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_ALPHA_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>
86 # include <afs/afsutil.h>
88 # include <sys/socket.h>
89 # include <sys/file.h>
91 # include <sys/stat.h>
92 # include <netinet/in.h>
93 # include <sys/time.h>
103 # include "rx_user.h"
104 # include "rx_clock.h"
105 # include "rx_queue.h"
106 # include "rx_globals.h"
107 # include "rx_trace.h"
108 # include <afs/rxgen_consts.h>
111 int (*registerProgram) () = 0;
112 int (*swapNameProgram) () = 0;
114 /* Local static routines */
115 static void rxi_DestroyConnectionNoLock(register struct rx_connection *conn);
116 #ifdef RX_ENABLE_LOCKS
117 static void rxi_SetAcksInTransmitQueue(register struct rx_call *call);
120 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
122 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
123 afs_int32 rxi_start_in_error;
125 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
128 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
129 * currently allocated within rx. This number is used to allocate the
130 * memory required to return the statistics when queried.
133 static unsigned int rxi_rpc_peer_stat_cnt;
136 * rxi_rpc_process_stat_cnt counts the total number of local process stat
137 * structures currently allocated within rx. The number is used to allocate
138 * the memory required to return the statistics when queried.
141 static unsigned int rxi_rpc_process_stat_cnt;
143 #if !defined(offsetof)
144 #include <stddef.h> /* for definition of offsetof() */
147 #ifdef AFS_PTHREAD_ENV
151 * Use procedural initialization of mutexes/condition variables
155 extern pthread_mutex_t rx_stats_mutex;
156 extern pthread_mutex_t rxkad_stats_mutex;
157 extern pthread_mutex_t des_init_mutex;
158 extern pthread_mutex_t des_random_mutex;
159 extern pthread_mutex_t rx_clock_mutex;
160 extern pthread_mutex_t rxi_connCacheMutex;
161 extern pthread_mutex_t rx_event_mutex;
162 extern pthread_mutex_t osi_malloc_mutex;
163 extern pthread_mutex_t event_handler_mutex;
164 extern pthread_mutex_t listener_mutex;
165 extern pthread_mutex_t rx_if_init_mutex;
166 extern pthread_mutex_t rx_if_mutex;
167 extern pthread_mutex_t rxkad_client_uid_mutex;
168 extern pthread_mutex_t rxkad_random_mutex;
170 extern pthread_cond_t rx_event_handler_cond;
171 extern pthread_cond_t rx_listener_cond;
173 static pthread_mutex_t epoch_mutex;
174 static pthread_mutex_t rx_init_mutex;
175 static pthread_mutex_t rx_debug_mutex;
178 rxi_InitPthread(void)
180 assert(pthread_mutex_init(&rx_clock_mutex, (const pthread_mutexattr_t *)0)
182 assert(pthread_mutex_init(&rx_stats_mutex, (const pthread_mutexattr_t *)0)
184 assert(pthread_mutex_init
185 (&rxi_connCacheMutex, (const pthread_mutexattr_t *)0) == 0);
186 assert(pthread_mutex_init(&rx_init_mutex, (const pthread_mutexattr_t *)0)
188 assert(pthread_mutex_init(&epoch_mutex, (const pthread_mutexattr_t *)0) ==
190 assert(pthread_mutex_init(&rx_event_mutex, (const pthread_mutexattr_t *)0)
192 assert(pthread_mutex_init(&des_init_mutex, (const pthread_mutexattr_t *)0)
194 assert(pthread_mutex_init
195 (&des_random_mutex, (const pthread_mutexattr_t *)0) == 0);
196 assert(pthread_mutex_init
197 (&osi_malloc_mutex, (const pthread_mutexattr_t *)0) == 0);
198 assert(pthread_mutex_init
199 (&event_handler_mutex, (const pthread_mutexattr_t *)0) == 0);
200 assert(pthread_mutex_init(&listener_mutex, (const pthread_mutexattr_t *)0)
202 assert(pthread_mutex_init
203 (&rx_if_init_mutex, (const pthread_mutexattr_t *)0) == 0);
204 assert(pthread_mutex_init(&rx_if_mutex, (const pthread_mutexattr_t *)0) ==
206 assert(pthread_mutex_init
207 (&rxkad_client_uid_mutex, (const pthread_mutexattr_t *)0) == 0);
208 assert(pthread_mutex_init
209 (&rxkad_random_mutex, (const pthread_mutexattr_t *)0) == 0);
210 assert(pthread_mutex_init
211 (&rxkad_stats_mutex, (const pthread_mutexattr_t *)0) == 0);
212 assert(pthread_mutex_init(&rx_debug_mutex, (const pthread_mutexattr_t *)0)
215 assert(pthread_cond_init
216 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
217 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
219 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
222 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
223 #define INIT_PTHREAD_LOCKS \
224 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0);
226 * The rx_stats_mutex mutex protects the following global variables:
231 * rxi_lowConnRefCount
232 * rxi_lowPeerRefCount
241 #define INIT_PTHREAD_LOCKS
245 /* Variables for handling the minProcs implementation. availProcs gives the
246 * number of threads available in the pool at this moment (not counting dudes
247 * executing right now). totalMin gives the total number of procs required
248 * for handling all minProcs requests. minDeficit is a dynamic variable
249 * tracking the # of procs required to satisfy all of the remaining minProcs
251 * For fine grain locking to work, the quota check and the reservation of
252 * a server thread has to come while rxi_availProcs and rxi_minDeficit
253 * are locked. To this end, the code has been modified under #ifdef
254 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
255 * same time. A new function, ReturnToServerPool() returns the allocation.
257 * A call can be on several queue's (but only one at a time). When
258 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
259 * that no one else is touching the queue. To this end, we store the address
260 * of the queue lock in the call structure (under the call lock) when we
261 * put the call on a queue, and we clear the call_queue_lock when the
262 * call is removed from a queue (once the call lock has been obtained).
263 * This allows rxi_ResetCall to safely synchronize with others wishing
264 * to manipulate the queue.
267 #ifdef RX_ENABLE_LOCKS
268 static afs_kmutex_t rx_rpc_stats;
269 void rxi_StartUnlocked();
272 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
273 ** pretty good that the next packet coming in is from the same connection
274 ** as the last packet, since we're send multiple packets in a transmit window.
276 struct rx_connection *rxLastConn = 0;
278 #ifdef RX_ENABLE_LOCKS
279 /* The locking hierarchy for rx fine grain locking is composed of these
282 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
283 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
284 * call->lock - locks call data fields.
285 * These are independent of each other:
286 * rx_freeCallQueue_lock
291 * serverQueueEntry->lock
293 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
294 * peer->lock - locks peer data fields.
295 * conn_data_lock - that more than one thread is not updating a conn data
296 * field at the same time.
304 * Do we need a lock to protect the peer field in the conn structure?
305 * conn->peer was previously a constant for all intents and so has no
306 * lock protecting this field. The multihomed client delta introduced
307 * a RX code change : change the peer field in the connection structure
308 * to that remote inetrface from which the last packet for this
309 * connection was sent out. This may become an issue if further changes
312 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
313 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
315 /* rxdb_fileID is used to identify the lock location, along with line#. */
316 static int rxdb_fileID = RXDB_FILE_RX;
317 #endif /* RX_LOCKS_DB */
318 #else /* RX_ENABLE_LOCKS */
319 #define SET_CALL_QUEUE_LOCK(C, L)
320 #define CLEAR_CALL_QUEUE_LOCK(C)
321 #endif /* RX_ENABLE_LOCKS */
322 struct rx_serverQueueEntry *rx_waitForPacket = 0;
323 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
325 /* ------------Exported Interfaces------------- */
327 /* This function allows rxkad to set the epoch to a suitably random number
328 * which rx_NewConnection will use in the future. The principle purpose is to
329 * get rxnull connections to use the same epoch as the rxkad connections do, at
330 * least once the first rxkad connection is established. This is important now
331 * that the host/port addresses aren't used in FindConnection: the uniqueness
332 * of epoch/cid matters and the start time won't do. */
334 #ifdef AFS_PTHREAD_ENV
336 * This mutex protects the following global variables:
340 #define LOCK_EPOCH assert(pthread_mutex_lock(&epoch_mutex)==0);
341 #define UNLOCK_EPOCH assert(pthread_mutex_unlock(&epoch_mutex)==0);
345 #endif /* AFS_PTHREAD_ENV */
348 rx_SetEpoch(afs_uint32 epoch)
350 LOCK_EPOCH rx_epoch = epoch;
353 /* Initialize rx. A port number may be mentioned, in which case this
354 * becomes the default port number for any service installed later.
355 * If 0 is provided for the port number, a random port will be chosen
356 * by the kernel. Whether this will ever overlap anything in
357 * /etc/services is anybody's guess... Returns 0 on success, -1 on
359 static int rxinit_status = 1;
360 #ifdef AFS_PTHREAD_ENV
362 * This mutex protects the following global variables:
366 #define LOCK_RX_INIT assert(pthread_mutex_lock(&rx_init_mutex)==0);
367 #define UNLOCK_RX_INIT assert(pthread_mutex_unlock(&rx_init_mutex)==0);
370 #define UNLOCK_RX_INIT
381 char *htable, *ptable;
384 #if defined(AFS_DJGPP_ENV) && !defined(DEBUG)
385 __djgpp_set_quiet_socket(1);
390 INIT_PTHREAD_LOCKS LOCK_RX_INIT if (rxinit_status == 0) {
391 tmp_status = rxinit_status;
392 UNLOCK_RX_INIT return tmp_status; /* Already started; return previous error code. */
395 if (afs_winsockInit() < 0)
401 * Initialize anything necessary to provide a non-premptive threading
404 rxi_InitializeThreadSupport();
407 /* Allocate and initialize a socket for client and perhaps server
410 rx_socket = rxi_GetUDPSocket((u_short) port);
411 if (rx_socket == OSI_NULLSOCKET) {
412 UNLOCK_RX_INIT return RX_ADDRINUSE;
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 #else /* RX_ENABLE_LOCKS */
439 #if defined(KERNEL) && defined(AFS_GLOBAL_SUNLOCK) && !defined(AFS_HPUX_ENV) && !defined(AFS_OBSD_ENV)
440 mutex_init(&afs_rxglobal_lock, "afs_rxglobal_lock", MUTEX_DEFAULT, NULL);
441 #endif /* AFS_GLOBAL_SUNLOCK */
442 #endif /* RX_ENABLE_LOCKS */
445 rx_connDeadTime = 12;
446 rx_tranquil = 0; /* reset flag */
447 memset((char *)&rx_stats, 0, sizeof(struct rx_stats));
449 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
450 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
451 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
452 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
453 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
454 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
456 /* Malloc up a bunch of packets & buffers */
458 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
459 queue_Init(&rx_freePacketQueue);
460 rxi_NeedMorePackets = FALSE;
461 rxi_MorePackets(rx_nPackets);
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) */
533 tmp_status = rxinit_status = 0;
534 UNLOCK_RX_INIT return tmp_status;
537 /* called with unincremented nRequestsRunning to see if it is OK to start
538 * a new thread in this service. Could be "no" for two reasons: over the
539 * max quota, or would prevent others from reaching their min quota.
541 #ifdef RX_ENABLE_LOCKS
542 /* This verion of QuotaOK reserves quota if it's ok while the
543 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
546 QuotaOK(register struct rx_service *aservice)
548 /* check if over max quota */
549 if (aservice->nRequestsRunning >= aservice->maxProcs) {
553 /* under min quota, we're OK */
554 /* otherwise, can use only if there are enough to allow everyone
555 * to go to their min quota after this guy starts.
557 MUTEX_ENTER(&rx_stats_mutex);
558 if ((aservice->nRequestsRunning < aservice->minProcs)
559 || (rxi_availProcs > rxi_minDeficit)) {
560 aservice->nRequestsRunning++;
561 /* just started call in minProcs pool, need fewer to maintain
563 if (aservice->nRequestsRunning <= aservice->minProcs)
566 MUTEX_EXIT(&rx_stats_mutex);
569 MUTEX_EXIT(&rx_stats_mutex);
575 ReturnToServerPool(register struct rx_service *aservice)
577 aservice->nRequestsRunning--;
578 MUTEX_ENTER(&rx_stats_mutex);
579 if (aservice->nRequestsRunning < aservice->minProcs)
582 MUTEX_EXIT(&rx_stats_mutex);
585 #else /* RX_ENABLE_LOCKS */
587 QuotaOK(register struct rx_service *aservice)
590 /* under min quota, we're OK */
591 if (aservice->nRequestsRunning < aservice->minProcs)
594 /* check if over max quota */
595 if (aservice->nRequestsRunning >= aservice->maxProcs)
598 /* otherwise, can use only if there are enough to allow everyone
599 * to go to their min quota after this guy starts.
601 if (rxi_availProcs > rxi_minDeficit)
605 #endif /* RX_ENABLE_LOCKS */
608 /* Called by rx_StartServer to start up lwp's to service calls.
609 NExistingProcs gives the number of procs already existing, and which
610 therefore needn't be created. */
612 rxi_StartServerProcs(int nExistingProcs)
614 register struct rx_service *service;
619 /* For each service, reserve N processes, where N is the "minimum"
620 * number of processes that MUST be able to execute a request in parallel,
621 * at any time, for that process. Also compute the maximum difference
622 * between any service's maximum number of processes that can run
623 * (i.e. the maximum number that ever will be run, and a guarantee
624 * that this number will run if other services aren't running), and its
625 * minimum number. The result is the extra number of processes that
626 * we need in order to provide the latter guarantee */
627 for (i = 0; i < RX_MAX_SERVICES; i++) {
629 service = rx_services[i];
630 if (service == (struct rx_service *)0)
632 nProcs += service->minProcs;
633 diff = service->maxProcs - service->minProcs;
637 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
638 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
639 for (i = 0; i < nProcs; i++) {
640 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
645 /* This routine must be called if any services are exported. If the
646 * donateMe flag is set, the calling process is donated to the server
649 rx_StartServer(int donateMe)
651 register struct rx_service *service;
652 register int i, nProcs = 0;
658 /* Start server processes, if necessary (exact function is dependent
659 * on the implementation environment--kernel or user space). DonateMe
660 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
661 * case, one less new proc will be created rx_StartServerProcs.
663 rxi_StartServerProcs(donateMe);
665 /* count up the # of threads in minProcs, and add set the min deficit to
666 * be that value, too.
668 for (i = 0; i < RX_MAX_SERVICES; i++) {
669 service = rx_services[i];
670 if (service == (struct rx_service *)0)
672 MUTEX_ENTER(&rx_stats_mutex);
673 rxi_totalMin += service->minProcs;
674 /* below works even if a thread is running, since minDeficit would
675 * still have been decremented and later re-incremented.
677 rxi_minDeficit += service->minProcs;
678 MUTEX_EXIT(&rx_stats_mutex);
681 /* Turn on reaping of idle server connections */
682 rxi_ReapConnections();
691 #ifdef AFS_PTHREAD_ENV
693 pid = (pid_t) pthread_self();
694 #else /* AFS_PTHREAD_ENV */
696 LWP_CurrentProcess(&pid);
697 #endif /* AFS_PTHREAD_ENV */
699 sprintf(name, "srv_%d", ++nProcs);
701 (*registerProgram) (pid, name);
703 #endif /* AFS_NT40_ENV */
704 rx_ServerProc(); /* Never returns */
709 /* Create a new client connection to the specified service, using the
710 * specified security object to implement the security model for this
712 struct rx_connection *
713 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
714 register struct rx_securityClass *securityObject,
715 int serviceSecurityIndex)
719 register struct rx_connection *conn;
724 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", shost, sport, sservice, securityObject, serviceSecurityIndex));
726 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
727 * the case of kmem_alloc? */
728 conn = rxi_AllocConnection();
729 #ifdef RX_ENABLE_LOCKS
730 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
731 MUTEX_INIT(&conn->conn_data_lock, "conn call lock", MUTEX_DEFAULT, 0);
732 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
736 MUTEX_ENTER(&rx_connHashTable_lock);
737 cid = (rx_nextCid += RX_MAXCALLS);
738 conn->type = RX_CLIENT_CONNECTION;
740 conn->epoch = rx_epoch;
741 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
742 conn->serviceId = sservice;
743 conn->securityObject = securityObject;
744 /* This doesn't work in all compilers with void (they're buggy), so fake it
746 conn->securityData = (VOID *) 0;
747 conn->securityIndex = serviceSecurityIndex;
748 rx_SetConnDeadTime(conn, rx_connDeadTime);
749 conn->ackRate = RX_FAST_ACK_RATE;
751 conn->specific = NULL;
752 conn->challengeEvent = NULL;
753 conn->delayedAbortEvent = NULL;
754 conn->abortCount = 0;
757 RXS_NewConnection(securityObject, conn);
759 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
761 conn->refCount++; /* no lock required since only this thread knows... */
762 conn->next = rx_connHashTable[hashindex];
763 rx_connHashTable[hashindex] = conn;
764 MUTEX_ENTER(&rx_stats_mutex);
765 rx_stats.nClientConns++;
766 MUTEX_EXIT(&rx_stats_mutex);
768 MUTEX_EXIT(&rx_connHashTable_lock);
775 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
777 /* The idea is to set the dead time to a value that allows several
778 * keepalives to be dropped without timing out the connection. */
779 conn->secondsUntilDead = MAX(seconds, 6);
780 conn->secondsUntilPing = conn->secondsUntilDead / 6;
783 int rxi_lowPeerRefCount = 0;
784 int rxi_lowConnRefCount = 0;
787 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
788 * NOTE: must not be called with rx_connHashTable_lock held.
791 rxi_CleanupConnection(struct rx_connection *conn)
793 /* Notify the service exporter, if requested, that this connection
794 * is being destroyed */
795 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
796 (*conn->service->destroyConnProc) (conn);
798 /* Notify the security module that this connection is being destroyed */
799 RXS_DestroyConnection(conn->securityObject, conn);
801 /* If this is the last connection using the rx_peer struct, set its
802 * idle time to now. rxi_ReapConnections will reap it if it's still
803 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
805 MUTEX_ENTER(&rx_peerHashTable_lock);
806 if (--conn->peer->refCount <= 0) {
807 conn->peer->idleWhen = clock_Sec();
808 if (conn->peer->refCount < 0) {
809 conn->peer->refCount = 0;
810 MUTEX_ENTER(&rx_stats_mutex);
811 rxi_lowPeerRefCount++;
812 MUTEX_EXIT(&rx_stats_mutex);
815 MUTEX_EXIT(&rx_peerHashTable_lock);
817 MUTEX_ENTER(&rx_stats_mutex);
818 if (conn->type == RX_SERVER_CONNECTION)
819 rx_stats.nServerConns--;
821 rx_stats.nClientConns--;
822 MUTEX_EXIT(&rx_stats_mutex);
825 if (conn->specific) {
827 for (i = 0; i < conn->nSpecific; i++) {
828 if (conn->specific[i] && rxi_keyCreate_destructor[i])
829 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
830 conn->specific[i] = NULL;
832 free(conn->specific);
834 conn->specific = NULL;
838 MUTEX_DESTROY(&conn->conn_call_lock);
839 MUTEX_DESTROY(&conn->conn_data_lock);
840 CV_DESTROY(&conn->conn_call_cv);
842 rxi_FreeConnection(conn);
845 /* Destroy the specified connection */
847 rxi_DestroyConnection(register struct rx_connection *conn)
849 MUTEX_ENTER(&rx_connHashTable_lock);
850 rxi_DestroyConnectionNoLock(conn);
851 /* conn should be at the head of the cleanup list */
852 if (conn == rx_connCleanup_list) {
853 rx_connCleanup_list = rx_connCleanup_list->next;
854 MUTEX_EXIT(&rx_connHashTable_lock);
855 rxi_CleanupConnection(conn);
857 #ifdef RX_ENABLE_LOCKS
859 MUTEX_EXIT(&rx_connHashTable_lock);
861 #endif /* RX_ENABLE_LOCKS */
865 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
867 register struct rx_connection **conn_ptr;
868 register int havecalls = 0;
869 struct rx_packet *packet;
876 MUTEX_ENTER(&conn->conn_data_lock);
877 if (conn->refCount > 0)
880 MUTEX_ENTER(&rx_stats_mutex);
881 rxi_lowConnRefCount++;
882 MUTEX_EXIT(&rx_stats_mutex);
885 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
886 /* Busy; wait till the last guy before proceeding */
887 MUTEX_EXIT(&conn->conn_data_lock);
892 /* If the client previously called rx_NewCall, but it is still
893 * waiting, treat this as a running call, and wait to destroy the
894 * connection later when the call completes. */
895 if ((conn->type == RX_CLIENT_CONNECTION)
896 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
897 conn->flags |= RX_CONN_DESTROY_ME;
898 MUTEX_EXIT(&conn->conn_data_lock);
902 MUTEX_EXIT(&conn->conn_data_lock);
904 /* Check for extant references to this connection */
905 for (i = 0; i < RX_MAXCALLS; i++) {
906 register struct rx_call *call = conn->call[i];
909 if (conn->type == RX_CLIENT_CONNECTION) {
910 MUTEX_ENTER(&call->lock);
911 if (call->delayedAckEvent) {
912 /* Push the final acknowledgment out now--there
913 * won't be a subsequent call to acknowledge the
914 * last reply packets */
915 rxevent_Cancel(call->delayedAckEvent, call,
916 RX_CALL_REFCOUNT_DELAY);
917 if (call->state == RX_STATE_PRECALL
918 || call->state == RX_STATE_ACTIVE) {
919 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
921 rxi_AckAll(NULL, call, 0);
924 MUTEX_EXIT(&call->lock);
928 #ifdef RX_ENABLE_LOCKS
930 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
931 MUTEX_EXIT(&conn->conn_data_lock);
933 /* Someone is accessing a packet right now. */
937 #endif /* RX_ENABLE_LOCKS */
940 /* Don't destroy the connection if there are any call
941 * structures still in use */
942 MUTEX_ENTER(&conn->conn_data_lock);
943 conn->flags |= RX_CONN_DESTROY_ME;
944 MUTEX_EXIT(&conn->conn_data_lock);
949 if (conn->delayedAbortEvent) {
950 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
951 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
953 MUTEX_ENTER(&conn->conn_data_lock);
954 rxi_SendConnectionAbort(conn, packet, 0, 1);
955 MUTEX_EXIT(&conn->conn_data_lock);
956 rxi_FreePacket(packet);
960 /* Remove from connection hash table before proceeding */
962 &rx_connHashTable[CONN_HASH
963 (peer->host, peer->port, conn->cid, conn->epoch,
965 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
966 if (*conn_ptr == conn) {
967 *conn_ptr = conn->next;
971 /* if the conn that we are destroying was the last connection, then we
972 * clear rxLastConn as well */
973 if (rxLastConn == conn)
976 /* Make sure the connection is completely reset before deleting it. */
977 /* get rid of pending events that could zap us later */
978 if (conn->challengeEvent)
979 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
980 if (conn->checkReachEvent)
981 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
983 /* Add the connection to the list of destroyed connections that
984 * need to be cleaned up. This is necessary to avoid deadlocks
985 * in the routines we call to inform others that this connection is
986 * being destroyed. */
987 conn->next = rx_connCleanup_list;
988 rx_connCleanup_list = conn;
991 /* Externally available version */
993 rx_DestroyConnection(register struct rx_connection *conn)
999 rxi_DestroyConnection(conn);
1004 /* Start a new rx remote procedure call, on the specified connection.
1005 * If wait is set to 1, wait for a free call channel; otherwise return
1006 * 0. Maxtime gives the maximum number of seconds this call may take,
1007 * after rx_MakeCall returns. After this time interval, a call to any
1008 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1009 * For fine grain locking, we hold the conn_call_lock in order to
1010 * to ensure that we don't get signalle after we found a call in an active
1011 * state and before we go to sleep.
1014 rx_NewCall(register struct rx_connection *conn)
1017 register struct rx_call *call;
1018 struct clock queueTime;
1022 dpf(("rx_MakeCall(conn %x)\n", conn));
1025 clock_GetTime(&queueTime);
1027 MUTEX_ENTER(&conn->conn_call_lock);
1030 * Check if there are others waiting for a new call.
1031 * If so, let them go first to avoid starving them.
1032 * This is a fairly simple scheme, and might not be
1033 * a complete solution for large numbers of waiters.
1035 if (conn->makeCallWaiters) {
1036 #ifdef RX_ENABLE_LOCKS
1037 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1044 for (i = 0; i < RX_MAXCALLS; i++) {
1045 call = conn->call[i];
1047 MUTEX_ENTER(&call->lock);
1048 if (call->state == RX_STATE_DALLY) {
1049 rxi_ResetCall(call, 0);
1050 (*call->callNumber)++;
1053 MUTEX_EXIT(&call->lock);
1055 call = rxi_NewCall(conn, i);
1059 if (i < RX_MAXCALLS) {
1062 MUTEX_ENTER(&conn->conn_data_lock);
1063 conn->flags |= RX_CONN_MAKECALL_WAITING;
1064 MUTEX_EXIT(&conn->conn_data_lock);
1066 conn->makeCallWaiters++;
1067 #ifdef RX_ENABLE_LOCKS
1068 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1072 conn->makeCallWaiters--;
1075 * Wake up anyone else who might be giving us a chance to
1076 * run (see code above that avoids resource starvation).
1078 #ifdef RX_ENABLE_LOCKS
1079 CV_BROADCAST(&conn->conn_call_cv);
1084 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1086 /* Client is initially in send mode */
1087 call->state = RX_STATE_ACTIVE;
1088 call->mode = RX_MODE_SENDING;
1090 /* remember start time for call in case we have hard dead time limit */
1091 call->queueTime = queueTime;
1092 clock_GetTime(&call->startTime);
1093 hzero(call->bytesSent);
1094 hzero(call->bytesRcvd);
1096 /* Turn on busy protocol. */
1097 rxi_KeepAliveOn(call);
1099 MUTEX_EXIT(&call->lock);
1100 MUTEX_EXIT(&conn->conn_call_lock);
1104 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1105 /* Now, if TQ wasn't cleared earlier, do it now. */
1107 MUTEX_ENTER(&call->lock);
1108 while (call->flags & RX_CALL_TQ_BUSY) {
1109 call->flags |= RX_CALL_TQ_WAIT;
1110 #ifdef RX_ENABLE_LOCKS
1111 CV_WAIT(&call->cv_tq, &call->lock);
1112 #else /* RX_ENABLE_LOCKS */
1113 osi_rxSleep(&call->tq);
1114 #endif /* RX_ENABLE_LOCKS */
1116 if (call->flags & RX_CALL_TQ_CLEARME) {
1117 rxi_ClearTransmitQueue(call, 0);
1118 queue_Init(&call->tq);
1120 MUTEX_EXIT(&call->lock);
1122 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1128 rxi_HasActiveCalls(register struct rx_connection *aconn)
1131 register struct rx_call *tcall;
1135 for (i = 0; i < RX_MAXCALLS; i++) {
1136 if ((tcall = aconn->call[i])) {
1137 if ((tcall->state == RX_STATE_ACTIVE)
1138 || (tcall->state == RX_STATE_PRECALL)) {
1149 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1150 register afs_int32 * aint32s)
1153 register struct rx_call *tcall;
1157 for (i = 0; i < RX_MAXCALLS; i++) {
1158 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1159 aint32s[i] = aconn->callNumber[i] + 1;
1161 aint32s[i] = aconn->callNumber[i];
1168 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1169 register afs_int32 * aint32s)
1172 register struct rx_call *tcall;
1176 for (i = 0; i < RX_MAXCALLS; i++) {
1177 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1178 aconn->callNumber[i] = aint32s[i] - 1;
1180 aconn->callNumber[i] = aint32s[i];
1186 /* Advertise a new service. A service is named locally by a UDP port
1187 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1190 char *serviceName; Name for identification purposes (e.g. the
1191 service name might be used for probing for
1194 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1195 struct rx_securityClass **securityObjects, int nSecurityObjects,
1196 afs_int32(*serviceProc) (struct rx_call * acall))
1198 osi_socket socket = OSI_NULLSOCKET;
1199 register struct rx_service *tservice;
1205 if (serviceId == 0) {
1207 "rx_NewService: service id for service %s is not non-zero.\n",
1214 "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",
1222 tservice = rxi_AllocService();
1225 for (i = 0; i < RX_MAX_SERVICES; i++) {
1226 register struct rx_service *service = rx_services[i];
1228 if (port == service->servicePort) {
1229 if (service->serviceId == serviceId) {
1230 /* The identical service has already been
1231 * installed; if the caller was intending to
1232 * change the security classes used by this
1233 * service, he/she loses. */
1235 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1236 serviceName, serviceId, service->serviceName);
1239 rxi_FreeService(tservice);
1242 /* Different service, same port: re-use the socket
1243 * which is bound to the same port */
1244 socket = service->socket;
1247 if (socket == OSI_NULLSOCKET) {
1248 /* If we don't already have a socket (from another
1249 * service on same port) get a new one */
1250 socket = rxi_GetUDPSocket(port);
1251 if (socket == OSI_NULLSOCKET) {
1254 rxi_FreeService(tservice);
1259 service->socket = socket;
1260 service->servicePort = port;
1261 service->serviceId = serviceId;
1262 service->serviceName = serviceName;
1263 service->nSecurityObjects = nSecurityObjects;
1264 service->securityObjects = securityObjects;
1265 service->minProcs = 0;
1266 service->maxProcs = 1;
1267 service->idleDeadTime = 60;
1268 service->connDeadTime = rx_connDeadTime;
1269 service->executeRequestProc = serviceProc;
1270 service->checkReach = 0;
1271 rx_services[i] = service; /* not visible until now */
1279 rxi_FreeService(tservice);
1280 (osi_Msg "rx_NewService: cannot support > %d services\n",
1285 /* Generic request processing loop. This routine should be called
1286 * by the implementation dependent rx_ServerProc. If socketp is
1287 * non-null, it will be set to the file descriptor that this thread
1288 * is now listening on. If socketp is null, this routine will never
1291 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1293 register struct rx_call *call;
1294 register afs_int32 code;
1295 register struct rx_service *tservice = NULL;
1302 call = rx_GetCall(threadID, tservice, socketp);
1303 if (socketp && *socketp != OSI_NULLSOCKET) {
1304 /* We are now a listener thread */
1309 /* if server is restarting( typically smooth shutdown) then do not
1310 * allow any new calls.
1313 if (rx_tranquil && (call != NULL)) {
1318 MUTEX_ENTER(&call->lock);
1320 rxi_CallError(call, RX_RESTARTING);
1321 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1323 MUTEX_EXIT(&call->lock);
1328 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1329 #ifdef RX_ENABLE_LOCKS
1331 #endif /* RX_ENABLE_LOCKS */
1332 afs_termState = AFSOP_STOP_AFS;
1333 afs_osi_Wakeup(&afs_termState);
1334 #ifdef RX_ENABLE_LOCKS
1336 #endif /* RX_ENABLE_LOCKS */
1341 tservice = call->conn->service;
1343 if (tservice->beforeProc)
1344 (*tservice->beforeProc) (call);
1346 code = call->conn->service->executeRequestProc(call);
1348 if (tservice->afterProc)
1349 (*tservice->afterProc) (call, code);
1351 rx_EndCall(call, code);
1352 MUTEX_ENTER(&rx_stats_mutex);
1354 MUTEX_EXIT(&rx_stats_mutex);
1360 rx_WakeupServerProcs(void)
1362 struct rx_serverQueueEntry *np, *tqp;
1367 MUTEX_ENTER(&rx_serverPool_lock);
1369 #ifdef RX_ENABLE_LOCKS
1370 if (rx_waitForPacket)
1371 CV_BROADCAST(&rx_waitForPacket->cv);
1372 #else /* RX_ENABLE_LOCKS */
1373 if (rx_waitForPacket)
1374 osi_rxWakeup(rx_waitForPacket);
1375 #endif /* RX_ENABLE_LOCKS */
1376 MUTEX_ENTER(&freeSQEList_lock);
1377 for (np = rx_FreeSQEList; np; np = tqp) {
1378 tqp = *(struct rx_serverQueueEntry **)np;
1379 #ifdef RX_ENABLE_LOCKS
1380 CV_BROADCAST(&np->cv);
1381 #else /* RX_ENABLE_LOCKS */
1383 #endif /* RX_ENABLE_LOCKS */
1385 MUTEX_EXIT(&freeSQEList_lock);
1386 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1387 #ifdef RX_ENABLE_LOCKS
1388 CV_BROADCAST(&np->cv);
1389 #else /* RX_ENABLE_LOCKS */
1391 #endif /* RX_ENABLE_LOCKS */
1393 MUTEX_EXIT(&rx_serverPool_lock);
1399 * One thing that seems to happen is that all the server threads get
1400 * tied up on some empty or slow call, and then a whole bunch of calls
1401 * arrive at once, using up the packet pool, so now there are more
1402 * empty calls. The most critical resources here are server threads
1403 * and the free packet pool. The "doreclaim" code seems to help in
1404 * general. I think that eventually we arrive in this state: there
1405 * are lots of pending calls which do have all their packets present,
1406 * so they won't be reclaimed, are multi-packet calls, so they won't
1407 * be scheduled until later, and thus are tying up most of the free
1408 * packet pool for a very long time.
1410 * 1. schedule multi-packet calls if all the packets are present.
1411 * Probably CPU-bound operation, useful to return packets to pool.
1412 * Do what if there is a full window, but the last packet isn't here?
1413 * 3. preserve one thread which *only* runs "best" calls, otherwise
1414 * it sleeps and waits for that type of call.
1415 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1416 * the current dataquota business is badly broken. The quota isn't adjusted
1417 * to reflect how many packets are presently queued for a running call.
1418 * So, when we schedule a queued call with a full window of packets queued
1419 * up for it, that *should* free up a window full of packets for other 2d-class
1420 * calls to be able to use from the packet pool. But it doesn't.
1422 * NB. Most of the time, this code doesn't run -- since idle server threads
1423 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1424 * as a new call arrives.
1426 /* Sleep until a call arrives. Returns a pointer to the call, ready
1427 * for an rx_Read. */
1428 #ifdef RX_ENABLE_LOCKS
1430 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1432 struct rx_serverQueueEntry *sq;
1433 register struct rx_call *call = (struct rx_call *)0;
1434 struct rx_service *service = NULL;
1437 MUTEX_ENTER(&freeSQEList_lock);
1439 if ((sq = rx_FreeSQEList)) {
1440 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1441 MUTEX_EXIT(&freeSQEList_lock);
1442 } else { /* otherwise allocate a new one and return that */
1443 MUTEX_EXIT(&freeSQEList_lock);
1444 sq = (struct rx_serverQueueEntry *)
1445 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1446 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1447 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1450 MUTEX_ENTER(&rx_serverPool_lock);
1451 if (cur_service != NULL) {
1452 ReturnToServerPool(cur_service);
1455 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1456 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1458 /* Scan for eligible incoming calls. A call is not eligible
1459 * if the maximum number of calls for its service type are
1460 * already executing */
1461 /* One thread will process calls FCFS (to prevent starvation),
1462 * while the other threads may run ahead looking for calls which
1463 * have all their input data available immediately. This helps
1464 * keep threads from blocking, waiting for data from the client. */
1465 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1466 service = tcall->conn->service;
1467 if (!QuotaOK(service)) {
1470 if (tno == rxi_fcfs_thread_num
1471 || !tcall->queue_item_header.next) {
1472 /* If we're the fcfs thread , then we'll just use
1473 * this call. If we haven't been able to find an optimal
1474 * choice, and we're at the end of the list, then use a
1475 * 2d choice if one has been identified. Otherwise... */
1476 call = (choice2 ? choice2 : tcall);
1477 service = call->conn->service;
1478 } else if (!queue_IsEmpty(&tcall->rq)) {
1479 struct rx_packet *rp;
1480 rp = queue_First(&tcall->rq, rx_packet);
1481 if (rp->header.seq == 1) {
1483 || (rp->header.flags & RX_LAST_PACKET)) {
1485 } else if (rxi_2dchoice && !choice2
1486 && !(tcall->flags & RX_CALL_CLEARED)
1487 && (tcall->rprev > rxi_HardAckRate)) {
1496 ReturnToServerPool(service);
1503 MUTEX_EXIT(&rx_serverPool_lock);
1504 MUTEX_ENTER(&call->lock);
1506 if (call->flags & RX_CALL_WAIT_PROC) {
1507 call->flags &= ~RX_CALL_WAIT_PROC;
1508 MUTEX_ENTER(&rx_stats_mutex);
1510 MUTEX_EXIT(&rx_stats_mutex);
1513 if (call->state != RX_STATE_PRECALL || call->error) {
1514 MUTEX_EXIT(&call->lock);
1515 MUTEX_ENTER(&rx_serverPool_lock);
1516 ReturnToServerPool(service);
1521 if (queue_IsEmpty(&call->rq)
1522 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1523 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1525 CLEAR_CALL_QUEUE_LOCK(call);
1528 /* If there are no eligible incoming calls, add this process
1529 * to the idle server queue, to wait for one */
1533 *socketp = OSI_NULLSOCKET;
1535 sq->socketp = socketp;
1536 queue_Append(&rx_idleServerQueue, sq);
1537 #ifndef AFS_AIX41_ENV
1538 rx_waitForPacket = sq;
1540 rx_waitingForPacket = sq;
1541 #endif /* AFS_AIX41_ENV */
1543 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1545 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1546 MUTEX_EXIT(&rx_serverPool_lock);
1547 return (struct rx_call *)0;
1550 } while (!(call = sq->newcall)
1551 && !(socketp && *socketp != OSI_NULLSOCKET));
1552 MUTEX_EXIT(&rx_serverPool_lock);
1554 MUTEX_ENTER(&call->lock);
1560 MUTEX_ENTER(&freeSQEList_lock);
1561 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1562 rx_FreeSQEList = sq;
1563 MUTEX_EXIT(&freeSQEList_lock);
1566 clock_GetTime(&call->startTime);
1567 call->state = RX_STATE_ACTIVE;
1568 call->mode = RX_MODE_RECEIVING;
1569 #ifdef RX_KERNEL_TRACE
1570 if (ICL_SETACTIVE(afs_iclSetp)) {
1571 int glockOwner = ISAFS_GLOCK();
1574 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1575 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1582 rxi_calltrace(RX_CALL_START, call);
1583 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1584 call->conn->service->servicePort, call->conn->service->serviceId,
1587 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1588 MUTEX_EXIT(&call->lock);
1590 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1595 #else /* RX_ENABLE_LOCKS */
1597 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1599 struct rx_serverQueueEntry *sq;
1600 register struct rx_call *call = (struct rx_call *)0, *choice2;
1601 struct rx_service *service = NULL;
1606 MUTEX_ENTER(&freeSQEList_lock);
1608 if ((sq = rx_FreeSQEList)) {
1609 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1610 MUTEX_EXIT(&freeSQEList_lock);
1611 } else { /* otherwise allocate a new one and return that */
1612 MUTEX_EXIT(&freeSQEList_lock);
1613 sq = (struct rx_serverQueueEntry *)
1614 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1615 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1616 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1618 MUTEX_ENTER(&sq->lock);
1620 if (cur_service != NULL) {
1621 cur_service->nRequestsRunning--;
1622 if (cur_service->nRequestsRunning < cur_service->minProcs)
1626 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1627 register struct rx_call *tcall, *ncall;
1628 /* Scan for eligible incoming calls. A call is not eligible
1629 * if the maximum number of calls for its service type are
1630 * already executing */
1631 /* One thread will process calls FCFS (to prevent starvation),
1632 * while the other threads may run ahead looking for calls which
1633 * have all their input data available immediately. This helps
1634 * keep threads from blocking, waiting for data from the client. */
1635 choice2 = (struct rx_call *)0;
1636 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1637 service = tcall->conn->service;
1638 if (QuotaOK(service)) {
1639 if (tno == rxi_fcfs_thread_num
1640 || !tcall->queue_item_header.next) {
1641 /* If we're the fcfs thread, then we'll just use
1642 * this call. If we haven't been able to find an optimal
1643 * choice, and we're at the end of the list, then use a
1644 * 2d choice if one has been identified. Otherwise... */
1645 call = (choice2 ? choice2 : tcall);
1646 service = call->conn->service;
1647 } else if (!queue_IsEmpty(&tcall->rq)) {
1648 struct rx_packet *rp;
1649 rp = queue_First(&tcall->rq, rx_packet);
1650 if (rp->header.seq == 1
1652 || (rp->header.flags & RX_LAST_PACKET))) {
1654 } else if (rxi_2dchoice && !choice2
1655 && !(tcall->flags & RX_CALL_CLEARED)
1656 && (tcall->rprev > rxi_HardAckRate)) {
1669 /* we can't schedule a call if there's no data!!! */
1670 /* send an ack if there's no data, if we're missing the
1671 * first packet, or we're missing something between first
1672 * and last -- there's a "hole" in the incoming data. */
1673 if (queue_IsEmpty(&call->rq)
1674 || queue_First(&call->rq, rx_packet)->header.seq != 1
1675 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1676 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1678 call->flags &= (~RX_CALL_WAIT_PROC);
1679 service->nRequestsRunning++;
1680 /* just started call in minProcs pool, need fewer to maintain
1682 if (service->nRequestsRunning <= service->minProcs)
1686 /* MUTEX_EXIT(&call->lock); */
1688 /* If there are no eligible incoming calls, add this process
1689 * to the idle server queue, to wait for one */
1692 *socketp = OSI_NULLSOCKET;
1694 sq->socketp = socketp;
1695 queue_Append(&rx_idleServerQueue, sq);
1699 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1702 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1703 return (struct rx_call *)0;
1706 } while (!(call = sq->newcall)
1707 && !(socketp && *socketp != OSI_NULLSOCKET));
1709 MUTEX_EXIT(&sq->lock);
1711 MUTEX_ENTER(&freeSQEList_lock);
1712 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1713 rx_FreeSQEList = sq;
1714 MUTEX_EXIT(&freeSQEList_lock);
1717 clock_GetTime(&call->startTime);
1718 call->state = RX_STATE_ACTIVE;
1719 call->mode = RX_MODE_RECEIVING;
1720 #ifdef RX_KERNEL_TRACE
1721 if (ICL_SETACTIVE(afs_iclSetp)) {
1722 int glockOwner = ISAFS_GLOCK();
1725 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1726 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1733 rxi_calltrace(RX_CALL_START, call);
1734 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1735 call->conn->service->servicePort, call->conn->service->serviceId,
1738 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1746 #endif /* RX_ENABLE_LOCKS */
1750 /* Establish a procedure to be called when a packet arrives for a
1751 * call. This routine will be called at most once after each call,
1752 * and will also be called if there is an error condition on the or
1753 * the call is complete. Used by multi rx to build a selection
1754 * function which determines which of several calls is likely to be a
1755 * good one to read from.
1756 * NOTE: the way this is currently implemented it is probably only a
1757 * good idea to (1) use it immediately after a newcall (clients only)
1758 * and (2) only use it once. Other uses currently void your warranty
1761 rx_SetArrivalProc(register struct rx_call *call,
1762 register VOID(*proc) (register struct rx_call * call,
1763 register struct multi_handle * mh,
1764 register int index),
1765 register VOID * handle, register VOID * arg)
1767 call->arrivalProc = proc;
1768 call->arrivalProcHandle = handle;
1769 call->arrivalProcArg = arg;
1772 /* Call is finished (possibly prematurely). Return rc to the peer, if
1773 * appropriate, and return the final error code from the conversation
1777 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1779 register struct rx_connection *conn = call->conn;
1780 register struct rx_service *service;
1781 register struct rx_packet *tp; /* Temporary packet pointer */
1782 register struct rx_packet *nxp; /* Next packet pointer, for queue_Scan */
1786 dpf(("rx_EndCall(call %x)\n", call));
1790 MUTEX_ENTER(&call->lock);
1792 if (rc == 0 && call->error == 0) {
1793 call->abortCode = 0;
1794 call->abortCount = 0;
1797 call->arrivalProc = (VOID(*)())0;
1798 if (rc && call->error == 0) {
1799 rxi_CallError(call, rc);
1800 /* Send an abort message to the peer if this error code has
1801 * only just been set. If it was set previously, assume the
1802 * peer has already been sent the error code or will request it
1804 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1806 if (conn->type == RX_SERVER_CONNECTION) {
1807 /* Make sure reply or at least dummy reply is sent */
1808 if (call->mode == RX_MODE_RECEIVING) {
1809 rxi_WriteProc(call, 0, 0);
1811 if (call->mode == RX_MODE_SENDING) {
1812 rxi_FlushWrite(call);
1814 service = conn->service;
1815 rxi_calltrace(RX_CALL_END, call);
1816 /* Call goes to hold state until reply packets are acknowledged */
1817 if (call->tfirst + call->nSoftAcked < call->tnext) {
1818 call->state = RX_STATE_HOLD;
1820 call->state = RX_STATE_DALLY;
1821 rxi_ClearTransmitQueue(call, 0);
1822 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1823 rxevent_Cancel(call->keepAliveEvent, call,
1824 RX_CALL_REFCOUNT_ALIVE);
1826 } else { /* Client connection */
1828 /* Make sure server receives input packets, in the case where
1829 * no reply arguments are expected */
1830 if ((call->mode == RX_MODE_SENDING)
1831 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1832 (void)rxi_ReadProc(call, &dummy, 1);
1835 /* If we had an outstanding delayed ack, be nice to the server
1836 * and force-send it now.
1838 if (call->delayedAckEvent) {
1839 rxevent_Cancel(call->delayedAckEvent, call,
1840 RX_CALL_REFCOUNT_DELAY);
1841 call->delayedAckEvent = NULL;
1842 rxi_SendDelayedAck(NULL, call, NULL);
1845 /* We need to release the call lock since it's lower than the
1846 * conn_call_lock and we don't want to hold the conn_call_lock
1847 * over the rx_ReadProc call. The conn_call_lock needs to be held
1848 * here for the case where rx_NewCall is perusing the calls on
1849 * the connection structure. We don't want to signal until
1850 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1851 * have checked this call, found it active and by the time it
1852 * goes to sleep, will have missed the signal.
1854 MUTEX_EXIT(&call->lock);
1855 MUTEX_ENTER(&conn->conn_call_lock);
1856 MUTEX_ENTER(&call->lock);
1857 MUTEX_ENTER(&conn->conn_data_lock);
1858 conn->flags |= RX_CONN_BUSY;
1859 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1860 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1861 MUTEX_EXIT(&conn->conn_data_lock);
1862 #ifdef RX_ENABLE_LOCKS
1863 CV_BROADCAST(&conn->conn_call_cv);
1868 #ifdef RX_ENABLE_LOCKS
1870 MUTEX_EXIT(&conn->conn_data_lock);
1872 #endif /* RX_ENABLE_LOCKS */
1873 call->state = RX_STATE_DALLY;
1875 error = call->error;
1877 /* currentPacket, nLeft, and NFree must be zeroed here, because
1878 * ResetCall cannot: ResetCall may be called at splnet(), in the
1879 * kernel version, and may interrupt the macros rx_Read or
1880 * rx_Write, which run at normal priority for efficiency. */
1881 if (call->currentPacket) {
1882 rxi_FreePacket(call->currentPacket);
1883 call->currentPacket = (struct rx_packet *)0;
1884 call->nLeft = call->nFree = call->curlen = 0;
1886 call->nLeft = call->nFree = call->curlen = 0;
1888 /* Free any packets from the last call to ReadvProc/WritevProc */
1889 for (queue_Scan(&call->iovq, tp, nxp, rx_packet)) {
1894 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1895 MUTEX_EXIT(&call->lock);
1896 if (conn->type == RX_CLIENT_CONNECTION) {
1897 MUTEX_EXIT(&conn->conn_call_lock);
1898 conn->flags &= ~RX_CONN_BUSY;
1903 * Map errors to the local host's errno.h format.
1905 error = ntoh_syserr_conv(error);
1909 #if !defined(KERNEL)
1911 /* Call this routine when shutting down a server or client (especially
1912 * clients). This will allow Rx to gracefully garbage collect server
1913 * connections, and reduce the number of retries that a server might
1914 * make to a dead client.
1915 * This is not quite right, since some calls may still be ongoing and
1916 * we can't lock them to destroy them. */
1920 register struct rx_connection **conn_ptr, **conn_end;
1922 INIT_PTHREAD_LOCKS LOCK_RX_INIT if (rxinit_status == 1) {
1923 UNLOCK_RX_INIT return; /* Already shutdown. */
1925 rxi_DeleteCachedConnections();
1926 if (rx_connHashTable) {
1927 MUTEX_ENTER(&rx_connHashTable_lock);
1928 for (conn_ptr = &rx_connHashTable[0], conn_end =
1929 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
1931 struct rx_connection *conn, *next;
1932 for (conn = *conn_ptr; conn; conn = next) {
1934 if (conn->type == RX_CLIENT_CONNECTION) {
1935 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
1937 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
1938 #ifdef RX_ENABLE_LOCKS
1939 rxi_DestroyConnectionNoLock(conn);
1940 #else /* RX_ENABLE_LOCKS */
1941 rxi_DestroyConnection(conn);
1942 #endif /* RX_ENABLE_LOCKS */
1946 #ifdef RX_ENABLE_LOCKS
1947 while (rx_connCleanup_list) {
1948 struct rx_connection *conn;
1949 conn = rx_connCleanup_list;
1950 rx_connCleanup_list = rx_connCleanup_list->next;
1951 MUTEX_EXIT(&rx_connHashTable_lock);
1952 rxi_CleanupConnection(conn);
1953 MUTEX_ENTER(&rx_connHashTable_lock);
1955 MUTEX_EXIT(&rx_connHashTable_lock);
1956 #endif /* RX_ENABLE_LOCKS */
1964 /* if we wakeup packet waiter too often, can get in loop with two
1965 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
1967 rxi_PacketsUnWait(void)
1969 if (!rx_waitingForPackets) {
1973 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
1974 return; /* still over quota */
1977 rx_waitingForPackets = 0;
1978 #ifdef RX_ENABLE_LOCKS
1979 CV_BROADCAST(&rx_waitingForPackets_cv);
1981 osi_rxWakeup(&rx_waitingForPackets);
1987 /* ------------------Internal interfaces------------------------- */
1989 /* Return this process's service structure for the
1990 * specified socket and service */
1992 rxi_FindService(register osi_socket socket, register u_short serviceId)
1994 register struct rx_service **sp;
1995 for (sp = &rx_services[0]; *sp; sp++) {
1996 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2002 /* Allocate a call structure, for the indicated channel of the
2003 * supplied connection. The mode and state of the call must be set by
2004 * the caller. Returns the call with mutex locked. */
2006 rxi_NewCall(register struct rx_connection *conn, register int channel)
2008 register struct rx_call *call;
2009 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2010 register struct rx_call *cp; /* Call pointer temp */
2011 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2012 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2014 /* Grab an existing call structure, or allocate a new one.
2015 * Existing call structures are assumed to have been left reset by
2017 MUTEX_ENTER(&rx_freeCallQueue_lock);
2019 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2021 * EXCEPT that the TQ might not yet be cleared out.
2022 * Skip over those with in-use TQs.
2025 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2026 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2032 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2033 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2034 call = queue_First(&rx_freeCallQueue, rx_call);
2035 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2037 MUTEX_ENTER(&rx_stats_mutex);
2038 rx_stats.nFreeCallStructs--;
2039 MUTEX_EXIT(&rx_stats_mutex);
2040 MUTEX_EXIT(&rx_freeCallQueue_lock);
2041 MUTEX_ENTER(&call->lock);
2042 CLEAR_CALL_QUEUE_LOCK(call);
2043 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2044 /* Now, if TQ wasn't cleared earlier, do it now. */
2045 if (call->flags & RX_CALL_TQ_CLEARME) {
2046 rxi_ClearTransmitQueue(call, 0);
2047 queue_Init(&call->tq);
2049 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2050 /* Bind the call to its connection structure */
2052 rxi_ResetCall(call, 1);
2054 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2056 MUTEX_EXIT(&rx_freeCallQueue_lock);
2057 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2058 MUTEX_ENTER(&call->lock);
2059 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2060 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2061 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2063 MUTEX_ENTER(&rx_stats_mutex);
2064 rx_stats.nCallStructs++;
2065 MUTEX_EXIT(&rx_stats_mutex);
2066 /* Initialize once-only items */
2067 queue_Init(&call->tq);
2068 queue_Init(&call->rq);
2069 queue_Init(&call->iovq);
2070 /* Bind the call to its connection structure (prereq for reset) */
2072 rxi_ResetCall(call, 1);
2074 call->channel = channel;
2075 call->callNumber = &conn->callNumber[channel];
2076 /* Note that the next expected call number is retained (in
2077 * conn->callNumber[i]), even if we reallocate the call structure
2079 conn->call[channel] = call;
2080 /* if the channel's never been used (== 0), we should start at 1, otherwise
2081 * the call number is valid from the last time this channel was used */
2082 if (*call->callNumber == 0)
2083 *call->callNumber = 1;
2088 /* A call has been inactive long enough that so we can throw away
2089 * state, including the call structure, which is placed on the call
2091 * Call is locked upon entry.
2092 * haveCTLock set if called from rxi_ReapConnections
2094 #ifdef RX_ENABLE_LOCKS
2096 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2097 #else /* RX_ENABLE_LOCKS */
2099 rxi_FreeCall(register struct rx_call *call)
2100 #endif /* RX_ENABLE_LOCKS */
2102 register int channel = call->channel;
2103 register struct rx_connection *conn = call->conn;
2106 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2107 (*call->callNumber)++;
2108 rxi_ResetCall(call, 0);
2109 call->conn->call[channel] = (struct rx_call *)0;
2111 MUTEX_ENTER(&rx_freeCallQueue_lock);
2112 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2113 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2114 /* A call may be free even though its transmit queue is still in use.
2115 * Since we search the call list from head to tail, put busy calls at
2116 * the head of the list, and idle calls at the tail.
2118 if (call->flags & RX_CALL_TQ_BUSY)
2119 queue_Prepend(&rx_freeCallQueue, call);
2121 queue_Append(&rx_freeCallQueue, call);
2122 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2123 queue_Append(&rx_freeCallQueue, call);
2124 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2125 MUTEX_ENTER(&rx_stats_mutex);
2126 rx_stats.nFreeCallStructs++;
2127 MUTEX_EXIT(&rx_stats_mutex);
2129 MUTEX_EXIT(&rx_freeCallQueue_lock);
2131 /* Destroy the connection if it was previously slated for
2132 * destruction, i.e. the Rx client code previously called
2133 * rx_DestroyConnection (client connections), or
2134 * rxi_ReapConnections called the same routine (server
2135 * connections). Only do this, however, if there are no
2136 * outstanding calls. Note that for fine grain locking, there appears
2137 * to be a deadlock in that rxi_FreeCall has a call locked and
2138 * DestroyConnectionNoLock locks each call in the conn. But note a
2139 * few lines up where we have removed this call from the conn.
2140 * If someone else destroys a connection, they either have no
2141 * call lock held or are going through this section of code.
2143 if (conn->flags & RX_CONN_DESTROY_ME) {
2144 MUTEX_ENTER(&conn->conn_data_lock);
2146 MUTEX_EXIT(&conn->conn_data_lock);
2147 #ifdef RX_ENABLE_LOCKS
2149 rxi_DestroyConnectionNoLock(conn);
2151 rxi_DestroyConnection(conn);
2152 #else /* RX_ENABLE_LOCKS */
2153 rxi_DestroyConnection(conn);
2154 #endif /* RX_ENABLE_LOCKS */
2158 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2160 rxi_Alloc(register size_t size)
2164 #if defined(AFS_AIX41_ENV) && defined(KERNEL)
2165 /* Grab the AFS filesystem lock. See afs/osi.h for the lock
2168 int glockOwner = ISAFS_GLOCK();
2172 MUTEX_ENTER(&rx_stats_mutex);
2174 rxi_Allocsize += size;
2175 MUTEX_EXIT(&rx_stats_mutex);
2176 #if (defined(AFS_AIX32_ENV) || defined(AFS_HPUX_ENV)) && !defined(AFS_HPUX100_ENV) && defined(KERNEL)
2177 if (size > AFS_SMALLOCSIZ) {
2178 p = (char *)osi_AllocMediumSpace(size);
2180 p = (char *)osi_AllocSmall(size, 1);
2181 #if defined(AFS_AIX41_ENV) && defined(KERNEL)
2186 p = (char *)osi_Alloc(size);
2189 osi_Panic("rxi_Alloc error");
2195 rxi_Free(void *addr, register size_t size)
2197 #if defined(AFS_AIX41_ENV) && defined(KERNEL)
2198 /* Grab the AFS filesystem lock. See afs/osi.h for the lock
2201 int glockOwner = ISAFS_GLOCK();
2205 MUTEX_ENTER(&rx_stats_mutex);
2207 rxi_Allocsize -= size;
2208 MUTEX_EXIT(&rx_stats_mutex);
2209 #if (defined(AFS_AIX32_ENV) || defined(AFS_HPUX_ENV)) && !defined(AFS_HPUX100_ENV) && defined(KERNEL)
2210 if (size > AFS_SMALLOCSIZ)
2211 osi_FreeMediumSpace(addr);
2213 osi_FreeSmall(addr);
2214 #if defined(AFS_AIX41_ENV) && defined(KERNEL)
2219 osi_Free(addr, size);
2223 /* Find the peer process represented by the supplied (host,port)
2224 * combination. If there is no appropriate active peer structure, a
2225 * new one will be allocated and initialized
2226 * The origPeer, if set, is a pointer to a peer structure on which the
2227 * refcount will be be decremented. This is used to replace the peer
2228 * structure hanging off a connection structure */
2230 rxi_FindPeer(register afs_uint32 host, register u_short port,
2231 struct rx_peer *origPeer, int create)
2233 register struct rx_peer *pp;
2235 hashIndex = PEER_HASH(host, port);
2236 MUTEX_ENTER(&rx_peerHashTable_lock);
2237 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2238 if ((pp->host == host) && (pp->port == port))
2243 pp = rxi_AllocPeer(); /* This bzero's *pp */
2244 pp->host = host; /* set here or in InitPeerParams is zero */
2246 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2247 queue_Init(&pp->congestionQueue);
2248 queue_Init(&pp->rpcStats);
2249 pp->next = rx_peerHashTable[hashIndex];
2250 rx_peerHashTable[hashIndex] = pp;
2251 rxi_InitPeerParams(pp);
2252 MUTEX_ENTER(&rx_stats_mutex);
2253 rx_stats.nPeerStructs++;
2254 MUTEX_EXIT(&rx_stats_mutex);
2261 origPeer->refCount--;
2262 MUTEX_EXIT(&rx_peerHashTable_lock);
2267 /* Find the connection at (host, port) started at epoch, and with the
2268 * given connection id. Creates the server connection if necessary.
2269 * The type specifies whether a client connection or a server
2270 * connection is desired. In both cases, (host, port) specify the
2271 * peer's (host, pair) pair. Client connections are not made
2272 * automatically by this routine. The parameter socket gives the
2273 * socket descriptor on which the packet was received. This is used,
2274 * in the case of server connections, to check that *new* connections
2275 * come via a valid (port, serviceId). Finally, the securityIndex
2276 * parameter must match the existing index for the connection. If a
2277 * server connection is created, it will be created using the supplied
2278 * index, if the index is valid for this service */
2279 struct rx_connection *
2280 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2281 register u_short port, u_short serviceId, afs_uint32 cid,
2282 afs_uint32 epoch, int type, u_int securityIndex)
2284 int hashindex, flag;
2285 register struct rx_connection *conn;
2286 hashindex = CONN_HASH(host, port, cid, epoch, type);
2287 MUTEX_ENTER(&rx_connHashTable_lock);
2288 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2289 rx_connHashTable[hashindex],
2292 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2293 && (epoch == conn->epoch)) {
2294 register struct rx_peer *pp = conn->peer;
2295 if (securityIndex != conn->securityIndex) {
2296 /* this isn't supposed to happen, but someone could forge a packet
2297 * like this, and there seems to be some CM bug that makes this
2298 * happen from time to time -- in which case, the fileserver
2300 MUTEX_EXIT(&rx_connHashTable_lock);
2301 return (struct rx_connection *)0;
2303 if (pp->host == host && pp->port == port)
2305 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2307 if (type == RX_CLIENT_CONNECTION && (conn->epoch & 0x80000000))
2311 /* the connection rxLastConn that was used the last time is not the
2312 ** one we are looking for now. Hence, start searching in the hash */
2314 conn = rx_connHashTable[hashindex];
2319 struct rx_service *service;
2320 if (type == RX_CLIENT_CONNECTION) {
2321 MUTEX_EXIT(&rx_connHashTable_lock);
2322 return (struct rx_connection *)0;
2324 service = rxi_FindService(socket, serviceId);
2325 if (!service || (securityIndex >= service->nSecurityObjects)
2326 || (service->securityObjects[securityIndex] == 0)) {
2327 MUTEX_EXIT(&rx_connHashTable_lock);
2328 return (struct rx_connection *)0;
2330 conn = rxi_AllocConnection(); /* This bzero's the connection */
2331 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2332 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2333 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2334 conn->next = rx_connHashTable[hashindex];
2335 rx_connHashTable[hashindex] = conn;
2336 conn->peer = rxi_FindPeer(host, port, 0, 1);
2337 conn->type = RX_SERVER_CONNECTION;
2338 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2339 conn->epoch = epoch;
2340 conn->cid = cid & RX_CIDMASK;
2341 /* conn->serial = conn->lastSerial = 0; */
2342 /* conn->timeout = 0; */
2343 conn->ackRate = RX_FAST_ACK_RATE;
2344 conn->service = service;
2345 conn->serviceId = serviceId;
2346 conn->securityIndex = securityIndex;
2347 conn->securityObject = service->securityObjects[securityIndex];
2348 conn->nSpecific = 0;
2349 conn->specific = NULL;
2350 rx_SetConnDeadTime(conn, service->connDeadTime);
2351 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2352 /* Notify security object of the new connection */
2353 RXS_NewConnection(conn->securityObject, conn);
2354 /* XXXX Connection timeout? */
2355 if (service->newConnProc)
2356 (*service->newConnProc) (conn);
2357 MUTEX_ENTER(&rx_stats_mutex);
2358 rx_stats.nServerConns++;
2359 MUTEX_EXIT(&rx_stats_mutex);
2362 MUTEX_ENTER(&conn->conn_data_lock);
2364 MUTEX_EXIT(&conn->conn_data_lock);
2366 rxLastConn = conn; /* store this connection as the last conn used */
2367 MUTEX_EXIT(&rx_connHashTable_lock);
2371 /* There are two packet tracing routines available for testing and monitoring
2372 * Rx. One is called just after every packet is received and the other is
2373 * called just before every packet is sent. Received packets, have had their
2374 * headers decoded, and packets to be sent have not yet had their headers
2375 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2376 * containing the network address. Both can be modified. The return value, if
2377 * non-zero, indicates that the packet should be dropped. */
2379 int (*rx_justReceived) () = 0;
2380 int (*rx_almostSent) () = 0;
2382 /* A packet has been received off the interface. Np is the packet, socket is
2383 * the socket number it was received from (useful in determining which service
2384 * this packet corresponds to), and (host, port) reflect the host,port of the
2385 * sender. This call returns the packet to the caller if it is finished with
2386 * it, rather than de-allocating it, just as a small performance hack */
2389 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2390 afs_uint32 host, u_short port, int *tnop,
2391 struct rx_call **newcallp)
2393 register struct rx_call *call;
2394 register struct rx_connection *conn;
2396 afs_uint32 currentCallNumber;
2402 struct rx_packet *tnp;
2405 /* We don't print out the packet until now because (1) the time may not be
2406 * accurate enough until now in the lwp implementation (rx_Listener only gets
2407 * the time after the packet is read) and (2) from a protocol point of view,
2408 * this is the first time the packet has been seen */
2409 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2410 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2411 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2412 np->header.serial, packetType, host, port, np->header.serviceId,
2413 np->header.epoch, np->header.cid, np->header.callNumber,
2414 np->header.seq, np->header.flags, np));
2417 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2418 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2421 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2422 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2425 /* If an input tracer function is defined, call it with the packet and
2426 * network address. Note this function may modify its arguments. */
2427 if (rx_justReceived) {
2428 struct sockaddr_in addr;
2430 addr.sin_family = AF_INET;
2431 addr.sin_port = port;
2432 addr.sin_addr.s_addr = host;
2433 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2434 addr.sin_len = sizeof(addr);
2435 #endif /* AFS_OSF_ENV */
2436 drop = (*rx_justReceived) (np, &addr);
2437 /* drop packet if return value is non-zero */
2440 port = addr.sin_port; /* in case fcn changed addr */
2441 host = addr.sin_addr.s_addr;
2445 /* If packet was not sent by the client, then *we* must be the client */
2446 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2447 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2449 /* Find the connection (or fabricate one, if we're the server & if
2450 * necessary) associated with this packet */
2452 rxi_FindConnection(socket, host, port, np->header.serviceId,
2453 np->header.cid, np->header.epoch, type,
2454 np->header.securityIndex);
2457 /* If no connection found or fabricated, just ignore the packet.
2458 * (An argument could be made for sending an abort packet for
2463 MUTEX_ENTER(&conn->conn_data_lock);
2464 if (conn->maxSerial < np->header.serial)
2465 conn->maxSerial = np->header.serial;
2466 MUTEX_EXIT(&conn->conn_data_lock);
2468 /* If the connection is in an error state, send an abort packet and ignore
2469 * the incoming packet */
2471 /* Don't respond to an abort packet--we don't want loops! */
2472 MUTEX_ENTER(&conn->conn_data_lock);
2473 if (np->header.type != RX_PACKET_TYPE_ABORT)
2474 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2476 MUTEX_EXIT(&conn->conn_data_lock);
2480 /* Check for connection-only requests (i.e. not call specific). */
2481 if (np->header.callNumber == 0) {
2482 switch (np->header.type) {
2483 case RX_PACKET_TYPE_ABORT:
2484 /* What if the supplied error is zero? */
2485 rxi_ConnectionError(conn, ntohl(rx_GetInt32(np, 0)));
2486 MUTEX_ENTER(&conn->conn_data_lock);
2488 MUTEX_EXIT(&conn->conn_data_lock);
2490 case RX_PACKET_TYPE_CHALLENGE:
2491 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2492 MUTEX_ENTER(&conn->conn_data_lock);
2494 MUTEX_EXIT(&conn->conn_data_lock);
2496 case RX_PACKET_TYPE_RESPONSE:
2497 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2498 MUTEX_ENTER(&conn->conn_data_lock);
2500 MUTEX_EXIT(&conn->conn_data_lock);
2502 case RX_PACKET_TYPE_PARAMS:
2503 case RX_PACKET_TYPE_PARAMS + 1:
2504 case RX_PACKET_TYPE_PARAMS + 2:
2505 /* ignore these packet types for now */
2506 MUTEX_ENTER(&conn->conn_data_lock);
2508 MUTEX_EXIT(&conn->conn_data_lock);
2513 /* Should not reach here, unless the peer is broken: send an
2515 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2516 MUTEX_ENTER(&conn->conn_data_lock);
2517 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2519 MUTEX_EXIT(&conn->conn_data_lock);
2524 channel = np->header.cid & RX_CHANNELMASK;
2525 call = conn->call[channel];
2526 #ifdef RX_ENABLE_LOCKS
2528 MUTEX_ENTER(&call->lock);
2529 /* Test to see if call struct is still attached to conn. */
2530 if (call != conn->call[channel]) {
2532 MUTEX_EXIT(&call->lock);
2533 if (type == RX_SERVER_CONNECTION) {
2534 call = conn->call[channel];
2535 /* If we started with no call attached and there is one now,
2536 * another thread is also running this routine and has gotten
2537 * the connection channel. We should drop this packet in the tests
2538 * below. If there was a call on this connection and it's now
2539 * gone, then we'll be making a new call below.
2540 * If there was previously a call and it's now different then
2541 * the old call was freed and another thread running this routine
2542 * has created a call on this channel. One of these two threads
2543 * has a packet for the old call and the code below handles those
2547 MUTEX_ENTER(&call->lock);
2549 /* This packet can't be for this call. If the new call address is
2550 * 0 then no call is running on this channel. If there is a call
2551 * then, since this is a client connection we're getting data for
2552 * it must be for the previous call.
2554 MUTEX_ENTER(&rx_stats_mutex);
2555 rx_stats.spuriousPacketsRead++;
2556 MUTEX_EXIT(&rx_stats_mutex);
2557 MUTEX_ENTER(&conn->conn_data_lock);
2559 MUTEX_EXIT(&conn->conn_data_lock);
2564 currentCallNumber = conn->callNumber[channel];
2566 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2567 if (np->header.callNumber < currentCallNumber) {
2568 MUTEX_ENTER(&rx_stats_mutex);
2569 rx_stats.spuriousPacketsRead++;
2570 MUTEX_EXIT(&rx_stats_mutex);
2571 #ifdef RX_ENABLE_LOCKS
2573 MUTEX_EXIT(&call->lock);
2575 MUTEX_ENTER(&conn->conn_data_lock);
2577 MUTEX_EXIT(&conn->conn_data_lock);
2581 MUTEX_ENTER(&conn->conn_call_lock);
2582 call = rxi_NewCall(conn, channel);
2583 MUTEX_EXIT(&conn->conn_call_lock);
2584 *call->callNumber = np->header.callNumber;
2585 call->state = RX_STATE_PRECALL;
2586 clock_GetTime(&call->queueTime);
2587 hzero(call->bytesSent);
2588 hzero(call->bytesRcvd);
2589 rxi_KeepAliveOn(call);
2590 } else if (np->header.callNumber != currentCallNumber) {
2591 /* Wait until the transmit queue is idle before deciding
2592 * whether to reset the current call. Chances are that the
2593 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2596 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2597 while ((call->state == RX_STATE_ACTIVE)
2598 && (call->flags & RX_CALL_TQ_BUSY)) {
2599 call->flags |= RX_CALL_TQ_WAIT;
2600 #ifdef RX_ENABLE_LOCKS
2601 CV_WAIT(&call->cv_tq, &call->lock);
2602 #else /* RX_ENABLE_LOCKS */
2603 osi_rxSleep(&call->tq);
2604 #endif /* RX_ENABLE_LOCKS */
2606 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2607 /* If the new call cannot be taken right now send a busy and set
2608 * the error condition in this call, so that it terminates as
2609 * quickly as possible */
2610 if (call->state == RX_STATE_ACTIVE) {
2611 struct rx_packet *tp;
2613 rxi_CallError(call, RX_CALL_DEAD);
2614 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2616 MUTEX_EXIT(&call->lock);
2617 MUTEX_ENTER(&conn->conn_data_lock);
2619 MUTEX_EXIT(&conn->conn_data_lock);
2622 rxi_ResetCall(call, 0);
2623 *call->callNumber = np->header.callNumber;
2624 call->state = RX_STATE_PRECALL;
2625 clock_GetTime(&call->queueTime);
2626 hzero(call->bytesSent);
2627 hzero(call->bytesRcvd);
2629 * If the number of queued calls exceeds the overload
2630 * threshold then abort this call.
2632 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2633 struct rx_packet *tp;
2635 rxi_CallError(call, rx_BusyError);
2636 tp = rxi_SendCallAbort(call, np, 1, 0);
2637 MUTEX_EXIT(&call->lock);
2638 MUTEX_ENTER(&conn->conn_data_lock);
2640 MUTEX_EXIT(&conn->conn_data_lock);
2643 rxi_KeepAliveOn(call);
2645 /* Continuing call; do nothing here. */
2647 } else { /* we're the client */
2648 /* Ignore all incoming acknowledgements for calls in DALLY state */
2649 if (call && (call->state == RX_STATE_DALLY)
2650 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2651 MUTEX_ENTER(&rx_stats_mutex);
2652 rx_stats.ignorePacketDally++;
2653 MUTEX_EXIT(&rx_stats_mutex);
2654 #ifdef RX_ENABLE_LOCKS
2656 MUTEX_EXIT(&call->lock);
2659 MUTEX_ENTER(&conn->conn_data_lock);
2661 MUTEX_EXIT(&conn->conn_data_lock);
2665 /* Ignore anything that's not relevant to the current call. If there
2666 * isn't a current call, then no packet is relevant. */
2667 if (!call || (np->header.callNumber != currentCallNumber)) {
2668 MUTEX_ENTER(&rx_stats_mutex);
2669 rx_stats.spuriousPacketsRead++;
2670 MUTEX_EXIT(&rx_stats_mutex);
2671 #ifdef RX_ENABLE_LOCKS
2673 MUTEX_EXIT(&call->lock);
2676 MUTEX_ENTER(&conn->conn_data_lock);
2678 MUTEX_EXIT(&conn->conn_data_lock);
2681 /* If the service security object index stamped in the packet does not
2682 * match the connection's security index, ignore the packet */
2683 if (np->header.securityIndex != conn->securityIndex) {
2684 #ifdef RX_ENABLE_LOCKS
2685 MUTEX_EXIT(&call->lock);
2687 MUTEX_ENTER(&conn->conn_data_lock);
2689 MUTEX_EXIT(&conn->conn_data_lock);
2693 /* If we're receiving the response, then all transmit packets are
2694 * implicitly acknowledged. Get rid of them. */
2695 if (np->header.type == RX_PACKET_TYPE_DATA) {
2696 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2697 /* XXX Hack. Because we must release the global rx lock when
2698 * sending packets (osi_NetSend) we drop all acks while we're
2699 * traversing the tq in rxi_Start sending packets out because
2700 * packets may move to the freePacketQueue as result of being here!
2701 * So we drop these packets until we're safely out of the
2702 * traversing. Really ugly!
2703 * For fine grain RX locking, we set the acked field in the
2704 * packets and let rxi_Start remove them from the transmit queue.
2706 if (call->flags & RX_CALL_TQ_BUSY) {
2707 #ifdef RX_ENABLE_LOCKS
2708 rxi_SetAcksInTransmitQueue(call);
2711 return np; /* xmitting; drop packet */
2714 rxi_ClearTransmitQueue(call, 0);
2716 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2717 rxi_ClearTransmitQueue(call, 0);
2718 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2720 if (np->header.type == RX_PACKET_TYPE_ACK) {
2721 /* now check to see if this is an ack packet acknowledging that the
2722 * server actually *lost* some hard-acked data. If this happens we
2723 * ignore this packet, as it may indicate that the server restarted in
2724 * the middle of a call. It is also possible that this is an old ack
2725 * packet. We don't abort the connection in this case, because this
2726 * *might* just be an old ack packet. The right way to detect a server
2727 * restart in the midst of a call is to notice that the server epoch
2729 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2730 * XXX unacknowledged. I think that this is off-by-one, but
2731 * XXX I don't dare change it just yet, since it will
2732 * XXX interact badly with the server-restart detection
2733 * XXX code in receiveackpacket. */
2734 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2735 MUTEX_ENTER(&rx_stats_mutex);
2736 rx_stats.spuriousPacketsRead++;
2737 MUTEX_EXIT(&rx_stats_mutex);
2738 MUTEX_EXIT(&call->lock);
2739 MUTEX_ENTER(&conn->conn_data_lock);
2741 MUTEX_EXIT(&conn->conn_data_lock);
2745 } /* else not a data packet */
2748 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2749 /* Set remote user defined status from packet */
2750 call->remoteStatus = np->header.userStatus;
2752 /* Note the gap between the expected next packet and the actual
2753 * packet that arrived, when the new packet has a smaller serial number
2754 * than expected. Rioses frequently reorder packets all by themselves,
2755 * so this will be quite important with very large window sizes.
2756 * Skew is checked against 0 here to avoid any dependence on the type of
2757 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2759 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2760 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2761 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2763 MUTEX_ENTER(&conn->conn_data_lock);
2764 skew = conn->lastSerial - np->header.serial;
2765 conn->lastSerial = np->header.serial;
2766 MUTEX_EXIT(&conn->conn_data_lock);
2768 register struct rx_peer *peer;
2770 if (skew > peer->inPacketSkew) {
2771 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2773 peer->inPacketSkew = skew;
2777 /* Now do packet type-specific processing */
2778 switch (np->header.type) {
2779 case RX_PACKET_TYPE_DATA:
2780 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2783 case RX_PACKET_TYPE_ACK:
2784 /* Respond immediately to ack packets requesting acknowledgement
2786 if (np->header.flags & RX_REQUEST_ACK) {
2788 (void)rxi_SendCallAbort(call, 0, 1, 0);
2790 (void)rxi_SendAck(call, 0, np->header.serial,
2791 RX_ACK_PING_RESPONSE, 1);
2793 np = rxi_ReceiveAckPacket(call, np, 1);
2795 case RX_PACKET_TYPE_ABORT:
2796 /* An abort packet: reset the connection, passing the error up to
2798 /* What if error is zero? */
2799 rxi_CallError(call, ntohl(*(afs_int32 *) rx_DataOf(np)));
2801 case RX_PACKET_TYPE_BUSY:
2804 case RX_PACKET_TYPE_ACKALL:
2805 /* All packets acknowledged, so we can drop all packets previously
2806 * readied for sending */
2807 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2808 /* XXX Hack. We because we can't release the global rx lock when
2809 * sending packets (osi_NetSend) we drop all ack pkts while we're
2810 * traversing the tq in rxi_Start sending packets out because
2811 * packets may move to the freePacketQueue as result of being
2812 * here! So we drop these packets until we're safely out of the
2813 * traversing. Really ugly!
2814 * For fine grain RX locking, we set the acked field in the packets
2815 * and let rxi_Start remove the packets from the transmit queue.
2817 if (call->flags & RX_CALL_TQ_BUSY) {
2818 #ifdef RX_ENABLE_LOCKS
2819 rxi_SetAcksInTransmitQueue(call);
2821 #else /* RX_ENABLE_LOCKS */
2823 return np; /* xmitting; drop packet */
2824 #endif /* RX_ENABLE_LOCKS */
2826 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2827 rxi_ClearTransmitQueue(call, 0);
2830 /* Should not reach here, unless the peer is broken: send an abort
2832 rxi_CallError(call, RX_PROTOCOL_ERROR);
2833 np = rxi_SendCallAbort(call, np, 1, 0);
2836 /* Note when this last legitimate packet was received, for keep-alive
2837 * processing. Note, we delay getting the time until now in the hope that
2838 * the packet will be delivered to the user before any get time is required
2839 * (if not, then the time won't actually be re-evaluated here). */
2840 call->lastReceiveTime = clock_Sec();
2841 MUTEX_EXIT(&call->lock);
2842 MUTEX_ENTER(&conn->conn_data_lock);
2844 MUTEX_EXIT(&conn->conn_data_lock);
2848 /* return true if this is an "interesting" connection from the point of view
2849 of someone trying to debug the system */
2851 rxi_IsConnInteresting(struct rx_connection *aconn)
2854 register struct rx_call *tcall;
2856 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2858 for (i = 0; i < RX_MAXCALLS; i++) {
2859 tcall = aconn->call[i];
2861 if ((tcall->state == RX_STATE_PRECALL)
2862 || (tcall->state == RX_STATE_ACTIVE))
2864 if ((tcall->mode == RX_MODE_SENDING)
2865 || (tcall->mode == RX_MODE_RECEIVING))
2873 /* if this is one of the last few packets AND it wouldn't be used by the
2874 receiving call to immediately satisfy a read request, then drop it on
2875 the floor, since accepting it might prevent a lock-holding thread from
2876 making progress in its reading. If a call has been cleared while in
2877 the precall state then ignore all subsequent packets until the call
2878 is assigned to a thread. */
2881 TooLow(struct rx_packet *ap, struct rx_call *acall)
2884 MUTEX_ENTER(&rx_stats_mutex);
2885 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
2886 && (acall->state == RX_STATE_PRECALL))
2887 || ((rx_nFreePackets < rxi_dataQuota + 2)
2888 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
2889 && (acall->flags & RX_CALL_READER_WAIT)))) {
2892 MUTEX_EXIT(&rx_stats_mutex);
2898 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
2899 struct rx_call *acall)
2901 struct rx_call *call = acall;
2905 MUTEX_ENTER(&conn->conn_data_lock);
2906 conn->checkReachEvent = NULL;
2907 waiting = conn->flags & RX_CONN_ATTACHWAIT;
2910 MUTEX_EXIT(&conn->conn_data_lock);
2914 MUTEX_ENTER(&conn->conn_call_lock);
2915 MUTEX_ENTER(&conn->conn_data_lock);
2916 for (i = 0; i < RX_MAXCALLS; i++) {
2917 struct rx_call *tc = conn->call[i];
2918 if (tc && tc->state == RX_STATE_PRECALL) {
2924 /* Indicate that rxi_CheckReachEvent is no longer running by
2925 * clearing the flag. Must be atomic under conn_data_lock to
2926 * avoid a new call slipping by: rxi_CheckConnReach holds
2927 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
2929 conn->flags &= ~RX_CONN_ATTACHWAIT;
2930 MUTEX_EXIT(&conn->conn_data_lock);
2931 MUTEX_EXIT(&conn->conn_call_lock);
2936 MUTEX_ENTER(&call->lock);
2937 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
2939 MUTEX_EXIT(&call->lock);
2941 clock_GetTime(&when);
2942 when.sec += RX_CHECKREACH_TIMEOUT;
2943 MUTEX_ENTER(&conn->conn_data_lock);
2944 if (!conn->checkReachEvent) {
2946 conn->checkReachEvent =
2947 rxevent_Post(&when, rxi_CheckReachEvent, conn, NULL);
2949 MUTEX_EXIT(&conn->conn_data_lock);
2955 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
2957 struct rx_service *service = conn->service;
2958 struct rx_peer *peer = conn->peer;
2959 afs_uint32 now, lastReach;
2961 if (service->checkReach == 0)
2965 MUTEX_ENTER(&peer->peer_lock);
2966 lastReach = peer->lastReachTime;
2967 MUTEX_EXIT(&peer->peer_lock);
2968 if (now - lastReach < RX_CHECKREACH_TTL)
2971 MUTEX_ENTER(&conn->conn_data_lock);
2972 if (conn->flags & RX_CONN_ATTACHWAIT) {
2973 MUTEX_EXIT(&conn->conn_data_lock);
2976 conn->flags |= RX_CONN_ATTACHWAIT;
2977 MUTEX_EXIT(&conn->conn_data_lock);
2978 if (!conn->checkReachEvent)
2979 rxi_CheckReachEvent(NULL, conn, call);
2984 /* try to attach call, if authentication is complete */
2986 TryAttach(register struct rx_call *acall, register osi_socket socket,
2987 register int *tnop, register struct rx_call **newcallp,
2990 struct rx_connection *conn = acall->conn;
2992 if (conn->type == RX_SERVER_CONNECTION
2993 && acall->state == RX_STATE_PRECALL) {
2994 /* Don't attach until we have any req'd. authentication. */
2995 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
2996 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
2997 rxi_AttachServerProc(acall, socket, tnop, newcallp);
2998 /* Note: this does not necessarily succeed; there
2999 * may not any proc available
3002 rxi_ChallengeOn(acall->conn);
3007 /* A data packet has been received off the interface. This packet is
3008 * appropriate to the call (the call is in the right state, etc.). This
3009 * routine can return a packet to the caller, for re-use */
3012 rxi_ReceiveDataPacket(register struct rx_call *call,
3013 register struct rx_packet *np, int istack,
3014 osi_socket socket, afs_uint32 host, u_short port,
3015 int *tnop, struct rx_call **newcallp)
3017 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3021 afs_uint32 seq, serial, flags;
3023 struct rx_packet *tnp;
3025 MUTEX_ENTER(&rx_stats_mutex);
3026 rx_stats.dataPacketsRead++;
3027 MUTEX_EXIT(&rx_stats_mutex);
3030 /* If there are no packet buffers, drop this new packet, unless we can find
3031 * packet buffers from inactive calls */
3033 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3034 MUTEX_ENTER(&rx_freePktQ_lock);
3035 rxi_NeedMorePackets = TRUE;
3036 MUTEX_EXIT(&rx_freePktQ_lock);
3037 MUTEX_ENTER(&rx_stats_mutex);
3038 rx_stats.noPacketBuffersOnRead++;
3039 MUTEX_EXIT(&rx_stats_mutex);
3040 call->rprev = np->header.serial;
3041 rxi_calltrace(RX_TRACE_DROP, call);
3042 dpf(("packet %x dropped on receipt - quota problems", np));
3044 rxi_ClearReceiveQueue(call);
3045 clock_GetTime(&when);
3046 clock_Add(&when, &rx_softAckDelay);
3047 if (!call->delayedAckEvent
3048 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3049 rxevent_Cancel(call->delayedAckEvent, call,
3050 RX_CALL_REFCOUNT_DELAY);
3051 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3052 call->delayedAckEvent =
3053 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3055 /* we've damaged this call already, might as well do it in. */
3061 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3062 * packet is one of several packets transmitted as a single
3063 * datagram. Do not send any soft or hard acks until all packets
3064 * in a jumbogram have been processed. Send negative acks right away.
3066 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3067 /* tnp is non-null when there are more packets in the
3068 * current jumbo gram */
3075 seq = np->header.seq;
3076 serial = np->header.serial;
3077 flags = np->header.flags;
3079 /* If the call is in an error state, send an abort message */
3081 return rxi_SendCallAbort(call, np, istack, 0);
3083 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3084 * AFS 3.5 jumbogram. */
3085 if (flags & RX_JUMBO_PACKET) {
3086 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3091 if (np->header.spare != 0) {
3092 MUTEX_ENTER(&call->conn->conn_data_lock);
3093 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3094 MUTEX_EXIT(&call->conn->conn_data_lock);
3097 /* The usual case is that this is the expected next packet */
3098 if (seq == call->rnext) {
3100 /* Check to make sure it is not a duplicate of one already queued */
3101 if (queue_IsNotEmpty(&call->rq)
3102 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3103 MUTEX_ENTER(&rx_stats_mutex);
3104 rx_stats.dupPacketsRead++;
3105 MUTEX_EXIT(&rx_stats_mutex);
3106 dpf(("packet %x dropped on receipt - duplicate", np));
3107 rxevent_Cancel(call->delayedAckEvent, call,
3108 RX_CALL_REFCOUNT_DELAY);
3109 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3115 /* It's the next packet. Stick it on the receive queue
3116 * for this call. Set newPackets to make sure we wake
3117 * the reader once all packets have been processed */
3118 queue_Prepend(&call->rq, np);
3120 np = NULL; /* We can't use this anymore */
3123 /* If an ack is requested then set a flag to make sure we
3124 * send an acknowledgement for this packet */
3125 if (flags & RX_REQUEST_ACK) {
3126 ackNeeded = RX_ACK_REQUESTED;
3129 /* Keep track of whether we have received the last packet */
3130 if (flags & RX_LAST_PACKET) {
3131 call->flags |= RX_CALL_HAVE_LAST;
3135 /* Check whether we have all of the packets for this call */
3136 if (call->flags & RX_CALL_HAVE_LAST) {
3137 afs_uint32 tseq; /* temporary sequence number */
3138 struct rx_packet *tp; /* Temporary packet pointer */
3139 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3141 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3142 if (tseq != tp->header.seq)
3144 if (tp->header.flags & RX_LAST_PACKET) {
3145 call->flags |= RX_CALL_RECEIVE_DONE;
3152 /* Provide asynchronous notification for those who want it
3153 * (e.g. multi rx) */
3154 if (call->arrivalProc) {
3155 (*call->arrivalProc) (call, call->arrivalProcHandle,
3156 (int)call->arrivalProcArg);
3157 call->arrivalProc = (VOID(*)())0;
3160 /* Update last packet received */
3163 /* If there is no server process serving this call, grab
3164 * one, if available. We only need to do this once. If a
3165 * server thread is available, this thread becomes a server
3166 * thread and the server thread becomes a listener thread. */
3168 TryAttach(call, socket, tnop, newcallp, 0);
3171 /* This is not the expected next packet. */
3173 /* Determine whether this is a new or old packet, and if it's
3174 * a new one, whether it fits into the current receive window.
3175 * Also figure out whether the packet was delivered in sequence.
3176 * We use the prev variable to determine whether the new packet
3177 * is the successor of its immediate predecessor in the
3178 * receive queue, and the missing flag to determine whether
3179 * any of this packets predecessors are missing. */
3181 afs_uint32 prev; /* "Previous packet" sequence number */
3182 struct rx_packet *tp; /* Temporary packet pointer */
3183 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3184 int missing; /* Are any predecessors missing? */
3186 /* If the new packet's sequence number has been sent to the
3187 * application already, then this is a duplicate */
3188 if (seq < call->rnext) {
3189 MUTEX_ENTER(&rx_stats_mutex);
3190 rx_stats.dupPacketsRead++;
3191 MUTEX_EXIT(&rx_stats_mutex);
3192 rxevent_Cancel(call->delayedAckEvent, call,
3193 RX_CALL_REFCOUNT_DELAY);
3194 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3200 /* If the sequence number is greater than what can be
3201 * accomodated by the current window, then send a negative
3202 * acknowledge and drop the packet */
3203 if ((call->rnext + call->rwind) <= seq) {
3204 rxevent_Cancel(call->delayedAckEvent, call,
3205 RX_CALL_REFCOUNT_DELAY);
3206 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3213 /* Look for the packet in the queue of old received packets */
3214 for (prev = call->rnext - 1, missing =
3215 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3216 /*Check for duplicate packet */
3217 if (seq == tp->header.seq) {
3218 MUTEX_ENTER(&rx_stats_mutex);
3219 rx_stats.dupPacketsRead++;
3220 MUTEX_EXIT(&rx_stats_mutex);
3221 rxevent_Cancel(call->delayedAckEvent, call,
3222 RX_CALL_REFCOUNT_DELAY);
3223 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3229 /* If we find a higher sequence packet, break out and
3230 * insert the new packet here. */
3231 if (seq < tp->header.seq)
3233 /* Check for missing packet */
3234 if (tp->header.seq != prev + 1) {
3238 prev = tp->header.seq;
3241 /* Keep track of whether we have received the last packet. */
3242 if (flags & RX_LAST_PACKET) {
3243 call->flags |= RX_CALL_HAVE_LAST;
3246 /* It's within the window: add it to the the receive queue.
3247 * tp is left by the previous loop either pointing at the
3248 * packet before which to insert the new packet, or at the
3249 * queue head if the queue is empty or the packet should be
3251 queue_InsertBefore(tp, np);
3255 /* Check whether we have all of the packets for this call */
3256 if ((call->flags & RX_CALL_HAVE_LAST)
3257 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3258 afs_uint32 tseq; /* temporary sequence number */
3261 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3262 if (tseq != tp->header.seq)
3264 if (tp->header.flags & RX_LAST_PACKET) {
3265 call->flags |= RX_CALL_RECEIVE_DONE;
3272 /* We need to send an ack of the packet is out of sequence,
3273 * or if an ack was requested by the peer. */
3274 if (seq != prev + 1 || missing || (flags & RX_REQUEST_ACK)) {
3275 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3278 /* Acknowledge the last packet for each call */
3279 if (flags & RX_LAST_PACKET) {
3290 * If the receiver is waiting for an iovec, fill the iovec
3291 * using the data from the receive queue */
3292 if (call->flags & RX_CALL_IOVEC_WAIT) {
3293 didHardAck = rxi_FillReadVec(call, serial);
3294 /* the call may have been aborted */
3303 /* Wakeup the reader if any */
3304 if ((call->flags & RX_CALL_READER_WAIT)
3305 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3306 || (call->iovNext >= call->iovMax)
3307 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3308 call->flags &= ~RX_CALL_READER_WAIT;
3309 #ifdef RX_ENABLE_LOCKS
3310 CV_BROADCAST(&call->cv_rq);
3312 osi_rxWakeup(&call->rq);
3318 * Send an ack when requested by the peer, or once every
3319 * rxi_SoftAckRate packets until the last packet has been
3320 * received. Always send a soft ack for the last packet in
3321 * the server's reply. */
3323 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3324 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3325 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3326 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3327 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3328 } else if (call->nSoftAcks) {
3329 clock_GetTime(&when);
3330 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3331 clock_Add(&when, &rx_lastAckDelay);
3333 clock_Add(&when, &rx_softAckDelay);
3335 if (!call->delayedAckEvent
3336 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3337 rxevent_Cancel(call->delayedAckEvent, call,
3338 RX_CALL_REFCOUNT_DELAY);
3339 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3340 call->delayedAckEvent =
3341 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3343 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3344 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3351 static void rxi_ComputeRate();
3355 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3357 struct rx_peer *peer = conn->peer;
3359 MUTEX_ENTER(&peer->peer_lock);
3360 peer->lastReachTime = clock_Sec();
3361 MUTEX_EXIT(&peer->peer_lock);
3363 MUTEX_ENTER(&conn->conn_data_lock);
3364 if (conn->flags & RX_CONN_ATTACHWAIT) {
3367 conn->flags &= ~RX_CONN_ATTACHWAIT;
3368 MUTEX_EXIT(&conn->conn_data_lock);
3370 for (i = 0; i < RX_MAXCALLS; i++) {
3371 struct rx_call *call = conn->call[i];
3374 MUTEX_ENTER(&call->lock);
3375 /* tnop can be null if newcallp is null */
3376 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3378 MUTEX_EXIT(&call->lock);
3382 MUTEX_EXIT(&conn->conn_data_lock);
3385 /* rxi_ComputePeerNetStats
3387 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3388 * estimates (like RTT and throughput) based on ack packets. Caller
3389 * must ensure that the packet in question is the right one (i.e.
3390 * serial number matches).
3393 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3394 struct rx_ackPacket *ap, struct rx_packet *np)
3396 struct rx_peer *peer = call->conn->peer;
3398 /* Use RTT if not delayed by client. */
3399 if (ap->reason != RX_ACK_DELAY)
3400 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3402 rxi_ComputeRate(peer, call, p, np, ap->reason);
3406 /* The real smarts of the whole thing. */
3408 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3411 struct rx_ackPacket *ap;
3413 register struct rx_packet *tp;
3414 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3415 register struct rx_connection *conn = call->conn;
3416 struct rx_peer *peer = conn->peer;
3419 /* because there are CM's that are bogus, sending weird values for this. */
3420 afs_uint32 skew = 0;
3425 int newAckCount = 0;
3426 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3427 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3429 MUTEX_ENTER(&rx_stats_mutex);
3430 rx_stats.ackPacketsRead++;
3431 MUTEX_EXIT(&rx_stats_mutex);
3432 ap = (struct rx_ackPacket *)rx_DataOf(np);
3433 nbytes = rx_Contiguous(np) - ((ap->acks) - (u_char *) ap);
3435 return np; /* truncated ack packet */
3437 /* depends on ack packet struct */
3438 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3439 first = ntohl(ap->firstPacket);
3440 serial = ntohl(ap->serial);
3441 /* temporarily disabled -- needs to degrade over time
3442 * skew = ntohs(ap->maxSkew); */
3444 /* Ignore ack packets received out of order */
3445 if (first < call->tfirst) {
3449 if (np->header.flags & RX_SLOW_START_OK) {
3450 call->flags |= RX_CALL_SLOW_START_OK;
3453 if (ap->reason == RX_ACK_PING_RESPONSE)
3454 rxi_UpdatePeerReach(conn, call);
3459 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3460 ap->reason, ntohl(ap->previousPacket),
3461 (unsigned int)np->header.seq, (unsigned int)serial,
3462 (unsigned int)skew, ntohl(ap->firstPacket));
3465 for (offset = 0; offset < nAcks; offset++)
3466 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3473 /* Update the outgoing packet skew value to the latest value of
3474 * the peer's incoming packet skew value. The ack packet, of
3475 * course, could arrive out of order, but that won't affect things
3477 MUTEX_ENTER(&peer->peer_lock);
3478 peer->outPacketSkew = skew;
3480 /* Check for packets that no longer need to be transmitted, and
3481 * discard them. This only applies to packets positively
3482 * acknowledged as having been sent to the peer's upper level.
3483 * All other packets must be retained. So only packets with
3484 * sequence numbers < ap->firstPacket are candidates. */
3485 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3486 if (tp->header.seq >= first)
3488 call->tfirst = tp->header.seq + 1;
3490 && (tp->header.serial == serial || tp->firstSerial == serial))
3491 rxi_ComputePeerNetStats(call, tp, ap, np);
3492 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3493 /* XXX Hack. Because we have to release the global rx lock when sending
3494 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3495 * in rxi_Start sending packets out because packets may move to the
3496 * freePacketQueue as result of being here! So we drop these packets until
3497 * we're safely out of the traversing. Really ugly!
3498 * To make it even uglier, if we're using fine grain locking, we can
3499 * set the ack bits in the packets and have rxi_Start remove the packets
3500 * when it's done transmitting.
3502 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3505 if (call->flags & RX_CALL_TQ_BUSY) {
3506 #ifdef RX_ENABLE_LOCKS
3507 tp->flags |= RX_PKTFLAG_ACKED;
3508 call->flags |= RX_CALL_TQ_SOME_ACKED;
3509 #else /* RX_ENABLE_LOCKS */
3511 #endif /* RX_ENABLE_LOCKS */
3513 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3516 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3521 /* Give rate detector a chance to respond to ping requests */
3522 if (ap->reason == RX_ACK_PING_RESPONSE) {
3523 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3527 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3529 /* Now go through explicit acks/nacks and record the results in
3530 * the waiting packets. These are packets that can't be released
3531 * yet, even with a positive acknowledge. This positive
3532 * acknowledge only means the packet has been received by the
3533 * peer, not that it will be retained long enough to be sent to
3534 * the peer's upper level. In addition, reset the transmit timers
3535 * of any missing packets (those packets that must be missing
3536 * because this packet was out of sequence) */
3538 call->nSoftAcked = 0;
3539 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3540 /* Update round trip time if the ack was stimulated on receipt
3542 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3543 #ifdef RX_ENABLE_LOCKS
3544 if (tp->header.seq >= first)
3545 #endif /* RX_ENABLE_LOCKS */
3546 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3548 && (tp->header.serial == serial || tp->firstSerial == serial))
3549 rxi_ComputePeerNetStats(call, tp, ap, np);
3551 /* Set the acknowledge flag per packet based on the
3552 * information in the ack packet. An acknowlegded packet can
3553 * be downgraded when the server has discarded a packet it
3554 * soacked previously, or when an ack packet is received
3555 * out of sequence. */
3556 if (tp->header.seq < first) {
3557 /* Implicit ack information */
3558 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3561 tp->flags |= RX_PKTFLAG_ACKED;
3562 } else if (tp->header.seq < first + nAcks) {
3563 /* Explicit ack information: set it in the packet appropriately */
3564 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3565 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3567 tp->flags |= RX_PKTFLAG_ACKED;
3575 tp->flags &= ~RX_PKTFLAG_ACKED;
3579 tp->flags &= ~RX_PKTFLAG_ACKED;
3583 /* If packet isn't yet acked, and it has been transmitted at least
3584 * once, reset retransmit time using latest timeout
3585 * ie, this should readjust the retransmit timer for all outstanding
3586 * packets... So we don't just retransmit when we should know better*/
3588 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3589 tp->retryTime = tp->timeSent;
3590 clock_Add(&tp->retryTime, &peer->timeout);
3591 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3592 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3596 /* If the window has been extended by this acknowledge packet,
3597 * then wakeup a sender waiting in alloc for window space, or try
3598 * sending packets now, if he's been sitting on packets due to
3599 * lack of window space */
3600 if (call->tnext < (call->tfirst + call->twind)) {
3601 #ifdef RX_ENABLE_LOCKS
3602 CV_SIGNAL(&call->cv_twind);
3604 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3605 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3606 osi_rxWakeup(&call->twind);
3609 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3610 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3614 /* if the ack packet has a receivelen field hanging off it,
3615 * update our state */
3616 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3619 /* If the ack packet has a "recommended" size that is less than
3620 * what I am using now, reduce my size to match */
3621 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3622 sizeof(afs_int32), &tSize);
3623 tSize = (afs_uint32) ntohl(tSize);
3624 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3626 /* Get the maximum packet size to send to this peer */
3627 rx_packetread(np, rx_AckDataSize(ap->nAcks), sizeof(afs_int32),
3629 tSize = (afs_uint32) ntohl(tSize);
3630 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3631 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3633 /* sanity check - peer might have restarted with different params.
3634 * If peer says "send less", dammit, send less... Peer should never
3635 * be unable to accept packets of the size that prior AFS versions would
3636 * send without asking. */
3637 if (peer->maxMTU != tSize) {
3638 peer->maxMTU = tSize;
3639 peer->MTU = MIN(tSize, peer->MTU);
3640 call->MTU = MIN(call->MTU, tSize);
3644 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3647 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3648 sizeof(afs_int32), &tSize);
3649 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3650 if (tSize < call->twind) { /* smaller than our send */
3651 call->twind = tSize; /* window, we must send less... */
3652 call->ssthresh = MIN(call->twind, call->ssthresh);
3655 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3656 * network MTU confused with the loopback MTU. Calculate the
3657 * maximum MTU here for use in the slow start code below.
3659 maxMTU = peer->maxMTU;
3660 /* Did peer restart with older RX version? */
3661 if (peer->maxDgramPackets > 1) {
3662 peer->maxDgramPackets = 1;
3664 } else if (np->length >=
3665 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3668 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3669 sizeof(afs_int32), &tSize);
3670 tSize = (afs_uint32) ntohl(tSize);
3672 * As of AFS 3.5 we set the send window to match the receive window.
3674 if (tSize < call->twind) {
3675 call->twind = tSize;
3676 call->ssthresh = MIN(call->twind, call->ssthresh);
3677 } else if (tSize > call->twind) {
3678 call->twind = tSize;
3682 * As of AFS 3.5, a jumbogram is more than one fixed size
3683 * packet transmitted in a single UDP datagram. If the remote
3684 * MTU is smaller than our local MTU then never send a datagram
3685 * larger than the natural MTU.
3688 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3689 sizeof(afs_int32), &tSize);
3690 maxDgramPackets = (afs_uint32) ntohl(tSize);
3691 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3693 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
3694 maxDgramPackets = MIN(maxDgramPackets, tSize);
3695 if (maxDgramPackets > 1) {
3696 peer->maxDgramPackets = maxDgramPackets;
3697 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3699 peer->maxDgramPackets = 1;
3700 call->MTU = peer->natMTU;
3702 } else if (peer->maxDgramPackets > 1) {
3703 /* Restarted with lower version of RX */
3704 peer->maxDgramPackets = 1;
3706 } else if (peer->maxDgramPackets > 1
3707 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3708 /* Restarted with lower version of RX */
3709 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3710 peer->natMTU = OLD_MAX_PACKET_SIZE;
3711 peer->MTU = OLD_MAX_PACKET_SIZE;
3712 peer->maxDgramPackets = 1;
3713 peer->nDgramPackets = 1;
3715 call->MTU = OLD_MAX_PACKET_SIZE;
3720 * Calculate how many datagrams were successfully received after
3721 * the first missing packet and adjust the negative ack counter
3726 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3727 if (call->nNacks < nNacked) {
3728 call->nNacks = nNacked;
3737 if (call->flags & RX_CALL_FAST_RECOVER) {
3739 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3741 call->flags &= ~RX_CALL_FAST_RECOVER;
3742 call->cwind = call->nextCwind;
3743 call->nextCwind = 0;
3746 call->nCwindAcks = 0;
3747 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3748 /* Three negative acks in a row trigger congestion recovery */
3749 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3750 MUTEX_EXIT(&peer->peer_lock);
3751 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3752 /* someone else is waiting to start recovery */
3755 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3756 while (call->flags & RX_CALL_TQ_BUSY) {
3757 call->flags |= RX_CALL_TQ_WAIT;
3758 #ifdef RX_ENABLE_LOCKS
3759 CV_WAIT(&call->cv_tq, &call->lock);
3760 #else /* RX_ENABLE_LOCKS */
3761 osi_rxSleep(&call->tq);
3762 #endif /* RX_ENABLE_LOCKS */
3764 MUTEX_ENTER(&peer->peer_lock);
3765 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3766 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3767 call->flags |= RX_CALL_FAST_RECOVER;
3768 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3770 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3771 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3772 call->nextCwind = call->ssthresh;
3775 peer->MTU = call->MTU;
3776 peer->cwind = call->nextCwind;
3777 peer->nDgramPackets = call->nDgramPackets;
3779 call->congestSeq = peer->congestSeq;
3780 /* Reset the resend times on the packets that were nacked
3781 * so we will retransmit as soon as the window permits*/
3782 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3784 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3785 clock_Zero(&tp->retryTime);
3787 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3792 /* If cwind is smaller than ssthresh, then increase
3793 * the window one packet for each ack we receive (exponential
3795 * If cwind is greater than or equal to ssthresh then increase
3796 * the congestion window by one packet for each cwind acks we
3797 * receive (linear growth). */
3798 if (call->cwind < call->ssthresh) {
3800 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3801 call->nCwindAcks = 0;
3803 call->nCwindAcks += newAckCount;
3804 if (call->nCwindAcks >= call->cwind) {
3805 call->nCwindAcks = 0;
3806 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3810 * If we have received several acknowledgements in a row then
3811 * it is time to increase the size of our datagrams
3813 if ((int)call->nAcks > rx_nDgramThreshold) {
3814 if (peer->maxDgramPackets > 1) {
3815 if (call->nDgramPackets < peer->maxDgramPackets) {
3816 call->nDgramPackets++;
3818 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3819 } else if (call->MTU < peer->maxMTU) {
3820 call->MTU += peer->natMTU;
3821 call->MTU = MIN(call->MTU, peer->maxMTU);
3827 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3829 /* Servers need to hold the call until all response packets have
3830 * been acknowledged. Soft acks are good enough since clients
3831 * are not allowed to clear their receive queues. */
3832 if (call->state == RX_STATE_HOLD
3833 && call->tfirst + call->nSoftAcked >= call->tnext) {
3834 call->state = RX_STATE_DALLY;
3835 rxi_ClearTransmitQueue(call, 0);
3836 } else if (!queue_IsEmpty(&call->tq)) {
3837 rxi_Start(0, call, istack);
3842 /* Received a response to a challenge packet */
3844 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
3845 register struct rx_packet *np, int istack)
3849 /* Ignore the packet if we're the client */
3850 if (conn->type == RX_CLIENT_CONNECTION)
3853 /* If already authenticated, ignore the packet (it's probably a retry) */
3854 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
3857 /* Otherwise, have the security object evaluate the response packet */
3858 error = RXS_CheckResponse(conn->securityObject, conn, np);
3860 /* If the response is invalid, reset the connection, sending
3861 * an abort to the peer */
3865 rxi_ConnectionError(conn, error);
3866 MUTEX_ENTER(&conn->conn_data_lock);
3867 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3868 MUTEX_EXIT(&conn->conn_data_lock);
3871 /* If the response is valid, any calls waiting to attach
3872 * servers can now do so */
3875 for (i = 0; i < RX_MAXCALLS; i++) {
3876 struct rx_call *call = conn->call[i];
3878 MUTEX_ENTER(&call->lock);
3879 if (call->state == RX_STATE_PRECALL)
3880 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
3881 /* tnop can be null if newcallp is null */
3882 MUTEX_EXIT(&call->lock);
3886 /* Update the peer reachability information, just in case
3887 * some calls went into attach-wait while we were waiting
3888 * for authentication..
3890 rxi_UpdatePeerReach(conn, NULL);
3895 /* A client has received an authentication challenge: the security
3896 * object is asked to cough up a respectable response packet to send
3897 * back to the server. The server is responsible for retrying the
3898 * challenge if it fails to get a response. */
3901 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
3902 register struct rx_packet *np, int istack)
3906 /* Ignore the challenge if we're the server */
3907 if (conn->type == RX_SERVER_CONNECTION)
3910 /* Ignore the challenge if the connection is otherwise idle; someone's
3911 * trying to use us as an oracle. */
3912 if (!rxi_HasActiveCalls(conn))
3915 /* Send the security object the challenge packet. It is expected to fill
3916 * in the response. */
3917 error = RXS_GetResponse(conn->securityObject, conn, np);
3919 /* If the security object is unable to return a valid response, reset the
3920 * connection and send an abort to the peer. Otherwise send the response
3921 * packet to the peer connection. */
3923 rxi_ConnectionError(conn, error);
3924 MUTEX_ENTER(&conn->conn_data_lock);
3925 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3926 MUTEX_EXIT(&conn->conn_data_lock);
3928 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
3929 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
3935 /* Find an available server process to service the current request in
3936 * the given call structure. If one isn't available, queue up this
3937 * call so it eventually gets one */
3939 rxi_AttachServerProc(register struct rx_call *call,
3940 register osi_socket socket, register int *tnop,
3941 register struct rx_call **newcallp)
3943 register struct rx_serverQueueEntry *sq;
3944 register struct rx_service *service = call->conn->service;
3945 register int haveQuota = 0;
3947 /* May already be attached */
3948 if (call->state == RX_STATE_ACTIVE)
3951 MUTEX_ENTER(&rx_serverPool_lock);
3953 haveQuota = QuotaOK(service);
3954 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
3955 /* If there are no processes available to service this call,
3956 * put the call on the incoming call queue (unless it's
3957 * already on the queue).
3959 #ifdef RX_ENABLE_LOCKS
3961 ReturnToServerPool(service);
3962 #endif /* RX_ENABLE_LOCKS */
3964 if (!(call->flags & RX_CALL_WAIT_PROC)) {
3965 call->flags |= RX_CALL_WAIT_PROC;
3966 MUTEX_ENTER(&rx_stats_mutex);
3968 MUTEX_EXIT(&rx_stats_mutex);
3969 rxi_calltrace(RX_CALL_ARRIVAL, call);
3970 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
3971 queue_Append(&rx_incomingCallQueue, call);
3974 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
3976 /* If hot threads are enabled, and both newcallp and sq->socketp
3977 * are non-null, then this thread will process the call, and the
3978 * idle server thread will start listening on this threads socket.
3981 if (rx_enable_hot_thread && newcallp && sq->socketp) {
3984 *sq->socketp = socket;
3985 clock_GetTime(&call->startTime);
3986 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
3990 if (call->flags & RX_CALL_WAIT_PROC) {
3991 /* Conservative: I don't think this should happen */
3992 call->flags &= ~RX_CALL_WAIT_PROC;
3993 MUTEX_ENTER(&rx_stats_mutex);
3995 MUTEX_EXIT(&rx_stats_mutex);
3998 call->state = RX_STATE_ACTIVE;
3999 call->mode = RX_MODE_RECEIVING;
4000 #ifdef RX_KERNEL_TRACE
4002 int glockOwner = ISAFS_GLOCK();
4005 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4006 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4012 if (call->flags & RX_CALL_CLEARED) {
4013 /* send an ack now to start the packet flow up again */
4014 call->flags &= ~RX_CALL_CLEARED;
4015 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4017 #ifdef RX_ENABLE_LOCKS
4020 service->nRequestsRunning++;
4021 if (service->nRequestsRunning <= service->minProcs)
4027 MUTEX_EXIT(&rx_serverPool_lock);
4030 /* Delay the sending of an acknowledge event for a short while, while
4031 * a new call is being prepared (in the case of a client) or a reply
4032 * is being prepared (in the case of a server). Rather than sending
4033 * an ack packet, an ACKALL packet is sent. */
4035 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4037 #ifdef RX_ENABLE_LOCKS
4039 MUTEX_ENTER(&call->lock);
4040 call->delayedAckEvent = NULL;
4041 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4043 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4044 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4046 MUTEX_EXIT(&call->lock);
4047 #else /* RX_ENABLE_LOCKS */
4049 call->delayedAckEvent = NULL;
4050 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4051 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4052 #endif /* RX_ENABLE_LOCKS */
4056 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4059 #ifdef RX_ENABLE_LOCKS
4061 MUTEX_ENTER(&call->lock);
4062 if (event == call->delayedAckEvent)
4063 call->delayedAckEvent = NULL;
4064 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4066 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4068 MUTEX_EXIT(&call->lock);
4069 #else /* RX_ENABLE_LOCKS */
4071 call->delayedAckEvent = NULL;
4072 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4073 #endif /* RX_ENABLE_LOCKS */
4077 #ifdef RX_ENABLE_LOCKS
4078 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4079 * clearing them out.
4082 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4084 register struct rx_packet *p, *tp;
4087 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4090 p->flags |= RX_PKTFLAG_ACKED;
4094 call->flags |= RX_CALL_TQ_CLEARME;
4095 call->flags |= RX_CALL_TQ_SOME_ACKED;
4098 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4099 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4100 call->tfirst = call->tnext;
4101 call->nSoftAcked = 0;
4103 if (call->flags & RX_CALL_FAST_RECOVER) {
4104 call->flags &= ~RX_CALL_FAST_RECOVER;
4105 call->cwind = call->nextCwind;
4106 call->nextCwind = 0;
4109 CV_SIGNAL(&call->cv_twind);
4111 #endif /* RX_ENABLE_LOCKS */
4113 /* Clear out the transmit queue for the current call (all packets have
4114 * been received by peer) */
4116 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4118 register struct rx_packet *p, *tp;
4120 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4121 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4123 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4126 p->flags |= RX_PKTFLAG_ACKED;
4130 call->flags |= RX_CALL_TQ_CLEARME;
4131 call->flags |= RX_CALL_TQ_SOME_ACKED;
4134 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4135 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4141 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4142 call->flags &= ~RX_CALL_TQ_CLEARME;
4144 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4146 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4147 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4148 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4149 call->nSoftAcked = 0;
4151 if (call->flags & RX_CALL_FAST_RECOVER) {
4152 call->flags &= ~RX_CALL_FAST_RECOVER;
4153 call->cwind = call->nextCwind;
4155 #ifdef RX_ENABLE_LOCKS
4156 CV_SIGNAL(&call->cv_twind);
4158 osi_rxWakeup(&call->twind);
4163 rxi_ClearReceiveQueue(register struct rx_call *call)
4165 register struct rx_packet *p, *tp;
4166 if (queue_IsNotEmpty(&call->rq)) {
4167 for (queue_Scan(&call->rq, p, tp, rx_packet)) {
4172 rx_packetReclaims++;
4174 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4176 if (call->state == RX_STATE_PRECALL) {
4177 call->flags |= RX_CALL_CLEARED;
4181 /* Send an abort packet for the specified call */
4183 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4184 int istack, int force)
4192 /* Clients should never delay abort messages */
4193 if (rx_IsClientConn(call->conn))
4196 if (call->abortCode != call->error) {
4197 call->abortCode = call->error;
4198 call->abortCount = 0;
4201 if (force || rxi_callAbortThreshhold == 0
4202 || call->abortCount < rxi_callAbortThreshhold) {
4203 if (call->delayedAbortEvent) {
4204 rxevent_Cancel(call->delayedAbortEvent, call,
4205 RX_CALL_REFCOUNT_ABORT);
4207 error = htonl(call->error);
4210 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4211 (char *)&error, sizeof(error), istack);
4212 } else if (!call->delayedAbortEvent) {
4213 clock_GetTime(&when);
4214 clock_Addmsec(&when, rxi_callAbortDelay);
4215 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4216 call->delayedAbortEvent =
4217 rxevent_Post(&when, rxi_SendDelayedCallAbort, call, 0);
4222 /* Send an abort packet for the specified connection. Packet is an
4223 * optional pointer to a packet that can be used to send the abort.
4224 * Once the number of abort messages reaches the threshhold, an
4225 * event is scheduled to send the abort. Setting the force flag
4226 * overrides sending delayed abort messages.
4228 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4229 * to send the abort packet.
4232 rxi_SendConnectionAbort(register struct rx_connection *conn,
4233 struct rx_packet *packet, int istack, int force)
4241 /* Clients should never delay abort messages */
4242 if (rx_IsClientConn(conn))
4245 if (force || rxi_connAbortThreshhold == 0
4246 || conn->abortCount < rxi_connAbortThreshhold) {
4247 if (conn->delayedAbortEvent) {
4248 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4250 error = htonl(conn->error);
4252 MUTEX_EXIT(&conn->conn_data_lock);
4254 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4255 RX_PACKET_TYPE_ABORT, (char *)&error,
4256 sizeof(error), istack);
4257 MUTEX_ENTER(&conn->conn_data_lock);
4258 } else if (!conn->delayedAbortEvent) {
4259 clock_GetTime(&when);
4260 clock_Addmsec(&when, rxi_connAbortDelay);
4261 conn->delayedAbortEvent =
4262 rxevent_Post(&when, rxi_SendDelayedConnAbort, conn, 0);
4267 /* Associate an error all of the calls owned by a connection. Called
4268 * with error non-zero. This is only for really fatal things, like
4269 * bad authentication responses. The connection itself is set in
4270 * error at this point, so that future packets received will be
4273 rxi_ConnectionError(register struct rx_connection *conn,
4274 register afs_int32 error)
4278 MUTEX_ENTER(&conn->conn_data_lock);
4279 if (conn->challengeEvent)
4280 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4281 if (conn->checkReachEvent) {
4282 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4283 conn->checkReachEvent = 0;
4284 conn->flags &= ~RX_CONN_ATTACHWAIT;
4287 MUTEX_EXIT(&conn->conn_data_lock);
4288 for (i = 0; i < RX_MAXCALLS; i++) {
4289 struct rx_call *call = conn->call[i];
4291 MUTEX_ENTER(&call->lock);
4292 rxi_CallError(call, error);
4293 MUTEX_EXIT(&call->lock);
4296 conn->error = error;
4297 MUTEX_ENTER(&rx_stats_mutex);
4298 rx_stats.fatalErrors++;
4299 MUTEX_EXIT(&rx_stats_mutex);
4304 rxi_CallError(register struct rx_call *call, afs_int32 error)
4307 error = call->error;
4308 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4309 if (!(call->flags & RX_CALL_TQ_BUSY)) {
4310 rxi_ResetCall(call, 0);
4313 rxi_ResetCall(call, 0);
4315 call->error = error;
4316 call->mode = RX_MODE_ERROR;
4319 /* Reset various fields in a call structure, and wakeup waiting
4320 * processes. Some fields aren't changed: state & mode are not
4321 * touched (these must be set by the caller), and bufptr, nLeft, and
4322 * nFree are not reset, since these fields are manipulated by
4323 * unprotected macros, and may only be reset by non-interrupting code.
4326 /* this code requires that call->conn be set properly as a pre-condition. */
4327 #endif /* ADAPT_WINDOW */
4330 rxi_ResetCall(register struct rx_call *call, register int newcall)
4333 register struct rx_peer *peer;
4334 struct rx_packet *packet;
4336 /* Notify anyone who is waiting for asynchronous packet arrival */
4337 if (call->arrivalProc) {
4338 (*call->arrivalProc) (call, call->arrivalProcHandle,
4339 (int)call->arrivalProcArg);
4340 call->arrivalProc = (VOID(*)())0;
4343 if (call->delayedAbortEvent) {
4344 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4345 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4347 rxi_SendCallAbort(call, packet, 0, 1);
4348 rxi_FreePacket(packet);
4353 * Update the peer with the congestion information in this call
4354 * so other calls on this connection can pick up where this call
4355 * left off. If the congestion sequence numbers don't match then
4356 * another call experienced a retransmission.
4358 peer = call->conn->peer;
4359 MUTEX_ENTER(&peer->peer_lock);
4361 if (call->congestSeq == peer->congestSeq) {
4362 peer->cwind = MAX(peer->cwind, call->cwind);
4363 peer->MTU = MAX(peer->MTU, call->MTU);
4364 peer->nDgramPackets =
4365 MAX(peer->nDgramPackets, call->nDgramPackets);
4368 call->abortCode = 0;
4369 call->abortCount = 0;
4371 if (peer->maxDgramPackets > 1) {
4372 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4374 call->MTU = peer->MTU;
4376 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4377 call->ssthresh = rx_maxSendWindow;
4378 call->nDgramPackets = peer->nDgramPackets;
4379 call->congestSeq = peer->congestSeq;
4380 MUTEX_EXIT(&peer->peer_lock);
4382 flags = call->flags;
4383 rxi_ClearReceiveQueue(call);
4384 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4385 if (call->flags & RX_CALL_TQ_BUSY) {
4386 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4387 call->flags |= (flags & RX_CALL_TQ_WAIT);
4389 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4391 rxi_ClearTransmitQueue(call, 0);
4392 queue_Init(&call->tq);
4395 queue_Init(&call->rq);
4397 call->rwind = rx_initReceiveWindow;
4398 call->twind = rx_initSendWindow;
4399 call->nSoftAcked = 0;
4400 call->nextCwind = 0;
4403 call->nCwindAcks = 0;
4404 call->nSoftAcks = 0;
4405 call->nHardAcks = 0;
4407 call->tfirst = call->rnext = call->tnext = 1;
4409 call->lastAcked = 0;
4410 call->localStatus = call->remoteStatus = 0;
4412 if (flags & RX_CALL_READER_WAIT) {
4413 #ifdef RX_ENABLE_LOCKS
4414 CV_BROADCAST(&call->cv_rq);
4416 osi_rxWakeup(&call->rq);
4419 if (flags & RX_CALL_WAIT_PACKETS) {
4420 MUTEX_ENTER(&rx_freePktQ_lock);
4421 rxi_PacketsUnWait(); /* XXX */
4422 MUTEX_EXIT(&rx_freePktQ_lock);
4424 #ifdef RX_ENABLE_LOCKS
4425 CV_SIGNAL(&call->cv_twind);
4427 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4428 osi_rxWakeup(&call->twind);
4431 #ifdef RX_ENABLE_LOCKS
4432 /* The following ensures that we don't mess with any queue while some
4433 * other thread might also be doing so. The call_queue_lock field is
4434 * is only modified under the call lock. If the call is in the process
4435 * of being removed from a queue, the call is not locked until the
4436 * the queue lock is dropped and only then is the call_queue_lock field
4437 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4438 * Note that any other routine which removes a call from a queue has to
4439 * obtain the queue lock before examing the queue and removing the call.
4441 if (call->call_queue_lock) {
4442 MUTEX_ENTER(call->call_queue_lock);
4443 if (queue_IsOnQueue(call)) {
4445 if (flags & RX_CALL_WAIT_PROC) {
4446 MUTEX_ENTER(&rx_stats_mutex);
4448 MUTEX_EXIT(&rx_stats_mutex);
4451 MUTEX_EXIT(call->call_queue_lock);
4452 CLEAR_CALL_QUEUE_LOCK(call);
4454 #else /* RX_ENABLE_LOCKS */
4455 if (queue_IsOnQueue(call)) {
4457 if (flags & RX_CALL_WAIT_PROC)
4460 #endif /* RX_ENABLE_LOCKS */
4462 rxi_KeepAliveOff(call);
4463 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4466 /* Send an acknowledge for the indicated packet (seq,serial) of the
4467 * indicated call, for the indicated reason (reason). This
4468 * acknowledge will specifically acknowledge receiving the packet, and
4469 * will also specify which other packets for this call have been
4470 * received. This routine returns the packet that was used to the
4471 * caller. The caller is responsible for freeing it or re-using it.
4472 * This acknowledgement also returns the highest sequence number
4473 * actually read out by the higher level to the sender; the sender
4474 * promises to keep around packets that have not been read by the
4475 * higher level yet (unless, of course, the sender decides to abort
4476 * the call altogether). Any of p, seq, serial, pflags, or reason may
4477 * be set to zero without ill effect. That is, if they are zero, they
4478 * will not convey any information.
4479 * NOW there is a trailer field, after the ack where it will safely be
4480 * ignored by mundanes, which indicates the maximum size packet this
4481 * host can swallow. */
4483 register struct rx_packet *optionalPacket; use to send ack (or null)
4484 int seq; Sequence number of the packet we are acking
4485 int serial; Serial number of the packet
4486 int pflags; Flags field from packet header
4487 int reason; Reason an acknowledge was prompted
4491 rxi_SendAck(register struct rx_call *call,
4492 register struct rx_packet *optionalPacket, int serial, int reason,
4495 struct rx_ackPacket *ap;
4496 register struct rx_packet *rqp;
4497 register struct rx_packet *nxp; /* For queue_Scan */
4498 register struct rx_packet *p;
4503 * Open the receive window once a thread starts reading packets
4505 if (call->rnext > 1) {
4506 call->rwind = rx_maxReceiveWindow;
4509 call->nHardAcks = 0;
4510 call->nSoftAcks = 0;
4511 if (call->rnext > call->lastAcked)
4512 call->lastAcked = call->rnext;
4516 rx_computelen(p, p->length); /* reset length, you never know */
4517 } /* where that's been... */
4518 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4519 /* We won't send the ack, but don't panic. */
4520 return optionalPacket;
4524 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4527 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL)) {
4528 if (!optionalPacket)
4530 return optionalPacket;
4532 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4533 if (rx_Contiguous(p) < templ) {
4534 if (!optionalPacket)
4536 return optionalPacket;
4541 /* MTUXXX failing to send an ack is very serious. We should */
4542 /* try as hard as possible to send even a partial ack; it's */
4543 /* better than nothing. */
4544 ap = (struct rx_ackPacket *)rx_DataOf(p);
4545 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4546 ap->reason = reason;
4548 /* The skew computation used to be bogus, I think it's better now. */
4549 /* We should start paying attention to skew. XXX */
4550 ap->serial = htonl(serial);
4551 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4553 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4554 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4556 /* No fear of running out of ack packet here because there can only be at most
4557 * one window full of unacknowledged packets. The window size must be constrained
4558 * to be less than the maximum ack size, of course. Also, an ack should always
4559 * fit into a single packet -- it should not ever be fragmented. */
4560 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4561 if (!rqp || !call->rq.next
4562 || (rqp->header.seq > (call->rnext + call->rwind))) {
4563 if (!optionalPacket)
4565 rxi_CallError(call, RX_CALL_DEAD);
4566 return optionalPacket;
4569 while (rqp->header.seq > call->rnext + offset)
4570 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4571 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4573 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4574 if (!optionalPacket)
4576 rxi_CallError(call, RX_CALL_DEAD);
4577 return optionalPacket;
4582 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4584 /* these are new for AFS 3.3 */
4585 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4586 templ = htonl(templ);
4587 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4588 templ = htonl(call->conn->peer->ifMTU);
4589 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4590 sizeof(afs_int32), &templ);
4592 /* new for AFS 3.4 */
4593 templ = htonl(call->rwind);
4594 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4595 sizeof(afs_int32), &templ);
4597 /* new for AFS 3.5 */
4598 templ = htonl(call->conn->peer->ifDgramPackets);
4599 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4600 sizeof(afs_int32), &templ);
4602 p->header.serviceId = call->conn->serviceId;
4603 p->header.cid = (call->conn->cid | call->channel);
4604 p->header.callNumber = *call->callNumber;
4606 p->header.securityIndex = call->conn->securityIndex;
4607 p->header.epoch = call->conn->epoch;
4608 p->header.type = RX_PACKET_TYPE_ACK;
4609 p->header.flags = RX_SLOW_START_OK;
4610 if (reason == RX_ACK_PING) {
4611 p->header.flags |= RX_REQUEST_ACK;
4613 clock_GetTime(&call->pingRequestTime);
4616 if (call->conn->type == RX_CLIENT_CONNECTION)
4617 p->header.flags |= RX_CLIENT_INITIATED;
4621 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u",
4622 ap->reason, ntohl(ap->previousPacket),
4623 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4625 for (offset = 0; offset < ap->nAcks; offset++)
4626 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4634 register int i, nbytes = p->length;
4636 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4637 if (nbytes <= p->wirevec[i].iov_len) {
4638 register int savelen, saven;
4640 savelen = p->wirevec[i].iov_len;
4642 p->wirevec[i].iov_len = nbytes;
4644 rxi_Send(call, p, istack);
4645 p->wirevec[i].iov_len = savelen;
4649 nbytes -= p->wirevec[i].iov_len;
4652 MUTEX_ENTER(&rx_stats_mutex);
4653 rx_stats.ackPacketsSent++;
4654 MUTEX_EXIT(&rx_stats_mutex);
4655 if (!optionalPacket)
4657 return optionalPacket; /* Return packet for re-use by caller */
4660 /* Send all of the packets in the list in single datagram */
4662 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4663 int istack, int moreFlag, struct clock *now,
4664 struct clock *retryTime, int resending)
4669 struct rx_connection *conn = call->conn;
4670 struct rx_peer *peer = conn->peer;
4672 MUTEX_ENTER(&peer->peer_lock);
4675 peer->reSends += len;
4676 MUTEX_ENTER(&rx_stats_mutex);
4677 rx_stats.dataPacketsSent += len;
4678 MUTEX_EXIT(&rx_stats_mutex);
4679 MUTEX_EXIT(&peer->peer_lock);
4681 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4685 /* Set the packet flags and schedule the resend events */
4686 /* Only request an ack for the last packet in the list */
4687 for (i = 0; i < len; i++) {
4688 list[i]->retryTime = *retryTime;
4689 if (list[i]->header.serial) {
4690 /* Exponentially backoff retry times */
4691 if (list[i]->backoff < MAXBACKOFF) {
4692 /* so it can't stay == 0 */
4693 list[i]->backoff = (list[i]->backoff << 1) + 1;
4696 clock_Addmsec(&(list[i]->retryTime),
4697 ((afs_uint32) list[i]->backoff) << 8);
4700 /* Wait a little extra for the ack on the last packet */
4701 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4702 clock_Addmsec(&(list[i]->retryTime), 400);
4705 /* Record the time sent */
4706 list[i]->timeSent = *now;
4708 /* Ask for an ack on retransmitted packets, on every other packet
4709 * if the peer doesn't support slow start. Ask for an ack on every
4710 * packet until the congestion window reaches the ack rate. */
4711 if (list[i]->header.serial) {
4713 MUTEX_ENTER(&rx_stats_mutex);
4714 rx_stats.dataPacketsReSent++;
4715 MUTEX_EXIT(&rx_stats_mutex);
4717 /* improved RTO calculation- not Karn */
4718 list[i]->firstSent = *now;
4719 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4720 || (!(call->flags & RX_CALL_SLOW_START_OK)
4721 && (list[i]->header.seq & 1)))) {
4726 MUTEX_ENTER(&peer->peer_lock);
4730 MUTEX_ENTER(&rx_stats_mutex);
4731 rx_stats.dataPacketsSent++;
4732 MUTEX_EXIT(&rx_stats_mutex);
4733 MUTEX_EXIT(&peer->peer_lock);
4735 /* Tag this packet as not being the last in this group,
4736 * for the receiver's benefit */
4737 if (i < len - 1 || moreFlag) {
4738 list[i]->header.flags |= RX_MORE_PACKETS;
4741 /* Install the new retransmit time for the packet, and
4742 * record the time sent */
4743 list[i]->timeSent = *now;
4747 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4750 /* Since we're about to send a data packet to the peer, it's
4751 * safe to nuke any scheduled end-of-packets ack */
4752 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4754 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4755 MUTEX_EXIT(&call->lock);
4757 rxi_SendPacketList(call, conn, list, len, istack);
4759 rxi_SendPacket(call, conn, list[0], istack);
4761 MUTEX_ENTER(&call->lock);
4762 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4764 /* Update last send time for this call (for keep-alive
4765 * processing), and for the connection (so that we can discover
4766 * idle connections) */
4767 conn->lastSendTime = call->lastSendTime = clock_Sec();
4770 /* When sending packets we need to follow these rules:
4771 * 1. Never send more than maxDgramPackets in a jumbogram.
4772 * 2. Never send a packet with more than two iovecs in a jumbogram.
4773 * 3. Never send a retransmitted packet in a jumbogram.
4774 * 4. Never send more than cwind/4 packets in a jumbogram
4775 * We always keep the last list we should have sent so we
4776 * can set the RX_MORE_PACKETS flags correctly.
4779 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4780 int istack, struct clock *now, struct clock *retryTime,
4783 int i, cnt, lastCnt = 0;
4784 struct rx_packet **listP, **lastP = 0;
4785 struct rx_peer *peer = call->conn->peer;
4786 int morePackets = 0;
4788 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4789 /* Does the current packet force us to flush the current list? */
4791 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4792 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4794 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
4796 /* If the call enters an error state stop sending, or if
4797 * we entered congestion recovery mode, stop sending */
4798 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4806 /* Add the current packet to the list if it hasn't been acked.
4807 * Otherwise adjust the list pointer to skip the current packet. */
4808 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
4810 /* Do we need to flush the list? */
4811 if (cnt >= (int)peer->maxDgramPackets
4812 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
4813 || list[i]->header.serial
4814 || list[i]->length != RX_JUMBOBUFFERSIZE) {
4816 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
4817 retryTime, resending);
4818 /* If the call enters an error state stop sending, or if
4819 * we entered congestion recovery mode, stop sending */
4821 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4826 listP = &list[i + 1];
4831 osi_Panic("rxi_SendList error");
4833 listP = &list[i + 1];
4837 /* Send the whole list when the call is in receive mode, when
4838 * the call is in eof mode, when we are in fast recovery mode,
4839 * and when we have the last packet */
4840 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
4841 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
4842 || (call->flags & RX_CALL_FAST_RECOVER)) {
4843 /* Check for the case where the current list contains
4844 * an acked packet. Since we always send retransmissions
4845 * in a separate packet, we only need to check the first
4846 * packet in the list */
4847 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
4851 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
4852 retryTime, resending);
4853 /* If the call enters an error state stop sending, or if
4854 * we entered congestion recovery mode, stop sending */
4855 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4859 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
4862 } else if (lastCnt > 0) {
4863 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
4868 #ifdef RX_ENABLE_LOCKS
4869 /* Call rxi_Start, below, but with the call lock held. */
4871 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
4874 MUTEX_ENTER(&call->lock);
4875 rxi_Start(event, call, istack);
4876 MUTEX_EXIT(&call->lock);
4878 #endif /* RX_ENABLE_LOCKS */
4880 /* This routine is called when new packets are readied for
4881 * transmission and when retransmission may be necessary, or when the
4882 * transmission window or burst count are favourable. This should be
4883 * better optimized for new packets, the usual case, now that we've
4884 * got rid of queues of send packets. XXXXXXXXXXX */
4886 rxi_Start(struct rxevent *event, register struct rx_call *call, int istack)
4888 struct rx_packet *p;
4889 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
4890 struct rx_peer *peer = call->conn->peer;
4891 struct clock now, retryTime;
4895 struct rx_packet **xmitList;
4898 /* If rxi_Start is being called as a result of a resend event,
4899 * then make sure that the event pointer is removed from the call
4900 * structure, since there is no longer a per-call retransmission
4902 if (event && event == call->resendEvent) {
4903 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
4904 call->resendEvent = NULL;
4906 if (queue_IsEmpty(&call->tq)) {
4910 /* Timeouts trigger congestion recovery */
4911 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4912 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4913 /* someone else is waiting to start recovery */
4916 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4917 while (call->flags & RX_CALL_TQ_BUSY) {
4918 call->flags |= RX_CALL_TQ_WAIT;
4919 #ifdef RX_ENABLE_LOCKS
4920 CV_WAIT(&call->cv_tq, &call->lock);
4921 #else /* RX_ENABLE_LOCKS */
4922 osi_rxSleep(&call->tq);
4923 #endif /* RX_ENABLE_LOCKS */
4925 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4926 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4927 call->flags |= RX_CALL_FAST_RECOVER;
4928 if (peer->maxDgramPackets > 1) {
4929 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4931 call->MTU = MIN(peer->natMTU, peer->maxMTU);
4933 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4934 call->nDgramPackets = 1;
4936 call->nextCwind = 1;
4939 MUTEX_ENTER(&peer->peer_lock);
4940 peer->MTU = call->MTU;
4941 peer->cwind = call->cwind;
4942 peer->nDgramPackets = 1;
4944 call->congestSeq = peer->congestSeq;
4945 MUTEX_EXIT(&peer->peer_lock);
4946 /* Clear retry times on packets. Otherwise, it's possible for
4947 * some packets in the queue to force resends at rates faster
4948 * than recovery rates.
4950 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
4951 if (!(p->flags & RX_PKTFLAG_ACKED)) {
4952 clock_Zero(&p->retryTime);
4957 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4958 MUTEX_ENTER(&rx_stats_mutex);
4959 rx_tq_debug.rxi_start_in_error++;
4960 MUTEX_EXIT(&rx_stats_mutex);
4965 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
4966 /* Get clock to compute the re-transmit time for any packets
4967 * in this burst. Note, if we back off, it's reasonable to
4968 * back off all of the packets in the same manner, even if
4969 * some of them have been retransmitted more times than more
4970 * recent additions */
4971 clock_GetTime(&now);
4972 retryTime = now; /* initialize before use */
4973 MUTEX_ENTER(&peer->peer_lock);
4974 clock_Add(&retryTime, &peer->timeout);
4975 MUTEX_EXIT(&peer->peer_lock);
4977 /* Send (or resend) any packets that need it, subject to
4978 * window restrictions and congestion burst control
4979 * restrictions. Ask for an ack on the last packet sent in
4980 * this burst. For now, we're relying upon the window being
4981 * considerably bigger than the largest number of packets that
4982 * are typically sent at once by one initial call to
4983 * rxi_Start. This is probably bogus (perhaps we should ask
4984 * for an ack when we're half way through the current
4985 * window?). Also, for non file transfer applications, this
4986 * may end up asking for an ack for every packet. Bogus. XXXX
4989 * But check whether we're here recursively, and let the other guy
4992 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4993 if (!(call->flags & RX_CALL_TQ_BUSY)) {
4994 call->flags |= RX_CALL_TQ_BUSY;
4996 call->flags &= ~RX_CALL_NEED_START;
4997 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4999 maxXmitPackets = MIN(call->twind, call->cwind);
5000 xmitList = (struct rx_packet **)
5001 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5002 if (xmitList == NULL)
5003 osi_Panic("rxi_Start, failed to allocate xmit list");
5004 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5005 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5006 /* We shouldn't be sending packets if a thread is waiting
5007 * to initiate congestion recovery */
5011 && (call->flags & RX_CALL_FAST_RECOVER)) {
5012 /* Only send one packet during fast recovery */
5015 if ((p->flags & RX_PKTFLAG_FREE)
5016 || (!queue_IsEnd(&call->tq, nxp)
5017 && (nxp->flags & RX_PKTFLAG_FREE))
5018 || (p == (struct rx_packet *)&rx_freePacketQueue)
5019 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5020 osi_Panic("rxi_Start: xmit queue clobbered");
5022 if (p->flags & RX_PKTFLAG_ACKED) {
5023 MUTEX_ENTER(&rx_stats_mutex);
5024 rx_stats.ignoreAckedPacket++;
5025 MUTEX_EXIT(&rx_stats_mutex);
5026 continue; /* Ignore this packet if it has been acknowledged */
5029 /* Turn off all flags except these ones, which are the same
5030 * on each transmission */
5031 p->header.flags &= RX_PRESET_FLAGS;
5033 if (p->header.seq >=
5034 call->tfirst + MIN((int)call->twind,
5035 (int)(call->nSoftAcked +
5037 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5038 /* Note: if we're waiting for more window space, we can
5039 * still send retransmits; hence we don't return here, but
5040 * break out to schedule a retransmit event */
5041 dpf(("call %d waiting for window",
5042 *(call->callNumber)));
5046 /* Transmit the packet if it needs to be sent. */
5047 if (!clock_Lt(&now, &p->retryTime)) {
5048 if (nXmitPackets == maxXmitPackets) {
5049 osi_Panic("rxi_Start: xmit list overflowed");
5051 xmitList[nXmitPackets++] = p;
5055 /* xmitList now hold pointers to all of the packets that are
5056 * ready to send. Now we loop to send the packets */
5057 if (nXmitPackets > 0) {
5058 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5059 &now, &retryTime, resending);
5062 maxXmitPackets * sizeof(struct rx_packet *));
5064 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5066 * TQ references no longer protected by this flag; they must remain
5067 * protected by the global lock.
5069 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5070 call->flags &= ~RX_CALL_TQ_BUSY;
5071 if (call->flags & RX_CALL_TQ_WAIT) {
5072 call->flags &= ~RX_CALL_TQ_WAIT;
5073 #ifdef RX_ENABLE_LOCKS
5074 CV_BROADCAST(&call->cv_tq);
5075 #else /* RX_ENABLE_LOCKS */
5076 osi_rxWakeup(&call->tq);
5077 #endif /* RX_ENABLE_LOCKS */
5082 /* We went into the error state while sending packets. Now is
5083 * the time to reset the call. This will also inform the using
5084 * process that the call is in an error state.
5086 MUTEX_ENTER(&rx_stats_mutex);
5087 rx_tq_debug.rxi_start_aborted++;
5088 MUTEX_EXIT(&rx_stats_mutex);
5089 call->flags &= ~RX_CALL_TQ_BUSY;
5090 if (call->flags & RX_CALL_TQ_WAIT) {
5091 call->flags &= ~RX_CALL_TQ_WAIT;
5092 #ifdef RX_ENABLE_LOCKS
5093 CV_BROADCAST(&call->cv_tq);
5094 #else /* RX_ENABLE_LOCKS */
5095 osi_rxWakeup(&call->tq);
5096 #endif /* RX_ENABLE_LOCKS */
5098 rxi_CallError(call, call->error);
5101 #ifdef RX_ENABLE_LOCKS
5102 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5103 register int missing;
5104 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5105 /* Some packets have received acks. If they all have, we can clear
5106 * the transmit queue.
5109 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5110 if (p->header.seq < call->tfirst
5111 && (p->flags & RX_PKTFLAG_ACKED)) {
5118 call->flags |= RX_CALL_TQ_CLEARME;
5120 #endif /* RX_ENABLE_LOCKS */
5121 /* Don't bother doing retransmits if the TQ is cleared. */
5122 if (call->flags & RX_CALL_TQ_CLEARME) {
5123 rxi_ClearTransmitQueue(call, 1);
5125 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5128 /* Always post a resend event, if there is anything in the
5129 * queue, and resend is possible. There should be at least
5130 * one unacknowledged packet in the queue ... otherwise none
5131 * of these packets should be on the queue in the first place.
5133 if (call->resendEvent) {
5134 /* Cancel the existing event and post a new one */
5135 rxevent_Cancel(call->resendEvent, call,
5136 RX_CALL_REFCOUNT_RESEND);
5139 /* The retry time is the retry time on the first unacknowledged
5140 * packet inside the current window */
5142 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5143 /* Don't set timers for packets outside the window */
5144 if (p->header.seq >= call->tfirst + call->twind) {
5148 if (!(p->flags & RX_PKTFLAG_ACKED)
5149 && !clock_IsZero(&p->retryTime)) {
5151 retryTime = p->retryTime;
5156 /* Post a new event to re-run rxi_Start when retries may be needed */
5157 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5158 #ifdef RX_ENABLE_LOCKS
5159 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5161 rxevent_Post(&retryTime, rxi_StartUnlocked,
5162 (void *)call, (void *)istack);
5163 #else /* RX_ENABLE_LOCKS */
5165 rxevent_Post(&retryTime, rxi_Start, (void *)call,
5167 #endif /* RX_ENABLE_LOCKS */
5170 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5171 } while (call->flags & RX_CALL_NEED_START);
5173 * TQ references no longer protected by this flag; they must remain
5174 * protected by the global lock.
5176 call->flags &= ~RX_CALL_TQ_BUSY;
5177 if (call->flags & RX_CALL_TQ_WAIT) {
5178 call->flags &= ~RX_CALL_TQ_WAIT;
5179 #ifdef RX_ENABLE_LOCKS
5180 CV_BROADCAST(&call->cv_tq);
5181 #else /* RX_ENABLE_LOCKS */
5182 osi_rxWakeup(&call->tq);
5183 #endif /* RX_ENABLE_LOCKS */
5186 call->flags |= RX_CALL_NEED_START;
5188 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5190 if (call->resendEvent) {
5191 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5196 /* Also adjusts the keep alive parameters for the call, to reflect
5197 * that we have just sent a packet (so keep alives aren't sent
5200 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5203 register struct rx_connection *conn = call->conn;
5205 /* Stamp each packet with the user supplied status */
5206 p->header.userStatus = call->localStatus;
5208 /* Allow the security object controlling this call's security to
5209 * make any last-minute changes to the packet */
5210 RXS_SendPacket(conn->securityObject, call, p);
5212 /* Since we're about to send SOME sort of packet to the peer, it's
5213 * safe to nuke any scheduled end-of-packets ack */
5214 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5216 /* Actually send the packet, filling in more connection-specific fields */
5217 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5218 MUTEX_EXIT(&call->lock);
5219 rxi_SendPacket(call, conn, p, istack);
5220 MUTEX_ENTER(&call->lock);
5221 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5223 /* Update last send time for this call (for keep-alive
5224 * processing), and for the connection (so that we can discover
5225 * idle connections) */
5226 conn->lastSendTime = call->lastSendTime = clock_Sec();
5230 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5231 * that things are fine. Also called periodically to guarantee that nothing
5232 * falls through the cracks (e.g. (error + dally) connections have keepalive
5233 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5235 * haveCTLock Set if calling from rxi_ReapConnections
5237 #ifdef RX_ENABLE_LOCKS
5239 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5240 #else /* RX_ENABLE_LOCKS */
5242 rxi_CheckCall(register struct rx_call *call)
5243 #endif /* RX_ENABLE_LOCKS */
5245 register struct rx_connection *conn = call->conn;
5247 afs_uint32 deadTime;
5249 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5250 if (call->flags & RX_CALL_TQ_BUSY) {
5251 /* Call is active and will be reset by rxi_Start if it's
5252 * in an error state.
5257 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5259 (((afs_uint32) conn->secondsUntilDead << 10) +
5260 ((afs_uint32) conn->peer->rtt >> 3) +
5261 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5263 /* These are computed to the second (+- 1 second). But that's
5264 * good enough for these values, which should be a significant
5265 * number of seconds. */
5266 if (now > (call->lastReceiveTime + deadTime)) {
5267 if (call->state == RX_STATE_ACTIVE) {
5268 rxi_CallError(call, RX_CALL_DEAD);
5271 #ifdef RX_ENABLE_LOCKS
5272 /* Cancel pending events */
5273 rxevent_Cancel(call->delayedAckEvent, call,
5274 RX_CALL_REFCOUNT_DELAY);
5275 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5276 rxevent_Cancel(call->keepAliveEvent, call,
5277 RX_CALL_REFCOUNT_ALIVE);
5278 if (call->refCount == 0) {
5279 rxi_FreeCall(call, haveCTLock);
5283 #else /* RX_ENABLE_LOCKS */
5286 #endif /* RX_ENABLE_LOCKS */
5288 /* Non-active calls are destroyed if they are not responding
5289 * to pings; active calls are simply flagged in error, so the
5290 * attached process can die reasonably gracefully. */
5292 /* see if we have a non-activity timeout */
5293 if (call->startWait && conn->idleDeadTime
5294 && ((call->startWait + conn->idleDeadTime) < now)) {
5295 if (call->state == RX_STATE_ACTIVE) {
5296 rxi_CallError(call, RX_CALL_TIMEOUT);
5300 /* see if we have a hard timeout */
5301 if (conn->hardDeadTime
5302 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5303 if (call->state == RX_STATE_ACTIVE)
5304 rxi_CallError(call, RX_CALL_TIMEOUT);
5311 /* When a call is in progress, this routine is called occasionally to
5312 * make sure that some traffic has arrived (or been sent to) the peer.
5313 * If nothing has arrived in a reasonable amount of time, the call is
5314 * declared dead; if nothing has been sent for a while, we send a
5315 * keep-alive packet (if we're actually trying to keep the call alive)
5318 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5321 struct rx_connection *conn;
5324 MUTEX_ENTER(&call->lock);
5325 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5326 if (event == call->keepAliveEvent)
5327 call->keepAliveEvent = NULL;
5330 #ifdef RX_ENABLE_LOCKS
5331 if (rxi_CheckCall(call, 0)) {
5332 MUTEX_EXIT(&call->lock);
5335 #else /* RX_ENABLE_LOCKS */
5336 if (rxi_CheckCall(call))
5338 #endif /* RX_ENABLE_LOCKS */
5340 /* Don't try to keep alive dallying calls */
5341 if (call->state == RX_STATE_DALLY) {
5342 MUTEX_EXIT(&call->lock);
5347 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5348 /* Don't try to send keepalives if there is unacknowledged data */
5349 /* the rexmit code should be good enough, this little hack
5350 * doesn't quite work XXX */
5351 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5353 rxi_ScheduleKeepAliveEvent(call);
5354 MUTEX_EXIT(&call->lock);
5359 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5361 if (!call->keepAliveEvent) {
5363 clock_GetTime(&when);
5364 when.sec += call->conn->secondsUntilPing;
5365 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5366 call->keepAliveEvent =
5367 rxevent_Post(&when, rxi_KeepAliveEvent, call, 0);
5371 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5373 rxi_KeepAliveOn(register struct rx_call *call)
5375 /* Pretend last packet received was received now--i.e. if another
5376 * packet isn't received within the keep alive time, then the call
5377 * will die; Initialize last send time to the current time--even
5378 * if a packet hasn't been sent yet. This will guarantee that a
5379 * keep-alive is sent within the ping time */
5380 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5381 rxi_ScheduleKeepAliveEvent(call);
5384 /* This routine is called to send connection abort messages
5385 * that have been delayed to throttle looping clients. */
5387 rxi_SendDelayedConnAbort(struct rxevent *event,
5388 register struct rx_connection *conn, char *dummy)
5391 struct rx_packet *packet;
5393 MUTEX_ENTER(&conn->conn_data_lock);
5394 conn->delayedAbortEvent = NULL;
5395 error = htonl(conn->error);
5397 MUTEX_EXIT(&conn->conn_data_lock);
5398 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5401 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5402 RX_PACKET_TYPE_ABORT, (char *)&error,
5404 rxi_FreePacket(packet);
5408 /* This routine is called to send call abort messages
5409 * that have been delayed to throttle looping clients. */
5411 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5415 struct rx_packet *packet;
5417 MUTEX_ENTER(&call->lock);
5418 call->delayedAbortEvent = NULL;
5419 error = htonl(call->error);
5421 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5424 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5425 (char *)&error, sizeof(error), 0);
5426 rxi_FreePacket(packet);
5428 MUTEX_EXIT(&call->lock);
5431 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5432 * seconds) to ask the client to authenticate itself. The routine
5433 * issues a challenge to the client, which is obtained from the
5434 * security object associated with the connection */
5436 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5439 int tries = (int)atries;
5440 conn->challengeEvent = NULL;
5441 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5442 register struct rx_packet *packet;
5446 /* We've failed to authenticate for too long.
5447 * Reset any calls waiting for authentication;
5448 * they are all in RX_STATE_PRECALL.
5452 MUTEX_ENTER(&conn->conn_call_lock);
5453 for (i = 0; i < RX_MAXCALLS; i++) {
5454 struct rx_call *call = conn->call[i];
5456 MUTEX_ENTER(&call->lock);
5457 if (call->state == RX_STATE_PRECALL) {
5458 rxi_CallError(call, RX_CALL_DEAD);
5459 rxi_SendCallAbort(call, NULL, 0, 0);
5461 MUTEX_EXIT(&call->lock);
5464 MUTEX_EXIT(&conn->conn_call_lock);
5468 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5470 /* If there's no packet available, do this later. */
5471 RXS_GetChallenge(conn->securityObject, conn, packet);
5472 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5473 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5474 rxi_FreePacket(packet);
5476 clock_GetTime(&when);
5477 when.sec += RX_CHALLENGE_TIMEOUT;
5478 conn->challengeEvent =
5479 rxevent_Post(&when, rxi_ChallengeEvent, conn,
5480 (void *)(tries - 1));
5484 /* Call this routine to start requesting the client to authenticate
5485 * itself. This will continue until authentication is established,
5486 * the call times out, or an invalid response is returned. The
5487 * security object associated with the connection is asked to create
5488 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5489 * defined earlier. */
5491 rxi_ChallengeOn(register struct rx_connection *conn)
5493 if (!conn->challengeEvent) {
5494 RXS_CreateChallenge(conn->securityObject, conn);
5495 rxi_ChallengeEvent(NULL, conn, (void *)RX_CHALLENGE_MAXTRIES);
5500 /* Compute round trip time of the packet provided, in *rttp.
5503 /* rxi_ComputeRoundTripTime is called with peer locked. */
5504 /* sentp and/or peer may be null */
5506 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5507 register struct clock *sentp,
5508 register struct rx_peer *peer)
5510 struct clock thisRtt, *rttp = &thisRtt;
5512 #if defined(AFS_ALPHA_LINUX22_ENV) && defined(AFS_PTHREAD_ENV) && !defined(KERNEL)
5513 /* making year 2038 bugs to get this running now - stroucki */
5514 struct timeval temptime;
5516 register int rtt_timeout;
5518 #if defined(AFS_ALPHA_LINUX20_ENV) && defined(AFS_PTHREAD_ENV) && !defined(KERNEL)
5519 /* yet again. This was the worst Heisenbug of the port - stroucki */
5520 clock_GetTime(&temptime);
5521 rttp->sec = (afs_int32) temptime.tv_sec;
5522 rttp->usec = (afs_int32) temptime.tv_usec;
5524 clock_GetTime(rttp);
5526 if (clock_Lt(rttp, sentp)) {
5528 return; /* somebody set the clock back, don't count this time. */
5530 clock_Sub(rttp, sentp);
5531 MUTEX_ENTER(&rx_stats_mutex);
5532 if (clock_Lt(rttp, &rx_stats.minRtt))
5533 rx_stats.minRtt = *rttp;
5534 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5535 if (rttp->sec > 60) {
5536 MUTEX_EXIT(&rx_stats_mutex);
5537 return; /* somebody set the clock ahead */
5539 rx_stats.maxRtt = *rttp;
5541 clock_Add(&rx_stats.totalRtt, rttp);
5542 rx_stats.nRttSamples++;
5543 MUTEX_EXIT(&rx_stats_mutex);
5545 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5547 /* Apply VanJacobson round-trip estimations */
5552 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5553 * srtt is stored as fixed point with 3 bits after the binary
5554 * point (i.e., scaled by 8). The following magic is
5555 * equivalent to the smoothing algorithm in rfc793 with an
5556 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5557 * srtt*8 = srtt*8 + rtt - srtt
5558 * srtt = srtt + rtt/8 - srtt/8
5561 delta = MSEC(rttp) - (peer->rtt >> 3);
5565 * We accumulate a smoothed rtt variance (actually, a smoothed
5566 * mean difference), then set the retransmit timer to smoothed
5567 * rtt + 4 times the smoothed variance (was 2x in van's original
5568 * paper, but 4x works better for me, and apparently for him as
5570 * rttvar is stored as
5571 * fixed point with 2 bits after the binary point (scaled by
5572 * 4). The following is equivalent to rfc793 smoothing with
5573 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5574 * replaces rfc793's wired-in beta.
5575 * dev*4 = dev*4 + (|actual - expected| - dev)
5581 delta -= (peer->rtt_dev >> 2);
5582 peer->rtt_dev += delta;
5584 /* I don't have a stored RTT so I start with this value. Since I'm
5585 * probably just starting a call, and will be pushing more data down
5586 * this, I expect congestion to increase rapidly. So I fudge a
5587 * little, and I set deviance to half the rtt. In practice,
5588 * deviance tends to approach something a little less than
5589 * half the smoothed rtt. */
5590 peer->rtt = (MSEC(rttp) << 3) + 8;
5591 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5593 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5594 * the other of these connections is usually in a user process, and can
5595 * be switched and/or swapped out. So on fast, reliable networks, the
5596 * timeout would otherwise be too short.
5598 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5599 clock_Zero(&(peer->timeout));
5600 clock_Addmsec(&(peer->timeout), rtt_timeout);
5602 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)));
5606 /* Find all server connections that have not been active for a long time, and
5609 rxi_ReapConnections(void)
5612 clock_GetTime(&now);
5614 /* Find server connection structures that haven't been used for
5615 * greater than rx_idleConnectionTime */
5617 struct rx_connection **conn_ptr, **conn_end;
5618 int i, havecalls = 0;
5619 MUTEX_ENTER(&rx_connHashTable_lock);
5620 for (conn_ptr = &rx_connHashTable[0], conn_end =
5621 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5623 struct rx_connection *conn, *next;
5624 struct rx_call *call;
5628 for (conn = *conn_ptr; conn; conn = next) {
5629 /* XXX -- Shouldn't the connection be locked? */
5632 for (i = 0; i < RX_MAXCALLS; i++) {
5633 call = conn->call[i];
5636 MUTEX_ENTER(&call->lock);
5637 #ifdef RX_ENABLE_LOCKS
5638 result = rxi_CheckCall(call, 1);
5639 #else /* RX_ENABLE_LOCKS */
5640 result = rxi_CheckCall(call);
5641 #endif /* RX_ENABLE_LOCKS */
5642 MUTEX_EXIT(&call->lock);
5644 /* If CheckCall freed the call, it might
5645 * have destroyed the connection as well,
5646 * which screws up the linked lists.
5652 if (conn->type == RX_SERVER_CONNECTION) {
5653 /* This only actually destroys the connection if
5654 * there are no outstanding calls */
5655 MUTEX_ENTER(&conn->conn_data_lock);
5656 if (!havecalls && !conn->refCount
5657 && ((conn->lastSendTime + rx_idleConnectionTime) <
5659 conn->refCount++; /* it will be decr in rx_DestroyConn */
5660 MUTEX_EXIT(&conn->conn_data_lock);
5661 #ifdef RX_ENABLE_LOCKS
5662 rxi_DestroyConnectionNoLock(conn);
5663 #else /* RX_ENABLE_LOCKS */
5664 rxi_DestroyConnection(conn);
5665 #endif /* RX_ENABLE_LOCKS */
5667 #ifdef RX_ENABLE_LOCKS
5669 MUTEX_EXIT(&conn->conn_data_lock);
5671 #endif /* RX_ENABLE_LOCKS */
5675 #ifdef RX_ENABLE_LOCKS
5676 while (rx_connCleanup_list) {
5677 struct rx_connection *conn;
5678 conn = rx_connCleanup_list;
5679 rx_connCleanup_list = rx_connCleanup_list->next;
5680 MUTEX_EXIT(&rx_connHashTable_lock);
5681 rxi_CleanupConnection(conn);
5682 MUTEX_ENTER(&rx_connHashTable_lock);
5684 MUTEX_EXIT(&rx_connHashTable_lock);
5685 #endif /* RX_ENABLE_LOCKS */
5688 /* Find any peer structures that haven't been used (haven't had an
5689 * associated connection) for greater than rx_idlePeerTime */
5691 struct rx_peer **peer_ptr, **peer_end;
5693 MUTEX_ENTER(&rx_rpc_stats);
5694 MUTEX_ENTER(&rx_peerHashTable_lock);
5695 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5696 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5698 struct rx_peer *peer, *next, *prev;
5699 for (prev = peer = *peer_ptr; peer; peer = next) {
5701 code = MUTEX_TRYENTER(&peer->peer_lock);
5702 if ((code) && (peer->refCount == 0)
5703 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5704 rx_interface_stat_p rpc_stat, nrpc_stat;
5706 MUTEX_EXIT(&peer->peer_lock);
5707 MUTEX_DESTROY(&peer->peer_lock);
5709 (&peer->rpcStats, rpc_stat, nrpc_stat,
5710 rx_interface_stat)) {
5711 unsigned int num_funcs;
5714 queue_Remove(&rpc_stat->queue_header);
5715 queue_Remove(&rpc_stat->all_peers);
5716 num_funcs = rpc_stat->stats[0].func_total;
5718 sizeof(rx_interface_stat_t) +
5719 rpc_stat->stats[0].func_total *
5720 sizeof(rx_function_entry_v1_t);
5722 rxi_Free(rpc_stat, space);
5723 rxi_rpc_peer_stat_cnt -= num_funcs;
5726 MUTEX_ENTER(&rx_stats_mutex);
5727 rx_stats.nPeerStructs--;
5728 MUTEX_EXIT(&rx_stats_mutex);
5729 if (prev == *peer_ptr) {
5736 MUTEX_EXIT(&peer->peer_lock);
5742 MUTEX_EXIT(&rx_peerHashTable_lock);
5743 MUTEX_EXIT(&rx_rpc_stats);
5746 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5747 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5748 * GC, just below. Really, we shouldn't have to keep moving packets from
5749 * one place to another, but instead ought to always know if we can
5750 * afford to hold onto a packet in its particular use. */
5751 MUTEX_ENTER(&rx_freePktQ_lock);
5752 if (rx_waitingForPackets) {
5753 rx_waitingForPackets = 0;
5754 #ifdef RX_ENABLE_LOCKS
5755 CV_BROADCAST(&rx_waitingForPackets_cv);
5757 osi_rxWakeup(&rx_waitingForPackets);
5760 MUTEX_EXIT(&rx_freePktQ_lock);
5762 now.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5763 rxevent_Post(&now, rxi_ReapConnections, 0, 0);
5767 /* rxs_Release - This isn't strictly necessary but, since the macro name from
5768 * rx.h is sort of strange this is better. This is called with a security
5769 * object before it is discarded. Each connection using a security object has
5770 * its own refcount to the object so it won't actually be freed until the last
5771 * connection is destroyed.
5773 * This is the only rxs module call. A hold could also be written but no one
5777 rxs_Release(struct rx_securityClass *aobj)
5779 return RXS_Close(aobj);
5783 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
5784 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
5785 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
5786 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
5788 /* Adjust our estimate of the transmission rate to this peer, given
5789 * that the packet p was just acked. We can adjust peer->timeout and
5790 * call->twind. Pragmatically, this is called
5791 * only with packets of maximal length.
5792 * Called with peer and call locked.
5796 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
5797 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
5799 afs_int32 xferSize, xferMs;
5800 register afs_int32 minTime;
5803 /* Count down packets */
5804 if (peer->rateFlag > 0)
5806 /* Do nothing until we're enabled */
5807 if (peer->rateFlag != 0)
5812 /* Count only when the ack seems legitimate */
5813 switch (ackReason) {
5814 case RX_ACK_REQUESTED:
5816 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
5820 case RX_ACK_PING_RESPONSE:
5821 if (p) /* want the response to ping-request, not data send */
5823 clock_GetTime(&newTO);
5824 if (clock_Gt(&newTO, &call->pingRequestTime)) {
5825 clock_Sub(&newTO, &call->pingRequestTime);
5826 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
5830 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
5837 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));
5839 /* Track only packets that are big enough. */
5840 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
5844 /* absorb RTT data (in milliseconds) for these big packets */
5845 if (peer->smRtt == 0) {
5846 peer->smRtt = xferMs;
5848 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
5853 if (peer->countDown) {
5857 peer->countDown = 10; /* recalculate only every so often */
5859 /* In practice, we can measure only the RTT for full packets,
5860 * because of the way Rx acks the data that it receives. (If it's
5861 * smaller than a full packet, it often gets implicitly acked
5862 * either by the call response (from a server) or by the next call
5863 * (from a client), and either case confuses transmission times
5864 * with processing times.) Therefore, replace the above
5865 * more-sophisticated processing with a simpler version, where the
5866 * smoothed RTT is kept for full-size packets, and the time to
5867 * transmit a windowful of full-size packets is simply RTT *
5868 * windowSize. Again, we take two steps:
5869 - ensure the timeout is large enough for a single packet's RTT;
5870 - ensure that the window is small enough to fit in the desired timeout.*/
5872 /* First, the timeout check. */
5873 minTime = peer->smRtt;
5874 /* Get a reasonable estimate for a timeout period */
5876 newTO.sec = minTime / 1000;
5877 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
5879 /* Increase the timeout period so that we can always do at least
5880 * one packet exchange */
5881 if (clock_Gt(&newTO, &peer->timeout)) {
5883 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));
5885 peer->timeout = newTO;
5888 /* Now, get an estimate for the transmit window size. */
5889 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
5890 /* Now, convert to the number of full packets that could fit in a
5891 * reasonable fraction of that interval */
5892 minTime /= (peer->smRtt << 1);
5893 xferSize = minTime; /* (make a copy) */
5895 /* Now clamp the size to reasonable bounds. */
5898 else if (minTime > rx_Window)
5899 minTime = rx_Window;
5900 /* if (minTime != peer->maxWindow) {
5901 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
5902 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
5903 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
5905 peer->maxWindow = minTime;
5906 elide... call->twind = minTime;
5910 /* Cut back on the peer timeout if it had earlier grown unreasonably.
5911 * Discern this by calculating the timeout necessary for rx_Window
5913 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
5914 /* calculate estimate for transmission interval in milliseconds */
5915 minTime = rx_Window * peer->smRtt;
5916 if (minTime < 1000) {
5917 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
5918 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
5919 peer->timeout.usec, peer->smRtt, peer->packetSize));
5921 newTO.sec = 0; /* cut back on timeout by half a second */
5922 newTO.usec = 500000;
5923 clock_Sub(&peer->timeout, &newTO);
5928 } /* end of rxi_ComputeRate */
5929 #endif /* ADAPT_WINDOW */
5937 /* Don't call this debugging routine directly; use dpf */
5939 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
5940 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
5944 clock_GetTime(&now);
5945 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
5946 (unsigned int)now.usec / 1000);
5947 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
5955 * This function is used to process the rx_stats structure that is local
5956 * to a process as well as an rx_stats structure received from a remote
5957 * process (via rxdebug). Therefore, it needs to do minimal version
5961 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
5962 afs_int32 freePackets, char version)
5966 if (size != sizeof(struct rx_stats)) {
5968 "Unexpected size of stats structure: was %d, expected %d\n",
5969 size, sizeof(struct rx_stats));
5972 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
5975 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
5976 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
5977 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
5978 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
5979 s->specialPktAllocFailures);
5981 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
5982 s->receivePktAllocFailures, s->sendPktAllocFailures,
5983 s->specialPktAllocFailures);
5987 " greedy %d, " "bogusReads %d (last from host %x), "
5988 "noPackets %d, " "noBuffers %d, " "selects %d, "
5989 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
5990 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
5991 s->selects, s->sendSelects);
5993 fprintf(file, " packets read: ");
5994 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
5995 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
5997 fprintf(file, "\n");
6000 " other read counters: data %d, " "ack %d, " "dup %d "
6001 "spurious %d " "dally %d\n", s->dataPacketsRead,
6002 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6003 s->ignorePacketDally);
6005 fprintf(file, " packets sent: ");
6006 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6007 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6009 fprintf(file, "\n");
6012 " other send counters: ack %d, " "data %d (not resends), "
6013 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6014 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6015 s->dataPacketsPushed, s->ignoreAckedPacket);
6018 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6019 s->netSendFailures, (int)s->fatalErrors);
6021 if (s->nRttSamples) {
6022 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6023 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6025 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6026 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6030 " %d server connections, " "%d client connections, "
6031 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6032 s->nServerConns, s->nClientConns, s->nPeerStructs,
6033 s->nCallStructs, s->nFreeCallStructs);
6035 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6036 fprintf(file, " %d clock updates\n", clock_nUpdates);
6041 /* for backward compatibility */
6043 rx_PrintStats(FILE * file)
6045 MUTEX_ENTER(&rx_stats_mutex);
6046 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6048 MUTEX_EXIT(&rx_stats_mutex);
6052 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6054 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6055 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6056 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6059 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6060 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6061 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6064 " Packet size %d, " "max in packet skew %d, "
6065 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6066 (int)peer->outPacketSkew);
6069 #ifdef AFS_PTHREAD_ENV
6071 * This mutex protects the following static variables:
6075 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0);
6076 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0);
6078 #define LOCK_RX_DEBUG
6079 #define UNLOCK_RX_DEBUG
6080 #endif /* AFS_PTHREAD_ENV */
6083 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6084 u_char type, void *inputData, size_t inputLength,
6085 void *outputData, size_t outputLength)
6087 static afs_int32 counter = 100;
6089 struct rx_header theader;
6091 register afs_int32 code;
6093 struct sockaddr_in taddr, faddr;
6098 endTime = time(0) + 20; /* try for 20 seconds */
6099 LOCK_RX_DEBUG counter++;
6100 UNLOCK_RX_DEBUG tp = &tbuffer[sizeof(struct rx_header)];
6101 taddr.sin_family = AF_INET;
6102 taddr.sin_port = remotePort;
6103 taddr.sin_addr.s_addr = remoteAddr;
6104 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6105 taddr.sin_len = sizeof(struct sockaddr_in);
6108 memset(&theader, 0, sizeof(theader));
6109 theader.epoch = htonl(999);
6111 theader.callNumber = htonl(counter);
6114 theader.type = type;
6115 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6116 theader.serviceId = 0;
6118 memcpy(tbuffer, &theader, sizeof(theader));
6119 memcpy(tp, inputData, inputLength);
6121 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6122 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6124 /* see if there's a packet available */
6126 FD_SET(socket, &imask);
6129 code = select(socket + 1, &imask, 0, 0, &tv);
6130 if (code == 1 && FD_ISSET(socket,&imask)) {
6131 /* now receive a packet */
6132 faddrLen = sizeof(struct sockaddr_in);
6134 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6135 (struct sockaddr *)&faddr, &faddrLen);
6138 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6139 if (counter == ntohl(theader.callNumber))
6144 /* see if we've timed out */
6145 if (endTime < time(0))
6148 code -= sizeof(struct rx_header);
6149 if (code > outputLength)
6150 code = outputLength;
6151 memcpy(outputData, tp, code);
6156 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6157 afs_uint16 remotePort, struct rx_debugStats * stat,
6158 afs_uint32 * supportedValues)
6160 struct rx_debugIn in;
6163 *supportedValues = 0;
6164 in.type = htonl(RX_DEBUGI_GETSTATS);
6167 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6168 &in, sizeof(in), stat, sizeof(*stat));
6171 * If the call was successful, fixup the version and indicate
6172 * what contents of the stat structure are valid.
6173 * Also do net to host conversion of fields here.
6177 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6178 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6180 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6181 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6183 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6184 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6186 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6187 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6189 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6190 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6192 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6193 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6195 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6196 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6199 stat->nFreePackets = ntohl(stat->nFreePackets);
6200 stat->packetReclaims = ntohl(stat->packetReclaims);
6201 stat->callsExecuted = ntohl(stat->callsExecuted);
6202 stat->nWaiting = ntohl(stat->nWaiting);
6203 stat->idleThreads = ntohl(stat->idleThreads);
6210 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6211 afs_uint16 remotePort, struct rx_stats * stat,
6212 afs_uint32 * supportedValues)
6214 struct rx_debugIn in;
6215 afs_int32 *lp = (afs_int32 *) stat;
6220 * supportedValues is currently unused, but added to allow future
6221 * versioning of this function.
6224 *supportedValues = 0;
6225 in.type = htonl(RX_DEBUGI_RXSTATS);
6227 memset(stat, 0, sizeof(*stat));
6229 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6230 &in, sizeof(in), stat, sizeof(*stat));
6235 * Do net to host conversion here
6238 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6247 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6248 afs_uint16 remotePort, size_t version_length,
6252 return MakeDebugCall(socket, remoteAddr, remotePort,
6253 RX_PACKET_TYPE_VERSION, a, 1, version,
6258 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6259 afs_uint16 remotePort, afs_int32 * nextConnection,
6260 int allConnections, afs_uint32 debugSupportedValues,
6261 struct rx_debugConn * conn,
6262 afs_uint32 * supportedValues)
6264 struct rx_debugIn in;
6269 * supportedValues is currently unused, but added to allow future
6270 * versioning of this function.
6273 *supportedValues = 0;
6274 if (allConnections) {
6275 in.type = htonl(RX_DEBUGI_GETALLCONN);
6277 in.type = htonl(RX_DEBUGI_GETCONN);
6279 in.index = htonl(*nextConnection);
6280 memset(conn, 0, sizeof(*conn));
6282 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6283 &in, sizeof(in), conn, sizeof(*conn));
6286 *nextConnection += 1;
6289 * Convert old connection format to new structure.
6292 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6293 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6294 #define MOVEvL(a) (conn->a = vL->a)
6296 /* any old or unrecognized version... */
6297 for (i = 0; i < RX_MAXCALLS; i++) {
6298 MOVEvL(callState[i]);
6299 MOVEvL(callMode[i]);
6300 MOVEvL(callFlags[i]);
6301 MOVEvL(callOther[i]);
6303 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6304 MOVEvL(secStats.type);
6305 MOVEvL(secStats.level);
6306 MOVEvL(secStats.flags);
6307 MOVEvL(secStats.expires);
6308 MOVEvL(secStats.packetsReceived);
6309 MOVEvL(secStats.packetsSent);
6310 MOVEvL(secStats.bytesReceived);
6311 MOVEvL(secStats.bytesSent);
6316 * Do net to host conversion here
6318 * I don't convert host or port since we are most likely
6319 * going to want these in NBO.
6321 conn->cid = ntohl(conn->cid);
6322 conn->serial = ntohl(conn->serial);
6323 for (i = 0; i < RX_MAXCALLS; i++) {
6324 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6326 conn->error = ntohl(conn->error);
6327 conn->secStats.flags = ntohl(conn->secStats.flags);
6328 conn->secStats.expires = ntohl(conn->secStats.expires);
6329 conn->secStats.packetsReceived =
6330 ntohl(conn->secStats.packetsReceived);
6331 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6332 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6333 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6334 conn->epoch = ntohl(conn->epoch);
6335 conn->natMTU = ntohl(conn->natMTU);
6342 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6343 afs_uint16 remotePort, afs_int32 * nextPeer,
6344 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6345 afs_uint32 * supportedValues)
6347 struct rx_debugIn in;
6351 * supportedValues is currently unused, but added to allow future
6352 * versioning of this function.
6355 *supportedValues = 0;
6356 in.type = htonl(RX_DEBUGI_GETPEER);
6357 in.index = htonl(*nextPeer);
6358 memset(peer, 0, sizeof(*peer));
6360 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6361 &in, sizeof(in), peer, sizeof(*peer));
6367 * Do net to host conversion here
6369 * I don't convert host or port since we are most likely
6370 * going to want these in NBO.
6372 peer->ifMTU = ntohs(peer->ifMTU);
6373 peer->idleWhen = ntohl(peer->idleWhen);
6374 peer->refCount = ntohs(peer->refCount);
6375 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6376 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6377 peer->rtt = ntohl(peer->rtt);
6378 peer->rtt_dev = ntohl(peer->rtt_dev);
6379 peer->timeout.sec = ntohl(peer->timeout.sec);
6380 peer->timeout.usec = ntohl(peer->timeout.usec);
6381 peer->nSent = ntohl(peer->nSent);
6382 peer->reSends = ntohl(peer->reSends);
6383 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6384 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6385 peer->rateFlag = ntohl(peer->rateFlag);
6386 peer->natMTU = ntohs(peer->natMTU);
6387 peer->maxMTU = ntohs(peer->maxMTU);
6388 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6389 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6390 peer->MTU = ntohs(peer->MTU);
6391 peer->cwind = ntohs(peer->cwind);
6392 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6393 peer->congestSeq = ntohs(peer->congestSeq);
6394 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6395 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6396 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6397 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6402 #endif /* RXDEBUG */
6407 struct rx_serverQueueEntry *np;
6410 register struct rx_call *call;
6411 register struct rx_serverQueueEntry *sq;
6414 LOCK_RX_INIT if (rxinit_status == 1) {
6415 UNLOCK_RX_INIT return; /* Already shutdown. */
6419 #ifndef AFS_PTHREAD_ENV
6420 FD_ZERO(&rx_selectMask);
6421 #endif /* AFS_PTHREAD_ENV */
6422 rxi_dataQuota = RX_MAX_QUOTA;
6423 #ifndef AFS_PTHREAD_ENV
6425 #endif /* AFS_PTHREAD_ENV */
6428 #ifndef AFS_PTHREAD_ENV
6429 #ifndef AFS_USE_GETTIMEOFDAY
6431 #endif /* AFS_USE_GETTIMEOFDAY */
6432 #endif /* AFS_PTHREAD_ENV */
6434 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6435 call = queue_First(&rx_freeCallQueue, rx_call);
6437 rxi_Free(call, sizeof(struct rx_call));
6440 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6441 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6447 struct rx_peer **peer_ptr, **peer_end;
6448 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6449 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6451 struct rx_peer *peer, *next;
6452 for (peer = *peer_ptr; peer; peer = next) {
6453 rx_interface_stat_p rpc_stat, nrpc_stat;
6456 (&peer->rpcStats, rpc_stat, nrpc_stat,
6457 rx_interface_stat)) {
6458 unsigned int num_funcs;
6461 queue_Remove(&rpc_stat->queue_header);
6462 queue_Remove(&rpc_stat->all_peers);
6463 num_funcs = rpc_stat->stats[0].func_total;
6465 sizeof(rx_interface_stat_t) +
6466 rpc_stat->stats[0].func_total *
6467 sizeof(rx_function_entry_v1_t);
6469 rxi_Free(rpc_stat, space);
6470 MUTEX_ENTER(&rx_rpc_stats);
6471 rxi_rpc_peer_stat_cnt -= num_funcs;
6472 MUTEX_EXIT(&rx_rpc_stats);
6476 MUTEX_ENTER(&rx_stats_mutex);
6477 rx_stats.nPeerStructs--;
6478 MUTEX_EXIT(&rx_stats_mutex);
6482 for (i = 0; i < RX_MAX_SERVICES; i++) {
6484 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6486 for (i = 0; i < rx_hashTableSize; i++) {
6487 register struct rx_connection *tc, *ntc;
6488 MUTEX_ENTER(&rx_connHashTable_lock);
6489 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6491 for (j = 0; j < RX_MAXCALLS; j++) {
6493 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6496 rxi_Free(tc, sizeof(*tc));
6498 MUTEX_EXIT(&rx_connHashTable_lock);
6501 MUTEX_ENTER(&freeSQEList_lock);
6503 while ((np = rx_FreeSQEList)) {
6504 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6505 MUTEX_DESTROY(&np->lock);
6506 rxi_Free(np, sizeof(*np));
6509 MUTEX_EXIT(&freeSQEList_lock);
6510 MUTEX_DESTROY(&freeSQEList_lock);
6511 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6512 MUTEX_DESTROY(&rx_connHashTable_lock);
6513 MUTEX_DESTROY(&rx_peerHashTable_lock);
6514 MUTEX_DESTROY(&rx_serverPool_lock);
6516 osi_Free(rx_connHashTable,
6517 rx_hashTableSize * sizeof(struct rx_connection *));
6518 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6520 UNPIN(rx_connHashTable,
6521 rx_hashTableSize * sizeof(struct rx_connection *));
6522 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6524 rxi_FreeAllPackets();
6526 MUTEX_ENTER(&rx_stats_mutex);
6527 rxi_dataQuota = RX_MAX_QUOTA;
6528 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6529 MUTEX_EXIT(&rx_stats_mutex);
6534 #ifdef RX_ENABLE_LOCKS
6536 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6538 if (!MUTEX_ISMINE(lockaddr))
6539 osi_Panic("Lock not held: %s", msg);
6541 #endif /* RX_ENABLE_LOCKS */
6546 * Routines to implement connection specific data.
6550 rx_KeyCreate(rx_destructor_t rtn)
6553 MUTEX_ENTER(&rxi_keyCreate_lock);
6554 key = rxi_keyCreate_counter++;
6555 rxi_keyCreate_destructor = (rx_destructor_t *)
6556 realloc((void *)rxi_keyCreate_destructor,
6557 (key + 1) * sizeof(rx_destructor_t));
6558 rxi_keyCreate_destructor[key] = rtn;
6559 MUTEX_EXIT(&rxi_keyCreate_lock);
6564 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6567 MUTEX_ENTER(&conn->conn_data_lock);
6568 if (!conn->specific) {
6569 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6570 for (i = 0; i < key; i++)
6571 conn->specific[i] = NULL;
6572 conn->nSpecific = key + 1;
6573 conn->specific[key] = ptr;
6574 } else if (key >= conn->nSpecific) {
6575 conn->specific = (void **)
6576 realloc(conn->specific, (key + 1) * sizeof(void *));
6577 for (i = conn->nSpecific; i < key; i++)
6578 conn->specific[i] = NULL;
6579 conn->nSpecific = key + 1;
6580 conn->specific[key] = ptr;
6582 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6583 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6584 conn->specific[key] = ptr;
6586 MUTEX_EXIT(&conn->conn_data_lock);
6590 rx_GetSpecific(struct rx_connection *conn, int key)
6593 MUTEX_ENTER(&conn->conn_data_lock);
6594 if (key >= conn->nSpecific)
6597 ptr = conn->specific[key];
6598 MUTEX_EXIT(&conn->conn_data_lock);
6602 #endif /* !KERNEL */
6605 * processStats is a queue used to store the statistics for the local
6606 * process. Its contents are similar to the contents of the rpcStats
6607 * queue on a rx_peer structure, but the actual data stored within
6608 * this queue contains totals across the lifetime of the process (assuming
6609 * the stats have not been reset) - unlike the per peer structures
6610 * which can come and go based upon the peer lifetime.
6613 static struct rx_queue processStats = { &processStats, &processStats };
6616 * peerStats is a queue used to store the statistics for all peer structs.
6617 * Its contents are the union of all the peer rpcStats queues.
6620 static struct rx_queue peerStats = { &peerStats, &peerStats };
6623 * rxi_monitor_processStats is used to turn process wide stat collection
6627 static int rxi_monitor_processStats = 0;
6630 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6633 static int rxi_monitor_peerStats = 0;
6636 * rxi_AddRpcStat - given all of the information for a particular rpc
6637 * call, create (if needed) and update the stat totals for the rpc.
6641 * IN stats - the queue of stats that will be updated with the new value
6643 * IN rxInterface - a unique number that identifies the rpc interface
6645 * IN currentFunc - the index of the function being invoked
6647 * IN totalFunc - the total number of functions in this interface
6649 * IN queueTime - the amount of time this function waited for a thread
6651 * IN execTime - the amount of time this function invocation took to execute
6653 * IN bytesSent - the number bytes sent by this invocation
6655 * IN bytesRcvd - the number bytes received by this invocation
6657 * IN isServer - if true, this invocation was made to a server
6659 * IN remoteHost - the ip address of the remote host
6661 * IN remotePort - the port of the remote host
6663 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6665 * INOUT counter - if a new stats structure is allocated, the counter will
6666 * be updated with the new number of allocated stat structures
6674 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6675 afs_uint32 currentFunc, afs_uint32 totalFunc,
6676 struct clock *queueTime, struct clock *execTime,
6677 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6678 afs_uint32 remoteHost, afs_uint32 remotePort,
6679 int addToPeerList, unsigned int *counter)
6682 rx_interface_stat_p rpc_stat, nrpc_stat;
6685 * See if there's already a structure for this interface
6688 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6689 if ((rpc_stat->stats[0].interfaceId == rxInterface)
6690 && (rpc_stat->stats[0].remote_is_server == isServer))
6695 * Didn't find a match so allocate a new structure and add it to the
6699 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
6700 || (rpc_stat->stats[0].interfaceId != rxInterface)
6701 || (rpc_stat->stats[0].remote_is_server != isServer)) {
6706 sizeof(rx_interface_stat_t) +
6707 totalFunc * sizeof(rx_function_entry_v1_t);
6709 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
6710 if (rpc_stat == NULL) {
6714 *counter += totalFunc;
6715 for (i = 0; i < totalFunc; i++) {
6716 rpc_stat->stats[i].remote_peer = remoteHost;
6717 rpc_stat->stats[i].remote_port = remotePort;
6718 rpc_stat->stats[i].remote_is_server = isServer;
6719 rpc_stat->stats[i].interfaceId = rxInterface;
6720 rpc_stat->stats[i].func_total = totalFunc;
6721 rpc_stat->stats[i].func_index = i;
6722 hzero(rpc_stat->stats[i].invocations);
6723 hzero(rpc_stat->stats[i].bytes_sent);
6724 hzero(rpc_stat->stats[i].bytes_rcvd);
6725 rpc_stat->stats[i].queue_time_sum.sec = 0;
6726 rpc_stat->stats[i].queue_time_sum.usec = 0;
6727 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
6728 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
6729 rpc_stat->stats[i].queue_time_min.sec = 9999999;
6730 rpc_stat->stats[i].queue_time_min.usec = 9999999;
6731 rpc_stat->stats[i].queue_time_max.sec = 0;
6732 rpc_stat->stats[i].queue_time_max.usec = 0;
6733 rpc_stat->stats[i].execution_time_sum.sec = 0;
6734 rpc_stat->stats[i].execution_time_sum.usec = 0;
6735 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
6736 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
6737 rpc_stat->stats[i].execution_time_min.sec = 9999999;
6738 rpc_stat->stats[i].execution_time_min.usec = 9999999;
6739 rpc_stat->stats[i].execution_time_max.sec = 0;
6740 rpc_stat->stats[i].execution_time_max.usec = 0;
6742 queue_Prepend(stats, rpc_stat);
6743 if (addToPeerList) {
6744 queue_Prepend(&peerStats, &rpc_stat->all_peers);
6749 * Increment the stats for this function
6752 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
6753 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
6754 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
6755 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
6756 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
6757 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
6758 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
6760 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
6761 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
6763 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
6764 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
6766 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
6767 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
6769 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
6770 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
6778 * rx_IncrementTimeAndCount - increment the times and count for a particular
6783 * IN peer - the peer who invoked the rpc
6785 * IN rxInterface - a unique number that identifies the rpc interface
6787 * IN currentFunc - the index of the function being invoked
6789 * IN totalFunc - the total number of functions in this interface
6791 * IN queueTime - the amount of time this function waited for a thread
6793 * IN execTime - the amount of time this function invocation took to execute
6795 * IN bytesSent - the number bytes sent by this invocation
6797 * IN bytesRcvd - the number bytes received by this invocation
6799 * IN isServer - if true, this invocation was made to a server
6807 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
6808 afs_uint32 currentFunc, afs_uint32 totalFunc,
6809 struct clock *queueTime, struct clock *execTime,
6810 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
6814 MUTEX_ENTER(&rx_rpc_stats);
6815 MUTEX_ENTER(&peer->peer_lock);
6817 if (rxi_monitor_peerStats) {
6818 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
6819 queueTime, execTime, bytesSent, bytesRcvd, isServer,
6820 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
6823 if (rxi_monitor_processStats) {
6824 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
6825 queueTime, execTime, bytesSent, bytesRcvd, isServer,
6826 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
6829 MUTEX_EXIT(&peer->peer_lock);
6830 MUTEX_EXIT(&rx_rpc_stats);
6835 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
6839 * IN callerVersion - the rpc stat version of the caller.
6841 * IN count - the number of entries to marshall.
6843 * IN stats - pointer to stats to be marshalled.
6845 * OUT ptr - Where to store the marshalled data.
6852 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
6853 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
6859 * We only support the first version
6861 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
6862 *(ptr++) = stats->remote_peer;
6863 *(ptr++) = stats->remote_port;
6864 *(ptr++) = stats->remote_is_server;
6865 *(ptr++) = stats->interfaceId;
6866 *(ptr++) = stats->func_total;
6867 *(ptr++) = stats->func_index;
6868 *(ptr++) = hgethi(stats->invocations);
6869 *(ptr++) = hgetlo(stats->invocations);
6870 *(ptr++) = hgethi(stats->bytes_sent);
6871 *(ptr++) = hgetlo(stats->bytes_sent);
6872 *(ptr++) = hgethi(stats->bytes_rcvd);
6873 *(ptr++) = hgetlo(stats->bytes_rcvd);
6874 *(ptr++) = stats->queue_time_sum.sec;
6875 *(ptr++) = stats->queue_time_sum.usec;
6876 *(ptr++) = stats->queue_time_sum_sqr.sec;
6877 *(ptr++) = stats->queue_time_sum_sqr.usec;
6878 *(ptr++) = stats->queue_time_min.sec;
6879 *(ptr++) = stats->queue_time_min.usec;
6880 *(ptr++) = stats->queue_time_max.sec;
6881 *(ptr++) = stats->queue_time_max.usec;
6882 *(ptr++) = stats->execution_time_sum.sec;
6883 *(ptr++) = stats->execution_time_sum.usec;
6884 *(ptr++) = stats->execution_time_sum_sqr.sec;
6885 *(ptr++) = stats->execution_time_sum_sqr.usec;
6886 *(ptr++) = stats->execution_time_min.sec;
6887 *(ptr++) = stats->execution_time_min.usec;
6888 *(ptr++) = stats->execution_time_max.sec;
6889 *(ptr++) = stats->execution_time_max.usec;
6895 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
6900 * IN callerVersion - the rpc stat version of the caller
6902 * OUT myVersion - the rpc stat version of this function
6904 * OUT clock_sec - local time seconds
6906 * OUT clock_usec - local time microseconds
6908 * OUT allocSize - the number of bytes allocated to contain stats
6910 * OUT statCount - the number stats retrieved from this process.
6912 * OUT stats - the actual stats retrieved from this process.
6916 * Returns void. If successful, stats will != NULL.
6920 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
6921 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
6922 size_t * allocSize, afs_uint32 * statCount,
6923 afs_uint32 ** stats)
6933 *myVersion = RX_STATS_RETRIEVAL_VERSION;
6936 * Check to see if stats are enabled
6939 MUTEX_ENTER(&rx_rpc_stats);
6940 if (!rxi_monitor_processStats) {
6941 MUTEX_EXIT(&rx_rpc_stats);
6945 clock_GetTime(&now);
6946 *clock_sec = now.sec;
6947 *clock_usec = now.usec;
6950 * Allocate the space based upon the caller version
6952 * If the client is at an older version than we are,
6953 * we return the statistic data in the older data format, but
6954 * we still return our version number so the client knows we
6955 * are maintaining more data than it can retrieve.
6958 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
6959 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
6960 *statCount = rxi_rpc_process_stat_cnt;
6963 * This can't happen yet, but in the future version changes
6964 * can be handled by adding additional code here
6968 if (space > (size_t) 0) {
6970 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
6973 rx_interface_stat_p rpc_stat, nrpc_stat;
6977 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6979 * Copy the data based upon the caller version
6981 rx_MarshallProcessRPCStats(callerVersion,
6982 rpc_stat->stats[0].func_total,
6983 rpc_stat->stats, &ptr);
6989 MUTEX_EXIT(&rx_rpc_stats);
6994 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
6998 * IN callerVersion - the rpc stat version of the caller
7000 * OUT myVersion - the rpc stat version of this function
7002 * OUT clock_sec - local time seconds
7004 * OUT clock_usec - local time microseconds
7006 * OUT allocSize - the number of bytes allocated to contain stats
7008 * OUT statCount - the number of stats retrieved from the individual
7011 * OUT stats - the actual stats retrieved from the individual peer structures.
7015 * Returns void. If successful, stats will != NULL.
7019 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7020 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7021 size_t * allocSize, afs_uint32 * statCount,
7022 afs_uint32 ** stats)
7032 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7035 * Check to see if stats are enabled
7038 MUTEX_ENTER(&rx_rpc_stats);
7039 if (!rxi_monitor_peerStats) {
7040 MUTEX_EXIT(&rx_rpc_stats);
7044 clock_GetTime(&now);
7045 *clock_sec = now.sec;
7046 *clock_usec = now.usec;
7049 * Allocate the space based upon the caller version
7051 * If the client is at an older version than we are,
7052 * we return the statistic data in the older data format, but
7053 * we still return our version number so the client knows we
7054 * are maintaining more data than it can retrieve.
7057 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7058 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7059 *statCount = rxi_rpc_peer_stat_cnt;
7062 * This can't happen yet, but in the future version changes
7063 * can be handled by adding additional code here
7067 if (space > (size_t) 0) {
7069 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7072 rx_interface_stat_p rpc_stat, nrpc_stat;
7076 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7078 * We have to fix the offset of rpc_stat since we are
7079 * keeping this structure on two rx_queues. The rx_queue
7080 * package assumes that the rx_queue member is the first
7081 * member of the structure. That is, rx_queue assumes that
7082 * any one item is only on one queue at a time. We are
7083 * breaking that assumption and so we have to do a little
7084 * math to fix our pointers.
7087 fix_offset = (char *)rpc_stat;
7088 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7089 rpc_stat = (rx_interface_stat_p) fix_offset;
7092 * Copy the data based upon the caller version
7094 rx_MarshallProcessRPCStats(callerVersion,
7095 rpc_stat->stats[0].func_total,
7096 rpc_stat->stats, &ptr);
7102 MUTEX_EXIT(&rx_rpc_stats);
7107 * rx_FreeRPCStats - free memory allocated by
7108 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7112 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7113 * rx_RetrievePeerRPCStats
7115 * IN allocSize - the number of bytes in stats.
7123 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7125 rxi_Free(stats, allocSize);
7129 * rx_queryProcessRPCStats - see if process rpc stat collection is
7130 * currently enabled.
7136 * Returns 0 if stats are not enabled != 0 otherwise
7140 rx_queryProcessRPCStats(void)
7143 MUTEX_ENTER(&rx_rpc_stats);
7144 rc = rxi_monitor_processStats;
7145 MUTEX_EXIT(&rx_rpc_stats);
7150 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7156 * Returns 0 if stats are not enabled != 0 otherwise
7160 rx_queryPeerRPCStats(void)
7163 MUTEX_ENTER(&rx_rpc_stats);
7164 rc = rxi_monitor_peerStats;
7165 MUTEX_EXIT(&rx_rpc_stats);
7170 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7180 rx_enableProcessRPCStats(void)
7182 MUTEX_ENTER(&rx_rpc_stats);
7183 rx_enable_stats = 1;
7184 rxi_monitor_processStats = 1;
7185 MUTEX_EXIT(&rx_rpc_stats);
7189 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7199 rx_enablePeerRPCStats(void)
7201 MUTEX_ENTER(&rx_rpc_stats);
7202 rx_enable_stats = 1;
7203 rxi_monitor_peerStats = 1;
7204 MUTEX_EXIT(&rx_rpc_stats);
7208 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7218 rx_disableProcessRPCStats(void)
7220 rx_interface_stat_p rpc_stat, nrpc_stat;
7223 MUTEX_ENTER(&rx_rpc_stats);
7226 * Turn off process statistics and if peer stats is also off, turn
7230 rxi_monitor_processStats = 0;
7231 if (rxi_monitor_peerStats == 0) {
7232 rx_enable_stats = 0;
7235 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7236 unsigned int num_funcs = 0;
7239 queue_Remove(rpc_stat);
7240 num_funcs = rpc_stat->stats[0].func_total;
7242 sizeof(rx_interface_stat_t) +
7243 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7245 rxi_Free(rpc_stat, space);
7246 rxi_rpc_process_stat_cnt -= num_funcs;
7248 MUTEX_EXIT(&rx_rpc_stats);
7252 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7262 rx_disablePeerRPCStats(void)
7264 struct rx_peer **peer_ptr, **peer_end;
7267 MUTEX_ENTER(&rx_rpc_stats);
7270 * Turn off peer statistics and if process stats is also off, turn
7274 rxi_monitor_peerStats = 0;
7275 if (rxi_monitor_processStats == 0) {
7276 rx_enable_stats = 0;
7279 MUTEX_ENTER(&rx_peerHashTable_lock);
7280 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7281 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7283 struct rx_peer *peer, *next, *prev;
7284 for (prev = peer = *peer_ptr; peer; peer = next) {
7286 code = MUTEX_TRYENTER(&peer->peer_lock);
7288 rx_interface_stat_p rpc_stat, nrpc_stat;
7291 (&peer->rpcStats, rpc_stat, nrpc_stat,
7292 rx_interface_stat)) {
7293 unsigned int num_funcs = 0;
7296 queue_Remove(&rpc_stat->queue_header);
7297 queue_Remove(&rpc_stat->all_peers);
7298 num_funcs = rpc_stat->stats[0].func_total;
7300 sizeof(rx_interface_stat_t) +
7301 rpc_stat->stats[0].func_total *
7302 sizeof(rx_function_entry_v1_t);
7304 rxi_Free(rpc_stat, space);
7305 rxi_rpc_peer_stat_cnt -= num_funcs;
7307 MUTEX_EXIT(&peer->peer_lock);
7308 if (prev == *peer_ptr) {
7318 MUTEX_EXIT(&rx_peerHashTable_lock);
7319 MUTEX_EXIT(&rx_rpc_stats);
7323 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7328 * IN clearFlag - flag indicating which stats to clear
7336 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7338 rx_interface_stat_p rpc_stat, nrpc_stat;
7340 MUTEX_ENTER(&rx_rpc_stats);
7342 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7343 unsigned int num_funcs = 0, i;
7344 num_funcs = rpc_stat->stats[0].func_total;
7345 for (i = 0; i < num_funcs; i++) {
7346 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7347 hzero(rpc_stat->stats[i].invocations);
7349 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7350 hzero(rpc_stat->stats[i].bytes_sent);
7352 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7353 hzero(rpc_stat->stats[i].bytes_rcvd);
7355 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7356 rpc_stat->stats[i].queue_time_sum.sec = 0;
7357 rpc_stat->stats[i].queue_time_sum.usec = 0;
7359 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7360 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7361 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7363 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7364 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7365 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7367 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7368 rpc_stat->stats[i].queue_time_max.sec = 0;
7369 rpc_stat->stats[i].queue_time_max.usec = 0;
7371 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7372 rpc_stat->stats[i].execution_time_sum.sec = 0;
7373 rpc_stat->stats[i].execution_time_sum.usec = 0;
7375 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7376 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7377 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7379 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7380 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7381 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7383 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7384 rpc_stat->stats[i].execution_time_max.sec = 0;
7385 rpc_stat->stats[i].execution_time_max.usec = 0;
7390 MUTEX_EXIT(&rx_rpc_stats);
7394 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7399 * IN clearFlag - flag indicating which stats to clear
7407 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7409 rx_interface_stat_p rpc_stat, nrpc_stat;
7411 MUTEX_ENTER(&rx_rpc_stats);
7413 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7414 unsigned int num_funcs = 0, i;
7417 * We have to fix the offset of rpc_stat since we are
7418 * keeping this structure on two rx_queues. The rx_queue
7419 * package assumes that the rx_queue member is the first
7420 * member of the structure. That is, rx_queue assumes that
7421 * any one item is only on one queue at a time. We are
7422 * breaking that assumption and so we have to do a little
7423 * math to fix our pointers.
7426 fix_offset = (char *)rpc_stat;
7427 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7428 rpc_stat = (rx_interface_stat_p) fix_offset;
7430 num_funcs = rpc_stat->stats[0].func_total;
7431 for (i = 0; i < num_funcs; i++) {
7432 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7433 hzero(rpc_stat->stats[i].invocations);
7435 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7436 hzero(rpc_stat->stats[i].bytes_sent);
7438 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7439 hzero(rpc_stat->stats[i].bytes_rcvd);
7441 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7442 rpc_stat->stats[i].queue_time_sum.sec = 0;
7443 rpc_stat->stats[i].queue_time_sum.usec = 0;
7445 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7446 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7447 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7449 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7450 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7451 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7453 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7454 rpc_stat->stats[i].queue_time_max.sec = 0;
7455 rpc_stat->stats[i].queue_time_max.usec = 0;
7457 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7458 rpc_stat->stats[i].execution_time_sum.sec = 0;
7459 rpc_stat->stats[i].execution_time_sum.usec = 0;
7461 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7462 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7463 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7465 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7466 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7467 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7469 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7470 rpc_stat->stats[i].execution_time_max.sec = 0;
7471 rpc_stat->stats[i].execution_time_max.usec = 0;
7476 MUTEX_EXIT(&rx_rpc_stats);
7480 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7481 * is authorized to enable/disable/clear RX statistics.
7483 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7486 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7488 rxi_rxstat_userok = proc;
7492 rx_RxStatUserOk(struct rx_call *call)
7494 if (!rxi_rxstat_userok)
7496 return rxi_rxstat_userok(call);