 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5553 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5554 RX_PACKET_TYPE_ABORT, (char *)&error,
5556 rxi_FreePacket(packet);
5560 /* This routine is called to send call abort messages
5561 * that have been delayed to throttle looping clients. */
5563 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5567 struct rx_packet *packet;
5569 MUTEX_ENTER(&call->lock);
5570 call->delayedAbortEvent = NULL;
5571 error = htonl(call->error);
5573 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5576 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5577 (char *)&error, sizeof(error), 0);
5578 rxi_FreePacket(packet);
5580 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5581 MUTEX_EXIT(&call->lock);
5584 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5585 * seconds) to ask the client to authenticate itself. The routine
5586 * issues a challenge to the client, which is obtained from the
5587 * security object associated with the connection */
5589 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5590 void *arg1, int tries)
5592 conn->challengeEvent = NULL;
5593 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5594 register struct rx_packet *packet;
5595 struct clock when, now;
5598 /* We've failed to authenticate for too long.
5599 * Reset any calls waiting for authentication;
5600 * they are all in RX_STATE_PRECALL.
5604 MUTEX_ENTER(&conn->conn_call_lock);
5605 for (i = 0; i < RX_MAXCALLS; i++) {
5606 struct rx_call *call = conn->call[i];
5608 MUTEX_ENTER(&call->lock);
5609 if (call->state == RX_STATE_PRECALL) {
5610 rxi_CallError(call, RX_CALL_DEAD);
5611 rxi_SendCallAbort(call, NULL, 0, 0);
5613 MUTEX_EXIT(&call->lock);
5616 MUTEX_EXIT(&conn->conn_call_lock);
5620 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5622 /* If there's no packet available, do this later. */
5623 RXS_GetChallenge(conn->securityObject, conn, packet);
5624 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5625 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5626 rxi_FreePacket(packet);
5628 clock_GetTime(&now);
5630 when.sec += RX_CHALLENGE_TIMEOUT;
5631 conn->challengeEvent =
5632 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
5637 /* Call this routine to start requesting the client to authenticate
5638 * itself. This will continue until authentication is established,
5639 * the call times out, or an invalid response is returned. The
5640 * security object associated with the connection is asked to create
5641 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5642 * defined earlier. */
5644 rxi_ChallengeOn(register struct rx_connection *conn)
5646 if (!conn->challengeEvent) {
5647 RXS_CreateChallenge(conn->securityObject, conn);
5648 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5653 /* Compute round trip time of the packet provided, in *rttp.
5656 /* rxi_ComputeRoundTripTime is called with peer locked. */
5657 /* sentp and/or peer may be null */
5659 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5660 register struct clock *sentp,
5661 register struct rx_peer *peer)
5663 struct clock thisRtt, *rttp = &thisRtt;
5665 register int rtt_timeout;
5667 clock_GetTime(rttp);
5669 if (clock_Lt(rttp, sentp)) {
5671 return; /* somebody set the clock back, don't count this time. */
5673 clock_Sub(rttp, sentp);
5674 MUTEX_ENTER(&rx_stats_mutex);
5675 if (clock_Lt(rttp, &rx_stats.minRtt))
5676 rx_stats.minRtt = *rttp;
5677 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5678 if (rttp->sec > 60) {
5679 MUTEX_EXIT(&rx_stats_mutex);
5680 return; /* somebody set the clock ahead */
5682 rx_stats.maxRtt = *rttp;
5684 clock_Add(&rx_stats.totalRtt, rttp);
5685 rx_stats.nRttSamples++;
5686 MUTEX_EXIT(&rx_stats_mutex);
5688 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5690 /* Apply VanJacobson round-trip estimations */
5695 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5696 * srtt is stored as fixed point with 3 bits after the binary
5697 * point (i.e., scaled by 8). The following magic is
5698 * equivalent to the smoothing algorithm in rfc793 with an
5699 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5700 * srtt*8 = srtt*8 + rtt - srtt
5701 * srtt = srtt + rtt/8 - srtt/8
5704 delta = MSEC(rttp) - (peer->rtt >> 3);
5708 * We accumulate a smoothed rtt variance (actually, a smoothed
5709 * mean difference), then set the retransmit timer to smoothed
5710 * rtt + 4 times the smoothed variance (was 2x in van's original
5711 * paper, but 4x works better for me, and apparently for him as
5713 * rttvar is stored as
5714 * fixed point with 2 bits after the binary point (scaled by
5715 * 4). The following is equivalent to rfc793 smoothing with
5716 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5717 * replaces rfc793's wired-in beta.
5718 * dev*4 = dev*4 + (|actual - expected| - dev)
5724 delta -= (peer->rtt_dev >> 2);
5725 peer->rtt_dev += delta;
5727 /* I don't have a stored RTT so I start with this value. Since I'm
5728 * probably just starting a call, and will be pushing more data down
5729 * this, I expect congestion to increase rapidly. So I fudge a
5730 * little, and I set deviance to half the rtt. In practice,
5731 * deviance tends to approach something a little less than
5732 * half the smoothed rtt. */
5733 peer->rtt = (MSEC(rttp) << 3) + 8;
5734 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5736 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5737 * the other of these connections is usually in a user process, and can
5738 * be switched and/or swapped out. So on fast, reliable networks, the
5739 * timeout would otherwise be too short.
5741 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5742 clock_Zero(&(peer->timeout));
5743 clock_Addmsec(&(peer->timeout), rtt_timeout);
5745 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)));
5749 /* Find all server connections that have not been active for a long time, and
5752 rxi_ReapConnections(void)
5754 struct clock now, when;
5755 clock_GetTime(&now);
5757 /* Find server connection structures that haven't been used for
5758 * greater than rx_idleConnectionTime */
5760 struct rx_connection **conn_ptr, **conn_end;
5761 int i, havecalls = 0;
5762 MUTEX_ENTER(&rx_connHashTable_lock);
5763 for (conn_ptr = &rx_connHashTable[0], conn_end =
5764 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5766 struct rx_connection *conn, *next;
5767 struct rx_call *call;
5771 for (conn = *conn_ptr; conn; conn = next) {
5772 /* XXX -- Shouldn't the connection be locked? */
5775 for (i = 0; i < RX_MAXCALLS; i++) {
5776 call = conn->call[i];
5779 MUTEX_ENTER(&call->lock);
5780 #ifdef RX_ENABLE_LOCKS
5781 result = rxi_CheckCall(call, 1);
5782 #else /* RX_ENABLE_LOCKS */
5783 result = rxi_CheckCall(call);
5784 #endif /* RX_ENABLE_LOCKS */
5785 MUTEX_EXIT(&call->lock);
5787 /* If CheckCall freed the call, it might
5788 * have destroyed the connection as well,
5789 * which screws up the linked lists.
5795 if (conn->type == RX_SERVER_CONNECTION) {
5796 /* This only actually destroys the connection if
5797 * there are no outstanding calls */
5798 MUTEX_ENTER(&conn->conn_data_lock);
5799 if (!havecalls && !conn->refCount
5800 && ((conn->lastSendTime + rx_idleConnectionTime) <
5802 conn->refCount++; /* it will be decr in rx_DestroyConn */
5803 MUTEX_EXIT(&conn->conn_data_lock);
5804 #ifdef RX_ENABLE_LOCKS
5805 rxi_DestroyConnectionNoLock(conn);
5806 #else /* RX_ENABLE_LOCKS */
5807 rxi_DestroyConnection(conn);
5808 #endif /* RX_ENABLE_LOCKS */
5810 #ifdef RX_ENABLE_LOCKS
5812 MUTEX_EXIT(&conn->conn_data_lock);
5814 #endif /* RX_ENABLE_LOCKS */
5818 #ifdef RX_ENABLE_LOCKS
5819 while (rx_connCleanup_list) {
5820 struct rx_connection *conn;
5821 conn = rx_connCleanup_list;
5822 rx_connCleanup_list = rx_connCleanup_list->next;
5823 MUTEX_EXIT(&rx_connHashTable_lock);
5824 rxi_CleanupConnection(conn);
5825 MUTEX_ENTER(&rx_connHashTable_lock);
5827 MUTEX_EXIT(&rx_connHashTable_lock);
5828 #endif /* RX_ENABLE_LOCKS */
5831 /* Find any peer structures that haven't been used (haven't had an
5832 * associated connection) for greater than rx_idlePeerTime */
5834 struct rx_peer **peer_ptr, **peer_end;
5836 MUTEX_ENTER(&rx_rpc_stats);
5837 MUTEX_ENTER(&rx_peerHashTable_lock);
5838 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5839 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5841 struct rx_peer *peer, *next, *prev;
5842 for (prev = peer = *peer_ptr; peer; peer = next) {
5844 code = MUTEX_TRYENTER(&peer->peer_lock);
5845 if ((code) && (peer->refCount == 0)
5846 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5847 rx_interface_stat_p rpc_stat, nrpc_stat;
5849 MUTEX_EXIT(&peer->peer_lock);
5850 MUTEX_DESTROY(&peer->peer_lock);
5852 (&peer->rpcStats, rpc_stat, nrpc_stat,
5853 rx_interface_stat)) {
5854 unsigned int num_funcs;
5857 queue_Remove(&rpc_stat->queue_header);
5858 queue_Remove(&rpc_stat->all_peers);
5859 num_funcs = rpc_stat->stats[0].func_total;
5861 sizeof(rx_interface_stat_t) +
5862 rpc_stat->stats[0].func_total *
5863 sizeof(rx_function_entry_v1_t);
5865 rxi_Free(rpc_stat, space);
5866 rxi_rpc_peer_stat_cnt -= num_funcs;
5869 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
5870 if (peer == *peer_ptr) {
5877 MUTEX_EXIT(&peer->peer_lock);
5883 MUTEX_EXIT(&rx_peerHashTable_lock);
5884 MUTEX_EXIT(&rx_rpc_stats);
5887 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5888 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5889 * GC, just below. Really, we shouldn't have to keep moving packets from
5890 * one place to another, but instead ought to always know if we can
5891 * afford to hold onto a packet in its particular use. */
5892 MUTEX_ENTER(&rx_freePktQ_lock);
5893 if (rx_waitingForPackets) {
5894 rx_waitingForPackets = 0;
5895 #ifdef RX_ENABLE_LOCKS
5896 CV_BROADCAST(&rx_waitingForPackets_cv);
5898 osi_rxWakeup(&rx_waitingForPackets);
5901 MUTEX_EXIT(&rx_freePktQ_lock);
5904 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5905 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
5909 /* rxs_Release - This isn't strictly necessary but, since the macro name from
5910 * rx.h is sort of strange this is better. This is called with a security
5911 * object before it is discarded. Each connection using a security object has
5912 * its own refcount to the object so it won't actually be freed until the last
5913 * connection is destroyed.
5915 * This is the only rxs module call. A hold could also be written but no one
5919 rxs_Release(struct rx_securityClass *aobj)
5921 return RXS_Close(aobj);
5925 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
5926 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
5927 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
5928 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
5930 /* Adjust our estimate of the transmission rate to this peer, given
5931 * that the packet p was just acked. We can adjust peer->timeout and
5932 * call->twind. Pragmatically, this is called
5933 * only with packets of maximal length.
5934 * Called with peer and call locked.
5938 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
5939 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
5941 afs_int32 xferSize, xferMs;
5942 register afs_int32 minTime;
5945 /* Count down packets */
5946 if (peer->rateFlag > 0)
5948 /* Do nothing until we're enabled */
5949 if (peer->rateFlag != 0)
5954 /* Count only when the ack seems legitimate */
5955 switch (ackReason) {
5956 case RX_ACK_REQUESTED:
5958 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
5962 case RX_ACK_PING_RESPONSE:
5963 if (p) /* want the response to ping-request, not data send */
5965 clock_GetTime(&newTO);
5966 if (clock_Gt(&newTO, &call->pingRequestTime)) {
5967 clock_Sub(&newTO, &call->pingRequestTime);
5968 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
5972 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
5979 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));
5981 /* Track only packets that are big enough. */
5982 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
5986 /* absorb RTT data (in milliseconds) for these big packets */
5987 if (peer->smRtt == 0) {
5988 peer->smRtt = xferMs;
5990 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
5995 if (peer->countDown) {
5999 peer->countDown = 10; /* recalculate only every so often */
6001 /* In practice, we can measure only the RTT for full packets,
6002 * because of the way Rx acks the data that it receives. (If it's
6003 * smaller than a full packet, it often gets implicitly acked
6004 * either by the call response (from a server) or by the next call
6005 * (from a client), and either case confuses transmission times
6006 * with processing times.) Therefore, replace the above
6007 * more-sophisticated processing with a simpler version, where the
6008 * smoothed RTT is kept for full-size packets, and the time to
6009 * transmit a windowful of full-size packets is simply RTT *
6010 * windowSize. Again, we take two steps:
6011 - ensure the timeout is large enough for a single packet's RTT;
6012 - ensure that the window is small enough to fit in the desired timeout.*/
6014 /* First, the timeout check. */
6015 minTime = peer->smRtt;
6016 /* Get a reasonable estimate for a timeout period */
6018 newTO.sec = minTime / 1000;
6019 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6021 /* Increase the timeout period so that we can always do at least
6022 * one packet exchange */
6023 if (clock_Gt(&newTO, &peer->timeout)) {
6025 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));
6027 peer->timeout = newTO;
6030 /* Now, get an estimate for the transmit window size. */
6031 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6032 /* Now, convert to the number of full packets that could fit in a
6033 * reasonable fraction of that interval */
6034 minTime /= (peer->smRtt << 1);
6035 xferSize = minTime; /* (make a copy) */
6037 /* Now clamp the size to reasonable bounds. */
6040 else if (minTime > rx_Window)
6041 minTime = rx_Window;
6042 /* if (minTime != peer->maxWindow) {
6043 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6044 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6045 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6047 peer->maxWindow = minTime;
6048 elide... call->twind = minTime;
6052 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6053 * Discern this by calculating the timeout necessary for rx_Window
6055 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6056 /* calculate estimate for transmission interval in milliseconds */
6057 minTime = rx_Window * peer->smRtt;
6058 if (minTime < 1000) {
6059 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
6060 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6061 peer->timeout.usec, peer->smRtt, peer->packetSize));
6063 newTO.sec = 0; /* cut back on timeout by half a second */
6064 newTO.usec = 500000;
6065 clock_Sub(&peer->timeout, &newTO);
6070 } /* end of rxi_ComputeRate */
6071 #endif /* ADAPT_WINDOW */
6079 #define TRACE_OPTION_DEBUGLOG 4
6087 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6088 0, KEY_QUERY_VALUE, &parmKey);
6089 if (code != ERROR_SUCCESS)
6092 dummyLen = sizeof(TraceOption);
6093 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6094 (BYTE *) &TraceOption, &dummyLen);
6095 if (code == ERROR_SUCCESS) {
6096 rxdebug_active = (TraceOption & TRACE_OPTION_DEBUGLOG) ? 1 : 0;
6098 RegCloseKey (parmKey);
6099 #endif /* AFS_NT40_ENV */
6104 rx_DebugOnOff(int on)
6106 rxdebug_active = on;
6108 #endif /* AFS_NT40_ENV */
6111 /* Don't call this debugging routine directly; use dpf */
6113 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
6114 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
6122 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6125 len = _snprintf(msg, sizeof(msg)-2,
6126 tformat, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10,
6127 a11, a12, a13, a14, a15);
6129 if (msg[len-1] != '\n') {
6133 OutputDebugString(msg);
6138 clock_GetTime(&now);
6139 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
6140 (unsigned int)now.usec / 1000);
6141 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
6148 * This function is used to process the rx_stats structure that is local
6149 * to a process as well as an rx_stats structure received from a remote
6150 * process (via rxdebug). Therefore, it needs to do minimal version
6154 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
6155 afs_int32 freePackets, char version)
6159 if (size != sizeof(struct rx_stats)) {
6161 "Unexpected size of stats structure: was %d, expected %d\n",
6162 size, sizeof(struct rx_stats));
6165 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6168 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6169 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6170 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6171 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6172 s->specialPktAllocFailures);
6174 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6175 s->receivePktAllocFailures, s->sendPktAllocFailures,
6176 s->specialPktAllocFailures);
6180 " greedy %d, " "bogusReads %d (last from host %x), "
6181 "noPackets %d, " "noBuffers %d, " "selects %d, "
6182 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6183 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6184 s->selects, s->sendSelects);
6186 fprintf(file, " packets read: ");
6187 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6188 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6190 fprintf(file, "\n");
6193 " other read counters: data %d, " "ack %d, " "dup %d "
6194 "spurious %d " "dally %d\n", s->dataPacketsRead,
6195 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6196 s->ignorePacketDally);
6198 fprintf(file, " packets sent: ");
6199 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6200 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6202 fprintf(file, "\n");
6205 " other send counters: ack %d, " "data %d (not resends), "
6206 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6207 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6208 s->dataPacketsPushed, s->ignoreAckedPacket);
6211 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6212 s->netSendFailures, (int)s->fatalErrors);
6214 if (s->nRttSamples) {
6215 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6216 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6218 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6219 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6223 " %d server connections, " "%d client connections, "
6224 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6225 s->nServerConns, s->nClientConns, s->nPeerStructs,
6226 s->nCallStructs, s->nFreeCallStructs);
6228 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6229 fprintf(file, " %d clock updates\n", clock_nUpdates);
6234 /* for backward compatibility */
6236 rx_PrintStats(FILE * file)
6238 MUTEX_ENTER(&rx_stats_mutex);
6239 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6241 MUTEX_EXIT(&rx_stats_mutex);
6245 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6247 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6248 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6249 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6252 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6253 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6254 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6257 " Packet size %d, " "max in packet skew %d, "
6258 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6259 (int)peer->outPacketSkew);
6262 #ifdef AFS_PTHREAD_ENV
6264 * This mutex protects the following static variables:
6268 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6269 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6271 #define LOCK_RX_DEBUG
6272 #define UNLOCK_RX_DEBUG
6273 #endif /* AFS_PTHREAD_ENV */
6276 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6277 u_char type, void *inputData, size_t inputLength,
6278 void *outputData, size_t outputLength)
6280 static afs_int32 counter = 100;
6281 time_t waitTime, waitCount, startTime;
6282 struct rx_header theader;
6284 register afs_int32 code;
6285 struct timeval tv_now, tv_wake, tv_delta;
6286 struct sockaddr_in taddr, faddr;
6291 startTime = time(0);
6297 tp = &tbuffer[sizeof(struct rx_header)];
6298 taddr.sin_family = AF_INET;
6299 taddr.sin_port = remotePort;
6300 taddr.sin_addr.s_addr = remoteAddr;
6301 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6302 taddr.sin_len = sizeof(struct sockaddr_in);
6305 memset(&theader, 0, sizeof(theader));
6306 theader.epoch = htonl(999);
6308 theader.callNumber = htonl(counter);
6311 theader.type = type;
6312 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6313 theader.serviceId = 0;
6315 memcpy(tbuffer, &theader, sizeof(theader));
6316 memcpy(tp, inputData, inputLength);
6318 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6319 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6321 /* see if there's a packet available */
6322 gettimeofday(&tv_wake,0);
6323 tv_wake.tv_sec += waitTime;
6326 FD_SET(socket, &imask);
6327 tv_delta.tv_sec = tv_wake.tv_sec;
6328 tv_delta.tv_usec = tv_wake.tv_usec;
6329 gettimeofday(&tv_now, 0);
6331 if (tv_delta.tv_usec < tv_now.tv_usec) {
6333 tv_delta.tv_usec += 1000000;
6336 tv_delta.tv_usec -= tv_now.tv_usec;
6338 if (tv_delta.tv_sec < tv_now.tv_sec) {
6342 tv_delta.tv_sec -= tv_now.tv_sec;
6344 code = select(socket + 1, &imask, 0, 0, &tv_delta);
6345 if (code == 1 && FD_ISSET(socket, &imask)) {
6346 /* now receive a packet */
6347 faddrLen = sizeof(struct sockaddr_in);
6349 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6350 (struct sockaddr *)&faddr, &faddrLen);
6353 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6354 if (counter == ntohl(theader.callNumber))
6362 /* see if we've timed out */
6370 code -= sizeof(struct rx_header);
6371 if (code > outputLength)
6372 code = outputLength;
6373 memcpy(outputData, tp, code);
6378 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6379 afs_uint16 remotePort, struct rx_debugStats * stat,
6380 afs_uint32 * supportedValues)
6382 struct rx_debugIn in;
6385 *supportedValues = 0;
6386 in.type = htonl(RX_DEBUGI_GETSTATS);
6389 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6390 &in, sizeof(in), stat, sizeof(*stat));
6393 * If the call was successful, fixup the version and indicate
6394 * what contents of the stat structure are valid.
6395 * Also do net to host conversion of fields here.
6399 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6400 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6402 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6403 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6405 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6406 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6408 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6409 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6411 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6412 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6414 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6415 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6417 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6418 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6420 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6421 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6424 stat->nFreePackets = ntohl(stat->nFreePackets);
6425 stat->packetReclaims = ntohl(stat->packetReclaims);
6426 stat->callsExecuted = ntohl(stat->callsExecuted);
6427 stat->nWaiting = ntohl(stat->nWaiting);
6428 stat->idleThreads = ntohl(stat->idleThreads);
6435 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6436 afs_uint16 remotePort, struct rx_stats * stat,
6437 afs_uint32 * supportedValues)
6439 struct rx_debugIn in;
6440 afs_int32 *lp = (afs_int32 *) stat;
6445 * supportedValues is currently unused, but added to allow future
6446 * versioning of this function.
6449 *supportedValues = 0;
6450 in.type = htonl(RX_DEBUGI_RXSTATS);
6452 memset(stat, 0, sizeof(*stat));
6454 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6455 &in, sizeof(in), stat, sizeof(*stat));
6460 * Do net to host conversion here
6463 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6472 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6473 afs_uint16 remotePort, size_t version_length,
6477 return MakeDebugCall(socket, remoteAddr, remotePort,
6478 RX_PACKET_TYPE_VERSION, a, 1, version,
6483 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6484 afs_uint16 remotePort, afs_int32 * nextConnection,
6485 int allConnections, afs_uint32 debugSupportedValues,
6486 struct rx_debugConn * conn,
6487 afs_uint32 * supportedValues)
6489 struct rx_debugIn in;
6494 * supportedValues is currently unused, but added to allow future
6495 * versioning of this function.
6498 *supportedValues = 0;
6499 if (allConnections) {
6500 in.type = htonl(RX_DEBUGI_GETALLCONN);
6502 in.type = htonl(RX_DEBUGI_GETCONN);
6504 in.index = htonl(*nextConnection);
6505 memset(conn, 0, sizeof(*conn));
6507 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6508 &in, sizeof(in), conn, sizeof(*conn));
6511 *nextConnection += 1;
6514 * Convert old connection format to new structure.
6517 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6518 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6519 #define MOVEvL(a) (conn->a = vL->a)
6521 /* any old or unrecognized version... */
6522 for (i = 0; i < RX_MAXCALLS; i++) {
6523 MOVEvL(callState[i]);
6524 MOVEvL(callMode[i]);
6525 MOVEvL(callFlags[i]);
6526 MOVEvL(callOther[i]);
6528 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6529 MOVEvL(secStats.type);
6530 MOVEvL(secStats.level);
6531 MOVEvL(secStats.flags);
6532 MOVEvL(secStats.expires);
6533 MOVEvL(secStats.packetsReceived);
6534 MOVEvL(secStats.packetsSent);
6535 MOVEvL(secStats.bytesReceived);
6536 MOVEvL(secStats.bytesSent);
6541 * Do net to host conversion here
6543 * I don't convert host or port since we are most likely
6544 * going to want these in NBO.
6546 conn->cid = ntohl(conn->cid);
6547 conn->serial = ntohl(conn->serial);
6548 for (i = 0; i < RX_MAXCALLS; i++) {
6549 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6551 conn->error = ntohl(conn->error);
6552 conn->secStats.flags = ntohl(conn->secStats.flags);
6553 conn->secStats.expires = ntohl(conn->secStats.expires);
6554 conn->secStats.packetsReceived =
6555 ntohl(conn->secStats.packetsReceived);
6556 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6557 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6558 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6559 conn->epoch = ntohl(conn->epoch);
6560 conn->natMTU = ntohl(conn->natMTU);
6567 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6568 afs_uint16 remotePort, afs_int32 * nextPeer,
6569 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6570 afs_uint32 * supportedValues)
6572 struct rx_debugIn in;
6576 * supportedValues is currently unused, but added to allow future
6577 * versioning of this function.
6580 *supportedValues = 0;
6581 in.type = htonl(RX_DEBUGI_GETPEER);
6582 in.index = htonl(*nextPeer);
6583 memset(peer, 0, sizeof(*peer));
6585 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6586 &in, sizeof(in), peer, sizeof(*peer));
6592 * Do net to host conversion here
6594 * I don't convert host or port since we are most likely
6595 * going to want these in NBO.
6597 peer->ifMTU = ntohs(peer->ifMTU);
6598 peer->idleWhen = ntohl(peer->idleWhen);
6599 peer->refCount = ntohs(peer->refCount);
6600 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6601 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6602 peer->rtt = ntohl(peer->rtt);
6603 peer->rtt_dev = ntohl(peer->rtt_dev);
6604 peer->timeout.sec = ntohl(peer->timeout.sec);
6605 peer->timeout.usec = ntohl(peer->timeout.usec);
6606 peer->nSent = ntohl(peer->nSent);
6607 peer->reSends = ntohl(peer->reSends);
6608 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6609 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6610 peer->rateFlag = ntohl(peer->rateFlag);
6611 peer->natMTU = ntohs(peer->natMTU);
6612 peer->maxMTU = ntohs(peer->maxMTU);
6613 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6614 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6615 peer->MTU = ntohs(peer->MTU);
6616 peer->cwind = ntohs(peer->cwind);
6617 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6618 peer->congestSeq = ntohs(peer->congestSeq);
6619 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6620 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6621 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6622 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6627 #endif /* RXDEBUG */
6632 struct rx_serverQueueEntry *np;
6635 register struct rx_call *call;
6636 register struct rx_serverQueueEntry *sq;
6640 if (rxinit_status == 1) {
6642 return; /* Already shutdown. */
6646 #ifndef AFS_PTHREAD_ENV
6647 FD_ZERO(&rx_selectMask);
6648 #endif /* AFS_PTHREAD_ENV */
6649 rxi_dataQuota = RX_MAX_QUOTA;
6650 #ifndef AFS_PTHREAD_ENV
6652 #endif /* AFS_PTHREAD_ENV */
6655 #ifndef AFS_PTHREAD_ENV
6656 #ifndef AFS_USE_GETTIMEOFDAY
6658 #endif /* AFS_USE_GETTIMEOFDAY */
6659 #endif /* AFS_PTHREAD_ENV */
6661 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6662 call = queue_First(&rx_freeCallQueue, rx_call);
6664 rxi_Free(call, sizeof(struct rx_call));
6667 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6668 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6674 struct rx_peer **peer_ptr, **peer_end;
6675 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6676 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6678 struct rx_peer *peer, *next;
6679 for (peer = *peer_ptr; peer; peer = next) {
6680 rx_interface_stat_p rpc_stat, nrpc_stat;
6683 (&peer->rpcStats, rpc_stat, nrpc_stat,
6684 rx_interface_stat)) {
6685 unsigned int num_funcs;
6688 queue_Remove(&rpc_stat->queue_header);
6689 queue_Remove(&rpc_stat->all_peers);
6690 num_funcs = rpc_stat->stats[0].func_total;
6692 sizeof(rx_interface_stat_t) +
6693 rpc_stat->stats[0].func_total *
6694 sizeof(rx_function_entry_v1_t);
6696 rxi_Free(rpc_stat, space);
6697 MUTEX_ENTER(&rx_rpc_stats);
6698 rxi_rpc_peer_stat_cnt -= num_funcs;
6699 MUTEX_EXIT(&rx_rpc_stats);
6703 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6707 for (i = 0; i < RX_MAX_SERVICES; i++) {
6709 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6711 for (i = 0; i < rx_hashTableSize; i++) {
6712 register struct rx_connection *tc, *ntc;
6713 MUTEX_ENTER(&rx_connHashTable_lock);
6714 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6716 for (j = 0; j < RX_MAXCALLS; j++) {
6718 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6721 rxi_Free(tc, sizeof(*tc));
6723 MUTEX_EXIT(&rx_connHashTable_lock);
6726 MUTEX_ENTER(&freeSQEList_lock);
6728 while ((np = rx_FreeSQEList)) {
6729 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6730 MUTEX_DESTROY(&np->lock);
6731 rxi_Free(np, sizeof(*np));
6734 MUTEX_EXIT(&freeSQEList_lock);
6735 MUTEX_DESTROY(&freeSQEList_lock);
6736 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6737 MUTEX_DESTROY(&rx_connHashTable_lock);
6738 MUTEX_DESTROY(&rx_peerHashTable_lock);
6739 MUTEX_DESTROY(&rx_serverPool_lock);
6741 osi_Free(rx_connHashTable,
6742 rx_hashTableSize * sizeof(struct rx_connection *));
6743 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6745 UNPIN(rx_connHashTable,
6746 rx_hashTableSize * sizeof(struct rx_connection *));
6747 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6749 rxi_FreeAllPackets();
6751 MUTEX_ENTER(&rx_stats_mutex);
6752 rxi_dataQuota = RX_MAX_QUOTA;
6753 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6754 MUTEX_EXIT(&rx_stats_mutex);
6760 #ifdef RX_ENABLE_LOCKS
6762 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6764 if (!MUTEX_ISMINE(lockaddr))
6765 osi_Panic("Lock not held: %s", msg);
6767 #endif /* RX_ENABLE_LOCKS */
6772 * Routines to implement connection specific data.
6776 rx_KeyCreate(rx_destructor_t rtn)
6779 MUTEX_ENTER(&rxi_keyCreate_lock);
6780 key = rxi_keyCreate_counter++;
6781 rxi_keyCreate_destructor = (rx_destructor_t *)
6782 realloc((void *)rxi_keyCreate_destructor,
6783 (key + 1) * sizeof(rx_destructor_t));
6784 rxi_keyCreate_destructor[key] = rtn;
6785 MUTEX_EXIT(&rxi_keyCreate_lock);
6790 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6793 MUTEX_ENTER(&conn->conn_data_lock);
6794 if (!conn->specific) {
6795 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6796 for (i = 0; i < key; i++)
6797 conn->specific[i] = NULL;
6798 conn->nSpecific = key + 1;
6799 conn->specific[key] = ptr;
6800 } else if (key >= conn->nSpecific) {
6801 conn->specific = (void **)
6802 realloc(conn->specific, (key + 1) * sizeof(void *));
6803 for (i = conn->nSpecific; i < key; i++)
6804 conn->specific[i] = NULL;
6805 conn->nSpecific = key + 1;
6806 conn->specific[key] = ptr;
6808 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6809 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6810 conn->specific[key] = ptr;
6812 MUTEX_EXIT(&conn->conn_data_lock);
6816 rx_GetSpecific(struct rx_connection *conn, int key)
6819 MUTEX_ENTER(&conn->conn_data_lock);
6820 if (key >= conn->nSpecific)
6823 ptr = conn->specific[key];
6824 MUTEX_EXIT(&conn->conn_data_lock);
6828 #endif /* !KERNEL */
6831 * processStats is a queue used to store the statistics for the local
6832 * process. Its contents are similar to the contents of the rpcStats
6833 * queue on a rx_peer structure, but the actual data stored within
6834 * this queue contains totals across the lifetime of the process (assuming
6835 * the stats have not been reset) - unlike the per peer structures
6836 * which can come and go based upon the peer lifetime.
6839 static struct rx_queue processStats = { &processStats, &processStats };
6842 * peerStats is a queue used to store the statistics for all peer structs.
6843 * Its contents are the union of all the peer rpcStats queues.
6846 static struct rx_queue peerStats = { &peerStats, &peerStats };
6849 * rxi_monitor_processStats is used to turn process wide stat collection
6853 static int rxi_monitor_processStats = 0;
6856 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6859 static int rxi_monitor_peerStats = 0;
6862 * rxi_AddRpcStat - given all of the information for a particular rpc
6863 * call, create (if needed) and update the stat totals for the rpc.
6867 * IN stats - the queue of stats that will be updated with the new value
6869 * IN rxInterface - a unique number that identifies the rpc interface
6871 * IN currentFunc - the index of the function being invoked
6873 * IN totalFunc - the total number of functions in this interface
6875 * IN queueTime - the amount of time this function waited for a thread
6877 * IN execTime - the amount of time this function invocation took to execute
6879 * IN bytesSent - the number bytes sent by this invocation
6881 * IN bytesRcvd - the number bytes received by this invocation
6883 * IN isServer - if true, this invocation was made to a server
6885 * IN remoteHost - the ip address of the remote host
6887 * IN remotePort - the port of the remote host
6889 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6891 * INOUT counter - if a new stats structure is allocated, the counter will
6892 * be updated with the new number of allocated stat structures
6900 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6901 afs_uint32 currentFunc, afs_uint32 totalFunc,
6902 struct clock *queueTime, struct clock *execTime,
6903 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6904 afs_uint32 remoteHost, afs_uint32 remotePort,
6905 int addToPeerList, unsigned int *counter)
6908 rx_interface_stat_p rpc_stat, nrpc_stat;
6911 * See if there's already a structure for this interface
6914 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6915 if ((rpc_stat->stats[0].interfaceId == rxInterface)
6916 && (rpc_stat->stats[0].remote_is_server == isServer))
6921 * Didn't find a match so allocate a new structure and add it to the
6925 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
6926 || (rpc_stat->stats[0].interfaceId != rxInterface)
6927 || (rpc_stat->stats[0].remote_is_server != isServer)) {
6932 sizeof(rx_interface_stat_t) +
6933 totalFunc * sizeof(rx_function_entry_v1_t);
6935 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
6936 if (rpc_stat == NULL) {
6940 *counter += totalFunc;
6941 for (i = 0; i < totalFunc; i++) {
6942 rpc_stat->stats[i].remote_peer = remoteHost;
6943 rpc_stat->stats[i].remote_port = remotePort;
6944 rpc_stat->stats[i].remote_is_server = isServer;
6945 rpc_stat->stats[i].interfaceId = rxInterface;
6946 rpc_stat->stats[i].func_total = totalFunc;
6947 rpc_stat->stats[i].func_index = i;
6948 hzero(rpc_stat->stats[i].invocations);
6949 hzero(rpc_stat->stats[i].bytes_sent);
6950 hzero(rpc_stat->stats[i].bytes_rcvd);
6951 rpc_stat->stats[i].queue_time_sum.sec = 0;
6952 rpc_stat->stats[i].queue_time_sum.usec = 0;
6953 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
6954 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
6955 rpc_stat->stats[i].queue_time_min.sec = 9999999;
6956 rpc_stat->stats[i].queue_time_min.usec = 9999999;
6957 rpc_stat->stats[i].queue_time_max.sec = 0;
6958 rpc_stat->stats[i].queue_time_max.usec = 0;
6959 rpc_stat->stats[i].execution_time_sum.sec = 0;
6960 rpc_stat->stats[i].execution_time_sum.usec = 0;
6961 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
6962 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
6963 rpc_stat->stats[i].execution_time_min.sec = 9999999;
6964 rpc_stat->stats[i].execution_time_min.usec = 9999999;
6965 rpc_stat->stats[i].execution_time_max.sec = 0;
6966 rpc_stat->stats[i].execution_time_max.usec = 0;
6968 queue_Prepend(stats, rpc_stat);
6969 if (addToPeerList) {
6970 queue_Prepend(&peerStats, &rpc_stat->all_peers);
6975 * Increment the stats for this function
6978 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
6979 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
6980 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
6981 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
6982 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
6983 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
6984 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
6986 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
6987 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
6989 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
6990 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
6992 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
6993 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
6995 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
6996 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7004 * rx_IncrementTimeAndCount - increment the times and count for a particular
7009 * IN peer - the peer who invoked the rpc
7011 * IN rxInterface - a unique number that identifies the rpc interface
7013 * IN currentFunc - the index of the function being invoked
7015 * IN totalFunc - the total number of functions in this interface
7017 * IN queueTime - the amount of time this function waited for a thread
7019 * IN execTime - the amount of time this function invocation took to execute
7021 * IN bytesSent - the number bytes sent by this invocation
7023 * IN bytesRcvd - the number bytes received by this invocation
7025 * IN isServer - if true, this invocation was made to a server
7033 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7034 afs_uint32 currentFunc, afs_uint32 totalFunc,
7035 struct clock *queueTime, struct clock *execTime,
7036 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7040 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
7043 MUTEX_ENTER(&rx_rpc_stats);
7044 MUTEX_ENTER(&peer->peer_lock);
7046 if (rxi_monitor_peerStats) {
7047 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7048 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7049 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7052 if (rxi_monitor_processStats) {
7053 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7054 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7055 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7058 MUTEX_EXIT(&peer->peer_lock);
7059 MUTEX_EXIT(&rx_rpc_stats);
7064 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7068 * IN callerVersion - the rpc stat version of the caller.
7070 * IN count - the number of entries to marshall.
7072 * IN stats - pointer to stats to be marshalled.
7074 * OUT ptr - Where to store the marshalled data.
7081 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7082 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7088 * We only support the first version
7090 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7091 *(ptr++) = stats->remote_peer;
7092 *(ptr++) = stats->remote_port;
7093 *(ptr++) = stats->remote_is_server;
7094 *(ptr++) = stats->interfaceId;
7095 *(ptr++) = stats->func_total;
7096 *(ptr++) = stats->func_index;
7097 *(ptr++) = hgethi(stats->invocations);
7098 *(ptr++) = hgetlo(stats->invocations);
7099 *(ptr++) = hgethi(stats->bytes_sent);
7100 *(ptr++) = hgetlo(stats->bytes_sent);
7101 *(ptr++) = hgethi(stats->bytes_rcvd);
7102 *(ptr++) = hgetlo(stats->bytes_rcvd);
7103 *(ptr++) = stats->queue_time_sum.sec;
7104 *(ptr++) = stats->queue_time_sum.usec;
7105 *(ptr++) = stats->queue_time_sum_sqr.sec;
7106 *(ptr++) = stats->queue_time_sum_sqr.usec;
7107 *(ptr++) = stats->queue_time_min.sec;
7108 *(ptr++) = stats->queue_time_min.usec;
7109 *(ptr++) = stats->queue_time_max.sec;
7110 *(ptr++) = stats->queue_time_max.usec;
7111 *(ptr++) = stats->execution_time_sum.sec;
7112 *(ptr++) = stats->execution_time_sum.usec;
7113 *(ptr++) = stats->execution_time_sum_sqr.sec;
7114 *(ptr++) = stats->execution_time_sum_sqr.usec;
7115 *(ptr++) = stats->execution_time_min.sec;
7116 *(ptr++) = stats->execution_time_min.usec;
7117 *(ptr++) = stats->execution_time_max.sec;
7118 *(ptr++) = stats->execution_time_max.usec;
7124 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7129 * IN callerVersion - the rpc stat version of the caller
7131 * OUT myVersion - the rpc stat version of this function
7133 * OUT clock_sec - local time seconds
7135 * OUT clock_usec - local time microseconds
7137 * OUT allocSize - the number of bytes allocated to contain stats
7139 * OUT statCount - the number stats retrieved from this process.
7141 * OUT stats - the actual stats retrieved from this process.
7145 * Returns void. If successful, stats will != NULL.
7149 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7150 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7151 size_t * allocSize, afs_uint32 * statCount,
7152 afs_uint32 ** stats)
7162 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7165 * Check to see if stats are enabled
7168 MUTEX_ENTER(&rx_rpc_stats);
7169 if (!rxi_monitor_processStats) {
7170 MUTEX_EXIT(&rx_rpc_stats);
7174 clock_GetTime(&now);
7175 *clock_sec = now.sec;
7176 *clock_usec = now.usec;
7179 * Allocate the space based upon the caller version
7181 * If the client is at an older version than we are,
7182 * we return the statistic data in the older data format, but
7183 * we still return our version number so the client knows we
7184 * are maintaining more data than it can retrieve.
7187 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7188 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7189 *statCount = rxi_rpc_process_stat_cnt;
7192 * This can't happen yet, but in the future version changes
7193 * can be handled by adding additional code here
7197 if (space > (size_t) 0) {
7199 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7202 rx_interface_stat_p rpc_stat, nrpc_stat;
7206 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7208 * Copy the data based upon the caller version
7210 rx_MarshallProcessRPCStats(callerVersion,
7211 rpc_stat->stats[0].func_total,
7212 rpc_stat->stats, &ptr);
7218 MUTEX_EXIT(&rx_rpc_stats);
7223 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7227 * IN callerVersion - the rpc stat version of the caller
7229 * OUT myVersion - the rpc stat version of this function
7231 * OUT clock_sec - local time seconds
7233 * OUT clock_usec - local time microseconds
7235 * OUT allocSize - the number of bytes allocated to contain stats
7237 * OUT statCount - the number of stats retrieved from the individual
7240 * OUT stats - the actual stats retrieved from the individual peer structures.
7244 * Returns void. If successful, stats will != NULL.
7248 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7249 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7250 size_t * allocSize, afs_uint32 * statCount,
7251 afs_uint32 ** stats)
7261 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7264 * Check to see if stats are enabled
7267 MUTEX_ENTER(&rx_rpc_stats);
7268 if (!rxi_monitor_peerStats) {
7269 MUTEX_EXIT(&rx_rpc_stats);
7273 clock_GetTime(&now);
7274 *clock_sec = now.sec;
7275 *clock_usec = now.usec;
7278 * Allocate the space based upon the caller version
7280 * If the client is at an older version than we are,
7281 * we return the statistic data in the older data format, but
7282 * we still return our version number so the client knows we
7283 * are maintaining more data than it can retrieve.
7286 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7287 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7288 *statCount = rxi_rpc_peer_stat_cnt;
7291 * This can't happen yet, but in the future version changes
7292 * can be handled by adding additional code here
7296 if (space > (size_t) 0) {
7298 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7301 rx_interface_stat_p rpc_stat, nrpc_stat;
7305 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7307 * We have to fix the offset of rpc_stat since we are
7308 * keeping this structure on two rx_queues. The rx_queue
7309 * package assumes that the rx_queue member is the first
7310 * member of the structure. That is, rx_queue assumes that
7311 * any one item is only on one queue at a time. We are
7312 * breaking that assumption and so we have to do a little
7313 * math to fix our pointers.
7316 fix_offset = (char *)rpc_stat;
7317 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7318 rpc_stat = (rx_interface_stat_p) fix_offset;
7321 * Copy the data based upon the caller version
7323 rx_MarshallProcessRPCStats(callerVersion,
7324 rpc_stat->stats[0].func_total,
7325 rpc_stat->stats, &ptr);
7331 MUTEX_EXIT(&rx_rpc_stats);
7336 * rx_FreeRPCStats - free memory allocated by
7337 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7341 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7342 * rx_RetrievePeerRPCStats
7344 * IN allocSize - the number of bytes in stats.
7352 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7354 rxi_Free(stats, allocSize);
7358 * rx_queryProcessRPCStats - see if process rpc stat collection is
7359 * currently enabled.
7365 * Returns 0 if stats are not enabled != 0 otherwise
7369 rx_queryProcessRPCStats(void)
7372 MUTEX_ENTER(&rx_rpc_stats);
7373 rc = rxi_monitor_processStats;
7374 MUTEX_EXIT(&rx_rpc_stats);
7379 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7385 * Returns 0 if stats are not enabled != 0 otherwise
7389 rx_queryPeerRPCStats(void)
7392 MUTEX_ENTER(&rx_rpc_stats);
7393 rc = rxi_monitor_peerStats;
7394 MUTEX_EXIT(&rx_rpc_stats);
7399 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7409 rx_enableProcessRPCStats(void)
7411 MUTEX_ENTER(&rx_rpc_stats);
7412 rx_enable_stats = 1;
7413 rxi_monitor_processStats = 1;
7414 MUTEX_EXIT(&rx_rpc_stats);
7418 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7428 rx_enablePeerRPCStats(void)
7430 MUTEX_ENTER(&rx_rpc_stats);
7431 rx_enable_stats = 1;
7432 rxi_monitor_peerStats = 1;
7433 MUTEX_EXIT(&rx_rpc_stats);
7437 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7447 rx_disableProcessRPCStats(void)
7449 rx_interface_stat_p rpc_stat, nrpc_stat;
7452 MUTEX_ENTER(&rx_rpc_stats);
7455 * Turn off process statistics and if peer stats is also off, turn
7459 rxi_monitor_processStats = 0;
7460 if (rxi_monitor_peerStats == 0) {
7461 rx_enable_stats = 0;
7464 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7465 unsigned int num_funcs = 0;
7468 queue_Remove(rpc_stat);
7469 num_funcs = rpc_stat->stats[0].func_total;
7471 sizeof(rx_interface_stat_t) +
7472 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7474 rxi_Free(rpc_stat, space);
7475 rxi_rpc_process_stat_cnt -= num_funcs;
7477 MUTEX_EXIT(&rx_rpc_stats);
7481 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7491 rx_disablePeerRPCStats(void)
7493 struct rx_peer **peer_ptr, **peer_end;
7496 MUTEX_ENTER(&rx_rpc_stats);
7499 * Turn off peer statistics and if process stats is also off, turn
7503 rxi_monitor_peerStats = 0;
7504 if (rxi_monitor_processStats == 0) {
7505 rx_enable_stats = 0;
7508 MUTEX_ENTER(&rx_peerHashTable_lock);
7509 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7510 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7512 struct rx_peer *peer, *next, *prev;
7513 for (prev = peer = *peer_ptr; peer; peer = next) {
7515 code = MUTEX_TRYENTER(&peer->peer_lock);
7517 rx_interface_stat_p rpc_stat, nrpc_stat;
7520 (&peer->rpcStats, rpc_stat, nrpc_stat,
7521 rx_interface_stat)) {
7522 unsigned int num_funcs = 0;
7525 queue_Remove(&rpc_stat->queue_header);
7526 queue_Remove(&rpc_stat->all_peers);
7527 num_funcs = rpc_stat->stats[0].func_total;
7529 sizeof(rx_interface_stat_t) +
7530 rpc_stat->stats[0].func_total *
7531 sizeof(rx_function_entry_v1_t);
7533 rxi_Free(rpc_stat, space);
7534 rxi_rpc_peer_stat_cnt -= num_funcs;
7536 MUTEX_EXIT(&peer->peer_lock);
7537 if (prev == *peer_ptr) {
7547 MUTEX_EXIT(&rx_peerHashTable_lock);
7548 MUTEX_EXIT(&rx_rpc_stats);
7552 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7557 * IN clearFlag - flag indicating which stats to clear
7565 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7567 rx_interface_stat_p rpc_stat, nrpc_stat;
7569 MUTEX_ENTER(&rx_rpc_stats);
7571 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7572 unsigned int num_funcs = 0, i;
7573 num_funcs = rpc_stat->stats[0].func_total;
7574 for (i = 0; i < num_funcs; i++) {
7575 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7576 hzero(rpc_stat->stats[i].invocations);
7578 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7579 hzero(rpc_stat->stats[i].bytes_sent);
7581 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7582 hzero(rpc_stat->stats[i].bytes_rcvd);
7584 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7585 rpc_stat->stats[i].queue_time_sum.sec = 0;
7586 rpc_stat->stats[i].queue_time_sum.usec = 0;
7588 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7589 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7590 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7592 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7593 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7594 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7596 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7597 rpc_stat->stats[i].queue_time_max.sec = 0;
7598 rpc_stat->stats[i].queue_time_max.usec = 0;
7600 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7601 rpc_stat->stats[i].execution_time_sum.sec = 0;
7602 rpc_stat->stats[i].execution_time_sum.usec = 0;
7604 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7605 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7606 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7608 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7609 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7610 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7612 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7613 rpc_stat->stats[i].execution_time_max.sec = 0;
7614 rpc_stat->stats[i].execution_time_max.usec = 0;
7619 MUTEX_EXIT(&rx_rpc_stats);
7623 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7628 * IN clearFlag - flag indicating which stats to clear
7636 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7638 rx_interface_stat_p rpc_stat, nrpc_stat;
7640 MUTEX_ENTER(&rx_rpc_stats);
7642 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7643 unsigned int num_funcs = 0, i;
7646 * We have to fix the offset of rpc_stat since we are
7647 * keeping this structure on two rx_queues. The rx_queue
7648 * package assumes that the rx_queue member is the first
7649 * member of the structure. That is, rx_queue assumes that
7650 * any one item is only on one queue at a time. We are
7651 * breaking that assumption and so we have to do a little
7652 * math to fix our pointers.
7655 fix_offset = (char *)rpc_stat;
7656 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7657 rpc_stat = (rx_interface_stat_p) fix_offset;
7659 num_funcs = rpc_stat->stats[0].func_total;
7660 for (i = 0; i < num_funcs; i++) {
7661 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7662 hzero(rpc_stat->stats[i].invocations);
7664 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7665 hzero(rpc_stat->stats[i].bytes_sent);
7667 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7668 hzero(rpc_stat->stats[i].bytes_rcvd);
7670 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7671 rpc_stat->stats[i].queue_time_sum.sec = 0;
7672 rpc_stat->stats[i].queue_time_sum.usec = 0;
7674 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7675 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7676 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7678 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7679 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7680 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7682 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7683 rpc_stat->stats[i].queue_time_max.sec = 0;
7684 rpc_stat->stats[i].queue_time_max.usec = 0;
7686 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7687 rpc_stat->stats[i].execution_time_sum.sec = 0;
7688 rpc_stat->stats[i].execution_time_sum.usec = 0;
7690 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7691 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7692 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7694 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7695 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7696 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7698 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7699 rpc_stat->stats[i].execution_time_max.sec = 0;
7700 rpc_stat->stats[i].execution_time_max.usec = 0;
7705 MUTEX_EXIT(&rx_rpc_stats);
7709 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7710 * is authorized to enable/disable/clear RX statistics.
7712 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7715 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7717 rxi_rxstat_userok = proc;
7721 rx_RxStatUserOk(struct rx_call *call)
7723 if (!rxi_rxstat_userok)
7725 return rxi_rxstat_userok(call);
7730 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
7731 * function in the MSVC runtime DLL (msvcrt.dll).
7733 * Note: the system serializes calls to this function.
7736 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
7737 DWORD reason, /* reason function is being called */
7738 LPVOID reserved) /* reserved for future use */
7741 case DLL_PROCESS_ATTACH:
7742 /* library is being attached to a process */
7746 case DLL_PROCESS_DETACH: