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;
385 if (rxinit_status == 0) {
386 tmp_status = rxinit_status;
388 return tmp_status; /* Already started; return previous error code. */
394 if (afs_winsockInit() < 0)
400 * Initialize anything necessary to provide a non-premptive threading
403 rxi_InitializeThreadSupport();
406 /* Allocate and initialize a socket for client and perhaps server
409 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
410 if (rx_socket == OSI_NULLSOCKET) {
414 #ifdef RX_ENABLE_LOCKS
417 #endif /* RX_LOCKS_DB */
418 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
419 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
420 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
421 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
422 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
424 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
426 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
428 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
430 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
432 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
434 #if defined(KERNEL) && defined(AFS_HPUX110_ENV)
436 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
437 #endif /* KERNEL && AFS_HPUX110_ENV */
438 #endif /* RX_ENABLE_LOCKS */
441 rx_connDeadTime = 12;
442 rx_tranquil = 0; /* reset flag */
443 memset((char *)&rx_stats, 0, sizeof(struct rx_stats));
445 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
446 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
447 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
448 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
449 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
450 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
452 /* Malloc up a bunch of packets & buffers */
454 queue_Init(&rx_freePacketQueue);
455 rxi_NeedMorePackets = FALSE;
456 #ifdef RX_ENABLE_TSFPQ
457 rx_nPackets = 0; /* in TSFPQ version, rx_nPackets is managed by rxi_MorePackets* */
458 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
459 #else /* RX_ENABLE_TSFPQ */
460 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
461 rxi_MorePackets(rx_nPackets);
462 #endif /* RX_ENABLE_TSFPQ */
469 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
470 tv.tv_sec = clock_now.sec;
471 tv.tv_usec = clock_now.usec;
472 srand((unsigned int)tv.tv_usec);
479 #if defined(KERNEL) && !defined(UKERNEL)
480 /* Really, this should never happen in a real kernel */
483 struct sockaddr_in addr;
484 int addrlen = sizeof(addr);
485 if (getsockname((int)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
489 rx_port = addr.sin_port;
492 rx_stats.minRtt.sec = 9999999;
494 rx_SetEpoch(tv.tv_sec | 0x80000000);
496 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
497 * will provide a randomer value. */
499 MUTEX_ENTER(&rx_stats_mutex);
500 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
501 MUTEX_EXIT(&rx_stats_mutex);
502 /* *Slightly* random start time for the cid. This is just to help
503 * out with the hashing function at the peer */
504 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
505 rx_connHashTable = (struct rx_connection **)htable;
506 rx_peerHashTable = (struct rx_peer **)ptable;
508 rx_lastAckDelay.sec = 0;
509 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
510 rx_hardAckDelay.sec = 0;
511 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
512 rx_softAckDelay.sec = 0;
513 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
515 rxevent_Init(20, rxi_ReScheduleEvents);
517 /* Initialize various global queues */
518 queue_Init(&rx_idleServerQueue);
519 queue_Init(&rx_incomingCallQueue);
520 queue_Init(&rx_freeCallQueue);
522 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
523 /* Initialize our list of usable IP addresses. */
527 /* Start listener process (exact function is dependent on the
528 * implementation environment--kernel or user space) */
532 tmp_status = rxinit_status = 0;
540 return rx_InitHost(htonl(INADDR_ANY), port);
543 /* called with unincremented nRequestsRunning to see if it is OK to start
544 * a new thread in this service. Could be "no" for two reasons: over the
545 * max quota, or would prevent others from reaching their min quota.
547 #ifdef RX_ENABLE_LOCKS
548 /* This verion of QuotaOK reserves quota if it's ok while the
549 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
552 QuotaOK(register struct rx_service *aservice)
554 /* check if over max quota */
555 if (aservice->nRequestsRunning >= aservice->maxProcs) {
559 /* under min quota, we're OK */
560 /* otherwise, can use only if there are enough to allow everyone
561 * to go to their min quota after this guy starts.
563 MUTEX_ENTER(&rx_stats_mutex);
564 if ((aservice->nRequestsRunning < aservice->minProcs)
565 || (rxi_availProcs > rxi_minDeficit)) {
566 aservice->nRequestsRunning++;
567 /* just started call in minProcs pool, need fewer to maintain
569 if (aservice->nRequestsRunning <= aservice->minProcs)
572 MUTEX_EXIT(&rx_stats_mutex);
575 MUTEX_EXIT(&rx_stats_mutex);
581 ReturnToServerPool(register struct rx_service *aservice)
583 aservice->nRequestsRunning--;
584 MUTEX_ENTER(&rx_stats_mutex);
585 if (aservice->nRequestsRunning < aservice->minProcs)
588 MUTEX_EXIT(&rx_stats_mutex);
591 #else /* RX_ENABLE_LOCKS */
593 QuotaOK(register struct rx_service *aservice)
596 /* under min quota, we're OK */
597 if (aservice->nRequestsRunning < aservice->minProcs)
600 /* check if over max quota */
601 if (aservice->nRequestsRunning >= aservice->maxProcs)
604 /* otherwise, can use only if there are enough to allow everyone
605 * to go to their min quota after this guy starts.
607 if (rxi_availProcs > rxi_minDeficit)
611 #endif /* RX_ENABLE_LOCKS */
614 /* Called by rx_StartServer to start up lwp's to service calls.
615 NExistingProcs gives the number of procs already existing, and which
616 therefore needn't be created. */
618 rxi_StartServerProcs(int nExistingProcs)
620 register struct rx_service *service;
625 /* For each service, reserve N processes, where N is the "minimum"
626 * number of processes that MUST be able to execute a request in parallel,
627 * at any time, for that process. Also compute the maximum difference
628 * between any service's maximum number of processes that can run
629 * (i.e. the maximum number that ever will be run, and a guarantee
630 * that this number will run if other services aren't running), and its
631 * minimum number. The result is the extra number of processes that
632 * we need in order to provide the latter guarantee */
633 for (i = 0; i < RX_MAX_SERVICES; i++) {
635 service = rx_services[i];
636 if (service == (struct rx_service *)0)
638 nProcs += service->minProcs;
639 diff = service->maxProcs - service->minProcs;
643 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
644 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
645 for (i = 0; i < nProcs; i++) {
646 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
652 /* This routine is only required on Windows */
654 rx_StartClientThread(void)
656 #ifdef AFS_PTHREAD_ENV
658 pid = pthread_self();
659 #endif /* AFS_PTHREAD_ENV */
661 #endif /* AFS_NT40_ENV */
663 /* This routine must be called if any services are exported. If the
664 * donateMe flag is set, the calling process is donated to the server
667 rx_StartServer(int donateMe)
669 register struct rx_service *service;
675 /* Start server processes, if necessary (exact function is dependent
676 * on the implementation environment--kernel or user space). DonateMe
677 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
678 * case, one less new proc will be created rx_StartServerProcs.
680 rxi_StartServerProcs(donateMe);
682 /* count up the # of threads in minProcs, and add set the min deficit to
683 * be that value, too.
685 for (i = 0; i < RX_MAX_SERVICES; i++) {
686 service = rx_services[i];
687 if (service == (struct rx_service *)0)
689 MUTEX_ENTER(&rx_stats_mutex);
690 rxi_totalMin += service->minProcs;
691 /* below works even if a thread is running, since minDeficit would
692 * still have been decremented and later re-incremented.
694 rxi_minDeficit += service->minProcs;
695 MUTEX_EXIT(&rx_stats_mutex);
698 /* Turn on reaping of idle server connections */
699 rxi_ReapConnections();
708 #ifdef AFS_PTHREAD_ENV
710 pid = (pid_t) pthread_self();
711 #else /* AFS_PTHREAD_ENV */
713 LWP_CurrentProcess(&pid);
714 #endif /* AFS_PTHREAD_ENV */
716 sprintf(name, "srv_%d", ++nProcs);
718 (*registerProgram) (pid, name);
720 #endif /* AFS_NT40_ENV */
721 rx_ServerProc(NULL); /* Never returns */
723 #ifdef RX_ENABLE_TSFPQ
724 /* no use leaving packets around in this thread's local queue if
725 * it isn't getting donated to the server thread pool.
727 rxi_FlushLocalPacketsTSFPQ();
728 #endif /* RX_ENABLE_TSFPQ */
732 /* Create a new client connection to the specified service, using the
733 * specified security object to implement the security model for this
735 struct rx_connection *
736 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
737 register struct rx_securityClass *securityObject,
738 int serviceSecurityIndex)
742 register struct rx_connection *conn;
747 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", ntohl(shost), ntohs(sport), sservice, securityObject, serviceSecurityIndex));
749 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
750 * the case of kmem_alloc? */
751 conn = rxi_AllocConnection();
752 #ifdef RX_ENABLE_LOCKS
753 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
754 MUTEX_INIT(&conn->conn_data_lock, "conn call lock", MUTEX_DEFAULT, 0);
755 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
758 MUTEX_ENTER(&rx_connHashTable_lock);
759 cid = (rx_nextCid += RX_MAXCALLS);
760 conn->type = RX_CLIENT_CONNECTION;
762 conn->epoch = rx_epoch;
763 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
764 conn->serviceId = sservice;
765 conn->securityObject = securityObject;
766 conn->securityData = (void *) 0;
767 conn->securityIndex = serviceSecurityIndex;
768 rx_SetConnDeadTime(conn, rx_connDeadTime);
769 conn->ackRate = RX_FAST_ACK_RATE;
771 conn->specific = NULL;
772 conn->challengeEvent = NULL;
773 conn->delayedAbortEvent = NULL;
774 conn->abortCount = 0;
776 for (i = 0; i < RX_MAXCALLS; i++) {
777 conn->twind[i] = rx_initSendWindow;
778 conn->rwind[i] = rx_initReceiveWindow;
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(htonl(INADDR_ANY), 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 call->rwind = conn->rwind[channel];
2157 call->twind = conn->twind[channel];
2158 /* Note that the next expected call number is retained (in
2159 * conn->callNumber[i]), even if we reallocate the call structure
2161 conn->call[channel] = call;
2162 /* if the channel's never been used (== 0), we should start at 1, otherwise
2163 * the call number is valid from the last time this channel was used */
2164 if (*call->callNumber == 0)
2165 *call->callNumber = 1;
2170 /* A call has been inactive long enough that so we can throw away
2171 * state, including the call structure, which is placed on the call
2173 * Call is locked upon entry.
2174 * haveCTLock set if called from rxi_ReapConnections
2176 #ifdef RX_ENABLE_LOCKS
2178 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2179 #else /* RX_ENABLE_LOCKS */
2181 rxi_FreeCall(register struct rx_call *call)
2182 #endif /* RX_ENABLE_LOCKS */
2184 register int channel = call->channel;
2185 register struct rx_connection *conn = call->conn;
2188 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2189 (*call->callNumber)++;
2190 rxi_ResetCall(call, 0);
2191 call->conn->call[channel] = (struct rx_call *)0;
2193 MUTEX_ENTER(&rx_freeCallQueue_lock);
2194 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2195 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2196 /* A call may be free even though its transmit queue is still in use.
2197 * Since we search the call list from head to tail, put busy calls at
2198 * the head of the list, and idle calls at the tail.
2200 if (call->flags & RX_CALL_TQ_BUSY)
2201 queue_Prepend(&rx_freeCallQueue, call);
2203 queue_Append(&rx_freeCallQueue, call);
2204 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2205 queue_Append(&rx_freeCallQueue, call);
2206 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2207 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2208 MUTEX_EXIT(&rx_freeCallQueue_lock);
2210 /* Destroy the connection if it was previously slated for
2211 * destruction, i.e. the Rx client code previously called
2212 * rx_DestroyConnection (client connections), or
2213 * rxi_ReapConnections called the same routine (server
2214 * connections). Only do this, however, if there are no
2215 * outstanding calls. Note that for fine grain locking, there appears
2216 * to be a deadlock in that rxi_FreeCall has a call locked and
2217 * DestroyConnectionNoLock locks each call in the conn. But note a
2218 * few lines up where we have removed this call from the conn.
2219 * If someone else destroys a connection, they either have no
2220 * call lock held or are going through this section of code.
2222 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2223 MUTEX_ENTER(&conn->conn_data_lock);
2225 MUTEX_EXIT(&conn->conn_data_lock);
2226 #ifdef RX_ENABLE_LOCKS
2228 rxi_DestroyConnectionNoLock(conn);
2230 rxi_DestroyConnection(conn);
2231 #else /* RX_ENABLE_LOCKS */
2232 rxi_DestroyConnection(conn);
2233 #endif /* RX_ENABLE_LOCKS */
2237 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2239 rxi_Alloc(register size_t size)
2243 rx_MutexAdd1Increment2(rxi_Allocsize, (afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2244 p = (char *)osi_Alloc(size);
2247 osi_Panic("rxi_Alloc error");
2253 rxi_Free(void *addr, register size_t size)
2255 rx_MutexAdd1Decrement2(rxi_Allocsize, -(afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2256 osi_Free(addr, size);
2259 /* Find the peer process represented by the supplied (host,port)
2260 * combination. If there is no appropriate active peer structure, a
2261 * new one will be allocated and initialized
2262 * The origPeer, if set, is a pointer to a peer structure on which the
2263 * refcount will be be decremented. This is used to replace the peer
2264 * structure hanging off a connection structure */
2266 rxi_FindPeer(register afs_uint32 host, register u_short port,
2267 struct rx_peer *origPeer, int create)
2269 register struct rx_peer *pp;
2271 hashIndex = PEER_HASH(host, port);
2272 MUTEX_ENTER(&rx_peerHashTable_lock);
2273 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2274 if ((pp->host == host) && (pp->port == port))
2279 pp = rxi_AllocPeer(); /* This bzero's *pp */
2280 pp->host = host; /* set here or in InitPeerParams is zero */
2282 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2283 queue_Init(&pp->congestionQueue);
2284 queue_Init(&pp->rpcStats);
2285 pp->next = rx_peerHashTable[hashIndex];
2286 rx_peerHashTable[hashIndex] = pp;
2287 rxi_InitPeerParams(pp);
2288 rx_MutexIncrement(rx_stats.nPeerStructs, rx_stats_mutex);
2295 origPeer->refCount--;
2296 MUTEX_EXIT(&rx_peerHashTable_lock);
2301 /* Find the connection at (host, port) started at epoch, and with the
2302 * given connection id. Creates the server connection if necessary.
2303 * The type specifies whether a client connection or a server
2304 * connection is desired. In both cases, (host, port) specify the
2305 * peer's (host, pair) pair. Client connections are not made
2306 * automatically by this routine. The parameter socket gives the
2307 * socket descriptor on which the packet was received. This is used,
2308 * in the case of server connections, to check that *new* connections
2309 * come via a valid (port, serviceId). Finally, the securityIndex
2310 * parameter must match the existing index for the connection. If a
2311 * server connection is created, it will be created using the supplied
2312 * index, if the index is valid for this service */
2313 struct rx_connection *
2314 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2315 register u_short port, u_short serviceId, afs_uint32 cid,
2316 afs_uint32 epoch, int type, u_int securityIndex)
2318 int hashindex, flag, i;
2319 register struct rx_connection *conn;
2320 hashindex = CONN_HASH(host, port, cid, epoch, type);
2321 MUTEX_ENTER(&rx_connHashTable_lock);
2322 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2323 rx_connHashTable[hashindex],
2326 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2327 && (epoch == conn->epoch)) {
2328 register struct rx_peer *pp = conn->peer;
2329 if (securityIndex != conn->securityIndex) {
2330 /* this isn't supposed to happen, but someone could forge a packet
2331 * like this, and there seems to be some CM bug that makes this
2332 * happen from time to time -- in which case, the fileserver
2334 MUTEX_EXIT(&rx_connHashTable_lock);
2335 return (struct rx_connection *)0;
2337 if (pp->host == host && pp->port == port)
2339 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2341 /* So what happens when it's a callback connection? */
2342 if ( /*type == RX_CLIENT_CONNECTION && */
2343 (conn->epoch & 0x80000000))
2347 /* the connection rxLastConn that was used the last time is not the
2348 ** one we are looking for now. Hence, start searching in the hash */
2350 conn = rx_connHashTable[hashindex];
2355 struct rx_service *service;
2356 if (type == RX_CLIENT_CONNECTION) {
2357 MUTEX_EXIT(&rx_connHashTable_lock);
2358 return (struct rx_connection *)0;
2360 service = rxi_FindService(socket, serviceId);
2361 if (!service || (securityIndex >= service->nSecurityObjects)
2362 || (service->securityObjects[securityIndex] == 0)) {
2363 MUTEX_EXIT(&rx_connHashTable_lock);
2364 return (struct rx_connection *)0;
2366 conn = rxi_AllocConnection(); /* This bzero's the connection */
2367 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2368 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2369 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2370 conn->next = rx_connHashTable[hashindex];
2371 rx_connHashTable[hashindex] = conn;
2372 conn->peer = rxi_FindPeer(host, port, 0, 1);
2373 conn->type = RX_SERVER_CONNECTION;
2374 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2375 conn->epoch = epoch;
2376 conn->cid = cid & RX_CIDMASK;
2377 /* conn->serial = conn->lastSerial = 0; */
2378 /* conn->timeout = 0; */
2379 conn->ackRate = RX_FAST_ACK_RATE;
2380 conn->service = service;
2381 conn->serviceId = serviceId;
2382 conn->securityIndex = securityIndex;
2383 conn->securityObject = service->securityObjects[securityIndex];
2384 conn->nSpecific = 0;
2385 conn->specific = NULL;
2386 rx_SetConnDeadTime(conn, service->connDeadTime);
2387 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2388 for (i = 0; i < RX_MAXCALLS; i++) {
2389 conn->twind[i] = rx_initSendWindow;
2390 conn->rwind[i] = rx_initReceiveWindow;
2392 /* Notify security object of the new connection */
2393 RXS_NewConnection(conn->securityObject, conn);
2394 /* XXXX Connection timeout? */
2395 if (service->newConnProc)
2396 (*service->newConnProc) (conn);
2397 rx_MutexIncrement(rx_stats.nServerConns, rx_stats_mutex);
2400 MUTEX_ENTER(&conn->conn_data_lock);
2402 MUTEX_EXIT(&conn->conn_data_lock);
2404 rxLastConn = conn; /* store this connection as the last conn used */
2405 MUTEX_EXIT(&rx_connHashTable_lock);
2409 /* There are two packet tracing routines available for testing and monitoring
2410 * Rx. One is called just after every packet is received and the other is
2411 * called just before every packet is sent. Received packets, have had their
2412 * headers decoded, and packets to be sent have not yet had their headers
2413 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2414 * containing the network address. Both can be modified. The return value, if
2415 * non-zero, indicates that the packet should be dropped. */
2417 int (*rx_justReceived) () = 0;
2418 int (*rx_almostSent) () = 0;
2420 /* A packet has been received off the interface. Np is the packet, socket is
2421 * the socket number it was received from (useful in determining which service
2422 * this packet corresponds to), and (host, port) reflect the host,port of the
2423 * sender. This call returns the packet to the caller if it is finished with
2424 * it, rather than de-allocating it, just as a small performance hack */
2427 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2428 afs_uint32 host, u_short port, int *tnop,
2429 struct rx_call **newcallp)
2431 register struct rx_call *call;
2432 register struct rx_connection *conn;
2434 afs_uint32 currentCallNumber;
2440 struct rx_packet *tnp;
2443 /* We don't print out the packet until now because (1) the time may not be
2444 * accurate enough until now in the lwp implementation (rx_Listener only gets
2445 * the time after the packet is read) and (2) from a protocol point of view,
2446 * this is the first time the packet has been seen */
2447 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2448 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2449 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2450 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2451 np->header.epoch, np->header.cid, np->header.callNumber,
2452 np->header.seq, np->header.flags, np));
2455 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2456 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2459 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2460 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2463 /* If an input tracer function is defined, call it with the packet and
2464 * network address. Note this function may modify its arguments. */
2465 if (rx_justReceived) {
2466 struct sockaddr_in addr;
2468 addr.sin_family = AF_INET;
2469 addr.sin_port = port;
2470 addr.sin_addr.s_addr = host;
2471 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2472 addr.sin_len = sizeof(addr);
2473 #endif /* AFS_OSF_ENV */
2474 drop = (*rx_justReceived) (np, &addr);
2475 /* drop packet if return value is non-zero */
2478 port = addr.sin_port; /* in case fcn changed addr */
2479 host = addr.sin_addr.s_addr;
2483 /* If packet was not sent by the client, then *we* must be the client */
2484 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2485 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2487 /* Find the connection (or fabricate one, if we're the server & if
2488 * necessary) associated with this packet */
2490 rxi_FindConnection(socket, host, port, np->header.serviceId,
2491 np->header.cid, np->header.epoch, type,
2492 np->header.securityIndex);
2495 /* If no connection found or fabricated, just ignore the packet.
2496 * (An argument could be made for sending an abort packet for
2501 MUTEX_ENTER(&conn->conn_data_lock);
2502 if (conn->maxSerial < np->header.serial)
2503 conn->maxSerial = np->header.serial;
2504 MUTEX_EXIT(&conn->conn_data_lock);
2506 /* If the connection is in an error state, send an abort packet and ignore
2507 * the incoming packet */
2509 /* Don't respond to an abort packet--we don't want loops! */
2510 MUTEX_ENTER(&conn->conn_data_lock);
2511 if (np->header.type != RX_PACKET_TYPE_ABORT)
2512 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2514 MUTEX_EXIT(&conn->conn_data_lock);
2518 /* Check for connection-only requests (i.e. not call specific). */
2519 if (np->header.callNumber == 0) {
2520 switch (np->header.type) {
2521 case RX_PACKET_TYPE_ABORT: {
2522 /* What if the supplied error is zero? */
2523 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2524 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2525 rxi_ConnectionError(conn, errcode);
2526 MUTEX_ENTER(&conn->conn_data_lock);
2528 MUTEX_EXIT(&conn->conn_data_lock);
2531 case RX_PACKET_TYPE_CHALLENGE:
2532 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2533 MUTEX_ENTER(&conn->conn_data_lock);
2535 MUTEX_EXIT(&conn->conn_data_lock);
2537 case RX_PACKET_TYPE_RESPONSE:
2538 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2539 MUTEX_ENTER(&conn->conn_data_lock);
2541 MUTEX_EXIT(&conn->conn_data_lock);
2543 case RX_PACKET_TYPE_PARAMS:
2544 case RX_PACKET_TYPE_PARAMS + 1:
2545 case RX_PACKET_TYPE_PARAMS + 2:
2546 /* ignore these packet types for now */
2547 MUTEX_ENTER(&conn->conn_data_lock);
2549 MUTEX_EXIT(&conn->conn_data_lock);
2554 /* Should not reach here, unless the peer is broken: send an
2556 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2557 MUTEX_ENTER(&conn->conn_data_lock);
2558 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2560 MUTEX_EXIT(&conn->conn_data_lock);
2565 channel = np->header.cid & RX_CHANNELMASK;
2566 call = conn->call[channel];
2567 #ifdef RX_ENABLE_LOCKS
2569 MUTEX_ENTER(&call->lock);
2570 /* Test to see if call struct is still attached to conn. */
2571 if (call != conn->call[channel]) {
2573 MUTEX_EXIT(&call->lock);
2574 if (type == RX_SERVER_CONNECTION) {
2575 call = conn->call[channel];
2576 /* If we started with no call attached and there is one now,
2577 * another thread is also running this routine and has gotten
2578 * the connection channel. We should drop this packet in the tests
2579 * below. If there was a call on this connection and it's now
2580 * gone, then we'll be making a new call below.
2581 * If there was previously a call and it's now different then
2582 * the old call was freed and another thread running this routine
2583 * has created a call on this channel. One of these two threads
2584 * has a packet for the old call and the code below handles those
2588 MUTEX_ENTER(&call->lock);
2590 /* This packet can't be for this call. If the new call address is
2591 * 0 then no call is running on this channel. If there is a call
2592 * then, since this is a client connection we're getting data for
2593 * it must be for the previous call.
2595 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2596 MUTEX_ENTER(&conn->conn_data_lock);
2598 MUTEX_EXIT(&conn->conn_data_lock);
2603 currentCallNumber = conn->callNumber[channel];
2605 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2606 if (np->header.callNumber < currentCallNumber) {
2607 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2608 #ifdef RX_ENABLE_LOCKS
2610 MUTEX_EXIT(&call->lock);
2612 MUTEX_ENTER(&conn->conn_data_lock);
2614 MUTEX_EXIT(&conn->conn_data_lock);
2618 MUTEX_ENTER(&conn->conn_call_lock);
2619 call = rxi_NewCall(conn, channel);
2620 MUTEX_EXIT(&conn->conn_call_lock);
2621 *call->callNumber = np->header.callNumber;
2622 if (np->header.callNumber == 0)
2623 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));
2625 call->state = RX_STATE_PRECALL;
2626 clock_GetTime(&call->queueTime);
2627 hzero(call->bytesSent);
2628 hzero(call->bytesRcvd);
2630 * If the number of queued calls exceeds the overload
2631 * threshold then abort this call.
2633 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2634 struct rx_packet *tp;
2636 rxi_CallError(call, rx_BusyError);
2637 tp = rxi_SendCallAbort(call, np, 1, 0);
2638 MUTEX_EXIT(&call->lock);
2639 MUTEX_ENTER(&conn->conn_data_lock);
2641 MUTEX_EXIT(&conn->conn_data_lock);
2642 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2645 rxi_KeepAliveOn(call);
2646 } else if (np->header.callNumber != currentCallNumber) {
2647 /* Wait until the transmit queue is idle before deciding
2648 * whether to reset the current call. Chances are that the
2649 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2652 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2653 while ((call->state == RX_STATE_ACTIVE)
2654 && (call->flags & RX_CALL_TQ_BUSY)) {
2655 call->flags |= RX_CALL_TQ_WAIT;
2657 #ifdef RX_ENABLE_LOCKS
2658 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2659 CV_WAIT(&call->cv_tq, &call->lock);
2660 #else /* RX_ENABLE_LOCKS */
2661 osi_rxSleep(&call->tq);
2662 #endif /* RX_ENABLE_LOCKS */
2664 if (call->tqWaiters == 0)
2665 call->flags &= ~RX_CALL_TQ_WAIT;
2667 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2668 /* If the new call cannot be taken right now send a busy and set
2669 * the error condition in this call, so that it terminates as
2670 * quickly as possible */
2671 if (call->state == RX_STATE_ACTIVE) {
2672 struct rx_packet *tp;
2674 rxi_CallError(call, RX_CALL_DEAD);
2675 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2677 MUTEX_EXIT(&call->lock);
2678 MUTEX_ENTER(&conn->conn_data_lock);
2680 MUTEX_EXIT(&conn->conn_data_lock);
2683 rxi_ResetCall(call, 0);
2684 *call->callNumber = np->header.callNumber;
2685 if (np->header.callNumber == 0)
2686 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));
2688 call->state = RX_STATE_PRECALL;
2689 clock_GetTime(&call->queueTime);
2690 hzero(call->bytesSent);
2691 hzero(call->bytesRcvd);
2693 * If the number of queued calls exceeds the overload
2694 * threshold then abort this call.
2696 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2697 struct rx_packet *tp;
2699 rxi_CallError(call, rx_BusyError);
2700 tp = rxi_SendCallAbort(call, np, 1, 0);
2701 MUTEX_EXIT(&call->lock);
2702 MUTEX_ENTER(&conn->conn_data_lock);
2704 MUTEX_EXIT(&conn->conn_data_lock);
2705 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2708 rxi_KeepAliveOn(call);
2710 /* Continuing call; do nothing here. */
2712 } else { /* we're the client */
2713 /* Ignore all incoming acknowledgements for calls in DALLY state */
2714 if (call && (call->state == RX_STATE_DALLY)
2715 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2716 rx_MutexIncrement(rx_stats.ignorePacketDally, rx_stats_mutex);
2717 #ifdef RX_ENABLE_LOCKS
2719 MUTEX_EXIT(&call->lock);
2722 MUTEX_ENTER(&conn->conn_data_lock);
2724 MUTEX_EXIT(&conn->conn_data_lock);
2728 /* Ignore anything that's not relevant to the current call. If there
2729 * isn't a current call, then no packet is relevant. */
2730 if (!call || (np->header.callNumber != currentCallNumber)) {
2731 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2732 #ifdef RX_ENABLE_LOCKS
2734 MUTEX_EXIT(&call->lock);
2737 MUTEX_ENTER(&conn->conn_data_lock);
2739 MUTEX_EXIT(&conn->conn_data_lock);
2742 /* If the service security object index stamped in the packet does not
2743 * match the connection's security index, ignore the packet */
2744 if (np->header.securityIndex != conn->securityIndex) {
2745 #ifdef RX_ENABLE_LOCKS
2746 MUTEX_EXIT(&call->lock);
2748 MUTEX_ENTER(&conn->conn_data_lock);
2750 MUTEX_EXIT(&conn->conn_data_lock);
2754 /* If we're receiving the response, then all transmit packets are
2755 * implicitly acknowledged. Get rid of them. */
2756 if (np->header.type == RX_PACKET_TYPE_DATA) {
2757 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2758 /* XXX Hack. Because we must release the global rx lock when
2759 * sending packets (osi_NetSend) we drop all acks while we're
2760 * traversing the tq in rxi_Start sending packets out because
2761 * packets may move to the freePacketQueue as result of being here!
2762 * So we drop these packets until we're safely out of the
2763 * traversing. Really ugly!
2764 * For fine grain RX locking, we set the acked field in the
2765 * packets and let rxi_Start remove them from the transmit queue.
2767 if (call->flags & RX_CALL_TQ_BUSY) {
2768 #ifdef RX_ENABLE_LOCKS
2769 rxi_SetAcksInTransmitQueue(call);
2772 return np; /* xmitting; drop packet */
2775 rxi_ClearTransmitQueue(call, 0);
2777 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2778 rxi_ClearTransmitQueue(call, 0);
2779 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2781 if (np->header.type == RX_PACKET_TYPE_ACK) {
2782 /* now check to see if this is an ack packet acknowledging that the
2783 * server actually *lost* some hard-acked data. If this happens we
2784 * ignore this packet, as it may indicate that the server restarted in
2785 * the middle of a call. It is also possible that this is an old ack
2786 * packet. We don't abort the connection in this case, because this
2787 * *might* just be an old ack packet. The right way to detect a server
2788 * restart in the midst of a call is to notice that the server epoch
2790 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2791 * XXX unacknowledged. I think that this is off-by-one, but
2792 * XXX I don't dare change it just yet, since it will
2793 * XXX interact badly with the server-restart detection
2794 * XXX code in receiveackpacket. */
2795 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2796 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2797 MUTEX_EXIT(&call->lock);
2798 MUTEX_ENTER(&conn->conn_data_lock);
2800 MUTEX_EXIT(&conn->conn_data_lock);
2804 } /* else not a data packet */
2807 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2808 /* Set remote user defined status from packet */
2809 call->remoteStatus = np->header.userStatus;
2811 /* Note the gap between the expected next packet and the actual
2812 * packet that arrived, when the new packet has a smaller serial number
2813 * than expected. Rioses frequently reorder packets all by themselves,
2814 * so this will be quite important with very large window sizes.
2815 * Skew is checked against 0 here to avoid any dependence on the type of
2816 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2818 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2819 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2820 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2822 MUTEX_ENTER(&conn->conn_data_lock);
2823 skew = conn->lastSerial - np->header.serial;
2824 conn->lastSerial = np->header.serial;
2825 MUTEX_EXIT(&conn->conn_data_lock);
2827 register struct rx_peer *peer;
2829 if (skew > peer->inPacketSkew) {
2830 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2832 peer->inPacketSkew = skew;
2836 /* Now do packet type-specific processing */
2837 switch (np->header.type) {
2838 case RX_PACKET_TYPE_DATA:
2839 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2842 case RX_PACKET_TYPE_ACK:
2843 /* Respond immediately to ack packets requesting acknowledgement
2845 if (np->header.flags & RX_REQUEST_ACK) {
2847 (void)rxi_SendCallAbort(call, 0, 1, 0);
2849 (void)rxi_SendAck(call, 0, np->header.serial,
2850 RX_ACK_PING_RESPONSE, 1);
2852 np = rxi_ReceiveAckPacket(call, np, 1);
2854 case RX_PACKET_TYPE_ABORT: {
2855 /* An abort packet: reset the call, passing the error up to the user. */
2856 /* What if error is zero? */
2857 /* What if the error is -1? the application will treat it as a timeout. */
2858 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
2859 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
2860 rxi_CallError(call, errdata);
2861 MUTEX_EXIT(&call->lock);
2862 MUTEX_ENTER(&conn->conn_data_lock);
2864 MUTEX_EXIT(&conn->conn_data_lock);
2865 return np; /* xmitting; drop packet */
2867 case RX_PACKET_TYPE_BUSY:
2870 case RX_PACKET_TYPE_ACKALL:
2871 /* All packets acknowledged, so we can drop all packets previously
2872 * readied for sending */
2873 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2874 /* XXX Hack. We because we can't release the global rx lock when
2875 * sending packets (osi_NetSend) we drop all ack pkts while we're
2876 * traversing the tq in rxi_Start sending packets out because
2877 * packets may move to the freePacketQueue as result of being
2878 * here! So we drop these packets until we're safely out of the
2879 * traversing. Really ugly!
2880 * For fine grain RX locking, we set the acked field in the packets
2881 * and let rxi_Start remove the packets from the transmit queue.
2883 if (call->flags & RX_CALL_TQ_BUSY) {
2884 #ifdef RX_ENABLE_LOCKS
2885 rxi_SetAcksInTransmitQueue(call);
2887 #else /* RX_ENABLE_LOCKS */
2888 MUTEX_EXIT(&call->lock);
2889 MUTEX_ENTER(&conn->conn_data_lock);
2891 MUTEX_EXIT(&conn->conn_data_lock);
2892 return np; /* xmitting; drop packet */
2893 #endif /* RX_ENABLE_LOCKS */
2895 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2896 rxi_ClearTransmitQueue(call, 0);
2899 /* Should not reach here, unless the peer is broken: send an abort
2901 rxi_CallError(call, RX_PROTOCOL_ERROR);
2902 np = rxi_SendCallAbort(call, np, 1, 0);
2905 /* Note when this last legitimate packet was received, for keep-alive
2906 * processing. Note, we delay getting the time until now in the hope that
2907 * the packet will be delivered to the user before any get time is required
2908 * (if not, then the time won't actually be re-evaluated here). */
2909 call->lastReceiveTime = clock_Sec();
2910 MUTEX_EXIT(&call->lock);
2911 MUTEX_ENTER(&conn->conn_data_lock);
2913 MUTEX_EXIT(&conn->conn_data_lock);
2917 /* return true if this is an "interesting" connection from the point of view
2918 of someone trying to debug the system */
2920 rxi_IsConnInteresting(struct rx_connection *aconn)
2923 register struct rx_call *tcall;
2925 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2927 for (i = 0; i < RX_MAXCALLS; i++) {
2928 tcall = aconn->call[i];
2930 if ((tcall->state == RX_STATE_PRECALL)
2931 || (tcall->state == RX_STATE_ACTIVE))
2933 if ((tcall->mode == RX_MODE_SENDING)
2934 || (tcall->mode == RX_MODE_RECEIVING))
2942 /* if this is one of the last few packets AND it wouldn't be used by the
2943 receiving call to immediately satisfy a read request, then drop it on
2944 the floor, since accepting it might prevent a lock-holding thread from
2945 making progress in its reading. If a call has been cleared while in
2946 the precall state then ignore all subsequent packets until the call
2947 is assigned to a thread. */
2950 TooLow(struct rx_packet *ap, struct rx_call *acall)
2953 MUTEX_ENTER(&rx_stats_mutex);
2954 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
2955 && (acall->state == RX_STATE_PRECALL))
2956 || ((rx_nFreePackets < rxi_dataQuota + 2)
2957 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
2958 && (acall->flags & RX_CALL_READER_WAIT)))) {
2961 MUTEX_EXIT(&rx_stats_mutex);
2967 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
2968 struct rx_call *acall)
2970 struct rx_call *call = acall;
2971 struct clock when, now;
2974 MUTEX_ENTER(&conn->conn_data_lock);
2975 conn->checkReachEvent = NULL;
2976 waiting = conn->flags & RX_CONN_ATTACHWAIT;
2979 MUTEX_EXIT(&conn->conn_data_lock);
2983 MUTEX_ENTER(&conn->conn_call_lock);
2984 MUTEX_ENTER(&conn->conn_data_lock);
2985 for (i = 0; i < RX_MAXCALLS; i++) {
2986 struct rx_call *tc = conn->call[i];
2987 if (tc && tc->state == RX_STATE_PRECALL) {
2993 /* Indicate that rxi_CheckReachEvent is no longer running by
2994 * clearing the flag. Must be atomic under conn_data_lock to
2995 * avoid a new call slipping by: rxi_CheckConnReach holds
2996 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
2998 conn->flags &= ~RX_CONN_ATTACHWAIT;
2999 MUTEX_EXIT(&conn->conn_data_lock);
3000 MUTEX_EXIT(&conn->conn_call_lock);
3005 MUTEX_ENTER(&call->lock);
3006 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3008 MUTEX_EXIT(&call->lock);
3010 clock_GetTime(&now);
3012 when.sec += RX_CHECKREACH_TIMEOUT;
3013 MUTEX_ENTER(&conn->conn_data_lock);
3014 if (!conn->checkReachEvent) {
3016 conn->checkReachEvent =
3017 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3020 MUTEX_EXIT(&conn->conn_data_lock);
3026 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3028 struct rx_service *service = conn->service;
3029 struct rx_peer *peer = conn->peer;
3030 afs_uint32 now, lastReach;
3032 if (service->checkReach == 0)
3036 MUTEX_ENTER(&peer->peer_lock);
3037 lastReach = peer->lastReachTime;
3038 MUTEX_EXIT(&peer->peer_lock);
3039 if (now - lastReach < RX_CHECKREACH_TTL)
3042 MUTEX_ENTER(&conn->conn_data_lock);
3043 if (conn->flags & RX_CONN_ATTACHWAIT) {
3044 MUTEX_EXIT(&conn->conn_data_lock);
3047 conn->flags |= RX_CONN_ATTACHWAIT;
3048 MUTEX_EXIT(&conn->conn_data_lock);
3049 if (!conn->checkReachEvent)
3050 rxi_CheckReachEvent(NULL, conn, call);
3055 /* try to attach call, if authentication is complete */
3057 TryAttach(register struct rx_call *acall, register osi_socket socket,
3058 register int *tnop, register struct rx_call **newcallp,
3061 struct rx_connection *conn = acall->conn;
3063 if (conn->type == RX_SERVER_CONNECTION
3064 && acall->state == RX_STATE_PRECALL) {
3065 /* Don't attach until we have any req'd. authentication. */
3066 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3067 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3068 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3069 /* Note: this does not necessarily succeed; there
3070 * may not any proc available
3073 rxi_ChallengeOn(acall->conn);
3078 /* A data packet has been received off the interface. This packet is
3079 * appropriate to the call (the call is in the right state, etc.). This
3080 * routine can return a packet to the caller, for re-use */
3083 rxi_ReceiveDataPacket(register struct rx_call *call,
3084 register struct rx_packet *np, int istack,
3085 osi_socket socket, afs_uint32 host, u_short port,
3086 int *tnop, struct rx_call **newcallp)
3088 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3092 afs_uint32 seq, serial, flags;
3094 struct rx_packet *tnp;
3095 struct clock when, now;
3096 rx_MutexIncrement(rx_stats.dataPacketsRead, rx_stats_mutex);
3099 /* If there are no packet buffers, drop this new packet, unless we can find
3100 * packet buffers from inactive calls */
3102 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3103 MUTEX_ENTER(&rx_freePktQ_lock);
3104 rxi_NeedMorePackets = TRUE;
3105 MUTEX_EXIT(&rx_freePktQ_lock);
3106 rx_MutexIncrement(rx_stats.noPacketBuffersOnRead, rx_stats_mutex);
3107 call->rprev = np->header.serial;
3108 rxi_calltrace(RX_TRACE_DROP, call);
3109 dpf(("packet %x dropped on receipt - quota problems", np));
3111 rxi_ClearReceiveQueue(call);
3112 clock_GetTime(&now);
3114 clock_Add(&when, &rx_softAckDelay);
3115 if (!call->delayedAckEvent
3116 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3117 rxevent_Cancel(call->delayedAckEvent, call,
3118 RX_CALL_REFCOUNT_DELAY);
3119 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3120 call->delayedAckEvent =
3121 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3123 /* we've damaged this call already, might as well do it in. */
3129 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3130 * packet is one of several packets transmitted as a single
3131 * datagram. Do not send any soft or hard acks until all packets
3132 * in a jumbogram have been processed. Send negative acks right away.
3134 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3135 /* tnp is non-null when there are more packets in the
3136 * current jumbo gram */
3143 seq = np->header.seq;
3144 serial = np->header.serial;
3145 flags = np->header.flags;
3147 /* If the call is in an error state, send an abort message */
3149 return rxi_SendCallAbort(call, np, istack, 0);
3151 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3152 * AFS 3.5 jumbogram. */
3153 if (flags & RX_JUMBO_PACKET) {
3154 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3159 if (np->header.spare != 0) {
3160 MUTEX_ENTER(&call->conn->conn_data_lock);
3161 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3162 MUTEX_EXIT(&call->conn->conn_data_lock);
3165 /* The usual case is that this is the expected next packet */
3166 if (seq == call->rnext) {
3168 /* Check to make sure it is not a duplicate of one already queued */
3169 if (queue_IsNotEmpty(&call->rq)
3170 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3171 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3172 dpf(("packet %x dropped on receipt - duplicate", np));
3173 rxevent_Cancel(call->delayedAckEvent, call,
3174 RX_CALL_REFCOUNT_DELAY);
3175 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3181 /* It's the next packet. Stick it on the receive queue
3182 * for this call. Set newPackets to make sure we wake
3183 * the reader once all packets have been processed */
3184 queue_Prepend(&call->rq, np);
3186 np = NULL; /* We can't use this anymore */
3189 /* If an ack is requested then set a flag to make sure we
3190 * send an acknowledgement for this packet */
3191 if (flags & RX_REQUEST_ACK) {
3192 ackNeeded = RX_ACK_REQUESTED;
3195 /* Keep track of whether we have received the last packet */
3196 if (flags & RX_LAST_PACKET) {
3197 call->flags |= RX_CALL_HAVE_LAST;
3201 /* Check whether we have all of the packets for this call */
3202 if (call->flags & RX_CALL_HAVE_LAST) {
3203 afs_uint32 tseq; /* temporary sequence number */
3204 struct rx_packet *tp; /* Temporary packet pointer */
3205 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3207 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3208 if (tseq != tp->header.seq)
3210 if (tp->header.flags & RX_LAST_PACKET) {
3211 call->flags |= RX_CALL_RECEIVE_DONE;
3218 /* Provide asynchronous notification for those who want it
3219 * (e.g. multi rx) */
3220 if (call->arrivalProc) {
3221 (*call->arrivalProc) (call, call->arrivalProcHandle,
3222 call->arrivalProcArg);
3223 call->arrivalProc = (void (*)())0;
3226 /* Update last packet received */
3229 /* If there is no server process serving this call, grab
3230 * one, if available. We only need to do this once. If a
3231 * server thread is available, this thread becomes a server
3232 * thread and the server thread becomes a listener thread. */
3234 TryAttach(call, socket, tnop, newcallp, 0);
3237 /* This is not the expected next packet. */
3239 /* Determine whether this is a new or old packet, and if it's
3240 * a new one, whether it fits into the current receive window.
3241 * Also figure out whether the packet was delivered in sequence.
3242 * We use the prev variable to determine whether the new packet
3243 * is the successor of its immediate predecessor in the
3244 * receive queue, and the missing flag to determine whether
3245 * any of this packets predecessors are missing. */
3247 afs_uint32 prev; /* "Previous packet" sequence number */
3248 struct rx_packet *tp; /* Temporary packet pointer */
3249 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3250 int missing; /* Are any predecessors missing? */
3252 /* If the new packet's sequence number has been sent to the
3253 * application already, then this is a duplicate */
3254 if (seq < call->rnext) {
3255 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3256 rxevent_Cancel(call->delayedAckEvent, call,
3257 RX_CALL_REFCOUNT_DELAY);
3258 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3264 /* If the sequence number is greater than what can be
3265 * accomodated by the current window, then send a negative
3266 * acknowledge and drop the packet */
3267 if ((call->rnext + call->rwind) <= seq) {
3268 rxevent_Cancel(call->delayedAckEvent, call,
3269 RX_CALL_REFCOUNT_DELAY);
3270 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3277 /* Look for the packet in the queue of old received packets */
3278 for (prev = call->rnext - 1, missing =
3279 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3280 /*Check for duplicate packet */
3281 if (seq == tp->header.seq) {
3282 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3283 rxevent_Cancel(call->delayedAckEvent, call,
3284 RX_CALL_REFCOUNT_DELAY);
3285 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3291 /* If we find a higher sequence packet, break out and
3292 * insert the new packet here. */
3293 if (seq < tp->header.seq)
3295 /* Check for missing packet */
3296 if (tp->header.seq != prev + 1) {
3300 prev = tp->header.seq;
3303 /* Keep track of whether we have received the last packet. */
3304 if (flags & RX_LAST_PACKET) {
3305 call->flags |= RX_CALL_HAVE_LAST;
3308 /* It's within the window: add it to the the receive queue.
3309 * tp is left by the previous loop either pointing at the
3310 * packet before which to insert the new packet, or at the
3311 * queue head if the queue is empty or the packet should be
3313 queue_InsertBefore(tp, np);
3317 /* Check whether we have all of the packets for this call */
3318 if ((call->flags & RX_CALL_HAVE_LAST)
3319 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3320 afs_uint32 tseq; /* temporary sequence number */
3323 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3324 if (tseq != tp->header.seq)
3326 if (tp->header.flags & RX_LAST_PACKET) {
3327 call->flags |= RX_CALL_RECEIVE_DONE;
3334 /* We need to send an ack of the packet is out of sequence,
3335 * or if an ack was requested by the peer. */
3336 if (seq != prev + 1 || missing) {
3337 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3338 } else if (flags & RX_REQUEST_ACK) {
3339 ackNeeded = RX_ACK_REQUESTED;
3342 /* Acknowledge the last packet for each call */
3343 if (flags & RX_LAST_PACKET) {
3354 * If the receiver is waiting for an iovec, fill the iovec
3355 * using the data from the receive queue */
3356 if (call->flags & RX_CALL_IOVEC_WAIT) {
3357 didHardAck = rxi_FillReadVec(call, serial);
3358 /* the call may have been aborted */
3367 /* Wakeup the reader if any */
3368 if ((call->flags & RX_CALL_READER_WAIT)
3369 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3370 || (call->iovNext >= call->iovMax)
3371 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3372 call->flags &= ~RX_CALL_READER_WAIT;
3373 #ifdef RX_ENABLE_LOCKS
3374 CV_BROADCAST(&call->cv_rq);
3376 osi_rxWakeup(&call->rq);
3382 * Send an ack when requested by the peer, or once every
3383 * rxi_SoftAckRate packets until the last packet has been
3384 * received. Always send a soft ack for the last packet in
3385 * the server's reply. */
3387 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3388 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3389 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3390 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3391 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3392 } else if (call->nSoftAcks) {
3393 clock_GetTime(&now);
3395 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3396 clock_Add(&when, &rx_lastAckDelay);
3398 clock_Add(&when, &rx_softAckDelay);
3400 if (!call->delayedAckEvent
3401 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3402 rxevent_Cancel(call->delayedAckEvent, call,
3403 RX_CALL_REFCOUNT_DELAY);
3404 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3405 call->delayedAckEvent =
3406 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3408 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3409 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3416 static void rxi_ComputeRate();
3420 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3422 struct rx_peer *peer = conn->peer;
3424 MUTEX_ENTER(&peer->peer_lock);
3425 peer->lastReachTime = clock_Sec();
3426 MUTEX_EXIT(&peer->peer_lock);
3428 MUTEX_ENTER(&conn->conn_data_lock);
3429 if (conn->flags & RX_CONN_ATTACHWAIT) {
3432 conn->flags &= ~RX_CONN_ATTACHWAIT;
3433 MUTEX_EXIT(&conn->conn_data_lock);
3435 for (i = 0; i < RX_MAXCALLS; i++) {
3436 struct rx_call *call = conn->call[i];
3439 MUTEX_ENTER(&call->lock);
3440 /* tnop can be null if newcallp is null */
3441 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3443 MUTEX_EXIT(&call->lock);
3447 MUTEX_EXIT(&conn->conn_data_lock);
3451 rx_ack_reason(int reason)
3454 case RX_ACK_REQUESTED:
3456 case RX_ACK_DUPLICATE:
3458 case RX_ACK_OUT_OF_SEQUENCE:
3460 case RX_ACK_EXCEEDS_WINDOW:
3462 case RX_ACK_NOSPACE:
3466 case RX_ACK_PING_RESPONSE:
3478 /* rxi_ComputePeerNetStats
3480 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3481 * estimates (like RTT and throughput) based on ack packets. Caller
3482 * must ensure that the packet in question is the right one (i.e.
3483 * serial number matches).
3486 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3487 struct rx_ackPacket *ap, struct rx_packet *np)
3489 struct rx_peer *peer = call->conn->peer;
3491 /* Use RTT if not delayed by client. */
3492 if (ap->reason != RX_ACK_DELAY)
3493 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3495 rxi_ComputeRate(peer, call, p, np, ap->reason);
3499 /* The real smarts of the whole thing. */
3501 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3504 struct rx_ackPacket *ap;
3506 register struct rx_packet *tp;
3507 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3508 register struct rx_connection *conn = call->conn;
3509 struct rx_peer *peer = conn->peer;
3512 /* because there are CM's that are bogus, sending weird values for this. */
3513 afs_uint32 skew = 0;
3518 int newAckCount = 0;
3519 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3520 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3522 rx_MutexIncrement(rx_stats.ackPacketsRead, rx_stats_mutex);
3523 ap = (struct rx_ackPacket *)rx_DataOf(np);
3524 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3526 return np; /* truncated ack packet */
3528 /* depends on ack packet struct */
3529 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3530 first = ntohl(ap->firstPacket);
3531 serial = ntohl(ap->serial);
3532 /* temporarily disabled -- needs to degrade over time
3533 * skew = ntohs(ap->maxSkew); */
3535 /* Ignore ack packets received out of order */
3536 if (first < call->tfirst) {
3540 if (np->header.flags & RX_SLOW_START_OK) {
3541 call->flags |= RX_CALL_SLOW_START_OK;
3544 if (ap->reason == RX_ACK_PING_RESPONSE)
3545 rxi_UpdatePeerReach(conn, call);
3549 if (rxdebug_active) {
3553 len = _snprintf(msg, sizeof(msg),
3554 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3555 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3556 ntohl(ap->serial), ntohl(ap->previousPacket),
3557 (unsigned int)np->header.seq, (unsigned int)skew,
3558 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3562 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3563 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3567 OutputDebugString(msg);
3569 #else /* AFS_NT40_ENV */
3572 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3573 ap->reason, ntohl(ap->previousPacket),
3574 (unsigned int)np->header.seq, (unsigned int)serial,
3575 (unsigned int)skew, ntohl(ap->firstPacket));
3578 for (offset = 0; offset < nAcks; offset++)
3579 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3584 #endif /* AFS_NT40_ENV */
3587 /* Update the outgoing packet skew value to the latest value of
3588 * the peer's incoming packet skew value. The ack packet, of
3589 * course, could arrive out of order, but that won't affect things
3591 MUTEX_ENTER(&peer->peer_lock);
3592 peer->outPacketSkew = skew;
3594 /* Check for packets that no longer need to be transmitted, and
3595 * discard them. This only applies to packets positively
3596 * acknowledged as having been sent to the peer's upper level.
3597 * All other packets must be retained. So only packets with
3598 * sequence numbers < ap->firstPacket are candidates. */
3599 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3600 if (tp->header.seq >= first)
3602 call->tfirst = tp->header.seq + 1;
3604 && (tp->header.serial == serial || tp->firstSerial == serial))
3605 rxi_ComputePeerNetStats(call, tp, ap, np);
3606 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3609 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3610 /* XXX Hack. Because we have to release the global rx lock when sending
3611 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3612 * in rxi_Start sending packets out because packets may move to the
3613 * freePacketQueue as result of being here! So we drop these packets until
3614 * we're safely out of the traversing. Really ugly!
3615 * To make it even uglier, if we're using fine grain locking, we can
3616 * set the ack bits in the packets and have rxi_Start remove the packets
3617 * when it's done transmitting.
3619 if (call->flags & RX_CALL_TQ_BUSY) {
3620 #ifdef RX_ENABLE_LOCKS
3621 tp->flags |= RX_PKTFLAG_ACKED;
3622 call->flags |= RX_CALL_TQ_SOME_ACKED;
3623 #else /* RX_ENABLE_LOCKS */
3625 #endif /* RX_ENABLE_LOCKS */
3627 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3630 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3635 /* Give rate detector a chance to respond to ping requests */
3636 if (ap->reason == RX_ACK_PING_RESPONSE) {
3637 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3641 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3643 /* Now go through explicit acks/nacks and record the results in
3644 * the waiting packets. These are packets that can't be released
3645 * yet, even with a positive acknowledge. This positive
3646 * acknowledge only means the packet has been received by the
3647 * peer, not that it will be retained long enough to be sent to
3648 * the peer's upper level. In addition, reset the transmit timers
3649 * of any missing packets (those packets that must be missing
3650 * because this packet was out of sequence) */
3652 call->nSoftAcked = 0;
3653 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3654 /* Update round trip time if the ack was stimulated on receipt
3656 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3657 #ifdef RX_ENABLE_LOCKS
3658 if (tp->header.seq >= first)
3659 #endif /* RX_ENABLE_LOCKS */
3660 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3662 && (tp->header.serial == serial || tp->firstSerial == serial))
3663 rxi_ComputePeerNetStats(call, tp, ap, np);
3665 /* Set the acknowledge flag per packet based on the
3666 * information in the ack packet. An acknowlegded packet can
3667 * be downgraded when the server has discarded a packet it
3668 * soacked previously, or when an ack packet is received
3669 * out of sequence. */
3670 if (tp->header.seq < first) {
3671 /* Implicit ack information */
3672 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3675 tp->flags |= RX_PKTFLAG_ACKED;
3676 } else if (tp->header.seq < first + nAcks) {
3677 /* Explicit ack information: set it in the packet appropriately */
3678 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3679 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3681 tp->flags |= RX_PKTFLAG_ACKED;
3688 } else /* RX_ACK_TYPE_NACK */ {
3689 tp->flags &= ~RX_PKTFLAG_ACKED;
3693 tp->flags &= ~RX_PKTFLAG_ACKED;
3697 /* If packet isn't yet acked, and it has been transmitted at least
3698 * once, reset retransmit time using latest timeout
3699 * ie, this should readjust the retransmit timer for all outstanding
3700 * packets... So we don't just retransmit when we should know better*/
3702 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3703 tp->retryTime = tp->timeSent;
3704 clock_Add(&tp->retryTime, &peer->timeout);
3705 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3706 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3710 /* If the window has been extended by this acknowledge packet,
3711 * then wakeup a sender waiting in alloc for window space, or try
3712 * sending packets now, if he's been sitting on packets due to
3713 * lack of window space */
3714 if (call->tnext < (call->tfirst + call->twind)) {
3715 #ifdef RX_ENABLE_LOCKS
3716 CV_SIGNAL(&call->cv_twind);
3718 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3719 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3720 osi_rxWakeup(&call->twind);
3723 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3724 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3728 /* if the ack packet has a receivelen field hanging off it,
3729 * update our state */
3730 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3733 /* If the ack packet has a "recommended" size that is less than
3734 * what I am using now, reduce my size to match */
3735 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3736 (int)sizeof(afs_int32), &tSize);
3737 tSize = (afs_uint32) ntohl(tSize);
3738 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3740 /* Get the maximum packet size to send to this peer */
3741 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
3743 tSize = (afs_uint32) ntohl(tSize);
3744 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3745 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3747 /* sanity check - peer might have restarted with different params.
3748 * If peer says "send less", dammit, send less... Peer should never
3749 * be unable to accept packets of the size that prior AFS versions would
3750 * send without asking. */
3751 if (peer->maxMTU != tSize) {
3752 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
3754 peer->maxMTU = tSize;
3755 peer->MTU = MIN(tSize, peer->MTU);
3756 call->MTU = MIN(call->MTU, tSize);
3759 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3762 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3763 (int)sizeof(afs_int32), &tSize);
3764 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3765 if (tSize < call->twind) { /* smaller than our send */
3766 call->twind = tSize; /* window, we must send less... */
3767 call->ssthresh = MIN(call->twind, call->ssthresh);
3768 call->conn->twind[call->channel] = call->twind;
3771 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3772 * network MTU confused with the loopback MTU. Calculate the
3773 * maximum MTU here for use in the slow start code below.
3775 maxMTU = peer->maxMTU;
3776 /* Did peer restart with older RX version? */
3777 if (peer->maxDgramPackets > 1) {
3778 peer->maxDgramPackets = 1;
3780 } else if (np->length >=
3781 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3784 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3785 sizeof(afs_int32), &tSize);
3786 tSize = (afs_uint32) ntohl(tSize);
3788 * As of AFS 3.5 we set the send window to match the receive window.
3790 if (tSize < call->twind) {
3791 call->twind = tSize;
3792 call->conn->twind[call->channel] = call->twind;
3793 call->ssthresh = MIN(call->twind, call->ssthresh);
3794 } else if (tSize > call->twind) {
3795 call->twind = tSize;
3796 call->conn->twind[call->channel] = call->twind;
3800 * As of AFS 3.5, a jumbogram is more than one fixed size
3801 * packet transmitted in a single UDP datagram. If the remote
3802 * MTU is smaller than our local MTU then never send a datagram
3803 * larger than the natural MTU.
3806 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3807 sizeof(afs_int32), &tSize);
3808 maxDgramPackets = (afs_uint32) ntohl(tSize);
3809 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3810 maxDgramPackets = MIN(maxDgramPackets, peer->ifDgramPackets);
3811 if (peer->natMTU < peer->ifMTU)
3812 maxDgramPackets = MIN(maxDgramPackets, rxi_AdjustDgramPackets(1, peer->natMTU));
3813 if (maxDgramPackets > 1) {
3814 peer->maxDgramPackets = maxDgramPackets;
3815 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3817 peer->maxDgramPackets = 1;
3818 call->MTU = peer->natMTU;
3820 } else if (peer->maxDgramPackets > 1) {
3821 /* Restarted with lower version of RX */
3822 peer->maxDgramPackets = 1;
3824 } else if (peer->maxDgramPackets > 1
3825 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3826 /* Restarted with lower version of RX */
3827 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3828 peer->natMTU = OLD_MAX_PACKET_SIZE;
3829 peer->MTU = OLD_MAX_PACKET_SIZE;
3830 peer->maxDgramPackets = 1;
3831 peer->nDgramPackets = 1;
3833 call->MTU = OLD_MAX_PACKET_SIZE;
3838 * Calculate how many datagrams were successfully received after
3839 * the first missing packet and adjust the negative ack counter
3844 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3845 if (call->nNacks < nNacked) {
3846 call->nNacks = nNacked;
3849 call->nAcks += newAckCount;
3853 if (call->flags & RX_CALL_FAST_RECOVER) {
3855 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3857 call->flags &= ~RX_CALL_FAST_RECOVER;
3858 call->cwind = call->nextCwind;
3859 call->nextCwind = 0;
3862 call->nCwindAcks = 0;
3863 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3864 /* Three negative acks in a row trigger congestion recovery */
3865 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3866 MUTEX_EXIT(&peer->peer_lock);
3867 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3868 /* someone else is waiting to start recovery */
3871 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3872 rxi_WaitforTQBusy(call);
3873 MUTEX_ENTER(&peer->peer_lock);
3874 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3875 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3876 call->flags |= RX_CALL_FAST_RECOVER;
3877 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3879 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3880 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3881 call->nextCwind = call->ssthresh;
3884 peer->MTU = call->MTU;
3885 peer->cwind = call->nextCwind;
3886 peer->nDgramPackets = call->nDgramPackets;
3888 call->congestSeq = peer->congestSeq;
3889 /* Reset the resend times on the packets that were nacked
3890 * so we will retransmit as soon as the window permits*/
3891 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3893 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3894 clock_Zero(&tp->retryTime);
3896 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3901 /* If cwind is smaller than ssthresh, then increase
3902 * the window one packet for each ack we receive (exponential
3904 * If cwind is greater than or equal to ssthresh then increase
3905 * the congestion window by one packet for each cwind acks we
3906 * receive (linear growth). */
3907 if (call->cwind < call->ssthresh) {
3909 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3910 call->nCwindAcks = 0;
3912 call->nCwindAcks += newAckCount;
3913 if (call->nCwindAcks >= call->cwind) {
3914 call->nCwindAcks = 0;
3915 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3919 * If we have received several acknowledgements in a row then
3920 * it is time to increase the size of our datagrams
3922 if ((int)call->nAcks > rx_nDgramThreshold) {
3923 if (peer->maxDgramPackets > 1) {
3924 if (call->nDgramPackets < peer->maxDgramPackets) {
3925 call->nDgramPackets++;
3927 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3928 } else if (call->MTU < peer->maxMTU) {
3929 call->MTU += peer->natMTU;
3930 call->MTU = MIN(call->MTU, peer->maxMTU);
3936 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3938 /* Servers need to hold the call until all response packets have
3939 * been acknowledged. Soft acks are good enough since clients
3940 * are not allowed to clear their receive queues. */
3941 if (call->state == RX_STATE_HOLD
3942 && call->tfirst + call->nSoftAcked >= call->tnext) {
3943 call->state = RX_STATE_DALLY;
3944 rxi_ClearTransmitQueue(call, 0);
3945 } else if (!queue_IsEmpty(&call->tq)) {
3946 rxi_Start(0, call, 0, istack);
3951 /* Received a response to a challenge packet */
3953 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
3954 register struct rx_packet *np, int istack)
3958 /* Ignore the packet if we're the client */
3959 if (conn->type == RX_CLIENT_CONNECTION)
3962 /* If already authenticated, ignore the packet (it's probably a retry) */
3963 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
3966 /* Otherwise, have the security object evaluate the response packet */
3967 error = RXS_CheckResponse(conn->securityObject, conn, np);
3969 /* If the response is invalid, reset the connection, sending
3970 * an abort to the peer */
3974 rxi_ConnectionError(conn, error);
3975 MUTEX_ENTER(&conn->conn_data_lock);
3976 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3977 MUTEX_EXIT(&conn->conn_data_lock);
3980 /* If the response is valid, any calls waiting to attach
3981 * servers can now do so */
3984 for (i = 0; i < RX_MAXCALLS; i++) {
3985 struct rx_call *call = conn->call[i];
3987 MUTEX_ENTER(&call->lock);
3988 if (call->state == RX_STATE_PRECALL)
3989 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
3990 /* tnop can be null if newcallp is null */
3991 MUTEX_EXIT(&call->lock);
3995 /* Update the peer reachability information, just in case
3996 * some calls went into attach-wait while we were waiting
3997 * for authentication..
3999 rxi_UpdatePeerReach(conn, NULL);
4004 /* A client has received an authentication challenge: the security
4005 * object is asked to cough up a respectable response packet to send
4006 * back to the server. The server is responsible for retrying the
4007 * challenge if it fails to get a response. */
4010 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
4011 register struct rx_packet *np, int istack)
4015 /* Ignore the challenge if we're the server */
4016 if (conn->type == RX_SERVER_CONNECTION)
4019 /* Ignore the challenge if the connection is otherwise idle; someone's
4020 * trying to use us as an oracle. */
4021 if (!rxi_HasActiveCalls(conn))
4024 /* Send the security object the challenge packet. It is expected to fill
4025 * in the response. */
4026 error = RXS_GetResponse(conn->securityObject, conn, np);
4028 /* If the security object is unable to return a valid response, reset the
4029 * connection and send an abort to the peer. Otherwise send the response
4030 * packet to the peer connection. */
4032 rxi_ConnectionError(conn, error);
4033 MUTEX_ENTER(&conn->conn_data_lock);
4034 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4035 MUTEX_EXIT(&conn->conn_data_lock);
4037 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4038 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4044 /* Find an available server process to service the current request in
4045 * the given call structure. If one isn't available, queue up this
4046 * call so it eventually gets one */
4048 rxi_AttachServerProc(register struct rx_call *call,
4049 register osi_socket socket, register int *tnop,
4050 register struct rx_call **newcallp)
4052 register struct rx_serverQueueEntry *sq;
4053 register struct rx_service *service = call->conn->service;
4054 register int haveQuota = 0;
4056 /* May already be attached */
4057 if (call->state == RX_STATE_ACTIVE)
4060 MUTEX_ENTER(&rx_serverPool_lock);
4062 haveQuota = QuotaOK(service);
4063 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4064 /* If there are no processes available to service this call,
4065 * put the call on the incoming call queue (unless it's
4066 * already on the queue).
4068 #ifdef RX_ENABLE_LOCKS
4070 ReturnToServerPool(service);
4071 #endif /* RX_ENABLE_LOCKS */
4073 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4074 call->flags |= RX_CALL_WAIT_PROC;
4075 MUTEX_ENTER(&rx_stats_mutex);
4078 MUTEX_EXIT(&rx_stats_mutex);
4079 rxi_calltrace(RX_CALL_ARRIVAL, call);
4080 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4081 queue_Append(&rx_incomingCallQueue, call);
4084 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4086 /* If hot threads are enabled, and both newcallp and sq->socketp
4087 * are non-null, then this thread will process the call, and the
4088 * idle server thread will start listening on this threads socket.
4091 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4094 *sq->socketp = socket;
4095 clock_GetTime(&call->startTime);
4096 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4100 if (call->flags & RX_CALL_WAIT_PROC) {
4101 /* Conservative: I don't think this should happen */
4102 call->flags &= ~RX_CALL_WAIT_PROC;
4103 if (queue_IsOnQueue(call)) {
4105 MUTEX_ENTER(&rx_stats_mutex);
4107 MUTEX_EXIT(&rx_stats_mutex);
4110 call->state = RX_STATE_ACTIVE;
4111 call->mode = RX_MODE_RECEIVING;
4112 #ifdef RX_KERNEL_TRACE
4114 int glockOwner = ISAFS_GLOCK();
4117 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4118 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4124 if (call->flags & RX_CALL_CLEARED) {
4125 /* send an ack now to start the packet flow up again */
4126 call->flags &= ~RX_CALL_CLEARED;
4127 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4129 #ifdef RX_ENABLE_LOCKS
4132 service->nRequestsRunning++;
4133 if (service->nRequestsRunning <= service->minProcs)
4139 MUTEX_EXIT(&rx_serverPool_lock);
4142 /* Delay the sending of an acknowledge event for a short while, while
4143 * a new call is being prepared (in the case of a client) or a reply
4144 * is being prepared (in the case of a server). Rather than sending
4145 * an ack packet, an ACKALL packet is sent. */
4147 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4149 #ifdef RX_ENABLE_LOCKS
4151 MUTEX_ENTER(&call->lock);
4152 call->delayedAckEvent = NULL;
4153 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4155 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4156 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4158 MUTEX_EXIT(&call->lock);
4159 #else /* RX_ENABLE_LOCKS */
4161 call->delayedAckEvent = NULL;
4162 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4163 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4164 #endif /* RX_ENABLE_LOCKS */
4168 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4171 #ifdef RX_ENABLE_LOCKS
4173 MUTEX_ENTER(&call->lock);
4174 if (event == call->delayedAckEvent)
4175 call->delayedAckEvent = NULL;
4176 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4178 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4180 MUTEX_EXIT(&call->lock);
4181 #else /* RX_ENABLE_LOCKS */
4183 call->delayedAckEvent = NULL;
4184 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4185 #endif /* RX_ENABLE_LOCKS */
4189 #ifdef RX_ENABLE_LOCKS
4190 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4191 * clearing them out.
4194 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4196 register struct rx_packet *p, *tp;
4199 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4200 p->flags |= RX_PKTFLAG_ACKED;
4204 call->flags |= RX_CALL_TQ_CLEARME;
4205 call->flags |= RX_CALL_TQ_SOME_ACKED;
4208 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4209 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4210 call->tfirst = call->tnext;
4211 call->nSoftAcked = 0;
4213 if (call->flags & RX_CALL_FAST_RECOVER) {
4214 call->flags &= ~RX_CALL_FAST_RECOVER;
4215 call->cwind = call->nextCwind;
4216 call->nextCwind = 0;
4219 CV_SIGNAL(&call->cv_twind);
4221 #endif /* RX_ENABLE_LOCKS */
4223 /* Clear out the transmit queue for the current call (all packets have
4224 * been received by peer) */
4226 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4228 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4229 register struct rx_packet *p, *tp;
4231 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4233 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4234 p->flags |= RX_PKTFLAG_ACKED;
4238 call->flags |= RX_CALL_TQ_CLEARME;
4239 call->flags |= RX_CALL_TQ_SOME_ACKED;
4242 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4243 rxi_FreePackets(0, &call->tq);
4244 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4245 call->flags &= ~RX_CALL_TQ_CLEARME;
4247 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4249 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4250 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4251 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4252 call->nSoftAcked = 0;
4254 if (call->flags & RX_CALL_FAST_RECOVER) {
4255 call->flags &= ~RX_CALL_FAST_RECOVER;
4256 call->cwind = call->nextCwind;
4258 #ifdef RX_ENABLE_LOCKS
4259 CV_SIGNAL(&call->cv_twind);
4261 osi_rxWakeup(&call->twind);
4266 rxi_ClearReceiveQueue(register struct rx_call *call)
4268 if (queue_IsNotEmpty(&call->rq)) {
4269 rx_packetReclaims += rxi_FreePackets(0, &call->rq);
4270 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4272 if (call->state == RX_STATE_PRECALL) {
4273 call->flags |= RX_CALL_CLEARED;
4277 /* Send an abort packet for the specified call */
4279 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4280 int istack, int force)
4283 struct clock when, now;
4288 /* Clients should never delay abort messages */
4289 if (rx_IsClientConn(call->conn))
4292 if (call->abortCode != call->error) {
4293 call->abortCode = call->error;
4294 call->abortCount = 0;
4297 if (force || rxi_callAbortThreshhold == 0
4298 || call->abortCount < rxi_callAbortThreshhold) {
4299 if (call->delayedAbortEvent) {
4300 rxevent_Cancel(call->delayedAbortEvent, call,
4301 RX_CALL_REFCOUNT_ABORT);
4303 error = htonl(call->error);
4306 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4307 (char *)&error, sizeof(error), istack);
4308 } else if (!call->delayedAbortEvent) {
4309 clock_GetTime(&now);
4311 clock_Addmsec(&when, rxi_callAbortDelay);
4312 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4313 call->delayedAbortEvent =
4314 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4319 /* Send an abort packet for the specified connection. Packet is an
4320 * optional pointer to a packet that can be used to send the abort.
4321 * Once the number of abort messages reaches the threshhold, an
4322 * event is scheduled to send the abort. Setting the force flag
4323 * overrides sending delayed abort messages.
4325 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4326 * to send the abort packet.
4329 rxi_SendConnectionAbort(register struct rx_connection *conn,
4330 struct rx_packet *packet, int istack, int force)
4333 struct clock when, now;
4338 /* Clients should never delay abort messages */
4339 if (rx_IsClientConn(conn))
4342 if (force || rxi_connAbortThreshhold == 0
4343 || conn->abortCount < rxi_connAbortThreshhold) {
4344 if (conn->delayedAbortEvent) {
4345 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4347 error = htonl(conn->error);
4349 MUTEX_EXIT(&conn->conn_data_lock);
4351 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4352 RX_PACKET_TYPE_ABORT, (char *)&error,
4353 sizeof(error), istack);
4354 MUTEX_ENTER(&conn->conn_data_lock);
4355 } else if (!conn->delayedAbortEvent) {
4356 clock_GetTime(&now);
4358 clock_Addmsec(&when, rxi_connAbortDelay);
4359 conn->delayedAbortEvent =
4360 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4365 /* Associate an error all of the calls owned by a connection. Called
4366 * with error non-zero. This is only for really fatal things, like
4367 * bad authentication responses. The connection itself is set in
4368 * error at this point, so that future packets received will be
4371 rxi_ConnectionError(register struct rx_connection *conn,
4372 register afs_int32 error)
4377 dpf(("rxi_ConnectionError conn %x error %d", conn, error));
4379 MUTEX_ENTER(&conn->conn_data_lock);
4380 if (conn->challengeEvent)
4381 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4382 if (conn->checkReachEvent) {
4383 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4384 conn->checkReachEvent = 0;
4385 conn->flags &= ~RX_CONN_ATTACHWAIT;
4388 MUTEX_EXIT(&conn->conn_data_lock);
4389 for (i = 0; i < RX_MAXCALLS; i++) {
4390 struct rx_call *call = conn->call[i];
4392 MUTEX_ENTER(&call->lock);
4393 rxi_CallError(call, error);
4394 MUTEX_EXIT(&call->lock);
4397 conn->error = error;
4398 rx_MutexIncrement(rx_stats.fatalErrors, rx_stats_mutex);
4403 rxi_CallError(register struct rx_call *call, afs_int32 error)
4405 dpf(("rxi_CallError call %x error %d call->error %d", call, error, call->error));
4407 error = call->error;
4409 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4410 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4411 rxi_ResetCall(call, 0);
4414 rxi_ResetCall(call, 0);
4416 call->error = error;
4417 call->mode = RX_MODE_ERROR;
4420 /* Reset various fields in a call structure, and wakeup waiting
4421 * processes. Some fields aren't changed: state & mode are not
4422 * touched (these must be set by the caller), and bufptr, nLeft, and
4423 * nFree are not reset, since these fields are manipulated by
4424 * unprotected macros, and may only be reset by non-interrupting code.
4427 /* this code requires that call->conn be set properly as a pre-condition. */
4428 #endif /* ADAPT_WINDOW */
4431 rxi_ResetCall(register struct rx_call *call, register int newcall)
4434 register struct rx_peer *peer;
4435 struct rx_packet *packet;
4437 dpf(("rxi_ResetCall(call %x, newcall %d)\n", call, newcall));
4439 /* Notify anyone who is waiting for asynchronous packet arrival */
4440 if (call->arrivalProc) {
4441 (*call->arrivalProc) (call, call->arrivalProcHandle,
4442 call->arrivalProcArg);
4443 call->arrivalProc = (void (*)())0;
4446 if (call->delayedAbortEvent) {
4447 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4448 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4450 rxi_SendCallAbort(call, packet, 0, 1);
4451 rxi_FreePacket(packet);
4456 * Update the peer with the congestion information in this call
4457 * so other calls on this connection can pick up where this call
4458 * left off. If the congestion sequence numbers don't match then
4459 * another call experienced a retransmission.
4461 peer = call->conn->peer;
4462 MUTEX_ENTER(&peer->peer_lock);
4464 if (call->congestSeq == peer->congestSeq) {
4465 peer->cwind = MAX(peer->cwind, call->cwind);
4466 peer->MTU = MAX(peer->MTU, call->MTU);
4467 peer->nDgramPackets =
4468 MAX(peer->nDgramPackets, call->nDgramPackets);
4471 call->abortCode = 0;
4472 call->abortCount = 0;
4474 if (peer->maxDgramPackets > 1) {
4475 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4477 call->MTU = peer->MTU;
4479 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4480 call->ssthresh = rx_maxSendWindow;
4481 call->nDgramPackets = peer->nDgramPackets;
4482 call->congestSeq = peer->congestSeq;
4483 MUTEX_EXIT(&peer->peer_lock);
4485 flags = call->flags;
4486 rxi_ClearReceiveQueue(call);
4487 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4488 if (flags & RX_CALL_TQ_BUSY) {
4489 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4490 call->flags |= (flags & RX_CALL_TQ_WAIT);
4492 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4494 rxi_ClearTransmitQueue(call, 0);
4495 queue_Init(&call->tq);
4496 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4497 dpf(("rcall %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4500 while (call->tqWaiters) {
4501 #ifdef RX_ENABLE_LOCKS
4502 CV_BROADCAST(&call->cv_tq);
4503 #else /* RX_ENABLE_LOCKS */
4504 osi_rxWakeup(&call->tq);
4505 #endif /* RX_ENABLE_LOCKS */
4509 queue_Init(&call->rq);
4511 call->twind = call->conn->twind[call->channel];
4512 call->rwind = call->conn->rwind[call->channel];
4513 call->nSoftAcked = 0;
4514 call->nextCwind = 0;
4517 call->nCwindAcks = 0;
4518 call->nSoftAcks = 0;
4519 call->nHardAcks = 0;
4521 call->tfirst = call->rnext = call->tnext = 1;
4523 call->lastAcked = 0;
4524 call->localStatus = call->remoteStatus = 0;
4526 if (flags & RX_CALL_READER_WAIT) {
4527 #ifdef RX_ENABLE_LOCKS
4528 CV_BROADCAST(&call->cv_rq);
4530 osi_rxWakeup(&call->rq);
4533 if (flags & RX_CALL_WAIT_PACKETS) {
4534 MUTEX_ENTER(&rx_freePktQ_lock);
4535 rxi_PacketsUnWait(); /* XXX */
4536 MUTEX_EXIT(&rx_freePktQ_lock);
4538 #ifdef RX_ENABLE_LOCKS
4539 CV_SIGNAL(&call->cv_twind);
4541 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4542 osi_rxWakeup(&call->twind);
4545 #ifdef RX_ENABLE_LOCKS
4546 /* The following ensures that we don't mess with any queue while some
4547 * other thread might also be doing so. The call_queue_lock field is
4548 * is only modified under the call lock. If the call is in the process
4549 * of being removed from a queue, the call is not locked until the
4550 * the queue lock is dropped and only then is the call_queue_lock field
4551 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4552 * Note that any other routine which removes a call from a queue has to
4553 * obtain the queue lock before examing the queue and removing the call.
4555 if (call->call_queue_lock) {
4556 MUTEX_ENTER(call->call_queue_lock);
4557 if (queue_IsOnQueue(call)) {
4559 if (flags & RX_CALL_WAIT_PROC) {
4560 MUTEX_ENTER(&rx_stats_mutex);
4562 MUTEX_EXIT(&rx_stats_mutex);
4565 MUTEX_EXIT(call->call_queue_lock);
4566 CLEAR_CALL_QUEUE_LOCK(call);
4568 #else /* RX_ENABLE_LOCKS */
4569 if (queue_IsOnQueue(call)) {
4571 if (flags & RX_CALL_WAIT_PROC)
4574 #endif /* RX_ENABLE_LOCKS */
4576 rxi_KeepAliveOff(call);
4577 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4580 /* Send an acknowledge for the indicated packet (seq,serial) of the
4581 * indicated call, for the indicated reason (reason). This
4582 * acknowledge will specifically acknowledge receiving the packet, and
4583 * will also specify which other packets for this call have been
4584 * received. This routine returns the packet that was used to the
4585 * caller. The caller is responsible for freeing it or re-using it.
4586 * This acknowledgement also returns the highest sequence number
4587 * actually read out by the higher level to the sender; the sender
4588 * promises to keep around packets that have not been read by the
4589 * higher level yet (unless, of course, the sender decides to abort
4590 * the call altogether). Any of p, seq, serial, pflags, or reason may
4591 * be set to zero without ill effect. That is, if they are zero, they
4592 * will not convey any information.
4593 * NOW there is a trailer field, after the ack where it will safely be
4594 * ignored by mundanes, which indicates the maximum size packet this
4595 * host can swallow. */
4597 register struct rx_packet *optionalPacket; use to send ack (or null)
4598 int seq; Sequence number of the packet we are acking
4599 int serial; Serial number of the packet
4600 int pflags; Flags field from packet header
4601 int reason; Reason an acknowledge was prompted
4605 rxi_SendAck(register struct rx_call *call,
4606 register struct rx_packet *optionalPacket, int serial, int reason,
4609 struct rx_ackPacket *ap;
4610 register struct rx_packet *rqp;
4611 register struct rx_packet *nxp; /* For queue_Scan */
4612 register struct rx_packet *p;
4615 #ifdef RX_ENABLE_TSFPQ
4616 struct rx_ts_info_t * rx_ts_info;
4620 * Open the receive window once a thread starts reading packets
4622 if (call->rnext > 1) {
4623 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
4626 call->nHardAcks = 0;
4627 call->nSoftAcks = 0;
4628 if (call->rnext > call->lastAcked)
4629 call->lastAcked = call->rnext;
4633 rx_computelen(p, p->length); /* reset length, you never know */
4634 } /* where that's been... */
4635 #ifdef RX_ENABLE_TSFPQ
4637 RX_TS_INFO_GET(rx_ts_info);
4638 if ((p = rx_ts_info->local_special_packet)) {
4639 rx_computelen(p, p->length);
4640 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4641 rx_ts_info->local_special_packet = p;
4642 } else { /* We won't send the ack, but don't panic. */
4643 return optionalPacket;
4647 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4648 /* We won't send the ack, but don't panic. */
4649 return optionalPacket;
4654 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4657 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4658 #ifndef RX_ENABLE_TSFPQ
4659 if (!optionalPacket)
4662 return optionalPacket;
4664 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4665 if (rx_Contiguous(p) < templ) {
4666 #ifndef RX_ENABLE_TSFPQ
4667 if (!optionalPacket)
4670 return optionalPacket;
4675 /* MTUXXX failing to send an ack is very serious. We should */
4676 /* try as hard as possible to send even a partial ack; it's */
4677 /* better than nothing. */
4678 ap = (struct rx_ackPacket *)rx_DataOf(p);
4679 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4680 ap->reason = reason;
4682 /* The skew computation used to be bogus, I think it's better now. */
4683 /* We should start paying attention to skew. XXX */
4684 ap->serial = htonl(serial);
4685 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4687 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4688 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4690 /* No fear of running out of ack packet here because there can only be at most
4691 * one window full of unacknowledged packets. The window size must be constrained
4692 * to be less than the maximum ack size, of course. Also, an ack should always
4693 * fit into a single packet -- it should not ever be fragmented. */
4694 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4695 if (!rqp || !call->rq.next
4696 || (rqp->header.seq > (call->rnext + call->rwind))) {
4697 #ifndef RX_ENABLE_TSFPQ
4698 if (!optionalPacket)
4701 rxi_CallError(call, RX_CALL_DEAD);
4702 return optionalPacket;
4705 while (rqp->header.seq > call->rnext + offset)
4706 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4707 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4709 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4710 #ifndef RX_ENABLE_TSFPQ
4711 if (!optionalPacket)
4714 rxi_CallError(call, RX_CALL_DEAD);
4715 return optionalPacket;
4720 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4722 /* these are new for AFS 3.3 */
4723 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4724 templ = htonl(templ);
4725 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4726 templ = htonl(call->conn->peer->ifMTU);
4727 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4728 sizeof(afs_int32), &templ);
4730 /* new for AFS 3.4 */
4731 templ = htonl(call->rwind);
4732 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4733 sizeof(afs_int32), &templ);
4735 /* new for AFS 3.5 */
4736 templ = htonl(call->conn->peer->ifDgramPackets);
4737 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4738 sizeof(afs_int32), &templ);
4740 p->header.serviceId = call->conn->serviceId;
4741 p->header.cid = (call->conn->cid | call->channel);
4742 p->header.callNumber = *call->callNumber;
4744 p->header.securityIndex = call->conn->securityIndex;
4745 p->header.epoch = call->conn->epoch;
4746 p->header.type = RX_PACKET_TYPE_ACK;
4747 p->header.flags = RX_SLOW_START_OK;
4748 if (reason == RX_ACK_PING) {
4749 p->header.flags |= RX_REQUEST_ACK;
4751 clock_GetTime(&call->pingRequestTime);
4754 if (call->conn->type == RX_CLIENT_CONNECTION)
4755 p->header.flags |= RX_CLIENT_INITIATED;
4759 if (rxdebug_active) {
4763 len = _snprintf(msg, sizeof(msg),
4764 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4765 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4766 ntohl(ap->serial), ntohl(ap->previousPacket),
4767 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
4768 ap->nAcks, ntohs(ap->bufferSpace) );
4772 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
4773 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4777 OutputDebugString(msg);
4779 #else /* AFS_NT40_ENV */
4781 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
4782 ap->reason, ntohl(ap->previousPacket),
4783 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4785 for (offset = 0; offset < ap->nAcks; offset++)
4786 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4791 #endif /* AFS_NT40_ENV */
4794 register int i, nbytes = p->length;
4796 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4797 if (nbytes <= p->wirevec[i].iov_len) {
4798 register int savelen, saven;
4800 savelen = p->wirevec[i].iov_len;
4802 p->wirevec[i].iov_len = nbytes;
4804 rxi_Send(call, p, istack);
4805 p->wirevec[i].iov_len = savelen;
4809 nbytes -= p->wirevec[i].iov_len;
4812 rx_MutexIncrement(rx_stats.ackPacketsSent, rx_stats_mutex);
4813 #ifndef RX_ENABLE_TSFPQ
4814 if (!optionalPacket)
4817 return optionalPacket; /* Return packet for re-use by caller */
4820 /* Send all of the packets in the list in single datagram */
4822 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4823 int istack, int moreFlag, struct clock *now,
4824 struct clock *retryTime, int resending)
4829 struct rx_connection *conn = call->conn;
4830 struct rx_peer *peer = conn->peer;
4832 MUTEX_ENTER(&peer->peer_lock);
4835 peer->reSends += len;
4836 rx_MutexIncrement(rx_stats.dataPacketsSent, rx_stats_mutex);
4837 MUTEX_EXIT(&peer->peer_lock);
4839 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4843 /* Set the packet flags and schedule the resend events */
4844 /* Only request an ack for the last packet in the list */
4845 for (i = 0; i < len; i++) {
4846 list[i]->retryTime = *retryTime;
4847 if (list[i]->header.serial) {
4848 /* Exponentially backoff retry times */
4849 if (list[i]->backoff < MAXBACKOFF) {
4850 /* so it can't stay == 0 */
4851 list[i]->backoff = (list[i]->backoff << 1) + 1;
4854 clock_Addmsec(&(list[i]->retryTime),
4855 ((afs_uint32) list[i]->backoff) << 8);
4858 /* Wait a little extra for the ack on the last packet */
4859 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4860 clock_Addmsec(&(list[i]->retryTime), 400);
4863 /* Record the time sent */
4864 list[i]->timeSent = *now;
4866 /* Ask for an ack on retransmitted packets, on every other packet
4867 * if the peer doesn't support slow start. Ask for an ack on every
4868 * packet until the congestion window reaches the ack rate. */
4869 if (list[i]->header.serial) {
4871 rx_MutexIncrement(rx_stats.dataPacketsReSent, rx_stats_mutex);
4873 /* improved RTO calculation- not Karn */
4874 list[i]->firstSent = *now;
4875 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4876 || (!(call->flags & RX_CALL_SLOW_START_OK)
4877 && (list[i]->header.seq & 1)))) {
4882 MUTEX_ENTER(&peer->peer_lock);
4886 rx_MutexIncrement(rx_stats.dataPacketsSent, rx_stats_mutex);
4887 MUTEX_EXIT(&peer->peer_lock);
4889 /* Tag this packet as not being the last in this group,
4890 * for the receiver's benefit */
4891 if (i < len - 1 || moreFlag) {
4892 list[i]->header.flags |= RX_MORE_PACKETS;
4895 /* Install the new retransmit time for the packet, and
4896 * record the time sent */
4897 list[i]->timeSent = *now;
4901 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4904 /* Since we're about to send a data packet to the peer, it's
4905 * safe to nuke any scheduled end-of-packets ack */
4906 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4908 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4909 MUTEX_EXIT(&call->lock);
4911 rxi_SendPacketList(call, conn, list, len, istack);
4913 rxi_SendPacket(call, conn, list[0], istack);
4915 MUTEX_ENTER(&call->lock);
4916 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4918 /* Update last send time for this call (for keep-alive
4919 * processing), and for the connection (so that we can discover
4920 * idle connections) */
4921 conn->lastSendTime = call->lastSendTime = clock_Sec();
4924 /* When sending packets we need to follow these rules:
4925 * 1. Never send more than maxDgramPackets in a jumbogram.
4926 * 2. Never send a packet with more than two iovecs in a jumbogram.
4927 * 3. Never send a retransmitted packet in a jumbogram.
4928 * 4. Never send more than cwind/4 packets in a jumbogram
4929 * We always keep the last list we should have sent so we
4930 * can set the RX_MORE_PACKETS flags correctly.
4933 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4934 int istack, struct clock *now, struct clock *retryTime,
4937 int i, cnt, lastCnt = 0;
4938 struct rx_packet **listP, **lastP = 0;
4939 struct rx_peer *peer = call->conn->peer;
4940 int morePackets = 0;
4942 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4943 /* Does the current packet force us to flush the current list? */
4945 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4946 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4948 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
4950 /* If the call enters an error state stop sending, or if
4951 * we entered congestion recovery mode, stop sending */
4952 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4960 /* Add the current packet to the list if it hasn't been acked.
4961 * Otherwise adjust the list pointer to skip the current packet. */
4962 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
4964 /* Do we need to flush the list? */
4965 if (cnt >= (int)peer->maxDgramPackets
4966 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
4967 || list[i]->header.serial
4968 || list[i]->length != RX_JUMBOBUFFERSIZE) {
4970 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
4971 retryTime, resending);
4972 /* If the call enters an error state stop sending, or if
4973 * we entered congestion recovery mode, stop sending */
4975 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4980 listP = &list[i + 1];
4985 osi_Panic("rxi_SendList error");
4987 listP = &list[i + 1];
4991 /* Send the whole list when the call is in receive mode, when
4992 * the call is in eof mode, when we are in fast recovery mode,
4993 * and when we have the last packet */
4994 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
4995 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
4996 || (call->flags & RX_CALL_FAST_RECOVER)) {
4997 /* Check for the case where the current list contains
4998 * an acked packet. Since we always send retransmissions
4999 * in a separate packet, we only need to check the first
5000 * packet in the list */
5001 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5005 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5006 retryTime, resending);
5007 /* If the call enters an error state stop sending, or if
5008 * we entered congestion recovery mode, stop sending */
5009 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5013 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5016 } else if (lastCnt > 0) {
5017 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5022 #ifdef RX_ENABLE_LOCKS
5023 /* Call rxi_Start, below, but with the call lock held. */
5025 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
5026 void *arg1, int istack)
5028 MUTEX_ENTER(&call->lock);
5029 rxi_Start(event, call, arg1, istack);
5030 MUTEX_EXIT(&call->lock);
5032 #endif /* RX_ENABLE_LOCKS */
5034 /* This routine is called when new packets are readied for
5035 * transmission and when retransmission may be necessary, or when the
5036 * transmission window or burst count are favourable. This should be
5037 * better optimized for new packets, the usual case, now that we've
5038 * got rid of queues of send packets. XXXXXXXXXXX */
5040 rxi_Start(struct rxevent *event, register struct rx_call *call,
5041 void *arg1, int istack)
5043 struct rx_packet *p;
5044 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
5045 struct rx_peer *peer = call->conn->peer;
5046 struct clock now, usenow, retryTime;
5050 struct rx_packet **xmitList;
5053 /* If rxi_Start is being called as a result of a resend event,
5054 * then make sure that the event pointer is removed from the call
5055 * structure, since there is no longer a per-call retransmission
5057 if (event && event == call->resendEvent) {
5058 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5059 call->resendEvent = NULL;
5061 if (queue_IsEmpty(&call->tq)) {
5065 /* Timeouts trigger congestion recovery */
5066 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5067 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5068 /* someone else is waiting to start recovery */
5071 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5072 rxi_WaitforTQBusy(call);
5073 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5074 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5075 call->flags |= RX_CALL_FAST_RECOVER;
5076 if (peer->maxDgramPackets > 1) {
5077 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5079 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5081 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5082 call->nDgramPackets = 1;
5084 call->nextCwind = 1;
5087 MUTEX_ENTER(&peer->peer_lock);
5088 peer->MTU = call->MTU;
5089 peer->cwind = call->cwind;
5090 peer->nDgramPackets = 1;
5092 call->congestSeq = peer->congestSeq;
5093 MUTEX_EXIT(&peer->peer_lock);
5094 /* Clear retry times on packets. Otherwise, it's possible for
5095 * some packets in the queue to force resends at rates faster
5096 * than recovery rates.
5098 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5099 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5100 clock_Zero(&p->retryTime);
5105 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5106 rx_MutexIncrement(rx_tq_debug.rxi_start_in_error, rx_stats_mutex);
5111 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5112 /* Get clock to compute the re-transmit time for any packets
5113 * in this burst. Note, if we back off, it's reasonable to
5114 * back off all of the packets in the same manner, even if
5115 * some of them have been retransmitted more times than more
5117 * Do a dance to avoid blocking after setting now. */
5118 clock_Zero(&retryTime);
5119 MUTEX_ENTER(&peer->peer_lock);
5120 clock_Add(&retryTime, &peer->timeout);
5121 MUTEX_EXIT(&peer->peer_lock);
5122 clock_GetTime(&now);
5123 clock_Add(&retryTime, &now);
5125 /* Send (or resend) any packets that need it, subject to
5126 * window restrictions and congestion burst control
5127 * restrictions. Ask for an ack on the last packet sent in
5128 * this burst. For now, we're relying upon the window being
5129 * considerably bigger than the largest number of packets that
5130 * are typically sent at once by one initial call to
5131 * rxi_Start. This is probably bogus (perhaps we should ask
5132 * for an ack when we're half way through the current
5133 * window?). Also, for non file transfer applications, this
5134 * may end up asking for an ack for every packet. Bogus. XXXX
5137 * But check whether we're here recursively, and let the other guy
5140 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5141 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5142 call->flags |= RX_CALL_TQ_BUSY;
5144 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5146 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5147 call->flags &= ~RX_CALL_NEED_START;
5148 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5150 maxXmitPackets = MIN(call->twind, call->cwind);
5151 xmitList = (struct rx_packet **)
5152 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5153 if (xmitList == NULL)
5154 osi_Panic("rxi_Start, failed to allocate xmit list");
5155 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5156 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5157 /* We shouldn't be sending packets if a thread is waiting
5158 * to initiate congestion recovery */
5162 && (call->flags & RX_CALL_FAST_RECOVER)) {
5163 /* Only send one packet during fast recovery */
5166 if ((p->flags & RX_PKTFLAG_FREE)
5167 || (!queue_IsEnd(&call->tq, nxp)
5168 && (nxp->flags & RX_PKTFLAG_FREE))
5169 || (p == (struct rx_packet *)&rx_freePacketQueue)
5170 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5171 osi_Panic("rxi_Start: xmit queue clobbered");
5173 if (p->flags & RX_PKTFLAG_ACKED) {
5174 /* Since we may block, don't trust this */
5175 usenow.sec = usenow.usec = 0;
5176 rx_MutexIncrement(rx_stats.ignoreAckedPacket, rx_stats_mutex);
5177 continue; /* Ignore this packet if it has been acknowledged */
5180 /* Turn off all flags except these ones, which are the same
5181 * on each transmission */
5182 p->header.flags &= RX_PRESET_FLAGS;
5184 if (p->header.seq >=
5185 call->tfirst + MIN((int)call->twind,
5186 (int)(call->nSoftAcked +
5188 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5189 /* Note: if we're waiting for more window space, we can
5190 * still send retransmits; hence we don't return here, but
5191 * break out to schedule a retransmit event */
5192 dpf(("call %d waiting for window",
5193 *(call->callNumber)));
5197 /* Transmit the packet if it needs to be sent. */
5198 if (!clock_Lt(&now, &p->retryTime)) {
5199 if (nXmitPackets == maxXmitPackets) {
5200 rxi_SendXmitList(call, xmitList, nXmitPackets,
5201 istack, &now, &retryTime,
5203 osi_Free(xmitList, maxXmitPackets *
5204 sizeof(struct rx_packet *));
5207 xmitList[nXmitPackets++] = p;
5211 /* xmitList now hold pointers to all of the packets that are
5212 * ready to send. Now we loop to send the packets */
5213 if (nXmitPackets > 0) {
5214 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5215 &now, &retryTime, resending);
5218 maxXmitPackets * sizeof(struct rx_packet *));
5220 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5222 * TQ references no longer protected by this flag; they must remain
5223 * protected by the global lock.
5225 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5226 call->flags &= ~RX_CALL_TQ_BUSY;
5227 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5228 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5229 #ifdef RX_ENABLE_LOCKS
5230 osirx_AssertMine(&call->lock, "rxi_Start start");
5231 CV_BROADCAST(&call->cv_tq);
5232 #else /* RX_ENABLE_LOCKS */
5233 osi_rxWakeup(&call->tq);
5234 #endif /* RX_ENABLE_LOCKS */
5239 /* We went into the error state while sending packets. Now is
5240 * the time to reset the call. This will also inform the using
5241 * process that the call is in an error state.
5243 rx_MutexIncrement(rx_tq_debug.rxi_start_aborted, rx_stats_mutex);
5244 call->flags &= ~RX_CALL_TQ_BUSY;
5245 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5246 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5247 #ifdef RX_ENABLE_LOCKS
5248 osirx_AssertMine(&call->lock, "rxi_Start middle");
5249 CV_BROADCAST(&call->cv_tq);
5250 #else /* RX_ENABLE_LOCKS */
5251 osi_rxWakeup(&call->tq);
5252 #endif /* RX_ENABLE_LOCKS */
5254 rxi_CallError(call, call->error);
5257 #ifdef RX_ENABLE_LOCKS
5258 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5259 register int missing;
5260 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5261 /* Some packets have received acks. If they all have, we can clear
5262 * the transmit queue.
5265 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5266 if (p->header.seq < call->tfirst
5267 && (p->flags & RX_PKTFLAG_ACKED)) {
5274 call->flags |= RX_CALL_TQ_CLEARME;
5276 #endif /* RX_ENABLE_LOCKS */
5277 /* Don't bother doing retransmits if the TQ is cleared. */
5278 if (call->flags & RX_CALL_TQ_CLEARME) {
5279 rxi_ClearTransmitQueue(call, 1);
5281 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5284 /* Always post a resend event, if there is anything in the
5285 * queue, and resend is possible. There should be at least
5286 * one unacknowledged packet in the queue ... otherwise none
5287 * of these packets should be on the queue in the first place.
5289 if (call->resendEvent) {
5290 /* Cancel the existing event and post a new one */
5291 rxevent_Cancel(call->resendEvent, call,
5292 RX_CALL_REFCOUNT_RESEND);
5295 /* The retry time is the retry time on the first unacknowledged
5296 * packet inside the current window */
5298 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5299 /* Don't set timers for packets outside the window */
5300 if (p->header.seq >= call->tfirst + call->twind) {
5304 if (!(p->flags & RX_PKTFLAG_ACKED)
5305 && !clock_IsZero(&p->retryTime)) {
5307 retryTime = p->retryTime;
5312 /* Post a new event to re-run rxi_Start when retries may be needed */
5313 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5314 #ifdef RX_ENABLE_LOCKS
5315 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5317 rxevent_PostNow2(&retryTime, &usenow,
5319 (void *)call, 0, istack);
5320 #else /* RX_ENABLE_LOCKS */
5322 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5323 (void *)call, 0, istack);
5324 #endif /* RX_ENABLE_LOCKS */
5327 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5328 } while (call->flags & RX_CALL_NEED_START);
5330 * TQ references no longer protected by this flag; they must remain
5331 * protected by the global lock.
5333 call->flags &= ~RX_CALL_TQ_BUSY;
5334 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5335 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5336 #ifdef RX_ENABLE_LOCKS
5337 osirx_AssertMine(&call->lock, "rxi_Start end");
5338 CV_BROADCAST(&call->cv_tq);
5339 #else /* RX_ENABLE_LOCKS */
5340 osi_rxWakeup(&call->tq);
5341 #endif /* RX_ENABLE_LOCKS */
5344 call->flags |= RX_CALL_NEED_START;
5346 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5348 if (call->resendEvent) {
5349 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5354 /* Also adjusts the keep alive parameters for the call, to reflect
5355 * that we have just sent a packet (so keep alives aren't sent
5358 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5361 register struct rx_connection *conn = call->conn;
5363 /* Stamp each packet with the user supplied status */
5364 p->header.userStatus = call->localStatus;
5366 /* Allow the security object controlling this call's security to
5367 * make any last-minute changes to the packet */
5368 RXS_SendPacket(conn->securityObject, call, p);
5370 /* Since we're about to send SOME sort of packet to the peer, it's
5371 * safe to nuke any scheduled end-of-packets ack */
5372 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5374 /* Actually send the packet, filling in more connection-specific fields */
5375 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5376 MUTEX_EXIT(&call->lock);
5377 rxi_SendPacket(call, conn, p, istack);
5378 MUTEX_ENTER(&call->lock);
5379 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5381 /* Update last send time for this call (for keep-alive
5382 * processing), and for the connection (so that we can discover
5383 * idle connections) */
5384 conn->lastSendTime = call->lastSendTime = clock_Sec();
5388 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5389 * that things are fine. Also called periodically to guarantee that nothing
5390 * falls through the cracks (e.g. (error + dally) connections have keepalive
5391 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5393 * haveCTLock Set if calling from rxi_ReapConnections
5395 #ifdef RX_ENABLE_LOCKS
5397 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5398 #else /* RX_ENABLE_LOCKS */
5400 rxi_CheckCall(register struct rx_call *call)
5401 #endif /* RX_ENABLE_LOCKS */
5403 register struct rx_connection *conn = call->conn;
5405 afs_uint32 deadTime;
5407 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5408 if (call->flags & RX_CALL_TQ_BUSY) {
5409 /* Call is active and will be reset by rxi_Start if it's
5410 * in an error state.
5415 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5417 (((afs_uint32) conn->secondsUntilDead << 10) +
5418 ((afs_uint32) conn->peer->rtt >> 3) +
5419 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5421 /* These are computed to the second (+- 1 second). But that's
5422 * good enough for these values, which should be a significant
5423 * number of seconds. */
5424 if (now > (call->lastReceiveTime + deadTime)) {
5425 if (call->state == RX_STATE_ACTIVE) {
5426 rxi_CallError(call, RX_CALL_DEAD);
5429 #ifdef RX_ENABLE_LOCKS
5430 /* Cancel pending events */
5431 rxevent_Cancel(call->delayedAckEvent, call,
5432 RX_CALL_REFCOUNT_DELAY);
5433 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5434 rxevent_Cancel(call->keepAliveEvent, call,
5435 RX_CALL_REFCOUNT_ALIVE);
5436 if (call->refCount == 0) {
5437 rxi_FreeCall(call, haveCTLock);
5441 #else /* RX_ENABLE_LOCKS */
5444 #endif /* RX_ENABLE_LOCKS */
5446 /* Non-active calls are destroyed if they are not responding
5447 * to pings; active calls are simply flagged in error, so the
5448 * attached process can die reasonably gracefully. */
5450 /* see if we have a non-activity timeout */
5451 if (call->startWait && conn->idleDeadTime
5452 && ((call->startWait + conn->idleDeadTime) < now)) {
5453 if (call->state == RX_STATE_ACTIVE) {
5454 rxi_CallError(call, RX_CALL_TIMEOUT);
5458 /* see if we have a hard timeout */
5459 if (conn->hardDeadTime
5460 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5461 if (call->state == RX_STATE_ACTIVE)
5462 rxi_CallError(call, RX_CALL_TIMEOUT);
5469 /* When a call is in progress, this routine is called occasionally to
5470 * make sure that some traffic has arrived (or been sent to) the peer.
5471 * If nothing has arrived in a reasonable amount of time, the call is
5472 * declared dead; if nothing has been sent for a while, we send a
5473 * keep-alive packet (if we're actually trying to keep the call alive)
5476 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5479 struct rx_connection *conn;
5482 MUTEX_ENTER(&call->lock);
5483 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5484 if (event == call->keepAliveEvent)
5485 call->keepAliveEvent = NULL;
5488 #ifdef RX_ENABLE_LOCKS
5489 if (rxi_CheckCall(call, 0)) {
5490 MUTEX_EXIT(&call->lock);
5493 #else /* RX_ENABLE_LOCKS */
5494 if (rxi_CheckCall(call))
5496 #endif /* RX_ENABLE_LOCKS */
5498 /* Don't try to keep alive dallying calls */
5499 if (call->state == RX_STATE_DALLY) {
5500 MUTEX_EXIT(&call->lock);
5505 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5506 /* Don't try to send keepalives if there is unacknowledged data */
5507 /* the rexmit code should be good enough, this little hack
5508 * doesn't quite work XXX */
5509 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5511 rxi_ScheduleKeepAliveEvent(call);
5512 MUTEX_EXIT(&call->lock);
5517 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5519 if (!call->keepAliveEvent) {
5520 struct clock when, now;
5521 clock_GetTime(&now);
5523 when.sec += call->conn->secondsUntilPing;
5524 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5525 call->keepAliveEvent =
5526 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
5530 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5532 rxi_KeepAliveOn(register struct rx_call *call)
5534 /* Pretend last packet received was received now--i.e. if another
5535 * packet isn't received within the keep alive time, then the call
5536 * will die; Initialize last send time to the current time--even
5537 * if a packet hasn't been sent yet. This will guarantee that a
5538 * keep-alive is sent within the ping time */
5539 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5540 rxi_ScheduleKeepAliveEvent(call);
5543 /* This routine is called to send connection abort messages
5544 * that have been delayed to throttle looping clients. */
5546 rxi_SendDelayedConnAbort(struct rxevent *event,
5547 register struct rx_connection *conn, char *dummy)
5550 struct rx_packet *packet;
5552 MUTEX_ENTER(&conn->conn_data_lock);
5553 conn->delayedAbortEvent = NULL;
5554 error = htonl(conn->error);
5556 MUTEX_EXIT(&conn->conn_data_lock);
5557 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5560 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5561 RX_PACKET_TYPE_ABORT, (char *)&error,
5563 rxi_FreePacket(packet);
5567 /* This routine is called to send call abort messages
5568 * that have been delayed to throttle looping clients. */
5570 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5574 struct rx_packet *packet;
5576 MUTEX_ENTER(&call->lock);
5577 call->delayedAbortEvent = NULL;
5578 error = htonl(call->error);
5580 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5583 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5584 (char *)&error, sizeof(error), 0);
5585 rxi_FreePacket(packet);
5587 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5588 MUTEX_EXIT(&call->lock);
5591 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5592 * seconds) to ask the client to authenticate itself. The routine
5593 * issues a challenge to the client, which is obtained from the
5594 * security object associated with the connection */
5596 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5597 void *arg1, int tries)
5599 conn->challengeEvent = NULL;
5600 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5601 register struct rx_packet *packet;
5602 struct clock when, now;
5605 /* We've failed to authenticate for too long.
5606 * Reset any calls waiting for authentication;
5607 * they are all in RX_STATE_PRECALL.
5611 MUTEX_ENTER(&conn->conn_call_lock);
5612 for (i = 0; i < RX_MAXCALLS; i++) {
5613 struct rx_call *call = conn->call[i];
5615 MUTEX_ENTER(&call->lock);
5616 if (call->state == RX_STATE_PRECALL) {
5617 rxi_CallError(call, RX_CALL_DEAD);
5618 rxi_SendCallAbort(call, NULL, 0, 0);
5620 MUTEX_EXIT(&call->lock);
5623 MUTEX_EXIT(&conn->conn_call_lock);
5627 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5629 /* If there's no packet available, do this later. */
5630 RXS_GetChallenge(conn->securityObject, conn, packet);
5631 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5632 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5633 rxi_FreePacket(packet);
5635 clock_GetTime(&now);
5637 when.sec += RX_CHALLENGE_TIMEOUT;
5638 conn->challengeEvent =
5639 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
5644 /* Call this routine to start requesting the client to authenticate
5645 * itself. This will continue until authentication is established,
5646 * the call times out, or an invalid response is returned. The
5647 * security object associated with the connection is asked to create
5648 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5649 * defined earlier. */
5651 rxi_ChallengeOn(register struct rx_connection *conn)
5653 if (!conn->challengeEvent) {
5654 RXS_CreateChallenge(conn->securityObject, conn);
5655 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5660 /* Compute round trip time of the packet provided, in *rttp.
5663 /* rxi_ComputeRoundTripTime is called with peer locked. */
5664 /* sentp and/or peer may be null */
5666 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5667 register struct clock *sentp,
5668 register struct rx_peer *peer)
5670 struct clock thisRtt, *rttp = &thisRtt;
5672 register int rtt_timeout;
5674 clock_GetTime(rttp);
5676 if (clock_Lt(rttp, sentp)) {
5678 return; /* somebody set the clock back, don't count this time. */
5680 clock_Sub(rttp, sentp);
5681 MUTEX_ENTER(&rx_stats_mutex);
5682 if (clock_Lt(rttp, &rx_stats.minRtt))
5683 rx_stats.minRtt = *rttp;
5684 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5685 if (rttp->sec > 60) {
5686 MUTEX_EXIT(&rx_stats_mutex);
5687 return; /* somebody set the clock ahead */
5689 rx_stats.maxRtt = *rttp;
5691 clock_Add(&rx_stats.totalRtt, rttp);
5692 rx_stats.nRttSamples++;
5693 MUTEX_EXIT(&rx_stats_mutex);
5695 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5697 /* Apply VanJacobson round-trip estimations */
5702 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5703 * srtt is stored as fixed point with 3 bits after the binary
5704 * point (i.e., scaled by 8). The following magic is
5705 * equivalent to the smoothing algorithm in rfc793 with an
5706 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5707 * srtt*8 = srtt*8 + rtt - srtt
5708 * srtt = srtt + rtt/8 - srtt/8
5711 delta = MSEC(rttp) - (peer->rtt >> 3);
5715 * We accumulate a smoothed rtt variance (actually, a smoothed
5716 * mean difference), then set the retransmit timer to smoothed
5717 * rtt + 4 times the smoothed variance (was 2x in van's original
5718 * paper, but 4x works better for me, and apparently for him as
5720 * rttvar is stored as
5721 * fixed point with 2 bits after the binary point (scaled by
5722 * 4). The following is equivalent to rfc793 smoothing with
5723 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5724 * replaces rfc793's wired-in beta.
5725 * dev*4 = dev*4 + (|actual - expected| - dev)
5731 delta -= (peer->rtt_dev >> 2);
5732 peer->rtt_dev += delta;
5734 /* I don't have a stored RTT so I start with this value. Since I'm
5735 * probably just starting a call, and will be pushing more data down
5736 * this, I expect congestion to increase rapidly. So I fudge a
5737 * little, and I set deviance to half the rtt. In practice,
5738 * deviance tends to approach something a little less than
5739 * half the smoothed rtt. */
5740 peer->rtt = (MSEC(rttp) << 3) + 8;
5741 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5743 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5744 * the other of these connections is usually in a user process, and can
5745 * be switched and/or swapped out. So on fast, reliable networks, the
5746 * timeout would otherwise be too short.
5748 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5749 clock_Zero(&(peer->timeout));
5750 clock_Addmsec(&(peer->timeout), rtt_timeout);
5752 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)));
5756 /* Find all server connections that have not been active for a long time, and
5759 rxi_ReapConnections(void)
5761 struct clock now, when;
5762 clock_GetTime(&now);
5764 /* Find server connection structures that haven't been used for
5765 * greater than rx_idleConnectionTime */
5767 struct rx_connection **conn_ptr, **conn_end;
5768 int i, havecalls = 0;
5769 MUTEX_ENTER(&rx_connHashTable_lock);
5770 for (conn_ptr = &rx_connHashTable[0], conn_end =
5771 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5773 struct rx_connection *conn, *next;
5774 struct rx_call *call;
5778 for (conn = *conn_ptr; conn; conn = next) {
5779 /* XXX -- Shouldn't the connection be locked? */
5782 for (i = 0; i < RX_MAXCALLS; i++) {
5783 call = conn->call[i];
5786 MUTEX_ENTER(&call->lock);
5787 #ifdef RX_ENABLE_LOCKS
5788 result = rxi_CheckCall(call, 1);
5789 #else /* RX_ENABLE_LOCKS */
5790 result = rxi_CheckCall(call);
5791 #endif /* RX_ENABLE_LOCKS */
5792 MUTEX_EXIT(&call->lock);
5794 /* If CheckCall freed the call, it might
5795 * have destroyed the connection as well,
5796 * which screws up the linked lists.
5802 if (conn->type == RX_SERVER_CONNECTION) {
5803 /* This only actually destroys the connection if
5804 * there are no outstanding calls */
5805 MUTEX_ENTER(&conn->conn_data_lock);
5806 if (!havecalls && !conn->refCount
5807 && ((conn->lastSendTime + rx_idleConnectionTime) <
5809 conn->refCount++; /* it will be decr in rx_DestroyConn */
5810 MUTEX_EXIT(&conn->conn_data_lock);
5811 #ifdef RX_ENABLE_LOCKS
5812 rxi_DestroyConnectionNoLock(conn);
5813 #else /* RX_ENABLE_LOCKS */
5814 rxi_DestroyConnection(conn);
5815 #endif /* RX_ENABLE_LOCKS */
5817 #ifdef RX_ENABLE_LOCKS
5819 MUTEX_EXIT(&conn->conn_data_lock);
5821 #endif /* RX_ENABLE_LOCKS */
5825 #ifdef RX_ENABLE_LOCKS
5826 while (rx_connCleanup_list) {
5827 struct rx_connection *conn;
5828 conn = rx_connCleanup_list;
5829 rx_connCleanup_list = rx_connCleanup_list->next;
5830 MUTEX_EXIT(&rx_connHashTable_lock);
5831 rxi_CleanupConnection(conn);
5832 MUTEX_ENTER(&rx_connHashTable_lock);
5834 MUTEX_EXIT(&rx_connHashTable_lock);
5835 #endif /* RX_ENABLE_LOCKS */
5838 /* Find any peer structures that haven't been used (haven't had an
5839 * associated connection) for greater than rx_idlePeerTime */
5841 struct rx_peer **peer_ptr, **peer_end;
5843 MUTEX_ENTER(&rx_rpc_stats);
5844 MUTEX_ENTER(&rx_peerHashTable_lock);
5845 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5846 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5848 struct rx_peer *peer, *next, *prev;
5849 for (prev = peer = *peer_ptr; peer; peer = next) {
5851 code = MUTEX_TRYENTER(&peer->peer_lock);
5852 if ((code) && (peer->refCount == 0)
5853 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5854 rx_interface_stat_p rpc_stat, nrpc_stat;
5856 MUTEX_EXIT(&peer->peer_lock);
5857 MUTEX_DESTROY(&peer->peer_lock);
5859 (&peer->rpcStats, rpc_stat, nrpc_stat,
5860 rx_interface_stat)) {
5861 unsigned int num_funcs;
5864 queue_Remove(&rpc_stat->queue_header);
5865 queue_Remove(&rpc_stat->all_peers);
5866 num_funcs = rpc_stat->stats[0].func_total;
5868 sizeof(rx_interface_stat_t) +
5869 rpc_stat->stats[0].func_total *
5870 sizeof(rx_function_entry_v1_t);
5872 rxi_Free(rpc_stat, space);
5873 rxi_rpc_peer_stat_cnt -= num_funcs;
5876 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
5877 if (peer == *peer_ptr) {
5884 MUTEX_EXIT(&peer->peer_lock);
5890 MUTEX_EXIT(&rx_peerHashTable_lock);
5891 MUTEX_EXIT(&rx_rpc_stats);
5894 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5895 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5896 * GC, just below. Really, we shouldn't have to keep moving packets from
5897 * one place to another, but instead ought to always know if we can
5898 * afford to hold onto a packet in its particular use. */
5899 MUTEX_ENTER(&rx_freePktQ_lock);
5900 if (rx_waitingForPackets) {
5901 rx_waitingForPackets = 0;
5902 #ifdef RX_ENABLE_LOCKS
5903 CV_BROADCAST(&rx_waitingForPackets_cv);
5905 osi_rxWakeup(&rx_waitingForPackets);
5908 MUTEX_EXIT(&rx_freePktQ_lock);
5911 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5912 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
5916 /* rxs_Release - This isn't strictly necessary but, since the macro name from
5917 * rx.h is sort of strange this is better. This is called with a security
5918 * object before it is discarded. Each connection using a security object has
5919 * its own refcount to the object so it won't actually be freed until the last
5920 * connection is destroyed.
5922 * This is the only rxs module call. A hold could also be written but no one
5926 rxs_Release(struct rx_securityClass *aobj)
5928 return RXS_Close(aobj);
5932 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
5933 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
5934 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
5935 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
5937 /* Adjust our estimate of the transmission rate to this peer, given
5938 * that the packet p was just acked. We can adjust peer->timeout and
5939 * call->twind. Pragmatically, this is called
5940 * only with packets of maximal length.
5941 * Called with peer and call locked.
5945 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
5946 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
5948 afs_int32 xferSize, xferMs;
5949 register afs_int32 minTime;
5952 /* Count down packets */
5953 if (peer->rateFlag > 0)
5955 /* Do nothing until we're enabled */
5956 if (peer->rateFlag != 0)
5961 /* Count only when the ack seems legitimate */
5962 switch (ackReason) {
5963 case RX_ACK_REQUESTED:
5965 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
5969 case RX_ACK_PING_RESPONSE:
5970 if (p) /* want the response to ping-request, not data send */
5972 clock_GetTime(&newTO);
5973 if (clock_Gt(&newTO, &call->pingRequestTime)) {
5974 clock_Sub(&newTO, &call->pingRequestTime);
5975 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
5979 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
5986 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));
5988 /* Track only packets that are big enough. */
5989 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
5993 /* absorb RTT data (in milliseconds) for these big packets */
5994 if (peer->smRtt == 0) {
5995 peer->smRtt = xferMs;
5997 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6002 if (peer->countDown) {
6006 peer->countDown = 10; /* recalculate only every so often */
6008 /* In practice, we can measure only the RTT for full packets,
6009 * because of the way Rx acks the data that it receives. (If it's
6010 * smaller than a full packet, it often gets implicitly acked
6011 * either by the call response (from a server) or by the next call
6012 * (from a client), and either case confuses transmission times
6013 * with processing times.) Therefore, replace the above
6014 * more-sophisticated processing with a simpler version, where the
6015 * smoothed RTT is kept for full-size packets, and the time to
6016 * transmit a windowful of full-size packets is simply RTT *
6017 * windowSize. Again, we take two steps:
6018 - ensure the timeout is large enough for a single packet's RTT;
6019 - ensure that the window is small enough to fit in the desired timeout.*/
6021 /* First, the timeout check. */
6022 minTime = peer->smRtt;
6023 /* Get a reasonable estimate for a timeout period */
6025 newTO.sec = minTime / 1000;
6026 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6028 /* Increase the timeout period so that we can always do at least
6029 * one packet exchange */
6030 if (clock_Gt(&newTO, &peer->timeout)) {
6032 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));
6034 peer->timeout = newTO;
6037 /* Now, get an estimate for the transmit window size. */
6038 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6039 /* Now, convert to the number of full packets that could fit in a
6040 * reasonable fraction of that interval */
6041 minTime /= (peer->smRtt << 1);
6042 xferSize = minTime; /* (make a copy) */
6044 /* Now clamp the size to reasonable bounds. */
6047 else if (minTime > rx_Window)
6048 minTime = rx_Window;
6049 /* if (minTime != peer->maxWindow) {
6050 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6051 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6052 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6054 peer->maxWindow = minTime;
6055 elide... call->twind = minTime;
6059 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6060 * Discern this by calculating the timeout necessary for rx_Window
6062 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6063 /* calculate estimate for transmission interval in milliseconds */
6064 minTime = rx_Window * peer->smRtt;
6065 if (minTime < 1000) {
6066 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
6067 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6068 peer->timeout.usec, peer->smRtt, peer->packetSize));
6070 newTO.sec = 0; /* cut back on timeout by half a second */
6071 newTO.usec = 500000;
6072 clock_Sub(&peer->timeout, &newTO);
6077 } /* end of rxi_ComputeRate */
6078 #endif /* ADAPT_WINDOW */
6086 #define TRACE_OPTION_DEBUGLOG 4
6094 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6095 0, KEY_QUERY_VALUE, &parmKey);
6096 if (code != ERROR_SUCCESS)
6099 dummyLen = sizeof(TraceOption);
6100 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6101 (BYTE *) &TraceOption, &dummyLen);
6102 if (code == ERROR_SUCCESS) {
6103 rxdebug_active = (TraceOption & TRACE_OPTION_DEBUGLOG) ? 1 : 0;
6105 RegCloseKey (parmKey);
6106 #endif /* AFS_NT40_ENV */
6111 rx_DebugOnOff(int on)
6113 rxdebug_active = on;
6115 #endif /* AFS_NT40_ENV */
6118 /* Don't call this debugging routine directly; use dpf */
6120 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
6121 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
6129 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6132 len = _snprintf(msg, sizeof(msg)-2,
6133 tformat, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10,
6134 a11, a12, a13, a14, a15);
6136 if (msg[len-1] != '\n') {
6140 OutputDebugString(msg);
6145 clock_GetTime(&now);
6146 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
6147 (unsigned int)now.usec / 1000);
6148 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
6155 * This function is used to process the rx_stats structure that is local
6156 * to a process as well as an rx_stats structure received from a remote
6157 * process (via rxdebug). Therefore, it needs to do minimal version
6161 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
6162 afs_int32 freePackets, char version)
6166 if (size != sizeof(struct rx_stats)) {
6168 "Unexpected size of stats structure: was %d, expected %d\n",
6169 size, sizeof(struct rx_stats));
6172 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6175 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6176 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6177 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6178 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6179 s->specialPktAllocFailures);
6181 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6182 s->receivePktAllocFailures, s->sendPktAllocFailures,
6183 s->specialPktAllocFailures);
6187 " greedy %d, " "bogusReads %d (last from host %x), "
6188 "noPackets %d, " "noBuffers %d, " "selects %d, "
6189 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6190 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6191 s->selects, s->sendSelects);
6193 fprintf(file, " packets read: ");
6194 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6195 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6197 fprintf(file, "\n");
6200 " other read counters: data %d, " "ack %d, " "dup %d "
6201 "spurious %d " "dally %d\n", s->dataPacketsRead,
6202 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6203 s->ignorePacketDally);
6205 fprintf(file, " packets sent: ");
6206 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6207 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6209 fprintf(file, "\n");
6212 " other send counters: ack %d, " "data %d (not resends), "
6213 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6214 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6215 s->dataPacketsPushed, s->ignoreAckedPacket);
6218 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6219 s->netSendFailures, (int)s->fatalErrors);
6221 if (s->nRttSamples) {
6222 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6223 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6225 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6226 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6230 " %d server connections, " "%d client connections, "
6231 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6232 s->nServerConns, s->nClientConns, s->nPeerStructs,
6233 s->nCallStructs, s->nFreeCallStructs);
6235 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6236 fprintf(file, " %d clock updates\n", clock_nUpdates);
6241 /* for backward compatibility */
6243 rx_PrintStats(FILE * file)
6245 MUTEX_ENTER(&rx_stats_mutex);
6246 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6248 MUTEX_EXIT(&rx_stats_mutex);
6252 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6254 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6255 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6256 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6259 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6260 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6261 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6264 " Packet size %d, " "max in packet skew %d, "
6265 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6266 (int)peer->outPacketSkew);
6269 #ifdef AFS_PTHREAD_ENV
6271 * This mutex protects the following static variables:
6275 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6276 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6278 #define LOCK_RX_DEBUG
6279 #define UNLOCK_RX_DEBUG
6280 #endif /* AFS_PTHREAD_ENV */
6283 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6284 u_char type, void *inputData, size_t inputLength,
6285 void *outputData, size_t outputLength)
6287 static afs_int32 counter = 100;
6288 time_t waitTime, waitCount, startTime;
6289 struct rx_header theader;
6291 register afs_int32 code;
6292 struct timeval tv_now, tv_wake, tv_delta;
6293 struct sockaddr_in taddr, faddr;
6298 startTime = time(0);
6304 tp = &tbuffer[sizeof(struct rx_header)];
6305 taddr.sin_family = AF_INET;
6306 taddr.sin_port = remotePort;
6307 taddr.sin_addr.s_addr = remoteAddr;
6308 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6309 taddr.sin_len = sizeof(struct sockaddr_in);
6312 memset(&theader, 0, sizeof(theader));
6313 theader.epoch = htonl(999);
6315 theader.callNumber = htonl(counter);
6318 theader.type = type;
6319 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6320 theader.serviceId = 0;
6322 memcpy(tbuffer, &theader, sizeof(theader));
6323 memcpy(tp, inputData, inputLength);
6325 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6326 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6328 /* see if there's a packet available */
6329 gettimeofday(&tv_wake,0);
6330 tv_wake.tv_sec += waitTime;
6333 FD_SET(socket, &imask);
6334 tv_delta.tv_sec = tv_wake.tv_sec;
6335 tv_delta.tv_usec = tv_wake.tv_usec;
6336 gettimeofday(&tv_now, 0);
6338 if (tv_delta.tv_usec < tv_now.tv_usec) {
6340 tv_delta.tv_usec += 1000000;
6343 tv_delta.tv_usec -= tv_now.tv_usec;
6345 if (tv_delta.tv_sec < tv_now.tv_sec) {
6349 tv_delta.tv_sec -= tv_now.tv_sec;
6351 code = select(socket + 1, &imask, 0, 0, &tv_delta);
6352 if (code == 1 && FD_ISSET(socket, &imask)) {
6353 /* now receive a packet */
6354 faddrLen = sizeof(struct sockaddr_in);
6356 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6357 (struct sockaddr *)&faddr, &faddrLen);
6360 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6361 if (counter == ntohl(theader.callNumber))
6369 /* see if we've timed out */
6377 code -= sizeof(struct rx_header);
6378 if (code > outputLength)
6379 code = outputLength;
6380 memcpy(outputData, tp, code);
6385 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6386 afs_uint16 remotePort, struct rx_debugStats * stat,
6387 afs_uint32 * supportedValues)
6389 struct rx_debugIn in;
6392 *supportedValues = 0;
6393 in.type = htonl(RX_DEBUGI_GETSTATS);
6396 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6397 &in, sizeof(in), stat, sizeof(*stat));
6400 * If the call was successful, fixup the version and indicate
6401 * what contents of the stat structure are valid.
6402 * Also do net to host conversion of fields here.
6406 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6407 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6409 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6410 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6412 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6413 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6415 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6416 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6418 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6419 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6421 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6422 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6424 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6425 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6427 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6428 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6431 stat->nFreePackets = ntohl(stat->nFreePackets);
6432 stat->packetReclaims = ntohl(stat->packetReclaims);
6433 stat->callsExecuted = ntohl(stat->callsExecuted);
6434 stat->nWaiting = ntohl(stat->nWaiting);
6435 stat->idleThreads = ntohl(stat->idleThreads);
6442 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6443 afs_uint16 remotePort, struct rx_stats * stat,
6444 afs_uint32 * supportedValues)
6446 struct rx_debugIn in;
6447 afs_int32 *lp = (afs_int32 *) stat;
6452 * supportedValues is currently unused, but added to allow future
6453 * versioning of this function.
6456 *supportedValues = 0;
6457 in.type = htonl(RX_DEBUGI_RXSTATS);
6459 memset(stat, 0, sizeof(*stat));
6461 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6462 &in, sizeof(in), stat, sizeof(*stat));
6467 * Do net to host conversion here
6470 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6479 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6480 afs_uint16 remotePort, size_t version_length,
6484 return MakeDebugCall(socket, remoteAddr, remotePort,
6485 RX_PACKET_TYPE_VERSION, a, 1, version,
6490 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6491 afs_uint16 remotePort, afs_int32 * nextConnection,
6492 int allConnections, afs_uint32 debugSupportedValues,
6493 struct rx_debugConn * conn,
6494 afs_uint32 * supportedValues)
6496 struct rx_debugIn in;
6501 * supportedValues is currently unused, but added to allow future
6502 * versioning of this function.
6505 *supportedValues = 0;
6506 if (allConnections) {
6507 in.type = htonl(RX_DEBUGI_GETALLCONN);
6509 in.type = htonl(RX_DEBUGI_GETCONN);
6511 in.index = htonl(*nextConnection);
6512 memset(conn, 0, sizeof(*conn));
6514 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6515 &in, sizeof(in), conn, sizeof(*conn));
6518 *nextConnection += 1;
6521 * Convert old connection format to new structure.
6524 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6525 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6526 #define MOVEvL(a) (conn->a = vL->a)
6528 /* any old or unrecognized version... */
6529 for (i = 0; i < RX_MAXCALLS; i++) {
6530 MOVEvL(callState[i]);
6531 MOVEvL(callMode[i]);
6532 MOVEvL(callFlags[i]);
6533 MOVEvL(callOther[i]);
6535 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6536 MOVEvL(secStats.type);
6537 MOVEvL(secStats.level);
6538 MOVEvL(secStats.flags);
6539 MOVEvL(secStats.expires);
6540 MOVEvL(secStats.packetsReceived);
6541 MOVEvL(secStats.packetsSent);
6542 MOVEvL(secStats.bytesReceived);
6543 MOVEvL(secStats.bytesSent);
6548 * Do net to host conversion here
6550 * I don't convert host or port since we are most likely
6551 * going to want these in NBO.
6553 conn->cid = ntohl(conn->cid);
6554 conn->serial = ntohl(conn->serial);
6555 for (i = 0; i < RX_MAXCALLS; i++) {
6556 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6558 conn->error = ntohl(conn->error);
6559 conn->secStats.flags = ntohl(conn->secStats.flags);
6560 conn->secStats.expires = ntohl(conn->secStats.expires);
6561 conn->secStats.packetsReceived =
6562 ntohl(conn->secStats.packetsReceived);
6563 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6564 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6565 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6566 conn->epoch = ntohl(conn->epoch);
6567 conn->natMTU = ntohl(conn->natMTU);
6574 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6575 afs_uint16 remotePort, afs_int32 * nextPeer,
6576 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6577 afs_uint32 * supportedValues)
6579 struct rx_debugIn in;
6583 * supportedValues is currently unused, but added to allow future
6584 * versioning of this function.
6587 *supportedValues = 0;
6588 in.type = htonl(RX_DEBUGI_GETPEER);
6589 in.index = htonl(*nextPeer);
6590 memset(peer, 0, sizeof(*peer));
6592 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6593 &in, sizeof(in), peer, sizeof(*peer));
6599 * Do net to host conversion here
6601 * I don't convert host or port since we are most likely
6602 * going to want these in NBO.
6604 peer->ifMTU = ntohs(peer->ifMTU);
6605 peer->idleWhen = ntohl(peer->idleWhen);
6606 peer->refCount = ntohs(peer->refCount);
6607 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6608 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6609 peer->rtt = ntohl(peer->rtt);
6610 peer->rtt_dev = ntohl(peer->rtt_dev);
6611 peer->timeout.sec = ntohl(peer->timeout.sec);
6612 peer->timeout.usec = ntohl(peer->timeout.usec);
6613 peer->nSent = ntohl(peer->nSent);
6614 peer->reSends = ntohl(peer->reSends);
6615 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6616 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6617 peer->rateFlag = ntohl(peer->rateFlag);
6618 peer->natMTU = ntohs(peer->natMTU);
6619 peer->maxMTU = ntohs(peer->maxMTU);
6620 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6621 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6622 peer->MTU = ntohs(peer->MTU);
6623 peer->cwind = ntohs(peer->cwind);
6624 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6625 peer->congestSeq = ntohs(peer->congestSeq);
6626 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6627 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6628 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6629 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6634 #endif /* RXDEBUG */
6639 struct rx_serverQueueEntry *np;
6642 register struct rx_call *call;
6643 register struct rx_serverQueueEntry *sq;
6647 if (rxinit_status == 1) {
6649 return; /* Already shutdown. */
6653 #ifndef AFS_PTHREAD_ENV
6654 FD_ZERO(&rx_selectMask);
6655 #endif /* AFS_PTHREAD_ENV */
6656 rxi_dataQuota = RX_MAX_QUOTA;
6657 #ifndef AFS_PTHREAD_ENV
6659 #endif /* AFS_PTHREAD_ENV */
6662 #ifndef AFS_PTHREAD_ENV
6663 #ifndef AFS_USE_GETTIMEOFDAY
6665 #endif /* AFS_USE_GETTIMEOFDAY */
6666 #endif /* AFS_PTHREAD_ENV */
6668 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6669 call = queue_First(&rx_freeCallQueue, rx_call);
6671 rxi_Free(call, sizeof(struct rx_call));
6674 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6675 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6681 struct rx_peer **peer_ptr, **peer_end;
6682 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6683 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6685 struct rx_peer *peer, *next;
6686 for (peer = *peer_ptr; peer; peer = next) {
6687 rx_interface_stat_p rpc_stat, nrpc_stat;
6690 (&peer->rpcStats, rpc_stat, nrpc_stat,
6691 rx_interface_stat)) {
6692 unsigned int num_funcs;
6695 queue_Remove(&rpc_stat->queue_header);
6696 queue_Remove(&rpc_stat->all_peers);
6697 num_funcs = rpc_stat->stats[0].func_total;
6699 sizeof(rx_interface_stat_t) +
6700 rpc_stat->stats[0].func_total *
6701 sizeof(rx_function_entry_v1_t);
6703 rxi_Free(rpc_stat, space);
6704 MUTEX_ENTER(&rx_rpc_stats);
6705 rxi_rpc_peer_stat_cnt -= num_funcs;
6706 MUTEX_EXIT(&rx_rpc_stats);
6710 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6714 for (i = 0; i < RX_MAX_SERVICES; i++) {
6716 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6718 for (i = 0; i < rx_hashTableSize; i++) {
6719 register struct rx_connection *tc, *ntc;
6720 MUTEX_ENTER(&rx_connHashTable_lock);
6721 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6723 for (j = 0; j < RX_MAXCALLS; j++) {
6725 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6728 rxi_Free(tc, sizeof(*tc));
6730 MUTEX_EXIT(&rx_connHashTable_lock);
6733 MUTEX_ENTER(&freeSQEList_lock);
6735 while ((np = rx_FreeSQEList)) {
6736 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6737 MUTEX_DESTROY(&np->lock);
6738 rxi_Free(np, sizeof(*np));
6741 MUTEX_EXIT(&freeSQEList_lock);
6742 MUTEX_DESTROY(&freeSQEList_lock);
6743 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6744 MUTEX_DESTROY(&rx_connHashTable_lock);
6745 MUTEX_DESTROY(&rx_peerHashTable_lock);
6746 MUTEX_DESTROY(&rx_serverPool_lock);
6748 osi_Free(rx_connHashTable,
6749 rx_hashTableSize * sizeof(struct rx_connection *));
6750 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6752 UNPIN(rx_connHashTable,
6753 rx_hashTableSize * sizeof(struct rx_connection *));
6754 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6756 rxi_FreeAllPackets();
6758 MUTEX_ENTER(&rx_stats_mutex);
6759 rxi_dataQuota = RX_MAX_QUOTA;
6760 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6761 MUTEX_EXIT(&rx_stats_mutex);
6767 #ifdef RX_ENABLE_LOCKS
6769 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6771 if (!MUTEX_ISMINE(lockaddr))
6772 osi_Panic("Lock not held: %s", msg);
6774 #endif /* RX_ENABLE_LOCKS */
6779 * Routines to implement connection specific data.
6783 rx_KeyCreate(rx_destructor_t rtn)
6786 MUTEX_ENTER(&rxi_keyCreate_lock);
6787 key = rxi_keyCreate_counter++;
6788 rxi_keyCreate_destructor = (rx_destructor_t *)
6789 realloc((void *)rxi_keyCreate_destructor,
6790 (key + 1) * sizeof(rx_destructor_t));
6791 rxi_keyCreate_destructor[key] = rtn;
6792 MUTEX_EXIT(&rxi_keyCreate_lock);
6797 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6800 MUTEX_ENTER(&conn->conn_data_lock);
6801 if (!conn->specific) {
6802 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6803 for (i = 0; i < key; i++)
6804 conn->specific[i] = NULL;
6805 conn->nSpecific = key + 1;
6806 conn->specific[key] = ptr;
6807 } else if (key >= conn->nSpecific) {
6808 conn->specific = (void **)
6809 realloc(conn->specific, (key + 1) * sizeof(void *));
6810 for (i = conn->nSpecific; i < key; i++)
6811 conn->specific[i] = NULL;
6812 conn->nSpecific = key + 1;
6813 conn->specific[key] = ptr;
6815 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6816 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6817 conn->specific[key] = ptr;
6819 MUTEX_EXIT(&conn->conn_data_lock);
6823 rx_GetSpecific(struct rx_connection *conn, int key)
6826 MUTEX_ENTER(&conn->conn_data_lock);
6827 if (key >= conn->nSpecific)
6830 ptr = conn->specific[key];
6831 MUTEX_EXIT(&conn->conn_data_lock);
6835 #endif /* !KERNEL */
6838 * processStats is a queue used to store the statistics for the local
6839 * process. Its contents are similar to the contents of the rpcStats
6840 * queue on a rx_peer structure, but the actual data stored within
6841 * this queue contains totals across the lifetime of the process (assuming
6842 * the stats have not been reset) - unlike the per peer structures
6843 * which can come and go based upon the peer lifetime.
6846 static struct rx_queue processStats = { &processStats, &processStats };
6849 * peerStats is a queue used to store the statistics for all peer structs.
6850 * Its contents are the union of all the peer rpcStats queues.
6853 static struct rx_queue peerStats = { &peerStats, &peerStats };
6856 * rxi_monitor_processStats is used to turn process wide stat collection
6860 static int rxi_monitor_processStats = 0;
6863 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6866 static int rxi_monitor_peerStats = 0;
6869 * rxi_AddRpcStat - given all of the information for a particular rpc
6870 * call, create (if needed) and update the stat totals for the rpc.
6874 * IN stats - the queue of stats that will be updated with the new value
6876 * IN rxInterface - a unique number that identifies the rpc interface
6878 * IN currentFunc - the index of the function being invoked
6880 * IN totalFunc - the total number of functions in this interface
6882 * IN queueTime - the amount of time this function waited for a thread
6884 * IN execTime - the amount of time this function invocation took to execute
6886 * IN bytesSent - the number bytes sent by this invocation
6888 * IN bytesRcvd - the number bytes received by this invocation
6890 * IN isServer - if true, this invocation was made to a server
6892 * IN remoteHost - the ip address of the remote host
6894 * IN remotePort - the port of the remote host
6896 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6898 * INOUT counter - if a new stats structure is allocated, the counter will
6899 * be updated with the new number of allocated stat structures
6907 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6908 afs_uint32 currentFunc, afs_uint32 totalFunc,
6909 struct clock *queueTime, struct clock *execTime,
6910 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6911 afs_uint32 remoteHost, afs_uint32 remotePort,
6912 int addToPeerList, unsigned int *counter)
6915 rx_interface_stat_p rpc_stat, nrpc_stat;
6918 * See if there's already a structure for this interface
6921 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6922 if ((rpc_stat->stats[0].interfaceId == rxInterface)
6923 && (rpc_stat->stats[0].remote_is_server == isServer))
6928 * Didn't find a match so allocate a new structure and add it to the
6932 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
6933 || (rpc_stat->stats[0].interfaceId != rxInterface)
6934 || (rpc_stat->stats[0].remote_is_server != isServer)) {
6939 sizeof(rx_interface_stat_t) +
6940 totalFunc * sizeof(rx_function_entry_v1_t);
6942 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
6943 if (rpc_stat == NULL) {
6947 *counter += totalFunc;
6948 for (i = 0; i < totalFunc; i++) {
6949 rpc_stat->stats[i].remote_peer = remoteHost;
6950 rpc_stat->stats[i].remote_port = remotePort;
6951 rpc_stat->stats[i].remote_is_server = isServer;
6952 rpc_stat->stats[i].interfaceId = rxInterface;
6953 rpc_stat->stats[i].func_total = totalFunc;
6954 rpc_stat->stats[i].func_index = i;
6955 hzero(rpc_stat->stats[i].invocations);
6956 hzero(rpc_stat->stats[i].bytes_sent);
6957 hzero(rpc_stat->stats[i].bytes_rcvd);
6958 rpc_stat->stats[i].queue_time_sum.sec = 0;
6959 rpc_stat->stats[i].queue_time_sum.usec = 0;
6960 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
6961 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
6962 rpc_stat->stats[i].queue_time_min.sec = 9999999;
6963 rpc_stat->stats[i].queue_time_min.usec = 9999999;
6964 rpc_stat->stats[i].queue_time_max.sec = 0;
6965 rpc_stat->stats[i].queue_time_max.usec = 0;
6966 rpc_stat->stats[i].execution_time_sum.sec = 0;
6967 rpc_stat->stats[i].execution_time_sum.usec = 0;
6968 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
6969 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
6970 rpc_stat->stats[i].execution_time_min.sec = 9999999;
6971 rpc_stat->stats[i].execution_time_min.usec = 9999999;
6972 rpc_stat->stats[i].execution_time_max.sec = 0;
6973 rpc_stat->stats[i].execution_time_max.usec = 0;
6975 queue_Prepend(stats, rpc_stat);
6976 if (addToPeerList) {
6977 queue_Prepend(&peerStats, &rpc_stat->all_peers);
6982 * Increment the stats for this function
6985 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
6986 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
6987 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
6988 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
6989 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
6990 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
6991 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
6993 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
6994 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
6996 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
6997 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
6999 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7000 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7002 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7003 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7011 * rx_IncrementTimeAndCount - increment the times and count for a particular
7016 * IN peer - the peer who invoked the rpc
7018 * IN rxInterface - a unique number that identifies the rpc interface
7020 * IN currentFunc - the index of the function being invoked
7022 * IN totalFunc - the total number of functions in this interface
7024 * IN queueTime - the amount of time this function waited for a thread
7026 * IN execTime - the amount of time this function invocation took to execute
7028 * IN bytesSent - the number bytes sent by this invocation
7030 * IN bytesRcvd - the number bytes received by this invocation
7032 * IN isServer - if true, this invocation was made to a server
7040 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7041 afs_uint32 currentFunc, afs_uint32 totalFunc,
7042 struct clock *queueTime, struct clock *execTime,
7043 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7047 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
7050 MUTEX_ENTER(&rx_rpc_stats);
7051 MUTEX_ENTER(&peer->peer_lock);
7053 if (rxi_monitor_peerStats) {
7054 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7055 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7056 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7059 if (rxi_monitor_processStats) {
7060 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7061 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7062 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7065 MUTEX_EXIT(&peer->peer_lock);
7066 MUTEX_EXIT(&rx_rpc_stats);
7071 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7075 * IN callerVersion - the rpc stat version of the caller.
7077 * IN count - the number of entries to marshall.
7079 * IN stats - pointer to stats to be marshalled.
7081 * OUT ptr - Where to store the marshalled data.
7088 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7089 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7095 * We only support the first version
7097 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7098 *(ptr++) = stats->remote_peer;
7099 *(ptr++) = stats->remote_port;
7100 *(ptr++) = stats->remote_is_server;
7101 *(ptr++) = stats->interfaceId;
7102 *(ptr++) = stats->func_total;
7103 *(ptr++) = stats->func_index;
7104 *(ptr++) = hgethi(stats->invocations);
7105 *(ptr++) = hgetlo(stats->invocations);
7106 *(ptr++) = hgethi(stats->bytes_sent);
7107 *(ptr++) = hgetlo(stats->bytes_sent);
7108 *(ptr++) = hgethi(stats->bytes_rcvd);
7109 *(ptr++) = hgetlo(stats->bytes_rcvd);
7110 *(ptr++) = stats->queue_time_sum.sec;
7111 *(ptr++) = stats->queue_time_sum.usec;
7112 *(ptr++) = stats->queue_time_sum_sqr.sec;
7113 *(ptr++) = stats->queue_time_sum_sqr.usec;
7114 *(ptr++) = stats->queue_time_min.sec;
7115 *(ptr++) = stats->queue_time_min.usec;
7116 *(ptr++) = stats->queue_time_max.sec;
7117 *(ptr++) = stats->queue_time_max.usec;
7118 *(ptr++) = stats->execution_time_sum.sec;
7119 *(ptr++) = stats->execution_time_sum.usec;
7120 *(ptr++) = stats->execution_time_sum_sqr.sec;
7121 *(ptr++) = stats->execution_time_sum_sqr.usec;
7122 *(ptr++) = stats->execution_time_min.sec;
7123 *(ptr++) = stats->execution_time_min.usec;
7124 *(ptr++) = stats->execution_time_max.sec;
7125 *(ptr++) = stats->execution_time_max.usec;
7131 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7136 * IN callerVersion - the rpc stat version of the caller
7138 * OUT myVersion - the rpc stat version of this function
7140 * OUT clock_sec - local time seconds
7142 * OUT clock_usec - local time microseconds
7144 * OUT allocSize - the number of bytes allocated to contain stats
7146 * OUT statCount - the number stats retrieved from this process.
7148 * OUT stats - the actual stats retrieved from this process.
7152 * Returns void. If successful, stats will != NULL.
7156 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7157 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7158 size_t * allocSize, afs_uint32 * statCount,
7159 afs_uint32 ** stats)
7169 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7172 * Check to see if stats are enabled
7175 MUTEX_ENTER(&rx_rpc_stats);
7176 if (!rxi_monitor_processStats) {
7177 MUTEX_EXIT(&rx_rpc_stats);
7181 clock_GetTime(&now);
7182 *clock_sec = now.sec;
7183 *clock_usec = now.usec;
7186 * Allocate the space based upon the caller version
7188 * If the client is at an older version than we are,
7189 * we return the statistic data in the older data format, but
7190 * we still return our version number so the client knows we
7191 * are maintaining more data than it can retrieve.
7194 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7195 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7196 *statCount = rxi_rpc_process_stat_cnt;
7199 * This can't happen yet, but in the future version changes
7200 * can be handled by adding additional code here
7204 if (space > (size_t) 0) {
7206 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7209 rx_interface_stat_p rpc_stat, nrpc_stat;
7213 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7215 * Copy the data based upon the caller version
7217 rx_MarshallProcessRPCStats(callerVersion,
7218 rpc_stat->stats[0].func_total,
7219 rpc_stat->stats, &ptr);
7225 MUTEX_EXIT(&rx_rpc_stats);
7230 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7234 * IN callerVersion - the rpc stat version of the caller
7236 * OUT myVersion - the rpc stat version of this function
7238 * OUT clock_sec - local time seconds
7240 * OUT clock_usec - local time microseconds
7242 * OUT allocSize - the number of bytes allocated to contain stats
7244 * OUT statCount - the number of stats retrieved from the individual
7247 * OUT stats - the actual stats retrieved from the individual peer structures.
7251 * Returns void. If successful, stats will != NULL.
7255 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7256 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7257 size_t * allocSize, afs_uint32 * statCount,
7258 afs_uint32 ** stats)
7268 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7271 * Check to see if stats are enabled
7274 MUTEX_ENTER(&rx_rpc_stats);
7275 if (!rxi_monitor_peerStats) {
7276 MUTEX_EXIT(&rx_rpc_stats);
7280 clock_GetTime(&now);
7281 *clock_sec = now.sec;
7282 *clock_usec = now.usec;
7285 * Allocate the space based upon the caller version
7287 * If the client is at an older version than we are,
7288 * we return the statistic data in the older data format, but
7289 * we still return our version number so the client knows we
7290 * are maintaining more data than it can retrieve.
7293 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7294 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7295 *statCount = rxi_rpc_peer_stat_cnt;
7298 * This can't happen yet, but in the future version changes
7299 * can be handled by adding additional code here
7303 if (space > (size_t) 0) {
7305 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7308 rx_interface_stat_p rpc_stat, nrpc_stat;
7312 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7314 * We have to fix the offset of rpc_stat since we are
7315 * keeping this structure on two rx_queues. The rx_queue
7316 * package assumes that the rx_queue member is the first
7317 * member of the structure. That is, rx_queue assumes that
7318 * any one item is only on one queue at a time. We are
7319 * breaking that assumption and so we have to do a little
7320 * math to fix our pointers.
7323 fix_offset = (char *)rpc_stat;
7324 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7325 rpc_stat = (rx_interface_stat_p) fix_offset;
7328 * Copy the data based upon the caller version
7330 rx_MarshallProcessRPCStats(callerVersion,
7331 rpc_stat->stats[0].func_total,
7332 rpc_stat->stats, &ptr);
7338 MUTEX_EXIT(&rx_rpc_stats);
7343 * rx_FreeRPCStats - free memory allocated by
7344 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7348 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7349 * rx_RetrievePeerRPCStats
7351 * IN allocSize - the number of bytes in stats.
7359 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7361 rxi_Free(stats, allocSize);
7365 * rx_queryProcessRPCStats - see if process rpc stat collection is
7366 * currently enabled.
7372 * Returns 0 if stats are not enabled != 0 otherwise
7376 rx_queryProcessRPCStats(void)
7379 MUTEX_ENTER(&rx_rpc_stats);
7380 rc = rxi_monitor_processStats;
7381 MUTEX_EXIT(&rx_rpc_stats);
7386 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7392 * Returns 0 if stats are not enabled != 0 otherwise
7396 rx_queryPeerRPCStats(void)
7399 MUTEX_ENTER(&rx_rpc_stats);
7400 rc = rxi_monitor_peerStats;
7401 MUTEX_EXIT(&rx_rpc_stats);
7406 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7416 rx_enableProcessRPCStats(void)
7418 MUTEX_ENTER(&rx_rpc_stats);
7419 rx_enable_stats = 1;
7420 rxi_monitor_processStats = 1;
7421 MUTEX_EXIT(&rx_rpc_stats);
7425 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7435 rx_enablePeerRPCStats(void)
7437 MUTEX_ENTER(&rx_rpc_stats);
7438 rx_enable_stats = 1;
7439 rxi_monitor_peerStats = 1;
7440 MUTEX_EXIT(&rx_rpc_stats);
7444 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7454 rx_disableProcessRPCStats(void)
7456 rx_interface_stat_p rpc_stat, nrpc_stat;
7459 MUTEX_ENTER(&rx_rpc_stats);
7462 * Turn off process statistics and if peer stats is also off, turn
7466 rxi_monitor_processStats = 0;
7467 if (rxi_monitor_peerStats == 0) {
7468 rx_enable_stats = 0;
7471 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7472 unsigned int num_funcs = 0;
7475 queue_Remove(rpc_stat);
7476 num_funcs = rpc_stat->stats[0].func_total;
7478 sizeof(rx_interface_stat_t) +
7479 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7481 rxi_Free(rpc_stat, space);
7482 rxi_rpc_process_stat_cnt -= num_funcs;
7484 MUTEX_EXIT(&rx_rpc_stats);
7488 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7498 rx_disablePeerRPCStats(void)
7500 struct rx_peer **peer_ptr, **peer_end;
7503 MUTEX_ENTER(&rx_rpc_stats);
7506 * Turn off peer statistics and if process stats is also off, turn
7510 rxi_monitor_peerStats = 0;
7511 if (rxi_monitor_processStats == 0) {
7512 rx_enable_stats = 0;
7515 MUTEX_ENTER(&rx_peerHashTable_lock);
7516 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7517 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7519 struct rx_peer *peer, *next, *prev;
7520 for (prev = peer = *peer_ptr; peer; peer = next) {
7522 code = MUTEX_TRYENTER(&peer->peer_lock);
7524 rx_interface_stat_p rpc_stat, nrpc_stat;
7527 (&peer->rpcStats, rpc_stat, nrpc_stat,
7528 rx_interface_stat)) {
7529 unsigned int num_funcs = 0;
7532 queue_Remove(&rpc_stat->queue_header);
7533 queue_Remove(&rpc_stat->all_peers);
7534 num_funcs = rpc_stat->stats[0].func_total;
7536 sizeof(rx_interface_stat_t) +
7537 rpc_stat->stats[0].func_total *
7538 sizeof(rx_function_entry_v1_t);
7540 rxi_Free(rpc_stat, space);
7541 rxi_rpc_peer_stat_cnt -= num_funcs;
7543 MUTEX_EXIT(&peer->peer_lock);
7544 if (prev == *peer_ptr) {
7554 MUTEX_EXIT(&rx_peerHashTable_lock);
7555 MUTEX_EXIT(&rx_rpc_stats);
7559 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7564 * IN clearFlag - flag indicating which stats to clear
7572 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7574 rx_interface_stat_p rpc_stat, nrpc_stat;
7576 MUTEX_ENTER(&rx_rpc_stats);
7578 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7579 unsigned int num_funcs = 0, i;
7580 num_funcs = rpc_stat->stats[0].func_total;
7581 for (i = 0; i < num_funcs; i++) {
7582 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7583 hzero(rpc_stat->stats[i].invocations);
7585 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7586 hzero(rpc_stat->stats[i].bytes_sent);
7588 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7589 hzero(rpc_stat->stats[i].bytes_rcvd);
7591 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7592 rpc_stat->stats[i].queue_time_sum.sec = 0;
7593 rpc_stat->stats[i].queue_time_sum.usec = 0;
7595 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7596 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7597 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7599 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7600 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7601 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7603 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7604 rpc_stat->stats[i].queue_time_max.sec = 0;
7605 rpc_stat->stats[i].queue_time_max.usec = 0;
7607 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7608 rpc_stat->stats[i].execution_time_sum.sec = 0;
7609 rpc_stat->stats[i].execution_time_sum.usec = 0;
7611 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7612 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7613 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7615 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7616 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7617 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7619 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7620 rpc_stat->stats[i].execution_time_max.sec = 0;
7621 rpc_stat->stats[i].execution_time_max.usec = 0;
7626 MUTEX_EXIT(&rx_rpc_stats);
7630 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7635 * IN clearFlag - flag indicating which stats to clear
7643 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7645 rx_interface_stat_p rpc_stat, nrpc_stat;
7647 MUTEX_ENTER(&rx_rpc_stats);
7649 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7650 unsigned int num_funcs = 0, i;
7653 * We have to fix the offset of rpc_stat since we are
7654 * keeping this structure on two rx_queues. The rx_queue
7655 * package assumes that the rx_queue member is the first
7656 * member of the structure. That is, rx_queue assumes that
7657 * any one item is only on one queue at a time. We are
7658 * breaking that assumption and so we have to do a little
7659 * math to fix our pointers.
7662 fix_offset = (char *)rpc_stat;
7663 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7664 rpc_stat = (rx_interface_stat_p) fix_offset;
7666 num_funcs = rpc_stat->stats[0].func_total;
7667 for (i = 0; i < num_funcs; i++) {
7668 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7669 hzero(rpc_stat->stats[i].invocations);
7671 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7672 hzero(rpc_stat->stats[i].bytes_sent);
7674 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7675 hzero(rpc_stat->stats[i].bytes_rcvd);
7677 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7678 rpc_stat->stats[i].queue_time_sum.sec = 0;
7679 rpc_stat->stats[i].queue_time_sum.usec = 0;
7681 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7682 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7683 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7685 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7686 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7687 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7689 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7690 rpc_stat->stats[i].queue_time_max.sec = 0;
7691 rpc_stat->stats[i].queue_time_max.usec = 0;
7693 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7694 rpc_stat->stats[i].execution_time_sum.sec = 0;
7695 rpc_stat->stats[i].execution_time_sum.usec = 0;
7697 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7698 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7699 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7701 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7702 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7703 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7705 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7706 rpc_stat->stats[i].execution_time_max.sec = 0;
7707 rpc_stat->stats[i].execution_time_max.usec = 0;
7712 MUTEX_EXIT(&rx_rpc_stats);
7716 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7717 * is authorized to enable/disable/clear RX statistics.
7719 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7722 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7724 rxi_rxstat_userok = proc;
7728 rx_RxStatUserOk(struct rx_call *call)
7730 if (!rxi_rxstat_userok)
7732 return rxi_rxstat_userok(call);
7737 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
7738 * function in the MSVC runtime DLL (msvcrt.dll).
7740 * Note: the system serializes calls to this function.
7743 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
7744 DWORD reason, /* reason function is being called */
7745 LPVOID reserved) /* reserved for future use */
7748 case DLL_PROCESS_ATTACH:
7749 /* library is being attached to a process */
7753 case DLL_PROCESS_DETACH: