2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
14 #include "afs/param.h"
16 #include <afs/param.h>
23 #include "afs/sysincludes.h"
24 #include "afsincludes.h"
30 #include <net/net_globals.h>
31 #endif /* AFS_OSF_ENV */
32 #ifdef AFS_LINUX20_ENV
35 #include "netinet/in.h"
36 #include "afs/afs_args.h"
37 #include "afs/afs_osi.h"
38 #ifdef RX_KERNEL_TRACE
39 #include "rx_kcommon.h"
41 #if (defined(AFS_AUX_ENV) || defined(AFS_AIX_ENV))
45 #undef RXDEBUG /* turn off debugging */
47 #if defined(AFS_SGI_ENV)
48 #include "sys/debug.h"
57 #endif /* AFS_OSF_ENV */
59 #include "afs/sysincludes.h"
60 #include "afsincludes.h"
63 #include "rx_kmutex.h"
64 #include "rx_kernel.h"
68 #include "rx_globals.h"
70 #define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
71 #define AFSOP_STOP_AFS 211 /* Stop AFS process */
72 #define AFSOP_STOP_BKG 212 /* Stop BKG process */
74 extern afs_int32 afs_termState;
76 #include "sys/lockl.h"
77 #include "sys/lock_def.h"
78 #endif /* AFS_AIX41_ENV */
79 # include "rxgen_consts.h"
81 # include <sys/types.h>
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);
220 assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
223 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
224 #define INIT_PTHREAD_LOCKS \
225 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
227 * The rx_stats_mutex mutex protects the following global variables:
232 * rxi_lowConnRefCount
233 * rxi_lowPeerRefCount
242 #define INIT_PTHREAD_LOCKS
246 /* Variables for handling the minProcs implementation. availProcs gives the
247 * number of threads available in the pool at this moment (not counting dudes
248 * executing right now). totalMin gives the total number of procs required
249 * for handling all minProcs requests. minDeficit is a dynamic variable
250 * tracking the # of procs required to satisfy all of the remaining minProcs
252 * For fine grain locking to work, the quota check and the reservation of
253 * a server thread has to come while rxi_availProcs and rxi_minDeficit
254 * are locked. To this end, the code has been modified under #ifdef
255 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
256 * same time. A new function, ReturnToServerPool() returns the allocation.
258 * A call can be on several queue's (but only one at a time). When
259 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
260 * that no one else is touching the queue. To this end, we store the address
261 * of the queue lock in the call structure (under the call lock) when we
262 * put the call on a queue, and we clear the call_queue_lock when the
263 * call is removed from a queue (once the call lock has been obtained).
264 * This allows rxi_ResetCall to safely synchronize with others wishing
265 * to manipulate the queue.
268 #ifdef RX_ENABLE_LOCKS
269 static afs_kmutex_t rx_rpc_stats;
270 void rxi_StartUnlocked();
273 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
274 ** pretty good that the next packet coming in is from the same connection
275 ** as the last packet, since we're send multiple packets in a transmit window.
277 struct rx_connection *rxLastConn = 0;
279 #ifdef RX_ENABLE_LOCKS
280 /* The locking hierarchy for rx fine grain locking is composed of these
283 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
284 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
285 * call->lock - locks call data fields.
286 * These are independent of each other:
287 * rx_freeCallQueue_lock
292 * serverQueueEntry->lock
294 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
295 * peer->lock - locks peer data fields.
296 * conn_data_lock - that more than one thread is not updating a conn data
297 * field at the same time.
305 * Do we need a lock to protect the peer field in the conn structure?
306 * conn->peer was previously a constant for all intents and so has no
307 * lock protecting this field. The multihomed client delta introduced
308 * a RX code change : change the peer field in the connection structure
309 * to that remote inetrface from which the last packet for this
310 * connection was sent out. This may become an issue if further changes
313 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
314 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
316 /* rxdb_fileID is used to identify the lock location, along with line#. */
317 static int rxdb_fileID = RXDB_FILE_RX;
318 #endif /* RX_LOCKS_DB */
319 #else /* RX_ENABLE_LOCKS */
320 #define SET_CALL_QUEUE_LOCK(C, L)
321 #define CLEAR_CALL_QUEUE_LOCK(C)
322 #endif /* RX_ENABLE_LOCKS */
323 struct rx_serverQueueEntry *rx_waitForPacket = 0;
324 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
326 /* ------------Exported Interfaces------------- */
328 /* This function allows rxkad to set the epoch to a suitably random number
329 * which rx_NewConnection will use in the future. The principle purpose is to
330 * get rxnull connections to use the same epoch as the rxkad connections do, at
331 * least once the first rxkad connection is established. This is important now
332 * that the host/port addresses aren't used in FindConnection: the uniqueness
333 * of epoch/cid matters and the start time won't do. */
335 #ifdef AFS_PTHREAD_ENV
337 * This mutex protects the following global variables:
341 #define LOCK_EPOCH assert(pthread_mutex_lock(&epoch_mutex)==0)
342 #define UNLOCK_EPOCH assert(pthread_mutex_unlock(&epoch_mutex)==0)
346 #endif /* AFS_PTHREAD_ENV */
349 rx_SetEpoch(afs_uint32 epoch)
356 /* Initialize rx. A port number may be mentioned, in which case this
357 * becomes the default port number for any service installed later.
358 * If 0 is provided for the port number, a random port will be chosen
359 * by the kernel. Whether this will ever overlap anything in
360 * /etc/services is anybody's guess... Returns 0 on success, -1 on
362 static int rxinit_status = 1;
363 #ifdef AFS_PTHREAD_ENV
365 * This mutex protects the following global variables:
369 #define LOCK_RX_INIT assert(pthread_mutex_lock(&rx_init_mutex)==0)
370 #define UNLOCK_RX_INIT assert(pthread_mutex_unlock(&rx_init_mutex)==0)
373 #define UNLOCK_RX_INIT
377 rx_InitHost(u_int host, u_int port)
384 char *htable, *ptable;
387 #if defined(AFS_DJGPP_ENV) && !defined(DEBUG)
388 __djgpp_set_quiet_socket(1);
395 if (rxinit_status == 0) {
396 tmp_status = rxinit_status;
398 return tmp_status; /* Already started; return previous error code. */
401 if (afs_winsockInit() < 0)
407 * Initialize anything necessary to provide a non-premptive threading
410 rxi_InitializeThreadSupport();
413 /* Allocate and initialize a socket for client and perhaps server
416 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
417 if (rx_socket == OSI_NULLSOCKET) {
421 #ifdef RX_ENABLE_LOCKS
424 #endif /* RX_LOCKS_DB */
425 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
426 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
427 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
428 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
429 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
431 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
433 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
435 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
437 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
439 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
441 #if defined(KERNEL) && defined(AFS_HPUX110_ENV)
443 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
444 #endif /* KERNEL && AFS_HPUX110_ENV */
445 #else /* RX_ENABLE_LOCKS */
446 #if defined(KERNEL) && defined(AFS_GLOBAL_SUNLOCK) && !defined(AFS_HPUX_ENV) && !defined(AFS_OBSD_ENV)
447 mutex_init(&afs_rxglobal_lock, "afs_rxglobal_lock", MUTEX_DEFAULT, NULL);
448 #endif /* AFS_GLOBAL_SUNLOCK */
449 #endif /* RX_ENABLE_LOCKS */
452 rx_connDeadTime = 12;
453 rx_tranquil = 0; /* reset flag */
454 memset((char *)&rx_stats, 0, sizeof(struct rx_stats));
456 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
457 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
458 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
459 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
460 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
461 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
463 /* Malloc up a bunch of packets & buffers */
465 queue_Init(&rx_freePacketQueue);
466 rxi_NeedMorePackets = FALSE;
467 #ifdef RX_ENABLE_TSFPQ
468 rx_nPackets = 0; /* in TSFPQ version, rx_nPackets is managed by rxi_MorePackets* */
469 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
470 #else /* RX_ENABLE_TSFPQ */
471 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
472 rxi_MorePackets(rx_nPackets);
473 #endif /* RX_ENABLE_TSFPQ */
480 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
481 tv.tv_sec = clock_now.sec;
482 tv.tv_usec = clock_now.usec;
483 srand((unsigned int)tv.tv_usec);
490 #if defined(KERNEL) && !defined(UKERNEL)
491 /* Really, this should never happen in a real kernel */
494 struct sockaddr_in addr;
495 int addrlen = sizeof(addr);
496 if (getsockname((int)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
500 rx_port = addr.sin_port;
503 rx_stats.minRtt.sec = 9999999;
505 rx_SetEpoch(tv.tv_sec | 0x80000000);
507 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
508 * will provide a randomer value. */
510 MUTEX_ENTER(&rx_stats_mutex);
511 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
512 MUTEX_EXIT(&rx_stats_mutex);
513 /* *Slightly* random start time for the cid. This is just to help
514 * out with the hashing function at the peer */
515 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
516 rx_connHashTable = (struct rx_connection **)htable;
517 rx_peerHashTable = (struct rx_peer **)ptable;
519 rx_lastAckDelay.sec = 0;
520 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
521 rx_hardAckDelay.sec = 0;
522 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
523 rx_softAckDelay.sec = 0;
524 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
526 rxevent_Init(20, rxi_ReScheduleEvents);
528 /* Initialize various global queues */
529 queue_Init(&rx_idleServerQueue);
530 queue_Init(&rx_incomingCallQueue);
531 queue_Init(&rx_freeCallQueue);
533 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
534 /* Initialize our list of usable IP addresses. */
538 /* Start listener process (exact function is dependent on the
539 * implementation environment--kernel or user space) */
543 tmp_status = rxinit_status = 0;
551 return rx_InitHost(htonl(INADDR_ANY), port);
554 /* called with unincremented nRequestsRunning to see if it is OK to start
555 * a new thread in this service. Could be "no" for two reasons: over the
556 * max quota, or would prevent others from reaching their min quota.
558 #ifdef RX_ENABLE_LOCKS
559 /* This verion of QuotaOK reserves quota if it's ok while the
560 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
563 QuotaOK(register struct rx_service *aservice)
565 /* check if over max quota */
566 if (aservice->nRequestsRunning >= aservice->maxProcs) {
570 /* under min quota, we're OK */
571 /* otherwise, can use only if there are enough to allow everyone
572 * to go to their min quota after this guy starts.
574 MUTEX_ENTER(&rx_stats_mutex);
575 if ((aservice->nRequestsRunning < aservice->minProcs)
576 || (rxi_availProcs > rxi_minDeficit)) {
577 aservice->nRequestsRunning++;
578 /* just started call in minProcs pool, need fewer to maintain
580 if (aservice->nRequestsRunning <= aservice->minProcs)
583 MUTEX_EXIT(&rx_stats_mutex);
586 MUTEX_EXIT(&rx_stats_mutex);
592 ReturnToServerPool(register struct rx_service *aservice)
594 aservice->nRequestsRunning--;
595 MUTEX_ENTER(&rx_stats_mutex);
596 if (aservice->nRequestsRunning < aservice->minProcs)
599 MUTEX_EXIT(&rx_stats_mutex);
602 #else /* RX_ENABLE_LOCKS */
604 QuotaOK(register struct rx_service *aservice)
607 /* under min quota, we're OK */
608 if (aservice->nRequestsRunning < aservice->minProcs)
611 /* check if over max quota */
612 if (aservice->nRequestsRunning >= aservice->maxProcs)
615 /* otherwise, can use only if there are enough to allow everyone
616 * to go to their min quota after this guy starts.
618 if (rxi_availProcs > rxi_minDeficit)
622 #endif /* RX_ENABLE_LOCKS */
625 /* Called by rx_StartServer to start up lwp's to service calls.
626 NExistingProcs gives the number of procs already existing, and which
627 therefore needn't be created. */
629 rxi_StartServerProcs(int nExistingProcs)
631 register struct rx_service *service;
636 /* For each service, reserve N processes, where N is the "minimum"
637 * number of processes that MUST be able to execute a request in parallel,
638 * at any time, for that process. Also compute the maximum difference
639 * between any service's maximum number of processes that can run
640 * (i.e. the maximum number that ever will be run, and a guarantee
641 * that this number will run if other services aren't running), and its
642 * minimum number. The result is the extra number of processes that
643 * we need in order to provide the latter guarantee */
644 for (i = 0; i < RX_MAX_SERVICES; i++) {
646 service = rx_services[i];
647 if (service == (struct rx_service *)0)
649 nProcs += service->minProcs;
650 diff = service->maxProcs - service->minProcs;
654 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
655 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
656 #ifdef RX_ENABLE_TSFPQ
657 rx_TSFPQMaxProcs += nProcs;
658 RX_TS_FPQ_COMPUTE_LIMITS;
659 #endif /* RX_ENABLE_TSFPQ */
660 for (i = 0; i < nProcs; i++) {
661 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
667 /* This routine is only required on Windows */
669 rx_StartClientThread(void)
671 #ifdef AFS_PTHREAD_ENV
673 pid = (int) pthread_self();
674 #endif /* AFS_PTHREAD_ENV */
675 #ifdef RX_ENABLE_TSFPQ
677 RX_TS_FPQ_COMPUTE_LIMITS;
678 #endif /* RX_ENABLE_TSFPQ */
680 #endif /* AFS_NT40_ENV */
682 /* This routine must be called if any services are exported. If the
683 * donateMe flag is set, the calling process is donated to the server
686 rx_StartServer(int donateMe)
688 register struct rx_service *service;
694 /* Start server processes, if necessary (exact function is dependent
695 * on the implementation environment--kernel or user space). DonateMe
696 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
697 * case, one less new proc will be created rx_StartServerProcs.
699 rxi_StartServerProcs(donateMe);
701 /* count up the # of threads in minProcs, and add set the min deficit to
702 * be that value, too.
704 for (i = 0; i < RX_MAX_SERVICES; i++) {
705 service = rx_services[i];
706 if (service == (struct rx_service *)0)
708 MUTEX_ENTER(&rx_stats_mutex);
709 rxi_totalMin += service->minProcs;
710 /* below works even if a thread is running, since minDeficit would
711 * still have been decremented and later re-incremented.
713 rxi_minDeficit += service->minProcs;
714 MUTEX_EXIT(&rx_stats_mutex);
717 /* Turn on reaping of idle server connections */
718 rxi_ReapConnections();
727 #ifdef AFS_PTHREAD_ENV
729 pid = (pid_t) pthread_self();
730 #else /* AFS_PTHREAD_ENV */
732 LWP_CurrentProcess(&pid);
733 #endif /* AFS_PTHREAD_ENV */
735 sprintf(name, "srv_%d", ++nProcs);
737 (*registerProgram) (pid, name);
739 #endif /* AFS_NT40_ENV */
740 rx_ServerProc(); /* Never returns */
745 /* Create a new client connection to the specified service, using the
746 * specified security object to implement the security model for this
748 struct rx_connection *
749 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
750 register struct rx_securityClass *securityObject,
751 int serviceSecurityIndex)
755 register struct rx_connection *conn;
760 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", shost, sport, sservice, securityObject, serviceSecurityIndex));
762 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
763 * the case of kmem_alloc? */
764 conn = rxi_AllocConnection();
765 #ifdef RX_ENABLE_LOCKS
766 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
767 MUTEX_INIT(&conn->conn_data_lock, "conn call lock", MUTEX_DEFAULT, 0);
768 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
771 MUTEX_ENTER(&rx_connHashTable_lock);
772 cid = (rx_nextCid += RX_MAXCALLS);
773 conn->type = RX_CLIENT_CONNECTION;
775 conn->epoch = rx_epoch;
776 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
777 conn->serviceId = sservice;
778 conn->securityObject = securityObject;
779 /* This doesn't work in all compilers with void (they're buggy), so fake it
781 conn->securityData = (VOID *) 0;
782 conn->securityIndex = serviceSecurityIndex;
783 rx_SetConnDeadTime(conn, rx_connDeadTime);
784 conn->ackRate = RX_FAST_ACK_RATE;
786 conn->specific = NULL;
787 conn->challengeEvent = NULL;
788 conn->delayedAbortEvent = NULL;
789 conn->abortCount = 0;
792 RXS_NewConnection(securityObject, conn);
794 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
796 conn->refCount++; /* no lock required since only this thread knows... */
797 conn->next = rx_connHashTable[hashindex];
798 rx_connHashTable[hashindex] = conn;
799 MUTEX_ENTER(&rx_stats_mutex);
800 rx_stats.nClientConns++;
801 MUTEX_EXIT(&rx_stats_mutex);
803 MUTEX_EXIT(&rx_connHashTable_lock);
809 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
811 /* The idea is to set the dead time to a value that allows several
812 * keepalives to be dropped without timing out the connection. */
813 conn->secondsUntilDead = MAX(seconds, 6);
814 conn->secondsUntilPing = conn->secondsUntilDead / 6;
817 int rxi_lowPeerRefCount = 0;
818 int rxi_lowConnRefCount = 0;
821 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
822 * NOTE: must not be called with rx_connHashTable_lock held.
825 rxi_CleanupConnection(struct rx_connection *conn)
827 /* Notify the service exporter, if requested, that this connection
828 * is being destroyed */
829 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
830 (*conn->service->destroyConnProc) (conn);
832 /* Notify the security module that this connection is being destroyed */
833 RXS_DestroyConnection(conn->securityObject, conn);
835 /* If this is the last connection using the rx_peer struct, set its
836 * idle time to now. rxi_ReapConnections will reap it if it's still
837 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
839 MUTEX_ENTER(&rx_peerHashTable_lock);
840 if (conn->peer->refCount < 2) {
841 conn->peer->idleWhen = clock_Sec();
842 if (conn->peer->refCount < 1) {
843 conn->peer->refCount = 1;
844 MUTEX_ENTER(&rx_stats_mutex);
845 rxi_lowPeerRefCount++;
846 MUTEX_EXIT(&rx_stats_mutex);
849 conn->peer->refCount--;
850 MUTEX_EXIT(&rx_peerHashTable_lock);
852 MUTEX_ENTER(&rx_stats_mutex);
853 if (conn->type == RX_SERVER_CONNECTION)
854 rx_stats.nServerConns--;
856 rx_stats.nClientConns--;
857 MUTEX_EXIT(&rx_stats_mutex);
860 if (conn->specific) {
862 for (i = 0; i < conn->nSpecific; i++) {
863 if (conn->specific[i] && rxi_keyCreate_destructor[i])
864 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
865 conn->specific[i] = NULL;
867 free(conn->specific);
869 conn->specific = NULL;
873 MUTEX_DESTROY(&conn->conn_call_lock);
874 MUTEX_DESTROY(&conn->conn_data_lock);
875 CV_DESTROY(&conn->conn_call_cv);
877 rxi_FreeConnection(conn);
880 /* Destroy the specified connection */
882 rxi_DestroyConnection(register struct rx_connection *conn)
884 MUTEX_ENTER(&rx_connHashTable_lock);
885 rxi_DestroyConnectionNoLock(conn);
886 /* conn should be at the head of the cleanup list */
887 if (conn == rx_connCleanup_list) {
888 rx_connCleanup_list = rx_connCleanup_list->next;
889 MUTEX_EXIT(&rx_connHashTable_lock);
890 rxi_CleanupConnection(conn);
892 #ifdef RX_ENABLE_LOCKS
894 MUTEX_EXIT(&rx_connHashTable_lock);
896 #endif /* RX_ENABLE_LOCKS */
900 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
902 register struct rx_connection **conn_ptr;
903 register int havecalls = 0;
904 struct rx_packet *packet;
911 MUTEX_ENTER(&conn->conn_data_lock);
912 if (conn->refCount > 0)
915 MUTEX_ENTER(&rx_stats_mutex);
916 rxi_lowConnRefCount++;
917 MUTEX_EXIT(&rx_stats_mutex);
920 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
921 /* Busy; wait till the last guy before proceeding */
922 MUTEX_EXIT(&conn->conn_data_lock);
927 /* If the client previously called rx_NewCall, but it is still
928 * waiting, treat this as a running call, and wait to destroy the
929 * connection later when the call completes. */
930 if ((conn->type == RX_CLIENT_CONNECTION)
931 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
932 conn->flags |= RX_CONN_DESTROY_ME;
933 MUTEX_EXIT(&conn->conn_data_lock);
937 MUTEX_EXIT(&conn->conn_data_lock);
939 /* Check for extant references to this connection */
940 for (i = 0; i < RX_MAXCALLS; i++) {
941 register struct rx_call *call = conn->call[i];
944 if (conn->type == RX_CLIENT_CONNECTION) {
945 MUTEX_ENTER(&call->lock);
946 if (call->delayedAckEvent) {
947 /* Push the final acknowledgment out now--there
948 * won't be a subsequent call to acknowledge the
949 * last reply packets */
950 rxevent_Cancel(call->delayedAckEvent, call,
951 RX_CALL_REFCOUNT_DELAY);
952 if (call->state == RX_STATE_PRECALL
953 || call->state == RX_STATE_ACTIVE) {
954 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
956 rxi_AckAll(NULL, call, 0);
959 MUTEX_EXIT(&call->lock);
963 #ifdef RX_ENABLE_LOCKS
965 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
966 MUTEX_EXIT(&conn->conn_data_lock);
968 /* Someone is accessing a packet right now. */
972 #endif /* RX_ENABLE_LOCKS */
975 /* Don't destroy the connection if there are any call
976 * structures still in use */
977 MUTEX_ENTER(&conn->conn_data_lock);
978 conn->flags |= RX_CONN_DESTROY_ME;
979 MUTEX_EXIT(&conn->conn_data_lock);
984 if (conn->delayedAbortEvent) {
985 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
986 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
988 MUTEX_ENTER(&conn->conn_data_lock);
989 rxi_SendConnectionAbort(conn, packet, 0, 1);
990 MUTEX_EXIT(&conn->conn_data_lock);
991 rxi_FreePacket(packet);
995 /* Remove from connection hash table before proceeding */
997 &rx_connHashTable[CONN_HASH
998 (peer->host, peer->port, conn->cid, conn->epoch,
1000 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1001 if (*conn_ptr == conn) {
1002 *conn_ptr = conn->next;
1006 /* if the conn that we are destroying was the last connection, then we
1007 * clear rxLastConn as well */
1008 if (rxLastConn == conn)
1011 /* Make sure the connection is completely reset before deleting it. */
1012 /* get rid of pending events that could zap us later */
1013 if (conn->challengeEvent)
1014 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1015 if (conn->checkReachEvent)
1016 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1018 /* Add the connection to the list of destroyed connections that
1019 * need to be cleaned up. This is necessary to avoid deadlocks
1020 * in the routines we call to inform others that this connection is
1021 * being destroyed. */
1022 conn->next = rx_connCleanup_list;
1023 rx_connCleanup_list = conn;
1026 /* Externally available version */
1028 rx_DestroyConnection(register struct rx_connection *conn)
1033 rxi_DestroyConnection(conn);
1038 rx_GetConnection(register struct rx_connection *conn)
1043 MUTEX_ENTER(&conn->conn_data_lock);
1045 MUTEX_EXIT(&conn->conn_data_lock);
1049 /* Start a new rx remote procedure call, on the specified connection.
1050 * If wait is set to 1, wait for a free call channel; otherwise return
1051 * 0. Maxtime gives the maximum number of seconds this call may take,
1052 * after rx_MakeCall returns. After this time interval, a call to any
1053 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1054 * For fine grain locking, we hold the conn_call_lock in order to
1055 * to ensure that we don't get signalle after we found a call in an active
1056 * state and before we go to sleep.
1059 rx_NewCall(register struct rx_connection *conn)
1062 register struct rx_call *call;
1063 struct clock queueTime;
1067 dpf(("rx_MakeCall(conn %x)\n", conn));
1070 clock_GetTime(&queueTime);
1071 MUTEX_ENTER(&conn->conn_call_lock);
1074 * Check if there are others waiting for a new call.
1075 * If so, let them go first to avoid starving them.
1076 * This is a fairly simple scheme, and might not be
1077 * a complete solution for large numbers of waiters.
1079 if (conn->makeCallWaiters) {
1080 #ifdef RX_ENABLE_LOCKS
1081 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1088 for (i = 0; i < RX_MAXCALLS; i++) {
1089 call = conn->call[i];
1091 MUTEX_ENTER(&call->lock);
1092 if (call->state == RX_STATE_DALLY) {
1093 rxi_ResetCall(call, 0);
1094 (*call->callNumber)++;
1097 MUTEX_EXIT(&call->lock);
1099 call = rxi_NewCall(conn, i);
1103 if (i < RX_MAXCALLS) {
1106 MUTEX_ENTER(&conn->conn_data_lock);
1107 conn->flags |= RX_CONN_MAKECALL_WAITING;
1108 MUTEX_EXIT(&conn->conn_data_lock);
1110 conn->makeCallWaiters++;
1111 #ifdef RX_ENABLE_LOCKS
1112 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1116 conn->makeCallWaiters--;
1119 * Wake up anyone else who might be giving us a chance to
1120 * run (see code above that avoids resource starvation).
1122 #ifdef RX_ENABLE_LOCKS
1123 CV_BROADCAST(&conn->conn_call_cv);
1128 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1130 /* Client is initially in send mode */
1131 call->state = RX_STATE_ACTIVE;
1132 call->mode = RX_MODE_SENDING;
1134 /* remember start time for call in case we have hard dead time limit */
1135 call->queueTime = queueTime;
1136 clock_GetTime(&call->startTime);
1137 hzero(call->bytesSent);
1138 hzero(call->bytesRcvd);
1140 /* Turn on busy protocol. */
1141 rxi_KeepAliveOn(call);
1143 MUTEX_EXIT(&call->lock);
1144 MUTEX_EXIT(&conn->conn_call_lock);
1147 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1148 /* Now, if TQ wasn't cleared earlier, do it now. */
1149 MUTEX_ENTER(&call->lock);
1150 while (call->flags & RX_CALL_TQ_BUSY) {
1151 call->flags |= RX_CALL_TQ_WAIT;
1152 #ifdef RX_ENABLE_LOCKS
1153 CV_WAIT(&call->cv_tq, &call->lock);
1154 #else /* RX_ENABLE_LOCKS */
1155 osi_rxSleep(&call->tq);
1156 #endif /* RX_ENABLE_LOCKS */
1158 if (call->flags & RX_CALL_TQ_CLEARME) {
1159 rxi_ClearTransmitQueue(call, 0);
1160 queue_Init(&call->tq);
1162 MUTEX_EXIT(&call->lock);
1163 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1169 rxi_HasActiveCalls(register struct rx_connection *aconn)
1172 register struct rx_call *tcall;
1176 for (i = 0; i < RX_MAXCALLS; i++) {
1177 if ((tcall = aconn->call[i])) {
1178 if ((tcall->state == RX_STATE_ACTIVE)
1179 || (tcall->state == RX_STATE_PRECALL)) {
1190 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1191 register afs_int32 * aint32s)
1194 register struct rx_call *tcall;
1198 for (i = 0; i < RX_MAXCALLS; i++) {
1199 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1200 aint32s[i] = aconn->callNumber[i] + 1;
1202 aint32s[i] = aconn->callNumber[i];
1209 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1210 register afs_int32 * aint32s)
1213 register struct rx_call *tcall;
1217 for (i = 0; i < RX_MAXCALLS; i++) {
1218 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1219 aconn->callNumber[i] = aint32s[i] - 1;
1221 aconn->callNumber[i] = aint32s[i];
1227 /* Advertise a new service. A service is named locally by a UDP port
1228 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1231 char *serviceName; Name for identification purposes (e.g. the
1232 service name might be used for probing for
1235 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1236 struct rx_securityClass **securityObjects, int nSecurityObjects,
1237 afs_int32(*serviceProc) (struct rx_call * acall))
1239 osi_socket socket = OSI_NULLSOCKET;
1240 register struct rx_service *tservice;
1246 if (serviceId == 0) {
1248 "rx_NewService: service id for service %s is not non-zero.\n",
1255 "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",
1263 tservice = rxi_AllocService();
1265 for (i = 0; i < RX_MAX_SERVICES; i++) {
1266 register struct rx_service *service = rx_services[i];
1268 if (port == service->servicePort) {
1269 if (service->serviceId == serviceId) {
1270 /* The identical service has already been
1271 * installed; if the caller was intending to
1272 * change the security classes used by this
1273 * service, he/she loses. */
1275 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1276 serviceName, serviceId, service->serviceName);
1278 rxi_FreeService(tservice);
1281 /* Different service, same port: re-use the socket
1282 * which is bound to the same port */
1283 socket = service->socket;
1286 if (socket == OSI_NULLSOCKET) {
1287 /* If we don't already have a socket (from another
1288 * service on same port) get a new one */
1289 socket = rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
1290 if (socket == OSI_NULLSOCKET) {
1292 rxi_FreeService(tservice);
1297 service->socket = socket;
1298 service->servicePort = port;
1299 service->serviceId = serviceId;
1300 service->serviceName = serviceName;
1301 service->nSecurityObjects = nSecurityObjects;
1302 service->securityObjects = securityObjects;
1303 service->minProcs = 0;
1304 service->maxProcs = 1;
1305 service->idleDeadTime = 60;
1306 service->connDeadTime = rx_connDeadTime;
1307 service->executeRequestProc = serviceProc;
1308 service->checkReach = 0;
1309 rx_services[i] = service; /* not visible until now */
1315 rxi_FreeService(tservice);
1316 (osi_Msg "rx_NewService: cannot support > %d services\n",
1321 /* Generic request processing loop. This routine should be called
1322 * by the implementation dependent rx_ServerProc. If socketp is
1323 * non-null, it will be set to the file descriptor that this thread
1324 * is now listening on. If socketp is null, this routine will never
1327 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1329 register struct rx_call *call;
1330 register afs_int32 code;
1331 register struct rx_service *tservice = NULL;
1338 call = rx_GetCall(threadID, tservice, socketp);
1339 if (socketp && *socketp != OSI_NULLSOCKET) {
1340 /* We are now a listener thread */
1345 /* if server is restarting( typically smooth shutdown) then do not
1346 * allow any new calls.
1349 if (rx_tranquil && (call != NULL)) {
1353 MUTEX_ENTER(&call->lock);
1355 rxi_CallError(call, RX_RESTARTING);
1356 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1358 MUTEX_EXIT(&call->lock);
1362 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1363 #ifdef RX_ENABLE_LOCKS
1365 #endif /* RX_ENABLE_LOCKS */
1366 afs_termState = AFSOP_STOP_AFS;
1367 afs_osi_Wakeup(&afs_termState);
1368 #ifdef RX_ENABLE_LOCKS
1370 #endif /* RX_ENABLE_LOCKS */
1375 tservice = call->conn->service;
1377 if (tservice->beforeProc)
1378 (*tservice->beforeProc) (call);
1380 code = call->conn->service->executeRequestProc(call);
1382 if (tservice->afterProc)
1383 (*tservice->afterProc) (call, code);
1385 rx_EndCall(call, code);
1386 MUTEX_ENTER(&rx_stats_mutex);
1388 MUTEX_EXIT(&rx_stats_mutex);
1394 rx_WakeupServerProcs(void)
1396 struct rx_serverQueueEntry *np, *tqp;
1400 MUTEX_ENTER(&rx_serverPool_lock);
1402 #ifdef RX_ENABLE_LOCKS
1403 if (rx_waitForPacket)
1404 CV_BROADCAST(&rx_waitForPacket->cv);
1405 #else /* RX_ENABLE_LOCKS */
1406 if (rx_waitForPacket)
1407 osi_rxWakeup(rx_waitForPacket);
1408 #endif /* RX_ENABLE_LOCKS */
1409 MUTEX_ENTER(&freeSQEList_lock);
1410 for (np = rx_FreeSQEList; np; np = tqp) {
1411 tqp = *(struct rx_serverQueueEntry **)np;
1412 #ifdef RX_ENABLE_LOCKS
1413 CV_BROADCAST(&np->cv);
1414 #else /* RX_ENABLE_LOCKS */
1416 #endif /* RX_ENABLE_LOCKS */
1418 MUTEX_EXIT(&freeSQEList_lock);
1419 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1420 #ifdef RX_ENABLE_LOCKS
1421 CV_BROADCAST(&np->cv);
1422 #else /* RX_ENABLE_LOCKS */
1424 #endif /* RX_ENABLE_LOCKS */
1426 MUTEX_EXIT(&rx_serverPool_lock);
1431 * One thing that seems to happen is that all the server threads get
1432 * tied up on some empty or slow call, and then a whole bunch of calls
1433 * arrive at once, using up the packet pool, so now there are more
1434 * empty calls. The most critical resources here are server threads
1435 * and the free packet pool. The "doreclaim" code seems to help in
1436 * general. I think that eventually we arrive in this state: there
1437 * are lots of pending calls which do have all their packets present,
1438 * so they won't be reclaimed, are multi-packet calls, so they won't
1439 * be scheduled until later, and thus are tying up most of the free
1440 * packet pool for a very long time.
1442 * 1. schedule multi-packet calls if all the packets are present.
1443 * Probably CPU-bound operation, useful to return packets to pool.
1444 * Do what if there is a full window, but the last packet isn't here?
1445 * 3. preserve one thread which *only* runs "best" calls, otherwise
1446 * it sleeps and waits for that type of call.
1447 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1448 * the current dataquota business is badly broken. The quota isn't adjusted
1449 * to reflect how many packets are presently queued for a running call.
1450 * So, when we schedule a queued call with a full window of packets queued
1451 * up for it, that *should* free up a window full of packets for other 2d-class
1452 * calls to be able to use from the packet pool. But it doesn't.
1454 * NB. Most of the time, this code doesn't run -- since idle server threads
1455 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1456 * as a new call arrives.
1458 /* Sleep until a call arrives. Returns a pointer to the call, ready
1459 * for an rx_Read. */
1460 #ifdef RX_ENABLE_LOCKS
1462 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1464 struct rx_serverQueueEntry *sq;
1465 register struct rx_call *call = (struct rx_call *)0;
1466 struct rx_service *service = NULL;
1469 MUTEX_ENTER(&freeSQEList_lock);
1471 if ((sq = rx_FreeSQEList)) {
1472 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1473 MUTEX_EXIT(&freeSQEList_lock);
1474 } else { /* otherwise allocate a new one and return that */
1475 MUTEX_EXIT(&freeSQEList_lock);
1476 sq = (struct rx_serverQueueEntry *)
1477 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1478 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1479 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1482 MUTEX_ENTER(&rx_serverPool_lock);
1483 if (cur_service != NULL) {
1484 ReturnToServerPool(cur_service);
1487 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1488 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1490 /* Scan for eligible incoming calls. A call is not eligible
1491 * if the maximum number of calls for its service type are
1492 * already executing */
1493 /* One thread will process calls FCFS (to prevent starvation),
1494 * while the other threads may run ahead looking for calls which
1495 * have all their input data available immediately. This helps
1496 * keep threads from blocking, waiting for data from the client. */
1497 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1498 service = tcall->conn->service;
1499 if (!QuotaOK(service)) {
1502 if (tno == rxi_fcfs_thread_num
1503 || !tcall->queue_item_header.next) {
1504 /* If we're the fcfs thread , then we'll just use
1505 * this call. If we haven't been able to find an optimal
1506 * choice, and we're at the end of the list, then use a
1507 * 2d choice if one has been identified. Otherwise... */
1508 call = (choice2 ? choice2 : tcall);
1509 service = call->conn->service;
1510 } else if (!queue_IsEmpty(&tcall->rq)) {
1511 struct rx_packet *rp;
1512 rp = queue_First(&tcall->rq, rx_packet);
1513 if (rp->header.seq == 1) {
1515 || (rp->header.flags & RX_LAST_PACKET)) {
1517 } else if (rxi_2dchoice && !choice2
1518 && !(tcall->flags & RX_CALL_CLEARED)
1519 && (tcall->rprev > rxi_HardAckRate)) {
1528 ReturnToServerPool(service);
1535 MUTEX_EXIT(&rx_serverPool_lock);
1536 MUTEX_ENTER(&call->lock);
1538 if (call->flags & RX_CALL_WAIT_PROC) {
1539 call->flags &= ~RX_CALL_WAIT_PROC;
1540 MUTEX_ENTER(&rx_stats_mutex);
1542 MUTEX_EXIT(&rx_stats_mutex);
1545 if (call->state != RX_STATE_PRECALL || call->error) {
1546 MUTEX_EXIT(&call->lock);
1547 MUTEX_ENTER(&rx_serverPool_lock);
1548 ReturnToServerPool(service);
1553 if (queue_IsEmpty(&call->rq)
1554 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1555 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1557 CLEAR_CALL_QUEUE_LOCK(call);
1560 /* If there are no eligible incoming calls, add this process
1561 * to the idle server queue, to wait for one */
1565 *socketp = OSI_NULLSOCKET;
1567 sq->socketp = socketp;
1568 queue_Append(&rx_idleServerQueue, sq);
1569 #ifndef AFS_AIX41_ENV
1570 rx_waitForPacket = sq;
1572 rx_waitingForPacket = sq;
1573 #endif /* AFS_AIX41_ENV */
1575 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1577 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1578 MUTEX_EXIT(&rx_serverPool_lock);
1579 return (struct rx_call *)0;
1582 } while (!(call = sq->newcall)
1583 && !(socketp && *socketp != OSI_NULLSOCKET));
1584 MUTEX_EXIT(&rx_serverPool_lock);
1586 MUTEX_ENTER(&call->lock);
1592 MUTEX_ENTER(&freeSQEList_lock);
1593 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1594 rx_FreeSQEList = sq;
1595 MUTEX_EXIT(&freeSQEList_lock);
1598 clock_GetTime(&call->startTime);
1599 call->state = RX_STATE_ACTIVE;
1600 call->mode = RX_MODE_RECEIVING;
1601 #ifdef RX_KERNEL_TRACE
1602 if (ICL_SETACTIVE(afs_iclSetp)) {
1603 int glockOwner = ISAFS_GLOCK();
1606 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1607 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1614 rxi_calltrace(RX_CALL_START, call);
1615 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1616 call->conn->service->servicePort, call->conn->service->serviceId,
1619 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1620 MUTEX_EXIT(&call->lock);
1622 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1627 #else /* RX_ENABLE_LOCKS */
1629 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1631 struct rx_serverQueueEntry *sq;
1632 register struct rx_call *call = (struct rx_call *)0, *choice2;
1633 struct rx_service *service = NULL;
1637 MUTEX_ENTER(&freeSQEList_lock);
1639 if ((sq = rx_FreeSQEList)) {
1640 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1641 MUTEX_EXIT(&freeSQEList_lock);
1642 } else { /* otherwise allocate a new one and return that */
1643 MUTEX_EXIT(&freeSQEList_lock);
1644 sq = (struct rx_serverQueueEntry *)
1645 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1646 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1647 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1649 MUTEX_ENTER(&sq->lock);
1651 if (cur_service != NULL) {
1652 cur_service->nRequestsRunning--;
1653 if (cur_service->nRequestsRunning < cur_service->minProcs)
1657 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1658 register struct rx_call *tcall, *ncall;
1659 /* Scan for eligible incoming calls. A call is not eligible
1660 * if the maximum number of calls for its service type are
1661 * already executing */
1662 /* One thread will process calls FCFS (to prevent starvation),
1663 * while the other threads may run ahead looking for calls which
1664 * have all their input data available immediately. This helps
1665 * keep threads from blocking, waiting for data from the client. */
1666 choice2 = (struct rx_call *)0;
1667 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1668 service = tcall->conn->service;
1669 if (QuotaOK(service)) {
1670 if (tno == rxi_fcfs_thread_num
1671 || !tcall->queue_item_header.next) {
1672 /* If we're the fcfs thread, then we'll just use
1673 * this call. If we haven't been able to find an optimal
1674 * choice, and we're at the end of the list, then use a
1675 * 2d choice if one has been identified. Otherwise... */
1676 call = (choice2 ? choice2 : tcall);
1677 service = call->conn->service;
1678 } else if (!queue_IsEmpty(&tcall->rq)) {
1679 struct rx_packet *rp;
1680 rp = queue_First(&tcall->rq, rx_packet);
1681 if (rp->header.seq == 1
1683 || (rp->header.flags & RX_LAST_PACKET))) {
1685 } else if (rxi_2dchoice && !choice2
1686 && !(tcall->flags & RX_CALL_CLEARED)
1687 && (tcall->rprev > rxi_HardAckRate)) {
1700 /* we can't schedule a call if there's no data!!! */
1701 /* send an ack if there's no data, if we're missing the
1702 * first packet, or we're missing something between first
1703 * and last -- there's a "hole" in the incoming data. */
1704 if (queue_IsEmpty(&call->rq)
1705 || queue_First(&call->rq, rx_packet)->header.seq != 1
1706 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1707 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1709 call->flags &= (~RX_CALL_WAIT_PROC);
1710 service->nRequestsRunning++;
1711 /* just started call in minProcs pool, need fewer to maintain
1713 if (service->nRequestsRunning <= service->minProcs)
1717 /* MUTEX_EXIT(&call->lock); */
1719 /* If there are no eligible incoming calls, add this process
1720 * to the idle server queue, to wait for one */
1723 *socketp = OSI_NULLSOCKET;
1725 sq->socketp = socketp;
1726 queue_Append(&rx_idleServerQueue, sq);
1730 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1732 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1733 return (struct rx_call *)0;
1736 } while (!(call = sq->newcall)
1737 && !(socketp && *socketp != OSI_NULLSOCKET));
1739 MUTEX_EXIT(&sq->lock);
1741 MUTEX_ENTER(&freeSQEList_lock);
1742 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1743 rx_FreeSQEList = sq;
1744 MUTEX_EXIT(&freeSQEList_lock);
1747 clock_GetTime(&call->startTime);
1748 call->state = RX_STATE_ACTIVE;
1749 call->mode = RX_MODE_RECEIVING;
1750 #ifdef RX_KERNEL_TRACE
1751 if (ICL_SETACTIVE(afs_iclSetp)) {
1752 int glockOwner = ISAFS_GLOCK();
1755 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1756 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1763 rxi_calltrace(RX_CALL_START, call);
1764 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1765 call->conn->service->servicePort, call->conn->service->serviceId,
1768 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1775 #endif /* RX_ENABLE_LOCKS */
1779 /* Establish a procedure to be called when a packet arrives for a
1780 * call. This routine will be called at most once after each call,
1781 * and will also be called if there is an error condition on the or
1782 * the call is complete. Used by multi rx to build a selection
1783 * function which determines which of several calls is likely to be a
1784 * good one to read from.
1785 * NOTE: the way this is currently implemented it is probably only a
1786 * good idea to (1) use it immediately after a newcall (clients only)
1787 * and (2) only use it once. Other uses currently void your warranty
1790 rx_SetArrivalProc(register struct rx_call *call,
1791 register void (*proc) (register struct rx_call * call,
1793 register int index),
1794 register VOID * handle, register int arg)
1796 call->arrivalProc = proc;
1797 call->arrivalProcHandle = handle;
1798 call->arrivalProcArg = arg;
1801 /* Call is finished (possibly prematurely). Return rc to the peer, if
1802 * appropriate, and return the final error code from the conversation
1806 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1808 register struct rx_connection *conn = call->conn;
1809 register struct rx_service *service;
1810 register struct rx_packet *tp; /* Temporary packet pointer */
1811 register struct rx_packet *nxp; /* Next packet pointer, for queue_Scan */
1815 dpf(("rx_EndCall(call %x)\n", call));
1818 MUTEX_ENTER(&call->lock);
1820 if (rc == 0 && call->error == 0) {
1821 call->abortCode = 0;
1822 call->abortCount = 0;
1825 call->arrivalProc = (void (*)())0;
1826 if (rc && call->error == 0) {
1827 rxi_CallError(call, rc);
1828 /* Send an abort message to the peer if this error code has
1829 * only just been set. If it was set previously, assume the
1830 * peer has already been sent the error code or will request it
1832 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1834 if (conn->type == RX_SERVER_CONNECTION) {
1835 /* Make sure reply or at least dummy reply is sent */
1836 if (call->mode == RX_MODE_RECEIVING) {
1837 rxi_WriteProc(call, 0, 0);
1839 if (call->mode == RX_MODE_SENDING) {
1840 rxi_FlushWrite(call);
1842 service = conn->service;
1843 rxi_calltrace(RX_CALL_END, call);
1844 /* Call goes to hold state until reply packets are acknowledged */
1845 if (call->tfirst + call->nSoftAcked < call->tnext) {
1846 call->state = RX_STATE_HOLD;
1848 call->state = RX_STATE_DALLY;
1849 rxi_ClearTransmitQueue(call, 0);
1850 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1851 rxevent_Cancel(call->keepAliveEvent, call,
1852 RX_CALL_REFCOUNT_ALIVE);
1854 } else { /* Client connection */
1856 /* Make sure server receives input packets, in the case where
1857 * no reply arguments are expected */
1858 if ((call->mode == RX_MODE_SENDING)
1859 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1860 (void)rxi_ReadProc(call, &dummy, 1);
1863 /* If we had an outstanding delayed ack, be nice to the server
1864 * and force-send it now.
1866 if (call->delayedAckEvent) {
1867 rxevent_Cancel(call->delayedAckEvent, call,
1868 RX_CALL_REFCOUNT_DELAY);
1869 call->delayedAckEvent = NULL;
1870 rxi_SendDelayedAck(NULL, call, NULL);
1873 /* We need to release the call lock since it's lower than the
1874 * conn_call_lock and we don't want to hold the conn_call_lock
1875 * over the rx_ReadProc call. The conn_call_lock needs to be held
1876 * here for the case where rx_NewCall is perusing the calls on
1877 * the connection structure. We don't want to signal until
1878 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1879 * have checked this call, found it active and by the time it
1880 * goes to sleep, will have missed the signal.
1882 MUTEX_EXIT(&call->lock);
1883 MUTEX_ENTER(&conn->conn_call_lock);
1884 MUTEX_ENTER(&call->lock);
1885 MUTEX_ENTER(&conn->conn_data_lock);
1886 conn->flags |= RX_CONN_BUSY;
1887 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1888 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1889 MUTEX_EXIT(&conn->conn_data_lock);
1890 #ifdef RX_ENABLE_LOCKS
1891 CV_BROADCAST(&conn->conn_call_cv);
1896 #ifdef RX_ENABLE_LOCKS
1898 MUTEX_EXIT(&conn->conn_data_lock);
1900 #endif /* RX_ENABLE_LOCKS */
1901 call->state = RX_STATE_DALLY;
1903 error = call->error;
1905 /* currentPacket, nLeft, and NFree must be zeroed here, because
1906 * ResetCall cannot: ResetCall may be called at splnet(), in the
1907 * kernel version, and may interrupt the macros rx_Read or
1908 * rx_Write, which run at normal priority for efficiency. */
1909 if (call->currentPacket) {
1910 rxi_FreePacket(call->currentPacket);
1911 call->currentPacket = (struct rx_packet *)0;
1912 call->nLeft = call->nFree = call->curlen = 0;
1914 call->nLeft = call->nFree = call->curlen = 0;
1916 /* Free any packets from the last call to ReadvProc/WritevProc */
1917 for (queue_Scan(&call->iovq, tp, nxp, rx_packet)) {
1922 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1923 MUTEX_EXIT(&call->lock);
1924 if (conn->type == RX_CLIENT_CONNECTION) {
1925 MUTEX_EXIT(&conn->conn_call_lock);
1926 conn->flags &= ~RX_CONN_BUSY;
1930 * Map errors to the local host's errno.h format.
1932 error = ntoh_syserr_conv(error);
1936 #if !defined(KERNEL)
1938 /* Call this routine when shutting down a server or client (especially
1939 * clients). This will allow Rx to gracefully garbage collect server
1940 * connections, and reduce the number of retries that a server might
1941 * make to a dead client.
1942 * This is not quite right, since some calls may still be ongoing and
1943 * we can't lock them to destroy them. */
1947 register struct rx_connection **conn_ptr, **conn_end;
1951 if (rxinit_status == 1) {
1953 return; /* Already shutdown. */
1955 rxi_DeleteCachedConnections();
1956 if (rx_connHashTable) {
1957 MUTEX_ENTER(&rx_connHashTable_lock);
1958 for (conn_ptr = &rx_connHashTable[0], conn_end =
1959 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
1961 struct rx_connection *conn, *next;
1962 for (conn = *conn_ptr; conn; conn = next) {
1964 if (conn->type == RX_CLIENT_CONNECTION) {
1965 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
1967 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
1968 #ifdef RX_ENABLE_LOCKS
1969 rxi_DestroyConnectionNoLock(conn);
1970 #else /* RX_ENABLE_LOCKS */
1971 rxi_DestroyConnection(conn);
1972 #endif /* RX_ENABLE_LOCKS */
1976 #ifdef RX_ENABLE_LOCKS
1977 while (rx_connCleanup_list) {
1978 struct rx_connection *conn;
1979 conn = rx_connCleanup_list;
1980 rx_connCleanup_list = rx_connCleanup_list->next;
1981 MUTEX_EXIT(&rx_connHashTable_lock);
1982 rxi_CleanupConnection(conn);
1983 MUTEX_ENTER(&rx_connHashTable_lock);
1985 MUTEX_EXIT(&rx_connHashTable_lock);
1986 #endif /* RX_ENABLE_LOCKS */
1995 /* if we wakeup packet waiter too often, can get in loop with two
1996 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
1998 rxi_PacketsUnWait(void)
2000 if (!rx_waitingForPackets) {
2004 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2005 return; /* still over quota */
2008 rx_waitingForPackets = 0;
2009 #ifdef RX_ENABLE_LOCKS
2010 CV_BROADCAST(&rx_waitingForPackets_cv);
2012 osi_rxWakeup(&rx_waitingForPackets);
2018 /* ------------------Internal interfaces------------------------- */
2020 /* Return this process's service structure for the
2021 * specified socket and service */
2023 rxi_FindService(register osi_socket socket, register u_short serviceId)
2025 register struct rx_service **sp;
2026 for (sp = &rx_services[0]; *sp; sp++) {
2027 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2033 /* Allocate a call structure, for the indicated channel of the
2034 * supplied connection. The mode and state of the call must be set by
2035 * the caller. Returns the call with mutex locked. */
2037 rxi_NewCall(register struct rx_connection *conn, register int channel)
2039 register struct rx_call *call;
2040 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2041 register struct rx_call *cp; /* Call pointer temp */
2042 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2043 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2045 /* Grab an existing call structure, or allocate a new one.
2046 * Existing call structures are assumed to have been left reset by
2048 MUTEX_ENTER(&rx_freeCallQueue_lock);
2050 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2052 * EXCEPT that the TQ might not yet be cleared out.
2053 * Skip over those with in-use TQs.
2056 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2057 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2063 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2064 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2065 call = queue_First(&rx_freeCallQueue, rx_call);
2066 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2068 MUTEX_ENTER(&rx_stats_mutex);
2069 rx_stats.nFreeCallStructs--;
2070 MUTEX_EXIT(&rx_stats_mutex);
2071 MUTEX_EXIT(&rx_freeCallQueue_lock);
2072 MUTEX_ENTER(&call->lock);
2073 CLEAR_CALL_QUEUE_LOCK(call);
2074 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2075 /* Now, if TQ wasn't cleared earlier, do it now. */
2076 if (call->flags & RX_CALL_TQ_CLEARME) {
2077 rxi_ClearTransmitQueue(call, 0);
2078 queue_Init(&call->tq);
2080 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2081 /* Bind the call to its connection structure */
2083 rxi_ResetCall(call, 1);
2085 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2087 MUTEX_EXIT(&rx_freeCallQueue_lock);
2088 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2089 MUTEX_ENTER(&call->lock);
2090 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2091 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2092 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2094 MUTEX_ENTER(&rx_stats_mutex);
2095 rx_stats.nCallStructs++;
2096 MUTEX_EXIT(&rx_stats_mutex);
2097 /* Initialize once-only items */
2098 queue_Init(&call->tq);
2099 queue_Init(&call->rq);
2100 queue_Init(&call->iovq);
2101 /* Bind the call to its connection structure (prereq for reset) */
2103 rxi_ResetCall(call, 1);
2105 call->channel = channel;
2106 call->callNumber = &conn->callNumber[channel];
2107 /* Note that the next expected call number is retained (in
2108 * conn->callNumber[i]), even if we reallocate the call structure
2110 conn->call[channel] = call;
2111 /* if the channel's never been used (== 0), we should start at 1, otherwise
2112 * the call number is valid from the last time this channel was used */
2113 if (*call->callNumber == 0)
2114 *call->callNumber = 1;
2119 /* A call has been inactive long enough that so we can throw away
2120 * state, including the call structure, which is placed on the call
2122 * Call is locked upon entry.
2123 * haveCTLock set if called from rxi_ReapConnections
2125 #ifdef RX_ENABLE_LOCKS
2127 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2128 #else /* RX_ENABLE_LOCKS */
2130 rxi_FreeCall(register struct rx_call *call)
2131 #endif /* RX_ENABLE_LOCKS */
2133 register int channel = call->channel;
2134 register struct rx_connection *conn = call->conn;
2137 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2138 (*call->callNumber)++;
2139 rxi_ResetCall(call, 0);
2140 call->conn->call[channel] = (struct rx_call *)0;
2142 MUTEX_ENTER(&rx_freeCallQueue_lock);
2143 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2144 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2145 /* A call may be free even though its transmit queue is still in use.
2146 * Since we search the call list from head to tail, put busy calls at
2147 * the head of the list, and idle calls at the tail.
2149 if (call->flags & RX_CALL_TQ_BUSY)
2150 queue_Prepend(&rx_freeCallQueue, call);
2152 queue_Append(&rx_freeCallQueue, call);
2153 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2154 queue_Append(&rx_freeCallQueue, call);
2155 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2156 MUTEX_ENTER(&rx_stats_mutex);
2157 rx_stats.nFreeCallStructs++;
2158 MUTEX_EXIT(&rx_stats_mutex);
2160 MUTEX_EXIT(&rx_freeCallQueue_lock);
2162 /* Destroy the connection if it was previously slated for
2163 * destruction, i.e. the Rx client code previously called
2164 * rx_DestroyConnection (client connections), or
2165 * rxi_ReapConnections called the same routine (server
2166 * connections). Only do this, however, if there are no
2167 * outstanding calls. Note that for fine grain locking, there appears
2168 * to be a deadlock in that rxi_FreeCall has a call locked and
2169 * DestroyConnectionNoLock locks each call in the conn. But note a
2170 * few lines up where we have removed this call from the conn.
2171 * If someone else destroys a connection, they either have no
2172 * call lock held or are going through this section of code.
2174 if (conn->flags & RX_CONN_DESTROY_ME) {
2175 MUTEX_ENTER(&conn->conn_data_lock);
2177 MUTEX_EXIT(&conn->conn_data_lock);
2178 #ifdef RX_ENABLE_LOCKS
2180 rxi_DestroyConnectionNoLock(conn);
2182 rxi_DestroyConnection(conn);
2183 #else /* RX_ENABLE_LOCKS */
2184 rxi_DestroyConnection(conn);
2185 #endif /* RX_ENABLE_LOCKS */
2189 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2191 rxi_Alloc(register size_t size)
2195 MUTEX_ENTER(&rx_stats_mutex);
2197 rxi_Allocsize += size;
2198 MUTEX_EXIT(&rx_stats_mutex);
2200 p = (char *)osi_Alloc(size);
2203 osi_Panic("rxi_Alloc error");
2209 rxi_Free(void *addr, register size_t size)
2211 MUTEX_ENTER(&rx_stats_mutex);
2213 rxi_Allocsize -= size;
2214 MUTEX_EXIT(&rx_stats_mutex);
2216 osi_Free(addr, size);
2219 /* Find the peer process represented by the supplied (host,port)
2220 * combination. If there is no appropriate active peer structure, a
2221 * new one will be allocated and initialized
2222 * The origPeer, if set, is a pointer to a peer structure on which the
2223 * refcount will be be decremented. This is used to replace the peer
2224 * structure hanging off a connection structure */
2226 rxi_FindPeer(register afs_uint32 host, register u_short port,
2227 struct rx_peer *origPeer, int create)
2229 register struct rx_peer *pp;
2231 hashIndex = PEER_HASH(host, port);
2232 MUTEX_ENTER(&rx_peerHashTable_lock);
2233 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2234 if ((pp->host == host) && (pp->port == port))
2239 pp = rxi_AllocPeer(); /* This bzero's *pp */
2240 pp->host = host; /* set here or in InitPeerParams is zero */
2242 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2243 queue_Init(&pp->congestionQueue);
2244 queue_Init(&pp->rpcStats);
2245 pp->next = rx_peerHashTable[hashIndex];
2246 rx_peerHashTable[hashIndex] = pp;
2247 rxi_InitPeerParams(pp);
2248 MUTEX_ENTER(&rx_stats_mutex);
2249 rx_stats.nPeerStructs++;
2250 MUTEX_EXIT(&rx_stats_mutex);
2257 origPeer->refCount--;
2258 MUTEX_EXIT(&rx_peerHashTable_lock);
2263 /* Find the connection at (host, port) started at epoch, and with the
2264 * given connection id. Creates the server connection if necessary.
2265 * The type specifies whether a client connection or a server
2266 * connection is desired. In both cases, (host, port) specify the
2267 * peer's (host, pair) pair. Client connections are not made
2268 * automatically by this routine. The parameter socket gives the
2269 * socket descriptor on which the packet was received. This is used,
2270 * in the case of server connections, to check that *new* connections
2271 * come via a valid (port, serviceId). Finally, the securityIndex
2272 * parameter must match the existing index for the connection. If a
2273 * server connection is created, it will be created using the supplied
2274 * index, if the index is valid for this service */
2275 struct rx_connection *
2276 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2277 register u_short port, u_short serviceId, afs_uint32 cid,
2278 afs_uint32 epoch, int type, u_int securityIndex)
2280 int hashindex, flag;
2281 register struct rx_connection *conn;
2282 hashindex = CONN_HASH(host, port, cid, epoch, type);
2283 MUTEX_ENTER(&rx_connHashTable_lock);
2284 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2285 rx_connHashTable[hashindex],
2288 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2289 && (epoch == conn->epoch)) {
2290 register struct rx_peer *pp = conn->peer;
2291 if (securityIndex != conn->securityIndex) {
2292 /* this isn't supposed to happen, but someone could forge a packet
2293 * like this, and there seems to be some CM bug that makes this
2294 * happen from time to time -- in which case, the fileserver
2296 MUTEX_EXIT(&rx_connHashTable_lock);
2297 return (struct rx_connection *)0;
2299 if (pp->host == host && pp->port == port)
2301 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2303 /* So what happens when it's a callback connection? */
2304 if ( /*type == RX_CLIENT_CONNECTION && */
2305 (conn->epoch & 0x80000000))
2309 /* the connection rxLastConn that was used the last time is not the
2310 ** one we are looking for now. Hence, start searching in the hash */
2312 conn = rx_connHashTable[hashindex];
2317 struct rx_service *service;
2318 if (type == RX_CLIENT_CONNECTION) {
2319 MUTEX_EXIT(&rx_connHashTable_lock);
2320 return (struct rx_connection *)0;
2322 service = rxi_FindService(socket, serviceId);
2323 if (!service || (securityIndex >= service->nSecurityObjects)
2324 || (service->securityObjects[securityIndex] == 0)) {
2325 MUTEX_EXIT(&rx_connHashTable_lock);
2326 return (struct rx_connection *)0;
2328 conn = rxi_AllocConnection(); /* This bzero's the connection */
2329 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2330 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2331 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2332 conn->next = rx_connHashTable[hashindex];
2333 rx_connHashTable[hashindex] = conn;
2334 conn->peer = rxi_FindPeer(host, port, 0, 1);
2335 conn->type = RX_SERVER_CONNECTION;
2336 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2337 conn->epoch = epoch;
2338 conn->cid = cid & RX_CIDMASK;
2339 /* conn->serial = conn->lastSerial = 0; */
2340 /* conn->timeout = 0; */
2341 conn->ackRate = RX_FAST_ACK_RATE;
2342 conn->service = service;
2343 conn->serviceId = serviceId;
2344 conn->securityIndex = securityIndex;
2345 conn->securityObject = service->securityObjects[securityIndex];
2346 conn->nSpecific = 0;
2347 conn->specific = NULL;
2348 rx_SetConnDeadTime(conn, service->connDeadTime);
2349 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2350 /* Notify security object of the new connection */
2351 RXS_NewConnection(conn->securityObject, conn);
2352 /* XXXX Connection timeout? */
2353 if (service->newConnProc)
2354 (*service->newConnProc) (conn);
2355 MUTEX_ENTER(&rx_stats_mutex);
2356 rx_stats.nServerConns++;
2357 MUTEX_EXIT(&rx_stats_mutex);
2360 MUTEX_ENTER(&conn->conn_data_lock);
2362 MUTEX_EXIT(&conn->conn_data_lock);
2364 rxLastConn = conn; /* store this connection as the last conn used */
2365 MUTEX_EXIT(&rx_connHashTable_lock);
2369 /* There are two packet tracing routines available for testing and monitoring
2370 * Rx. One is called just after every packet is received and the other is
2371 * called just before every packet is sent. Received packets, have had their
2372 * headers decoded, and packets to be sent have not yet had their headers
2373 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2374 * containing the network address. Both can be modified. The return value, if
2375 * non-zero, indicates that the packet should be dropped. */
2377 int (*rx_justReceived) () = 0;
2378 int (*rx_almostSent) () = 0;
2380 /* A packet has been received off the interface. Np is the packet, socket is
2381 * the socket number it was received from (useful in determining which service
2382 * this packet corresponds to), and (host, port) reflect the host,port of the
2383 * sender. This call returns the packet to the caller if it is finished with
2384 * it, rather than de-allocating it, just as a small performance hack */
2387 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2388 afs_uint32 host, u_short port, int *tnop,
2389 struct rx_call **newcallp)
2391 register struct rx_call *call;
2392 register struct rx_connection *conn;
2394 afs_uint32 currentCallNumber;
2400 struct rx_packet *tnp;
2403 /* We don't print out the packet until now because (1) the time may not be
2404 * accurate enough until now in the lwp implementation (rx_Listener only gets
2405 * the time after the packet is read) and (2) from a protocol point of view,
2406 * this is the first time the packet has been seen */
2407 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2408 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2409 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2410 np->header.serial, packetType, host, port, np->header.serviceId,
2411 np->header.epoch, np->header.cid, np->header.callNumber,
2412 np->header.seq, np->header.flags, np));
2415 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2416 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2419 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2420 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2423 /* If an input tracer function is defined, call it with the packet and
2424 * network address. Note this function may modify its arguments. */
2425 if (rx_justReceived) {
2426 struct sockaddr_in addr;
2428 addr.sin_family = AF_INET;
2429 addr.sin_port = port;
2430 addr.sin_addr.s_addr = host;
2431 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2432 addr.sin_len = sizeof(addr);
2433 #endif /* AFS_OSF_ENV */
2434 drop = (*rx_justReceived) (np, &addr);
2435 /* drop packet if return value is non-zero */
2438 port = addr.sin_port; /* in case fcn changed addr */
2439 host = addr.sin_addr.s_addr;
2443 /* If packet was not sent by the client, then *we* must be the client */
2444 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2445 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2447 /* Find the connection (or fabricate one, if we're the server & if
2448 * necessary) associated with this packet */
2450 rxi_FindConnection(socket, host, port, np->header.serviceId,
2451 np->header.cid, np->header.epoch, type,
2452 np->header.securityIndex);
2455 /* If no connection found or fabricated, just ignore the packet.
2456 * (An argument could be made for sending an abort packet for
2461 MUTEX_ENTER(&conn->conn_data_lock);
2462 if (conn->maxSerial < np->header.serial)
2463 conn->maxSerial = np->header.serial;
2464 MUTEX_EXIT(&conn->conn_data_lock);
2466 /* If the connection is in an error state, send an abort packet and ignore
2467 * the incoming packet */
2469 /* Don't respond to an abort packet--we don't want loops! */
2470 MUTEX_ENTER(&conn->conn_data_lock);
2471 if (np->header.type != RX_PACKET_TYPE_ABORT)
2472 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2474 MUTEX_EXIT(&conn->conn_data_lock);
2478 /* Check for connection-only requests (i.e. not call specific). */
2479 if (np->header.callNumber == 0) {
2480 switch (np->header.type) {
2481 case RX_PACKET_TYPE_ABORT:
2482 /* What if the supplied error is zero? */
2483 rxi_ConnectionError(conn, ntohl(rx_GetInt32(np, 0)));
2484 MUTEX_ENTER(&conn->conn_data_lock);
2486 MUTEX_EXIT(&conn->conn_data_lock);
2488 case RX_PACKET_TYPE_CHALLENGE:
2489 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2490 MUTEX_ENTER(&conn->conn_data_lock);
2492 MUTEX_EXIT(&conn->conn_data_lock);
2494 case RX_PACKET_TYPE_RESPONSE:
2495 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2496 MUTEX_ENTER(&conn->conn_data_lock);
2498 MUTEX_EXIT(&conn->conn_data_lock);
2500 case RX_PACKET_TYPE_PARAMS:
2501 case RX_PACKET_TYPE_PARAMS + 1:
2502 case RX_PACKET_TYPE_PARAMS + 2:
2503 /* ignore these packet types for now */
2504 MUTEX_ENTER(&conn->conn_data_lock);
2506 MUTEX_EXIT(&conn->conn_data_lock);
2511 /* Should not reach here, unless the peer is broken: send an
2513 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2514 MUTEX_ENTER(&conn->conn_data_lock);
2515 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2517 MUTEX_EXIT(&conn->conn_data_lock);
2522 channel = np->header.cid & RX_CHANNELMASK;
2523 call = conn->call[channel];
2524 #ifdef RX_ENABLE_LOCKS
2526 MUTEX_ENTER(&call->lock);
2527 /* Test to see if call struct is still attached to conn. */
2528 if (call != conn->call[channel]) {
2530 MUTEX_EXIT(&call->lock);
2531 if (type == RX_SERVER_CONNECTION) {
2532 call = conn->call[channel];
2533 /* If we started with no call attached and there is one now,
2534 * another thread is also running this routine and has gotten
2535 * the connection channel. We should drop this packet in the tests
2536 * below. If there was a call on this connection and it's now
2537 * gone, then we'll be making a new call below.
2538 * If there was previously a call and it's now different then
2539 * the old call was freed and another thread running this routine
2540 * has created a call on this channel. One of these two threads
2541 * has a packet for the old call and the code below handles those
2545 MUTEX_ENTER(&call->lock);
2547 /* This packet can't be for this call. If the new call address is
2548 * 0 then no call is running on this channel. If there is a call
2549 * then, since this is a client connection we're getting data for
2550 * it must be for the previous call.
2552 MUTEX_ENTER(&rx_stats_mutex);
2553 rx_stats.spuriousPacketsRead++;
2554 MUTEX_EXIT(&rx_stats_mutex);
2555 MUTEX_ENTER(&conn->conn_data_lock);
2557 MUTEX_EXIT(&conn->conn_data_lock);
2562 currentCallNumber = conn->callNumber[channel];
2564 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2565 if (np->header.callNumber < currentCallNumber) {
2566 MUTEX_ENTER(&rx_stats_mutex);
2567 rx_stats.spuriousPacketsRead++;
2568 MUTEX_EXIT(&rx_stats_mutex);
2569 #ifdef RX_ENABLE_LOCKS
2571 MUTEX_EXIT(&call->lock);
2573 MUTEX_ENTER(&conn->conn_data_lock);
2575 MUTEX_EXIT(&conn->conn_data_lock);
2579 MUTEX_ENTER(&conn->conn_call_lock);
2580 call = rxi_NewCall(conn, channel);
2581 MUTEX_EXIT(&conn->conn_call_lock);
2582 *call->callNumber = np->header.callNumber;
2583 call->state = RX_STATE_PRECALL;
2584 clock_GetTime(&call->queueTime);
2585 hzero(call->bytesSent);
2586 hzero(call->bytesRcvd);
2587 rxi_KeepAliveOn(call);
2588 } else if (np->header.callNumber != currentCallNumber) {
2589 /* Wait until the transmit queue is idle before deciding
2590 * whether to reset the current call. Chances are that the
2591 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2594 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2595 while ((call->state == RX_STATE_ACTIVE)
2596 && (call->flags & RX_CALL_TQ_BUSY)) {
2597 call->flags |= RX_CALL_TQ_WAIT;
2598 #ifdef RX_ENABLE_LOCKS
2599 CV_WAIT(&call->cv_tq, &call->lock);
2600 #else /* RX_ENABLE_LOCKS */
2601 osi_rxSleep(&call->tq);
2602 #endif /* RX_ENABLE_LOCKS */
2604 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2605 /* If the new call cannot be taken right now send a busy and set
2606 * the error condition in this call, so that it terminates as
2607 * quickly as possible */
2608 if (call->state == RX_STATE_ACTIVE) {
2609 struct rx_packet *tp;
2611 rxi_CallError(call, RX_CALL_DEAD);
2612 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2614 MUTEX_EXIT(&call->lock);
2615 MUTEX_ENTER(&conn->conn_data_lock);
2617 MUTEX_EXIT(&conn->conn_data_lock);
2620 rxi_ResetCall(call, 0);
2621 *call->callNumber = np->header.callNumber;
2622 call->state = RX_STATE_PRECALL;
2623 clock_GetTime(&call->queueTime);
2624 hzero(call->bytesSent);
2625 hzero(call->bytesRcvd);
2627 * If the number of queued calls exceeds the overload
2628 * threshold then abort this call.
2630 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2631 struct rx_packet *tp;
2633 rxi_CallError(call, rx_BusyError);
2634 tp = rxi_SendCallAbort(call, np, 1, 0);
2635 MUTEX_EXIT(&call->lock);
2636 MUTEX_ENTER(&conn->conn_data_lock);
2638 MUTEX_EXIT(&conn->conn_data_lock);
2639 MUTEX_ENTER(&rx_stats_mutex);
2641 MUTEX_EXIT(&rx_stats_mutex);
2644 rxi_KeepAliveOn(call);
2646 /* Continuing call; do nothing here. */
2648 } else { /* we're the client */
2649 /* Ignore all incoming acknowledgements for calls in DALLY state */
2650 if (call && (call->state == RX_STATE_DALLY)
2651 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2652 MUTEX_ENTER(&rx_stats_mutex);
2653 rx_stats.ignorePacketDally++;
2654 MUTEX_EXIT(&rx_stats_mutex);
2655 #ifdef RX_ENABLE_LOCKS
2657 MUTEX_EXIT(&call->lock);
2660 MUTEX_ENTER(&conn->conn_data_lock);
2662 MUTEX_EXIT(&conn->conn_data_lock);
2666 /* Ignore anything that's not relevant to the current call. If there
2667 * isn't a current call, then no packet is relevant. */
2668 if (!call || (np->header.callNumber != currentCallNumber)) {
2669 MUTEX_ENTER(&rx_stats_mutex);
2670 rx_stats.spuriousPacketsRead++;
2671 MUTEX_EXIT(&rx_stats_mutex);
2672 #ifdef RX_ENABLE_LOCKS
2674 MUTEX_EXIT(&call->lock);
2677 MUTEX_ENTER(&conn->conn_data_lock);
2679 MUTEX_EXIT(&conn->conn_data_lock);
2682 /* If the service security object index stamped in the packet does not
2683 * match the connection's security index, ignore the packet */
2684 if (np->header.securityIndex != conn->securityIndex) {
2685 #ifdef RX_ENABLE_LOCKS
2686 MUTEX_EXIT(&call->lock);
2688 MUTEX_ENTER(&conn->conn_data_lock);
2690 MUTEX_EXIT(&conn->conn_data_lock);
2694 /* If we're receiving the response, then all transmit packets are
2695 * implicitly acknowledged. Get rid of them. */
2696 if (np->header.type == RX_PACKET_TYPE_DATA) {
2697 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2698 /* XXX Hack. Because we must release the global rx lock when
2699 * sending packets (osi_NetSend) we drop all acks while we're
2700 * traversing the tq in rxi_Start sending packets out because
2701 * packets may move to the freePacketQueue as result of being here!
2702 * So we drop these packets until we're safely out of the
2703 * traversing. Really ugly!
2704 * For fine grain RX locking, we set the acked field in the
2705 * packets and let rxi_Start remove them from the transmit queue.
2707 if (call->flags & RX_CALL_TQ_BUSY) {
2708 #ifdef RX_ENABLE_LOCKS
2709 rxi_SetAcksInTransmitQueue(call);
2712 return np; /* xmitting; drop packet */
2715 rxi_ClearTransmitQueue(call, 0);
2717 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2718 rxi_ClearTransmitQueue(call, 0);
2719 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2721 if (np->header.type == RX_PACKET_TYPE_ACK) {
2722 /* now check to see if this is an ack packet acknowledging that the
2723 * server actually *lost* some hard-acked data. If this happens we
2724 * ignore this packet, as it may indicate that the server restarted in
2725 * the middle of a call. It is also possible that this is an old ack
2726 * packet. We don't abort the connection in this case, because this
2727 * *might* just be an old ack packet. The right way to detect a server
2728 * restart in the midst of a call is to notice that the server epoch
2730 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2731 * XXX unacknowledged. I think that this is off-by-one, but
2732 * XXX I don't dare change it just yet, since it will
2733 * XXX interact badly with the server-restart detection
2734 * XXX code in receiveackpacket. */
2735 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2736 MUTEX_ENTER(&rx_stats_mutex);
2737 rx_stats.spuriousPacketsRead++;
2738 MUTEX_EXIT(&rx_stats_mutex);
2739 MUTEX_EXIT(&call->lock);
2740 MUTEX_ENTER(&conn->conn_data_lock);
2742 MUTEX_EXIT(&conn->conn_data_lock);
2746 } /* else not a data packet */
2749 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2750 /* Set remote user defined status from packet */
2751 call->remoteStatus = np->header.userStatus;
2753 /* Note the gap between the expected next packet and the actual
2754 * packet that arrived, when the new packet has a smaller serial number
2755 * than expected. Rioses frequently reorder packets all by themselves,
2756 * so this will be quite important with very large window sizes.
2757 * Skew is checked against 0 here to avoid any dependence on the type of
2758 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2760 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2761 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2762 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2764 MUTEX_ENTER(&conn->conn_data_lock);
2765 skew = conn->lastSerial - np->header.serial;
2766 conn->lastSerial = np->header.serial;
2767 MUTEX_EXIT(&conn->conn_data_lock);
2769 register struct rx_peer *peer;
2771 if (skew > peer->inPacketSkew) {
2772 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2774 peer->inPacketSkew = skew;
2778 /* Now do packet type-specific processing */
2779 switch (np->header.type) {
2780 case RX_PACKET_TYPE_DATA:
2781 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2784 case RX_PACKET_TYPE_ACK:
2785 /* Respond immediately to ack packets requesting acknowledgement
2787 if (np->header.flags & RX_REQUEST_ACK) {
2789 (void)rxi_SendCallAbort(call, 0, 1, 0);
2791 (void)rxi_SendAck(call, 0, np->header.serial,
2792 RX_ACK_PING_RESPONSE, 1);
2794 np = rxi_ReceiveAckPacket(call, np, 1);
2796 case RX_PACKET_TYPE_ABORT:
2797 /* An abort packet: reset the connection, passing the error up to
2799 /* What if error is zero? */
2800 rxi_CallError(call, ntohl(*(afs_int32 *) rx_DataOf(np)));
2802 case RX_PACKET_TYPE_BUSY:
2805 case RX_PACKET_TYPE_ACKALL:
2806 /* All packets acknowledged, so we can drop all packets previously
2807 * readied for sending */
2808 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2809 /* XXX Hack. We because we can't release the global rx lock when
2810 * sending packets (osi_NetSend) we drop all ack pkts while we're
2811 * traversing the tq in rxi_Start sending packets out because
2812 * packets may move to the freePacketQueue as result of being
2813 * here! So we drop these packets until we're safely out of the
2814 * traversing. Really ugly!
2815 * For fine grain RX locking, we set the acked field in the packets
2816 * and let rxi_Start remove the packets from the transmit queue.
2818 if (call->flags & RX_CALL_TQ_BUSY) {
2819 #ifdef RX_ENABLE_LOCKS
2820 rxi_SetAcksInTransmitQueue(call);
2822 #else /* RX_ENABLE_LOCKS */
2824 return np; /* xmitting; drop packet */
2825 #endif /* RX_ENABLE_LOCKS */
2827 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2828 rxi_ClearTransmitQueue(call, 0);
2831 /* Should not reach here, unless the peer is broken: send an abort
2833 rxi_CallError(call, RX_PROTOCOL_ERROR);
2834 np = rxi_SendCallAbort(call, np, 1, 0);
2837 /* Note when this last legitimate packet was received, for keep-alive
2838 * processing. Note, we delay getting the time until now in the hope that
2839 * the packet will be delivered to the user before any get time is required
2840 * (if not, then the time won't actually be re-evaluated here). */
2841 call->lastReceiveTime = clock_Sec();
2842 MUTEX_EXIT(&call->lock);
2843 MUTEX_ENTER(&conn->conn_data_lock);
2845 MUTEX_EXIT(&conn->conn_data_lock);
2849 /* return true if this is an "interesting" connection from the point of view
2850 of someone trying to debug the system */
2852 rxi_IsConnInteresting(struct rx_connection *aconn)
2855 register struct rx_call *tcall;
2857 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2859 for (i = 0; i < RX_MAXCALLS; i++) {
2860 tcall = aconn->call[i];
2862 if ((tcall->state == RX_STATE_PRECALL)
2863 || (tcall->state == RX_STATE_ACTIVE))
2865 if ((tcall->mode == RX_MODE_SENDING)
2866 || (tcall->mode == RX_MODE_RECEIVING))
2874 /* if this is one of the last few packets AND it wouldn't be used by the
2875 receiving call to immediately satisfy a read request, then drop it on
2876 the floor, since accepting it might prevent a lock-holding thread from
2877 making progress in its reading. If a call has been cleared while in
2878 the precall state then ignore all subsequent packets until the call
2879 is assigned to a thread. */
2882 TooLow(struct rx_packet *ap, struct rx_call *acall)
2885 MUTEX_ENTER(&rx_stats_mutex);
2886 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
2887 && (acall->state == RX_STATE_PRECALL))
2888 || ((rx_nFreePackets < rxi_dataQuota + 2)
2889 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
2890 && (acall->flags & RX_CALL_READER_WAIT)))) {
2893 MUTEX_EXIT(&rx_stats_mutex);
2899 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
2900 struct rx_call *acall)
2902 struct rx_call *call = acall;
2906 MUTEX_ENTER(&conn->conn_data_lock);
2907 conn->checkReachEvent = NULL;
2908 waiting = conn->flags & RX_CONN_ATTACHWAIT;
2911 MUTEX_EXIT(&conn->conn_data_lock);
2915 MUTEX_ENTER(&conn->conn_call_lock);
2916 MUTEX_ENTER(&conn->conn_data_lock);
2917 for (i = 0; i < RX_MAXCALLS; i++) {
2918 struct rx_call *tc = conn->call[i];
2919 if (tc && tc->state == RX_STATE_PRECALL) {
2925 /* Indicate that rxi_CheckReachEvent is no longer running by
2926 * clearing the flag. Must be atomic under conn_data_lock to
2927 * avoid a new call slipping by: rxi_CheckConnReach holds
2928 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
2930 conn->flags &= ~RX_CONN_ATTACHWAIT;
2931 MUTEX_EXIT(&conn->conn_data_lock);
2932 MUTEX_EXIT(&conn->conn_call_lock);
2937 MUTEX_ENTER(&call->lock);
2938 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
2940 MUTEX_EXIT(&call->lock);
2942 clock_GetTime(&when);
2943 when.sec += RX_CHECKREACH_TIMEOUT;
2944 MUTEX_ENTER(&conn->conn_data_lock);
2945 if (!conn->checkReachEvent) {
2947 conn->checkReachEvent =
2948 rxevent_Post(&when, rxi_CheckReachEvent, conn, NULL);
2950 MUTEX_EXIT(&conn->conn_data_lock);
2956 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
2958 struct rx_service *service = conn->service;
2959 struct rx_peer *peer = conn->peer;
2960 afs_uint32 now, lastReach;
2962 if (service->checkReach == 0)
2966 MUTEX_ENTER(&peer->peer_lock);
2967 lastReach = peer->lastReachTime;
2968 MUTEX_EXIT(&peer->peer_lock);
2969 if (now - lastReach < RX_CHECKREACH_TTL)
2972 MUTEX_ENTER(&conn->conn_data_lock);
2973 if (conn->flags & RX_CONN_ATTACHWAIT) {
2974 MUTEX_EXIT(&conn->conn_data_lock);
2977 conn->flags |= RX_CONN_ATTACHWAIT;
2978 MUTEX_EXIT(&conn->conn_data_lock);
2979 if (!conn->checkReachEvent)
2980 rxi_CheckReachEvent(NULL, conn, call);
2985 /* try to attach call, if authentication is complete */
2987 TryAttach(register struct rx_call *acall, register osi_socket socket,
2988 register int *tnop, register struct rx_call **newcallp,
2991 struct rx_connection *conn = acall->conn;
2993 if (conn->type == RX_SERVER_CONNECTION
2994 && acall->state == RX_STATE_PRECALL) {
2995 /* Don't attach until we have any req'd. authentication. */
2996 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
2997 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
2998 rxi_AttachServerProc(acall, socket, tnop, newcallp);
2999 /* Note: this does not necessarily succeed; there
3000 * may not any proc available
3003 rxi_ChallengeOn(acall->conn);
3008 /* A data packet has been received off the interface. This packet is
3009 * appropriate to the call (the call is in the right state, etc.). This
3010 * routine can return a packet to the caller, for re-use */
3013 rxi_ReceiveDataPacket(register struct rx_call *call,
3014 register struct rx_packet *np, int istack,
3015 osi_socket socket, afs_uint32 host, u_short port,
3016 int *tnop, struct rx_call **newcallp)
3018 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3022 afs_uint32 seq, serial, flags;
3024 struct rx_packet *tnp;
3026 MUTEX_ENTER(&rx_stats_mutex);
3027 rx_stats.dataPacketsRead++;
3028 MUTEX_EXIT(&rx_stats_mutex);
3031 /* If there are no packet buffers, drop this new packet, unless we can find
3032 * packet buffers from inactive calls */
3034 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3035 MUTEX_ENTER(&rx_freePktQ_lock);
3036 rxi_NeedMorePackets = TRUE;
3037 MUTEX_EXIT(&rx_freePktQ_lock);
3038 MUTEX_ENTER(&rx_stats_mutex);
3039 rx_stats.noPacketBuffersOnRead++;
3040 MUTEX_EXIT(&rx_stats_mutex);
3041 call->rprev = np->header.serial;
3042 rxi_calltrace(RX_TRACE_DROP, call);
3043 dpf(("packet %x dropped on receipt - quota problems", np));
3045 rxi_ClearReceiveQueue(call);
3046 clock_GetTime(&when);
3047 clock_Add(&when, &rx_softAckDelay);
3048 if (!call->delayedAckEvent
3049 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3050 rxevent_Cancel(call->delayedAckEvent, call,
3051 RX_CALL_REFCOUNT_DELAY);
3052 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3053 call->delayedAckEvent =
3054 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3056 /* we've damaged this call already, might as well do it in. */
3062 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3063 * packet is one of several packets transmitted as a single
3064 * datagram. Do not send any soft or hard acks until all packets
3065 * in a jumbogram have been processed. Send negative acks right away.
3067 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3068 /* tnp is non-null when there are more packets in the
3069 * current jumbo gram */
3076 seq = np->header.seq;
3077 serial = np->header.serial;
3078 flags = np->header.flags;
3080 /* If the call is in an error state, send an abort message */
3082 return rxi_SendCallAbort(call, np, istack, 0);
3084 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3085 * AFS 3.5 jumbogram. */
3086 if (flags & RX_JUMBO_PACKET) {
3087 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3092 if (np->header.spare != 0) {
3093 MUTEX_ENTER(&call->conn->conn_data_lock);
3094 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3095 MUTEX_EXIT(&call->conn->conn_data_lock);
3098 /* The usual case is that this is the expected next packet */
3099 if (seq == call->rnext) {
3101 /* Check to make sure it is not a duplicate of one already queued */
3102 if (queue_IsNotEmpty(&call->rq)
3103 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3104 MUTEX_ENTER(&rx_stats_mutex);
3105 rx_stats.dupPacketsRead++;
3106 MUTEX_EXIT(&rx_stats_mutex);
3107 dpf(("packet %x dropped on receipt - duplicate", np));
3108 rxevent_Cancel(call->delayedAckEvent, call,
3109 RX_CALL_REFCOUNT_DELAY);
3110 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3116 /* It's the next packet. Stick it on the receive queue
3117 * for this call. Set newPackets to make sure we wake
3118 * the reader once all packets have been processed */
3119 queue_Prepend(&call->rq, np);
3121 np = NULL; /* We can't use this anymore */
3124 /* If an ack is requested then set a flag to make sure we
3125 * send an acknowledgement for this packet */
3126 if (flags & RX_REQUEST_ACK) {
3127 ackNeeded = RX_ACK_REQUESTED;
3130 /* Keep track of whether we have received the last packet */
3131 if (flags & RX_LAST_PACKET) {
3132 call->flags |= RX_CALL_HAVE_LAST;
3136 /* Check whether we have all of the packets for this call */
3137 if (call->flags & RX_CALL_HAVE_LAST) {
3138 afs_uint32 tseq; /* temporary sequence number */
3139 struct rx_packet *tp; /* Temporary packet pointer */
3140 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3142 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3143 if (tseq != tp->header.seq)
3145 if (tp->header.flags & RX_LAST_PACKET) {
3146 call->flags |= RX_CALL_RECEIVE_DONE;
3153 /* Provide asynchronous notification for those who want it
3154 * (e.g. multi rx) */
3155 if (call->arrivalProc) {
3156 (*call->arrivalProc) (call, call->arrivalProcHandle,
3157 call->arrivalProcArg);
3158 call->arrivalProc = (void (*)())0;
3161 /* Update last packet received */
3164 /* If there is no server process serving this call, grab
3165 * one, if available. We only need to do this once. If a
3166 * server thread is available, this thread becomes a server
3167 * thread and the server thread becomes a listener thread. */
3169 TryAttach(call, socket, tnop, newcallp, 0);
3172 /* This is not the expected next packet. */
3174 /* Determine whether this is a new or old packet, and if it's
3175 * a new one, whether it fits into the current receive window.
3176 * Also figure out whether the packet was delivered in sequence.
3177 * We use the prev variable to determine whether the new packet
3178 * is the successor of its immediate predecessor in the
3179 * receive queue, and the missing flag to determine whether
3180 * any of this packets predecessors are missing. */
3182 afs_uint32 prev; /* "Previous packet" sequence number */
3183 struct rx_packet *tp; /* Temporary packet pointer */
3184 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3185 int missing; /* Are any predecessors missing? */
3187 /* If the new packet's sequence number has been sent to the
3188 * application already, then this is a duplicate */
3189 if (seq < call->rnext) {
3190 MUTEX_ENTER(&rx_stats_mutex);
3191 rx_stats.dupPacketsRead++;
3192 MUTEX_EXIT(&rx_stats_mutex);
3193 rxevent_Cancel(call->delayedAckEvent, call,
3194 RX_CALL_REFCOUNT_DELAY);
3195 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3201 /* If the sequence number is greater than what can be
3202 * accomodated by the current window, then send a negative
3203 * acknowledge and drop the packet */
3204 if ((call->rnext + call->rwind) <= seq) {
3205 rxevent_Cancel(call->delayedAckEvent, call,
3206 RX_CALL_REFCOUNT_DELAY);
3207 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3214 /* Look for the packet in the queue of old received packets */
3215 for (prev = call->rnext - 1, missing =
3216 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3217 /*Check for duplicate packet */
3218 if (seq == tp->header.seq) {
3219 MUTEX_ENTER(&rx_stats_mutex);
3220 rx_stats.dupPacketsRead++;
3221 MUTEX_EXIT(&rx_stats_mutex);
3222 rxevent_Cancel(call->delayedAckEvent, call,
3223 RX_CALL_REFCOUNT_DELAY);
3224 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3230 /* If we find a higher sequence packet, break out and
3231 * insert the new packet here. */
3232 if (seq < tp->header.seq)
3234 /* Check for missing packet */
3235 if (tp->header.seq != prev + 1) {
3239 prev = tp->header.seq;
3242 /* Keep track of whether we have received the last packet. */
3243 if (flags & RX_LAST_PACKET) {
3244 call->flags |= RX_CALL_HAVE_LAST;
3247 /* It's within the window: add it to the the receive queue.
3248 * tp is left by the previous loop either pointing at the
3249 * packet before which to insert the new packet, or at the
3250 * queue head if the queue is empty or the packet should be
3252 queue_InsertBefore(tp, np);
3256 /* Check whether we have all of the packets for this call */
3257 if ((call->flags & RX_CALL_HAVE_LAST)
3258 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3259 afs_uint32 tseq; /* temporary sequence number */
3262 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3263 if (tseq != tp->header.seq)
3265 if (tp->header.flags & RX_LAST_PACKET) {
3266 call->flags |= RX_CALL_RECEIVE_DONE;
3273 /* We need to send an ack of the packet is out of sequence,
3274 * or if an ack was requested by the peer. */
3275 if (seq != prev + 1 || missing || (flags & RX_REQUEST_ACK)) {
3276 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3279 /* Acknowledge the last packet for each call */
3280 if (flags & RX_LAST_PACKET) {
3291 * If the receiver is waiting for an iovec, fill the iovec
3292 * using the data from the receive queue */
3293 if (call->flags & RX_CALL_IOVEC_WAIT) {
3294 didHardAck = rxi_FillReadVec(call, serial);
3295 /* the call may have been aborted */
3304 /* Wakeup the reader if any */
3305 if ((call->flags & RX_CALL_READER_WAIT)
3306 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3307 || (call->iovNext >= call->iovMax)
3308 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3309 call->flags &= ~RX_CALL_READER_WAIT;
3310 #ifdef RX_ENABLE_LOCKS
3311 CV_BROADCAST(&call->cv_rq);
3313 osi_rxWakeup(&call->rq);
3319 * Send an ack when requested by the peer, or once every
3320 * rxi_SoftAckRate packets until the last packet has been
3321 * received. Always send a soft ack for the last packet in
3322 * the server's reply. */
3324 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3325 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3326 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3327 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3328 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3329 } else if (call->nSoftAcks) {
3330 clock_GetTime(&when);
3331 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3332 clock_Add(&when, &rx_lastAckDelay);
3334 clock_Add(&when, &rx_softAckDelay);
3336 if (!call->delayedAckEvent
3337 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3338 rxevent_Cancel(call->delayedAckEvent, call,
3339 RX_CALL_REFCOUNT_DELAY);
3340 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3341 call->delayedAckEvent =
3342 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3344 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3345 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3352 static void rxi_ComputeRate();
3356 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3358 struct rx_peer *peer = conn->peer;
3360 MUTEX_ENTER(&peer->peer_lock);
3361 peer->lastReachTime = clock_Sec();
3362 MUTEX_EXIT(&peer->peer_lock);
3364 MUTEX_ENTER(&conn->conn_data_lock);
3365 if (conn->flags & RX_CONN_ATTACHWAIT) {
3368 conn->flags &= ~RX_CONN_ATTACHWAIT;
3369 MUTEX_EXIT(&conn->conn_data_lock);
3371 for (i = 0; i < RX_MAXCALLS; i++) {
3372 struct rx_call *call = conn->call[i];
3375 MUTEX_ENTER(&call->lock);
3376 /* tnop can be null if newcallp is null */
3377 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3379 MUTEX_EXIT(&call->lock);
3383 MUTEX_EXIT(&conn->conn_data_lock);
3386 /* rxi_ComputePeerNetStats
3388 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3389 * estimates (like RTT and throughput) based on ack packets. Caller
3390 * must ensure that the packet in question is the right one (i.e.
3391 * serial number matches).
3394 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3395 struct rx_ackPacket *ap, struct rx_packet *np)
3397 struct rx_peer *peer = call->conn->peer;
3399 /* Use RTT if not delayed by client. */
3400 if (ap->reason != RX_ACK_DELAY)
3401 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3403 rxi_ComputeRate(peer, call, p, np, ap->reason);
3407 /* The real smarts of the whole thing. */
3409 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3412 struct rx_ackPacket *ap;
3414 register struct rx_packet *tp;
3415 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3416 register struct rx_connection *conn = call->conn;
3417 struct rx_peer *peer = conn->peer;
3420 /* because there are CM's that are bogus, sending weird values for this. */
3421 afs_uint32 skew = 0;
3426 int newAckCount = 0;
3427 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3428 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3430 MUTEX_ENTER(&rx_stats_mutex);
3431 rx_stats.ackPacketsRead++;
3432 MUTEX_EXIT(&rx_stats_mutex);
3433 ap = (struct rx_ackPacket *)rx_DataOf(np);
3434 nbytes = rx_Contiguous(np) - ((ap->acks) - (u_char *) ap);
3436 return np; /* truncated ack packet */
3438 /* depends on ack packet struct */
3439 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3440 first = ntohl(ap->firstPacket);
3441 serial = ntohl(ap->serial);
3442 /* temporarily disabled -- needs to degrade over time
3443 * skew = ntohs(ap->maxSkew); */
3445 /* Ignore ack packets received out of order */
3446 if (first < call->tfirst) {
3450 if (np->header.flags & RX_SLOW_START_OK) {
3451 call->flags |= RX_CALL_SLOW_START_OK;
3454 if (ap->reason == RX_ACK_PING_RESPONSE)
3455 rxi_UpdatePeerReach(conn, call);
3460 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3461 ap->reason, ntohl(ap->previousPacket),
3462 (unsigned int)np->header.seq, (unsigned int)serial,
3463 (unsigned int)skew, ntohl(ap->firstPacket));
3466 for (offset = 0; offset < nAcks; offset++)
3467 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3474 /* Update the outgoing packet skew value to the latest value of
3475 * the peer's incoming packet skew value. The ack packet, of
3476 * course, could arrive out of order, but that won't affect things
3478 MUTEX_ENTER(&peer->peer_lock);
3479 peer->outPacketSkew = skew;
3481 /* Check for packets that no longer need to be transmitted, and
3482 * discard them. This only applies to packets positively
3483 * acknowledged as having been sent to the peer's upper level.
3484 * All other packets must be retained. So only packets with
3485 * sequence numbers < ap->firstPacket are candidates. */
3486 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3487 if (tp->header.seq >= first)
3489 call->tfirst = tp->header.seq + 1;
3491 && (tp->header.serial == serial || tp->firstSerial == serial))
3492 rxi_ComputePeerNetStats(call, tp, ap, np);
3493 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3494 /* XXX Hack. Because we have to release the global rx lock when sending
3495 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3496 * in rxi_Start sending packets out because packets may move to the
3497 * freePacketQueue as result of being here! So we drop these packets until
3498 * we're safely out of the traversing. Really ugly!
3499 * To make it even uglier, if we're using fine grain locking, we can
3500 * set the ack bits in the packets and have rxi_Start remove the packets
3501 * when it's done transmitting.
3503 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3506 if (call->flags & RX_CALL_TQ_BUSY) {
3507 #ifdef RX_ENABLE_LOCKS
3508 tp->flags |= RX_PKTFLAG_ACKED;
3509 call->flags |= RX_CALL_TQ_SOME_ACKED;
3510 #else /* RX_ENABLE_LOCKS */
3512 #endif /* RX_ENABLE_LOCKS */
3514 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3517 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3522 /* Give rate detector a chance to respond to ping requests */
3523 if (ap->reason == RX_ACK_PING_RESPONSE) {
3524 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3528 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3530 /* Now go through explicit acks/nacks and record the results in
3531 * the waiting packets. These are packets that can't be released
3532 * yet, even with a positive acknowledge. This positive
3533 * acknowledge only means the packet has been received by the
3534 * peer, not that it will be retained long enough to be sent to
3535 * the peer's upper level. In addition, reset the transmit timers
3536 * of any missing packets (those packets that must be missing
3537 * because this packet was out of sequence) */
3539 call->nSoftAcked = 0;
3540 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3541 /* Update round trip time if the ack was stimulated on receipt
3543 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3544 #ifdef RX_ENABLE_LOCKS
3545 if (tp->header.seq >= first)
3546 #endif /* RX_ENABLE_LOCKS */
3547 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3549 && (tp->header.serial == serial || tp->firstSerial == serial))
3550 rxi_ComputePeerNetStats(call, tp, ap, np);
3552 /* Set the acknowledge flag per packet based on the
3553 * information in the ack packet. An acknowlegded packet can
3554 * be downgraded when the server has discarded a packet it
3555 * soacked previously, or when an ack packet is received
3556 * out of sequence. */
3557 if (tp->header.seq < first) {
3558 /* Implicit ack information */
3559 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3562 tp->flags |= RX_PKTFLAG_ACKED;
3563 } else if (tp->header.seq < first + nAcks) {
3564 /* Explicit ack information: set it in the packet appropriately */
3565 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3566 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3568 tp->flags |= RX_PKTFLAG_ACKED;
3576 tp->flags &= ~RX_PKTFLAG_ACKED;
3580 tp->flags &= ~RX_PKTFLAG_ACKED;
3584 /* If packet isn't yet acked, and it has been transmitted at least
3585 * once, reset retransmit time using latest timeout
3586 * ie, this should readjust the retransmit timer for all outstanding
3587 * packets... So we don't just retransmit when we should know better*/
3589 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3590 tp->retryTime = tp->timeSent;
3591 clock_Add(&tp->retryTime, &peer->timeout);
3592 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3593 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3597 /* If the window has been extended by this acknowledge packet,
3598 * then wakeup a sender waiting in alloc for window space, or try
3599 * sending packets now, if he's been sitting on packets due to
3600 * lack of window space */
3601 if (call->tnext < (call->tfirst + call->twind)) {
3602 #ifdef RX_ENABLE_LOCKS
3603 CV_SIGNAL(&call->cv_twind);
3605 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3606 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3607 osi_rxWakeup(&call->twind);
3610 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3611 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3615 /* if the ack packet has a receivelen field hanging off it,
3616 * update our state */
3617 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3620 /* If the ack packet has a "recommended" size that is less than
3621 * what I am using now, reduce my size to match */
3622 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3623 sizeof(afs_int32), &tSize);
3624 tSize = (afs_uint32) ntohl(tSize);
3625 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3627 /* Get the maximum packet size to send to this peer */
3628 rx_packetread(np, rx_AckDataSize(ap->nAcks), sizeof(afs_int32),
3630 tSize = (afs_uint32) ntohl(tSize);
3631 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3632 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3634 /* sanity check - peer might have restarted with different params.
3635 * If peer says "send less", dammit, send less... Peer should never
3636 * be unable to accept packets of the size that prior AFS versions would
3637 * send without asking. */
3638 if (peer->maxMTU != tSize) {
3639 peer->maxMTU = tSize;
3640 peer->MTU = MIN(tSize, peer->MTU);
3641 call->MTU = MIN(call->MTU, tSize);
3645 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3648 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3649 sizeof(afs_int32), &tSize);
3650 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3651 if (tSize < call->twind) { /* smaller than our send */
3652 call->twind = tSize; /* window, we must send less... */
3653 call->ssthresh = MIN(call->twind, call->ssthresh);
3656 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3657 * network MTU confused with the loopback MTU. Calculate the
3658 * maximum MTU here for use in the slow start code below.
3660 maxMTU = peer->maxMTU;
3661 /* Did peer restart with older RX version? */
3662 if (peer->maxDgramPackets > 1) {
3663 peer->maxDgramPackets = 1;
3665 } else if (np->length >=
3666 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3669 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3670 sizeof(afs_int32), &tSize);
3671 tSize = (afs_uint32) ntohl(tSize);
3673 * As of AFS 3.5 we set the send window to match the receive window.
3675 if (tSize < call->twind) {
3676 call->twind = tSize;
3677 call->ssthresh = MIN(call->twind, call->ssthresh);
3678 } else if (tSize > call->twind) {
3679 call->twind = tSize;
3683 * As of AFS 3.5, a jumbogram is more than one fixed size
3684 * packet transmitted in a single UDP datagram. If the remote
3685 * MTU is smaller than our local MTU then never send a datagram
3686 * larger than the natural MTU.
3689 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3690 sizeof(afs_int32), &tSize);
3691 maxDgramPackets = (afs_uint32) ntohl(tSize);
3692 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3694 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
3695 maxDgramPackets = MIN(maxDgramPackets, tSize);
3696 if (maxDgramPackets > 1) {
3697 peer->maxDgramPackets = maxDgramPackets;
3698 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3700 peer->maxDgramPackets = 1;
3701 call->MTU = peer->natMTU;
3703 } else if (peer->maxDgramPackets > 1) {
3704 /* Restarted with lower version of RX */
3705 peer->maxDgramPackets = 1;
3707 } else if (peer->maxDgramPackets > 1
3708 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3709 /* Restarted with lower version of RX */
3710 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3711 peer->natMTU = OLD_MAX_PACKET_SIZE;
3712 peer->MTU = OLD_MAX_PACKET_SIZE;
3713 peer->maxDgramPackets = 1;
3714 peer->nDgramPackets = 1;
3716 call->MTU = OLD_MAX_PACKET_SIZE;
3721 * Calculate how many datagrams were successfully received after
3722 * the first missing packet and adjust the negative ack counter
3727 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3728 if (call->nNacks < nNacked) {
3729 call->nNacks = nNacked;
3738 if (call->flags & RX_CALL_FAST_RECOVER) {
3740 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3742 call->flags &= ~RX_CALL_FAST_RECOVER;
3743 call->cwind = call->nextCwind;
3744 call->nextCwind = 0;
3747 call->nCwindAcks = 0;
3748 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3749 /* Three negative acks in a row trigger congestion recovery */
3750 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3751 MUTEX_EXIT(&peer->peer_lock);
3752 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3753 /* someone else is waiting to start recovery */
3756 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3757 while (call->flags & RX_CALL_TQ_BUSY) {
3758 call->flags |= RX_CALL_TQ_WAIT;
3759 #ifdef RX_ENABLE_LOCKS
3760 CV_WAIT(&call->cv_tq, &call->lock);
3761 #else /* RX_ENABLE_LOCKS */
3762 osi_rxSleep(&call->tq);
3763 #endif /* RX_ENABLE_LOCKS */
3765 MUTEX_ENTER(&peer->peer_lock);
3766 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3767 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3768 call->flags |= RX_CALL_FAST_RECOVER;
3769 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3771 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3772 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3773 call->nextCwind = call->ssthresh;
3776 peer->MTU = call->MTU;
3777 peer->cwind = call->nextCwind;
3778 peer->nDgramPackets = call->nDgramPackets;
3780 call->congestSeq = peer->congestSeq;
3781 /* Reset the resend times on the packets that were nacked
3782 * so we will retransmit as soon as the window permits*/
3783 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3785 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3786 clock_Zero(&tp->retryTime);
3788 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3793 /* If cwind is smaller than ssthresh, then increase
3794 * the window one packet for each ack we receive (exponential
3796 * If cwind is greater than or equal to ssthresh then increase
3797 * the congestion window by one packet for each cwind acks we
3798 * receive (linear growth). */
3799 if (call->cwind < call->ssthresh) {
3801 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3802 call->nCwindAcks = 0;
3804 call->nCwindAcks += newAckCount;
3805 if (call->nCwindAcks >= call->cwind) {
3806 call->nCwindAcks = 0;
3807 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3811 * If we have received several acknowledgements in a row then
3812 * it is time to increase the size of our datagrams
3814 if ((int)call->nAcks > rx_nDgramThreshold) {
3815 if (peer->maxDgramPackets > 1) {
3816 if (call->nDgramPackets < peer->maxDgramPackets) {
3817 call->nDgramPackets++;
3819 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3820 } else if (call->MTU < peer->maxMTU) {
3821 call->MTU += peer->natMTU;
3822 call->MTU = MIN(call->MTU, peer->maxMTU);
3828 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3830 /* Servers need to hold the call until all response packets have
3831 * been acknowledged. Soft acks are good enough since clients
3832 * are not allowed to clear their receive queues. */
3833 if (call->state == RX_STATE_HOLD
3834 && call->tfirst + call->nSoftAcked >= call->tnext) {
3835 call->state = RX_STATE_DALLY;
3836 rxi_ClearTransmitQueue(call, 0);
3837 } else if (!queue_IsEmpty(&call->tq)) {
3838 rxi_Start(0, call, 0, istack);
3843 /* Received a response to a challenge packet */
3845 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
3846 register struct rx_packet *np, int istack)
3850 /* Ignore the packet if we're the client */
3851 if (conn->type == RX_CLIENT_CONNECTION)
3854 /* If already authenticated, ignore the packet (it's probably a retry) */
3855 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
3858 /* Otherwise, have the security object evaluate the response packet */
3859 error = RXS_CheckResponse(conn->securityObject, conn, np);
3861 /* If the response is invalid, reset the connection, sending
3862 * an abort to the peer */
3866 rxi_ConnectionError(conn, error);
3867 MUTEX_ENTER(&conn->conn_data_lock);
3868 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3869 MUTEX_EXIT(&conn->conn_data_lock);
3872 /* If the response is valid, any calls waiting to attach
3873 * servers can now do so */
3876 for (i = 0; i < RX_MAXCALLS; i++) {
3877 struct rx_call *call = conn->call[i];
3879 MUTEX_ENTER(&call->lock);
3880 if (call->state == RX_STATE_PRECALL)
3881 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
3882 /* tnop can be null if newcallp is null */
3883 MUTEX_EXIT(&call->lock);
3887 /* Update the peer reachability information, just in case
3888 * some calls went into attach-wait while we were waiting
3889 * for authentication..
3891 rxi_UpdatePeerReach(conn, NULL);
3896 /* A client has received an authentication challenge: the security
3897 * object is asked to cough up a respectable response packet to send
3898 * back to the server. The server is responsible for retrying the
3899 * challenge if it fails to get a response. */
3902 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
3903 register struct rx_packet *np, int istack)
3907 /* Ignore the challenge if we're the server */
3908 if (conn->type == RX_SERVER_CONNECTION)
3911 /* Ignore the challenge if the connection is otherwise idle; someone's
3912 * trying to use us as an oracle. */
3913 if (!rxi_HasActiveCalls(conn))
3916 /* Send the security object the challenge packet. It is expected to fill
3917 * in the response. */
3918 error = RXS_GetResponse(conn->securityObject, conn, np);
3920 /* If the security object is unable to return a valid response, reset the
3921 * connection and send an abort to the peer. Otherwise send the response
3922 * packet to the peer connection. */
3924 rxi_ConnectionError(conn, error);
3925 MUTEX_ENTER(&conn->conn_data_lock);
3926 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3927 MUTEX_EXIT(&conn->conn_data_lock);
3929 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
3930 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
3936 /* Find an available server process to service the current request in
3937 * the given call structure. If one isn't available, queue up this
3938 * call so it eventually gets one */
3940 rxi_AttachServerProc(register struct rx_call *call,
3941 register osi_socket socket, register int *tnop,
3942 register struct rx_call **newcallp)
3944 register struct rx_serverQueueEntry *sq;
3945 register struct rx_service *service = call->conn->service;
3946 register int haveQuota = 0;
3948 /* May already be attached */
3949 if (call->state == RX_STATE_ACTIVE)
3952 MUTEX_ENTER(&rx_serverPool_lock);
3954 haveQuota = QuotaOK(service);
3955 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
3956 /* If there are no processes available to service this call,
3957 * put the call on the incoming call queue (unless it's
3958 * already on the queue).
3960 #ifdef RX_ENABLE_LOCKS
3962 ReturnToServerPool(service);
3963 #endif /* RX_ENABLE_LOCKS */
3965 if (!(call->flags & RX_CALL_WAIT_PROC)) {
3966 call->flags |= RX_CALL_WAIT_PROC;
3967 MUTEX_ENTER(&rx_stats_mutex);
3970 MUTEX_EXIT(&rx_stats_mutex);
3971 rxi_calltrace(RX_CALL_ARRIVAL, call);
3972 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
3973 queue_Append(&rx_incomingCallQueue, call);
3976 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
3978 /* If hot threads are enabled, and both newcallp and sq->socketp
3979 * are non-null, then this thread will process the call, and the
3980 * idle server thread will start listening on this threads socket.
3983 if (rx_enable_hot_thread && newcallp && sq->socketp) {
3986 *sq->socketp = socket;
3987 clock_GetTime(&call->startTime);
3988 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
3992 if (call->flags & RX_CALL_WAIT_PROC) {
3993 /* Conservative: I don't think this should happen */
3994 call->flags &= ~RX_CALL_WAIT_PROC;
3995 if (queue_IsOnQueue(call)) {
3997 MUTEX_ENTER(&rx_stats_mutex);
3999 MUTEX_EXIT(&rx_stats_mutex);
4002 call->state = RX_STATE_ACTIVE;
4003 call->mode = RX_MODE_RECEIVING;
4004 #ifdef RX_KERNEL_TRACE
4006 int glockOwner = ISAFS_GLOCK();
4009 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4010 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4016 if (call->flags & RX_CALL_CLEARED) {
4017 /* send an ack now to start the packet flow up again */
4018 call->flags &= ~RX_CALL_CLEARED;
4019 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4021 #ifdef RX_ENABLE_LOCKS
4024 service->nRequestsRunning++;
4025 if (service->nRequestsRunning <= service->minProcs)
4031 MUTEX_EXIT(&rx_serverPool_lock);
4034 /* Delay the sending of an acknowledge event for a short while, while
4035 * a new call is being prepared (in the case of a client) or a reply
4036 * is being prepared (in the case of a server). Rather than sending
4037 * an ack packet, an ACKALL packet is sent. */
4039 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4041 #ifdef RX_ENABLE_LOCKS
4043 MUTEX_ENTER(&call->lock);
4044 call->delayedAckEvent = NULL;
4045 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4047 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4048 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4050 MUTEX_EXIT(&call->lock);
4051 #else /* RX_ENABLE_LOCKS */
4053 call->delayedAckEvent = NULL;
4054 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4055 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4056 #endif /* RX_ENABLE_LOCKS */
4060 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4063 #ifdef RX_ENABLE_LOCKS
4065 MUTEX_ENTER(&call->lock);
4066 if (event == call->delayedAckEvent)
4067 call->delayedAckEvent = NULL;
4068 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4070 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4072 MUTEX_EXIT(&call->lock);
4073 #else /* RX_ENABLE_LOCKS */
4075 call->delayedAckEvent = NULL;
4076 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4077 #endif /* RX_ENABLE_LOCKS */
4081 #ifdef RX_ENABLE_LOCKS
4082 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4083 * clearing them out.
4086 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4088 register struct rx_packet *p, *tp;
4091 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4094 p->flags |= RX_PKTFLAG_ACKED;
4098 call->flags |= RX_CALL_TQ_CLEARME;
4099 call->flags |= RX_CALL_TQ_SOME_ACKED;
4102 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4103 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4104 call->tfirst = call->tnext;
4105 call->nSoftAcked = 0;
4107 if (call->flags & RX_CALL_FAST_RECOVER) {
4108 call->flags &= ~RX_CALL_FAST_RECOVER;
4109 call->cwind = call->nextCwind;
4110 call->nextCwind = 0;
4113 CV_SIGNAL(&call->cv_twind);
4115 #endif /* RX_ENABLE_LOCKS */
4117 /* Clear out the transmit queue for the current call (all packets have
4118 * been received by peer) */
4120 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4122 register struct rx_packet *p, *tp;
4124 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4125 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4127 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4130 p->flags |= RX_PKTFLAG_ACKED;
4134 call->flags |= RX_CALL_TQ_CLEARME;
4135 call->flags |= RX_CALL_TQ_SOME_ACKED;
4138 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4139 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4145 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4146 call->flags &= ~RX_CALL_TQ_CLEARME;
4148 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4150 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4151 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4152 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4153 call->nSoftAcked = 0;
4155 if (call->flags & RX_CALL_FAST_RECOVER) {
4156 call->flags &= ~RX_CALL_FAST_RECOVER;
4157 call->cwind = call->nextCwind;
4159 #ifdef RX_ENABLE_LOCKS
4160 CV_SIGNAL(&call->cv_twind);
4162 osi_rxWakeup(&call->twind);
4167 rxi_ClearReceiveQueue(register struct rx_call *call)
4169 register struct rx_packet *p, *tp;
4170 if (queue_IsNotEmpty(&call->rq)) {
4171 for (queue_Scan(&call->rq, p, tp, rx_packet)) {
4176 rx_packetReclaims++;
4178 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4180 if (call->state == RX_STATE_PRECALL) {
4181 call->flags |= RX_CALL_CLEARED;
4185 /* Send an abort packet for the specified call */
4187 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4188 int istack, int force)
4196 /* Clients should never delay abort messages */
4197 if (rx_IsClientConn(call->conn))
4200 if (call->abortCode != call->error) {
4201 call->abortCode = call->error;
4202 call->abortCount = 0;
4205 if (force || rxi_callAbortThreshhold == 0
4206 || call->abortCount < rxi_callAbortThreshhold) {
4207 if (call->delayedAbortEvent) {
4208 rxevent_Cancel(call->delayedAbortEvent, call,
4209 RX_CALL_REFCOUNT_ABORT);
4211 error = htonl(call->error);
4214 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4215 (char *)&error, sizeof(error), istack);
4216 } else if (!call->delayedAbortEvent) {
4217 clock_GetTime(&when);
4218 clock_Addmsec(&when, rxi_callAbortDelay);
4219 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4220 call->delayedAbortEvent =
4221 rxevent_Post(&when, rxi_SendDelayedCallAbort, call, 0);
4226 /* Send an abort packet for the specified connection. Packet is an
4227 * optional pointer to a packet that can be used to send the abort.
4228 * Once the number of abort messages reaches the threshhold, an
4229 * event is scheduled to send the abort. Setting the force flag
4230 * overrides sending delayed abort messages.
4232 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4233 * to send the abort packet.
4236 rxi_SendConnectionAbort(register struct rx_connection *conn,
4237 struct rx_packet *packet, int istack, int force)
4245 /* Clients should never delay abort messages */
4246 if (rx_IsClientConn(conn))
4249 if (force || rxi_connAbortThreshhold == 0
4250 || conn->abortCount < rxi_connAbortThreshhold) {
4251 if (conn->delayedAbortEvent) {
4252 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4254 error = htonl(conn->error);
4256 MUTEX_EXIT(&conn->conn_data_lock);
4258 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4259 RX_PACKET_TYPE_ABORT, (char *)&error,
4260 sizeof(error), istack);
4261 MUTEX_ENTER(&conn->conn_data_lock);
4262 } else if (!conn->delayedAbortEvent) {
4263 clock_GetTime(&when);
4264 clock_Addmsec(&when, rxi_connAbortDelay);
4265 conn->delayedAbortEvent =
4266 rxevent_Post(&when, rxi_SendDelayedConnAbort, conn, 0);
4271 /* Associate an error all of the calls owned by a connection. Called
4272 * with error non-zero. This is only for really fatal things, like
4273 * bad authentication responses. The connection itself is set in
4274 * error at this point, so that future packets received will be
4277 rxi_ConnectionError(register struct rx_connection *conn,
4278 register afs_int32 error)
4282 MUTEX_ENTER(&conn->conn_data_lock);
4283 if (conn->challengeEvent)
4284 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4285 if (conn->checkReachEvent) {
4286 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4287 conn->checkReachEvent = 0;
4288 conn->flags &= ~RX_CONN_ATTACHWAIT;
4291 MUTEX_EXIT(&conn->conn_data_lock);
4292 for (i = 0; i < RX_MAXCALLS; i++) {
4293 struct rx_call *call = conn->call[i];
4295 MUTEX_ENTER(&call->lock);
4296 rxi_CallError(call, error);
4297 MUTEX_EXIT(&call->lock);
4300 conn->error = error;
4301 MUTEX_ENTER(&rx_stats_mutex);
4302 rx_stats.fatalErrors++;
4303 MUTEX_EXIT(&rx_stats_mutex);
4308 rxi_CallError(register struct rx_call *call, afs_int32 error)
4311 error = call->error;
4312 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4313 if (!(call->flags & RX_CALL_TQ_BUSY)) {
4314 rxi_ResetCall(call, 0);
4317 rxi_ResetCall(call, 0);
4319 call->error = error;
4320 call->mode = RX_MODE_ERROR;
4323 /* Reset various fields in a call structure, and wakeup waiting
4324 * processes. Some fields aren't changed: state & mode are not
4325 * touched (these must be set by the caller), and bufptr, nLeft, and
4326 * nFree are not reset, since these fields are manipulated by
4327 * unprotected macros, and may only be reset by non-interrupting code.
4330 /* this code requires that call->conn be set properly as a pre-condition. */
4331 #endif /* ADAPT_WINDOW */
4334 rxi_ResetCall(register struct rx_call *call, register int newcall)
4337 register struct rx_peer *peer;
4338 struct rx_packet *packet;
4340 /* Notify anyone who is waiting for asynchronous packet arrival */
4341 if (call->arrivalProc) {
4342 (*call->arrivalProc) (call, call->arrivalProcHandle,
4343 call->arrivalProcArg);
4344 call->arrivalProc = (void (*)())0;
4347 if (call->delayedAbortEvent) {
4348 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4349 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4351 rxi_SendCallAbort(call, packet, 0, 1);
4352 rxi_FreePacket(packet);
4357 * Update the peer with the congestion information in this call
4358 * so other calls on this connection can pick up where this call
4359 * left off. If the congestion sequence numbers don't match then
4360 * another call experienced a retransmission.
4362 peer = call->conn->peer;
4363 MUTEX_ENTER(&peer->peer_lock);
4365 if (call->congestSeq == peer->congestSeq) {
4366 peer->cwind = MAX(peer->cwind, call->cwind);
4367 peer->MTU = MAX(peer->MTU, call->MTU);
4368 peer->nDgramPackets =
4369 MAX(peer->nDgramPackets, call->nDgramPackets);
4372 call->abortCode = 0;
4373 call->abortCount = 0;
4375 if (peer->maxDgramPackets > 1) {
4376 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4378 call->MTU = peer->MTU;
4380 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4381 call->ssthresh = rx_maxSendWindow;
4382 call->nDgramPackets = peer->nDgramPackets;
4383 call->congestSeq = peer->congestSeq;
4384 MUTEX_EXIT(&peer->peer_lock);
4386 flags = call->flags;
4387 rxi_ClearReceiveQueue(call);
4388 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4389 if (call->flags & RX_CALL_TQ_BUSY) {
4390 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4391 call->flags |= (flags & RX_CALL_TQ_WAIT);
4393 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4395 rxi_ClearTransmitQueue(call, 0);
4396 queue_Init(&call->tq);
4399 queue_Init(&call->rq);
4401 call->rwind = rx_initReceiveWindow;
4402 call->twind = rx_initSendWindow;
4403 call->nSoftAcked = 0;
4404 call->nextCwind = 0;
4407 call->nCwindAcks = 0;
4408 call->nSoftAcks = 0;
4409 call->nHardAcks = 0;
4411 call->tfirst = call->rnext = call->tnext = 1;
4413 call->lastAcked = 0;
4414 call->localStatus = call->remoteStatus = 0;
4416 if (flags & RX_CALL_READER_WAIT) {
4417 #ifdef RX_ENABLE_LOCKS
4418 CV_BROADCAST(&call->cv_rq);
4420 osi_rxWakeup(&call->rq);
4423 if (flags & RX_CALL_WAIT_PACKETS) {
4424 MUTEX_ENTER(&rx_freePktQ_lock);
4425 rxi_PacketsUnWait(); /* XXX */
4426 MUTEX_EXIT(&rx_freePktQ_lock);
4428 #ifdef RX_ENABLE_LOCKS
4429 CV_SIGNAL(&call->cv_twind);
4431 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4432 osi_rxWakeup(&call->twind);
4435 #ifdef RX_ENABLE_LOCKS
4436 /* The following ensures that we don't mess with any queue while some
4437 * other thread might also be doing so. The call_queue_lock field is
4438 * is only modified under the call lock. If the call is in the process
4439 * of being removed from a queue, the call is not locked until the
4440 * the queue lock is dropped and only then is the call_queue_lock field
4441 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4442 * Note that any other routine which removes a call from a queue has to
4443 * obtain the queue lock before examing the queue and removing the call.
4445 if (call->call_queue_lock) {
4446 MUTEX_ENTER(call->call_queue_lock);
4447 if (queue_IsOnQueue(call)) {
4449 if (flags & RX_CALL_WAIT_PROC) {
4450 MUTEX_ENTER(&rx_stats_mutex);
4452 MUTEX_EXIT(&rx_stats_mutex);
4455 MUTEX_EXIT(call->call_queue_lock);
4456 CLEAR_CALL_QUEUE_LOCK(call);
4458 #else /* RX_ENABLE_LOCKS */
4459 if (queue_IsOnQueue(call)) {
4461 if (flags & RX_CALL_WAIT_PROC)
4464 #endif /* RX_ENABLE_LOCKS */
4466 rxi_KeepAliveOff(call);
4467 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4470 /* Send an acknowledge for the indicated packet (seq,serial) of the
4471 * indicated call, for the indicated reason (reason). This
4472 * acknowledge will specifically acknowledge receiving the packet, and
4473 * will also specify which other packets for this call have been
4474 * received. This routine returns the packet that was used to the
4475 * caller. The caller is responsible for freeing it or re-using it.
4476 * This acknowledgement also returns the highest sequence number
4477 * actually read out by the higher level to the sender; the sender
4478 * promises to keep around packets that have not been read by the
4479 * higher level yet (unless, of course, the sender decides to abort
4480 * the call altogether). Any of p, seq, serial, pflags, or reason may
4481 * be set to zero without ill effect. That is, if they are zero, they
4482 * will not convey any information.
4483 * NOW there is a trailer field, after the ack where it will safely be
4484 * ignored by mundanes, which indicates the maximum size packet this
4485 * host can swallow. */
4487 register struct rx_packet *optionalPacket; use to send ack (or null)
4488 int seq; Sequence number of the packet we are acking
4489 int serial; Serial number of the packet
4490 int pflags; Flags field from packet header
4491 int reason; Reason an acknowledge was prompted
4495 rxi_SendAck(register struct rx_call *call,
4496 register struct rx_packet *optionalPacket, int serial, int reason,
4499 struct rx_ackPacket *ap;
4500 register struct rx_packet *rqp;
4501 register struct rx_packet *nxp; /* For queue_Scan */
4502 register struct rx_packet *p;
4507 * Open the receive window once a thread starts reading packets
4509 if (call->rnext > 1) {
4510 call->rwind = rx_maxReceiveWindow;
4513 call->nHardAcks = 0;
4514 call->nSoftAcks = 0;
4515 if (call->rnext > call->lastAcked)
4516 call->lastAcked = call->rnext;
4520 rx_computelen(p, p->length); /* reset length, you never know */
4521 } /* where that's been... */
4522 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4523 /* We won't send the ack, but don't panic. */
4524 return optionalPacket;
4528 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4531 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL)) {
4532 if (!optionalPacket)
4534 return optionalPacket;
4536 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4537 if (rx_Contiguous(p) < templ) {
4538 if (!optionalPacket)
4540 return optionalPacket;
4545 /* MTUXXX failing to send an ack is very serious. We should */
4546 /* try as hard as possible to send even a partial ack; it's */
4547 /* better than nothing. */
4548 ap = (struct rx_ackPacket *)rx_DataOf(p);
4549 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4550 ap->reason = reason;
4552 /* The skew computation used to be bogus, I think it's better now. */
4553 /* We should start paying attention to skew. XXX */
4554 ap->serial = htonl(serial);
4555 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4557 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4558 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4560 /* No fear of running out of ack packet here because there can only be at most
4561 * one window full of unacknowledged packets. The window size must be constrained
4562 * to be less than the maximum ack size, of course. Also, an ack should always
4563 * fit into a single packet -- it should not ever be fragmented. */
4564 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4565 if (!rqp || !call->rq.next
4566 || (rqp->header.seq > (call->rnext + call->rwind))) {
4567 if (!optionalPacket)
4569 rxi_CallError(call, RX_CALL_DEAD);
4570 return optionalPacket;
4573 while (rqp->header.seq > call->rnext + offset)
4574 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4575 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4577 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4578 if (!optionalPacket)
4580 rxi_CallError(call, RX_CALL_DEAD);
4581 return optionalPacket;
4586 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4588 /* these are new for AFS 3.3 */
4589 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4590 templ = htonl(templ);
4591 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4592 templ = htonl(call->conn->peer->ifMTU);
4593 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4594 sizeof(afs_int32), &templ);
4596 /* new for AFS 3.4 */
4597 templ = htonl(call->rwind);
4598 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4599 sizeof(afs_int32), &templ);
4601 /* new for AFS 3.5 */
4602 templ = htonl(call->conn->peer->ifDgramPackets);
4603 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4604 sizeof(afs_int32), &templ);
4606 p->header.serviceId = call->conn->serviceId;
4607 p->header.cid = (call->conn->cid | call->channel);
4608 p->header.callNumber = *call->callNumber;
4610 p->header.securityIndex = call->conn->securityIndex;
4611 p->header.epoch = call->conn->epoch;
4612 p->header.type = RX_PACKET_TYPE_ACK;
4613 p->header.flags = RX_SLOW_START_OK;
4614 if (reason == RX_ACK_PING) {
4615 p->header.flags |= RX_REQUEST_ACK;
4617 clock_GetTime(&call->pingRequestTime);
4620 if (call->conn->type == RX_CLIENT_CONNECTION)
4621 p->header.flags |= RX_CLIENT_INITIATED;
4625 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u",
4626 ap->reason, ntohl(ap->previousPacket),
4627 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4629 for (offset = 0; offset < ap->nAcks; offset++)
4630 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4638 register int i, nbytes = p->length;
4640 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4641 if (nbytes <= p->wirevec[i].iov_len) {
4642 register int savelen, saven;
4644 savelen = p->wirevec[i].iov_len;
4646 p->wirevec[i].iov_len = nbytes;
4648 rxi_Send(call, p, istack);
4649 p->wirevec[i].iov_len = savelen;
4653 nbytes -= p->wirevec[i].iov_len;
4656 MUTEX_ENTER(&rx_stats_mutex);
4657 rx_stats.ackPacketsSent++;
4658 MUTEX_EXIT(&rx_stats_mutex);
4659 if (!optionalPacket)
4661 return optionalPacket; /* Return packet for re-use by caller */
4664 /* Send all of the packets in the list in single datagram */
4666 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4667 int istack, int moreFlag, struct clock *now,
4668 struct clock *retryTime, int resending)
4673 struct rx_connection *conn = call->conn;
4674 struct rx_peer *peer = conn->peer;
4676 MUTEX_ENTER(&peer->peer_lock);
4679 peer->reSends += len;
4680 MUTEX_ENTER(&rx_stats_mutex);
4681 rx_stats.dataPacketsSent += len;
4682 MUTEX_EXIT(&rx_stats_mutex);
4683 MUTEX_EXIT(&peer->peer_lock);
4685 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4689 /* Set the packet flags and schedule the resend events */
4690 /* Only request an ack for the last packet in the list */
4691 for (i = 0; i < len; i++) {
4692 list[i]->retryTime = *retryTime;
4693 if (list[i]->header.serial) {
4694 /* Exponentially backoff retry times */
4695 if (list[i]->backoff < MAXBACKOFF) {
4696 /* so it can't stay == 0 */
4697 list[i]->backoff = (list[i]->backoff << 1) + 1;
4700 clock_Addmsec(&(list[i]->retryTime),
4701 ((afs_uint32) list[i]->backoff) << 8);
4704 /* Wait a little extra for the ack on the last packet */
4705 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4706 clock_Addmsec(&(list[i]->retryTime), 400);
4709 /* Record the time sent */
4710 list[i]->timeSent = *now;
4712 /* Ask for an ack on retransmitted packets, on every other packet
4713 * if the peer doesn't support slow start. Ask for an ack on every
4714 * packet until the congestion window reaches the ack rate. */
4715 if (list[i]->header.serial) {
4717 MUTEX_ENTER(&rx_stats_mutex);
4718 rx_stats.dataPacketsReSent++;
4719 MUTEX_EXIT(&rx_stats_mutex);
4721 /* improved RTO calculation- not Karn */
4722 list[i]->firstSent = *now;
4723 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4724 || (!(call->flags & RX_CALL_SLOW_START_OK)
4725 && (list[i]->header.seq & 1)))) {
4730 MUTEX_ENTER(&peer->peer_lock);
4734 MUTEX_ENTER(&rx_stats_mutex);
4735 rx_stats.dataPacketsSent++;
4736 MUTEX_EXIT(&rx_stats_mutex);
4737 MUTEX_EXIT(&peer->peer_lock);
4739 /* Tag this packet as not being the last in this group,
4740 * for the receiver's benefit */
4741 if (i < len - 1 || moreFlag) {
4742 list[i]->header.flags |= RX_MORE_PACKETS;
4745 /* Install the new retransmit time for the packet, and
4746 * record the time sent */
4747 list[i]->timeSent = *now;
4751 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4754 /* Since we're about to send a data packet to the peer, it's
4755 * safe to nuke any scheduled end-of-packets ack */
4756 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4758 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4759 MUTEX_EXIT(&call->lock);
4761 rxi_SendPacketList(call, conn, list, len, istack);
4763 rxi_SendPacket(call, conn, list[0], istack);
4765 MUTEX_ENTER(&call->lock);
4766 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4768 /* Update last send time for this call (for keep-alive
4769 * processing), and for the connection (so that we can discover
4770 * idle connections) */
4771 conn->lastSendTime = call->lastSendTime = clock_Sec();
4774 /* When sending packets we need to follow these rules:
4775 * 1. Never send more than maxDgramPackets in a jumbogram.
4776 * 2. Never send a packet with more than two iovecs in a jumbogram.
4777 * 3. Never send a retransmitted packet in a jumbogram.
4778 * 4. Never send more than cwind/4 packets in a jumbogram
4779 * We always keep the last list we should have sent so we
4780 * can set the RX_MORE_PACKETS flags correctly.
4783 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4784 int istack, struct clock *now, struct clock *retryTime,
4787 int i, cnt, lastCnt = 0;
4788 struct rx_packet **listP, **lastP = 0;
4789 struct rx_peer *peer = call->conn->peer;
4790 int morePackets = 0;
4792 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4793 /* Does the current packet force us to flush the current list? */
4795 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4796 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4798 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
4800 /* If the call enters an error state stop sending, or if
4801 * we entered congestion recovery mode, stop sending */
4802 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4810 /* Add the current packet to the list if it hasn't been acked.
4811 * Otherwise adjust the list pointer to skip the current packet. */
4812 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
4814 /* Do we need to flush the list? */
4815 if (cnt >= (int)peer->maxDgramPackets
4816 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
4817 || list[i]->header.serial
4818 || list[i]->length != RX_JUMBOBUFFERSIZE) {
4820 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
4821 retryTime, resending);
4822 /* If the call enters an error state stop sending, or if
4823 * we entered congestion recovery mode, stop sending */
4825 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4830 listP = &list[i + 1];
4835 osi_Panic("rxi_SendList error");
4837 listP = &list[i + 1];
4841 /* Send the whole list when the call is in receive mode, when
4842 * the call is in eof mode, when we are in fast recovery mode,
4843 * and when we have the last packet */
4844 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
4845 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
4846 || (call->flags & RX_CALL_FAST_RECOVER)) {
4847 /* Check for the case where the current list contains
4848 * an acked packet. Since we always send retransmissions
4849 * in a separate packet, we only need to check the first
4850 * packet in the list */
4851 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
4855 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
4856 retryTime, resending);
4857 /* If the call enters an error state stop sending, or if
4858 * we entered congestion recovery mode, stop sending */
4859 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4863 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
4866 } else if (lastCnt > 0) {
4867 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
4872 #ifdef RX_ENABLE_LOCKS
4873 /* Call rxi_Start, below, but with the call lock held. */
4875 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
4876 void *arg1, int istack)
4878 MUTEX_ENTER(&call->lock);
4879 rxi_Start(event, call, arg1, istack);
4880 MUTEX_EXIT(&call->lock);
4882 #endif /* RX_ENABLE_LOCKS */
4884 /* This routine is called when new packets are readied for
4885 * transmission and when retransmission may be necessary, or when the
4886 * transmission window or burst count are favourable. This should be
4887 * better optimized for new packets, the usual case, now that we've
4888 * got rid of queues of send packets. XXXXXXXXXXX */
4890 rxi_Start(struct rxevent *event, register struct rx_call *call,
4891 void *arg1, int istack)
4893 struct rx_packet *p;
4894 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
4895 struct rx_peer *peer = call->conn->peer;
4896 struct clock now, retryTime;
4900 struct rx_packet **xmitList;
4903 /* If rxi_Start is being called as a result of a resend event,
4904 * then make sure that the event pointer is removed from the call
4905 * structure, since there is no longer a per-call retransmission
4907 if (event && event == call->resendEvent) {
4908 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
4909 call->resendEvent = NULL;
4911 if (queue_IsEmpty(&call->tq)) {
4915 /* Timeouts trigger congestion recovery */
4916 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4917 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4918 /* someone else is waiting to start recovery */
4921 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4922 while (call->flags & RX_CALL_TQ_BUSY) {
4923 call->flags |= RX_CALL_TQ_WAIT;
4924 #ifdef RX_ENABLE_LOCKS
4925 CV_WAIT(&call->cv_tq, &call->lock);
4926 #else /* RX_ENABLE_LOCKS */
4927 osi_rxSleep(&call->tq);
4928 #endif /* RX_ENABLE_LOCKS */
4930 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4931 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4932 call->flags |= RX_CALL_FAST_RECOVER;
4933 if (peer->maxDgramPackets > 1) {
4934 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4936 call->MTU = MIN(peer->natMTU, peer->maxMTU);
4938 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4939 call->nDgramPackets = 1;
4941 call->nextCwind = 1;
4944 MUTEX_ENTER(&peer->peer_lock);
4945 peer->MTU = call->MTU;
4946 peer->cwind = call->cwind;
4947 peer->nDgramPackets = 1;
4949 call->congestSeq = peer->congestSeq;
4950 MUTEX_EXIT(&peer->peer_lock);
4951 /* Clear retry times on packets. Otherwise, it's possible for
4952 * some packets in the queue to force resends at rates faster
4953 * than recovery rates.
4955 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
4956 if (!(p->flags & RX_PKTFLAG_ACKED)) {
4957 clock_Zero(&p->retryTime);
4962 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4963 MUTEX_ENTER(&rx_stats_mutex);
4964 rx_tq_debug.rxi_start_in_error++;
4965 MUTEX_EXIT(&rx_stats_mutex);
4970 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
4971 /* Get clock to compute the re-transmit time for any packets
4972 * in this burst. Note, if we back off, it's reasonable to
4973 * back off all of the packets in the same manner, even if
4974 * some of them have been retransmitted more times than more
4975 * recent additions */
4976 clock_GetTime(&now);
4977 retryTime = now; /* initialize before use */
4978 MUTEX_ENTER(&peer->peer_lock);
4979 clock_Add(&retryTime, &peer->timeout);
4980 MUTEX_EXIT(&peer->peer_lock);
4982 /* Send (or resend) any packets that need it, subject to
4983 * window restrictions and congestion burst control
4984 * restrictions. Ask for an ack on the last packet sent in
4985 * this burst. For now, we're relying upon the window being
4986 * considerably bigger than the largest number of packets that
4987 * are typically sent at once by one initial call to
4988 * rxi_Start. This is probably bogus (perhaps we should ask
4989 * for an ack when we're half way through the current
4990 * window?). Also, for non file transfer applications, this
4991 * may end up asking for an ack for every packet. Bogus. XXXX
4994 * But check whether we're here recursively, and let the other guy
4997 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4998 if (!(call->flags & RX_CALL_TQ_BUSY)) {
4999 call->flags |= RX_CALL_TQ_BUSY;
5001 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5003 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5004 call->flags &= ~RX_CALL_NEED_START;
5005 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5007 maxXmitPackets = MIN(call->twind, call->cwind);
5008 xmitList = (struct rx_packet **)
5009 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5010 if (xmitList == NULL)
5011 osi_Panic("rxi_Start, failed to allocate xmit list");
5012 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5013 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5014 /* We shouldn't be sending packets if a thread is waiting
5015 * to initiate congestion recovery */
5019 && (call->flags & RX_CALL_FAST_RECOVER)) {
5020 /* Only send one packet during fast recovery */
5023 if ((p->flags & RX_PKTFLAG_FREE)
5024 || (!queue_IsEnd(&call->tq, nxp)
5025 && (nxp->flags & RX_PKTFLAG_FREE))
5026 || (p == (struct rx_packet *)&rx_freePacketQueue)
5027 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5028 osi_Panic("rxi_Start: xmit queue clobbered");
5030 if (p->flags & RX_PKTFLAG_ACKED) {
5031 MUTEX_ENTER(&rx_stats_mutex);
5032 rx_stats.ignoreAckedPacket++;
5033 MUTEX_EXIT(&rx_stats_mutex);
5034 continue; /* Ignore this packet if it has been acknowledged */
5037 /* Turn off all flags except these ones, which are the same
5038 * on each transmission */
5039 p->header.flags &= RX_PRESET_FLAGS;
5041 if (p->header.seq >=
5042 call->tfirst + MIN((int)call->twind,
5043 (int)(call->nSoftAcked +
5045 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5046 /* Note: if we're waiting for more window space, we can
5047 * still send retransmits; hence we don't return here, but
5048 * break out to schedule a retransmit event */
5049 dpf(("call %d waiting for window",
5050 *(call->callNumber)));
5054 /* Transmit the packet if it needs to be sent. */
5055 if (!clock_Lt(&now, &p->retryTime)) {
5056 if (nXmitPackets == maxXmitPackets) {
5057 rxi_SendXmitList(call, xmitList, nXmitPackets,
5058 istack, &now, &retryTime,
5060 osi_Free(xmitList, maxXmitPackets *
5061 sizeof(struct rx_packet *));
5064 xmitList[nXmitPackets++] = p;
5068 /* xmitList now hold pointers to all of the packets that are
5069 * ready to send. Now we loop to send the packets */
5070 if (nXmitPackets > 0) {
5071 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5072 &now, &retryTime, resending);
5075 maxXmitPackets * sizeof(struct rx_packet *));
5077 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5079 * TQ references no longer protected by this flag; they must remain
5080 * protected by the global lock.
5082 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5083 call->flags &= ~RX_CALL_TQ_BUSY;
5084 if (call->flags & RX_CALL_TQ_WAIT) {
5085 call->flags &= ~RX_CALL_TQ_WAIT;
5086 #ifdef RX_ENABLE_LOCKS
5087 CV_BROADCAST(&call->cv_tq);
5088 #else /* RX_ENABLE_LOCKS */
5089 osi_rxWakeup(&call->tq);
5090 #endif /* RX_ENABLE_LOCKS */
5095 /* We went into the error state while sending packets. Now is
5096 * the time to reset the call. This will also inform the using
5097 * process that the call is in an error state.
5099 MUTEX_ENTER(&rx_stats_mutex);
5100 rx_tq_debug.rxi_start_aborted++;
5101 MUTEX_EXIT(&rx_stats_mutex);
5102 call->flags &= ~RX_CALL_TQ_BUSY;
5103 if (call->flags & RX_CALL_TQ_WAIT) {
5104 call->flags &= ~RX_CALL_TQ_WAIT;
5105 #ifdef RX_ENABLE_LOCKS
5106 CV_BROADCAST(&call->cv_tq);
5107 #else /* RX_ENABLE_LOCKS */
5108 osi_rxWakeup(&call->tq);
5109 #endif /* RX_ENABLE_LOCKS */
5111 rxi_CallError(call, call->error);
5114 #ifdef RX_ENABLE_LOCKS
5115 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5116 register int missing;
5117 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5118 /* Some packets have received acks. If they all have, we can clear
5119 * the transmit queue.
5122 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5123 if (p->header.seq < call->tfirst
5124 && (p->flags & RX_PKTFLAG_ACKED)) {
5131 call->flags |= RX_CALL_TQ_CLEARME;
5133 #endif /* RX_ENABLE_LOCKS */
5134 /* Don't bother doing retransmits if the TQ is cleared. */
5135 if (call->flags & RX_CALL_TQ_CLEARME) {
5136 rxi_ClearTransmitQueue(call, 1);
5138 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5141 /* Always post a resend event, if there is anything in the
5142 * queue, and resend is possible. There should be at least
5143 * one unacknowledged packet in the queue ... otherwise none
5144 * of these packets should be on the queue in the first place.
5146 if (call->resendEvent) {
5147 /* Cancel the existing event and post a new one */
5148 rxevent_Cancel(call->resendEvent, call,
5149 RX_CALL_REFCOUNT_RESEND);
5152 /* The retry time is the retry time on the first unacknowledged
5153 * packet inside the current window */
5155 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5156 /* Don't set timers for packets outside the window */
5157 if (p->header.seq >= call->tfirst + call->twind) {
5161 if (!(p->flags & RX_PKTFLAG_ACKED)
5162 && !clock_IsZero(&p->retryTime)) {
5164 retryTime = p->retryTime;
5169 /* Post a new event to re-run rxi_Start when retries may be needed */
5170 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5171 #ifdef RX_ENABLE_LOCKS
5172 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5174 rxevent_Post2(&retryTime, rxi_StartUnlocked,
5175 (void *)call, 0, istack);
5176 #else /* RX_ENABLE_LOCKS */
5178 rxevent_Post2(&retryTime, rxi_Start, (void *)call,
5180 #endif /* RX_ENABLE_LOCKS */
5183 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5184 } while (call->flags & RX_CALL_NEED_START);
5186 * TQ references no longer protected by this flag; they must remain
5187 * protected by the global lock.
5189 call->flags &= ~RX_CALL_TQ_BUSY;
5190 if (call->flags & RX_CALL_TQ_WAIT) {
5191 call->flags &= ~RX_CALL_TQ_WAIT;
5192 #ifdef RX_ENABLE_LOCKS
5193 CV_BROADCAST(&call->cv_tq);
5194 #else /* RX_ENABLE_LOCKS */
5195 osi_rxWakeup(&call->tq);
5196 #endif /* RX_ENABLE_LOCKS */
5199 call->flags |= RX_CALL_NEED_START;
5201 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5203 if (call->resendEvent) {
5204 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5209 /* Also adjusts the keep alive parameters for the call, to reflect
5210 * that we have just sent a packet (so keep alives aren't sent
5213 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5216 register struct rx_connection *conn = call->conn;
5218 /* Stamp each packet with the user supplied status */
5219 p->header.userStatus = call->localStatus;
5221 /* Allow the security object controlling this call's security to
5222 * make any last-minute changes to the packet */
5223 RXS_SendPacket(conn->securityObject, call, p);
5225 /* Since we're about to send SOME sort of packet to the peer, it's
5226 * safe to nuke any scheduled end-of-packets ack */
5227 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5229 /* Actually send the packet, filling in more connection-specific fields */
5230 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5231 MUTEX_EXIT(&call->lock);
5232 rxi_SendPacket(call, conn, p, istack);
5233 MUTEX_ENTER(&call->lock);
5234 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5236 /* Update last send time for this call (for keep-alive
5237 * processing), and for the connection (so that we can discover
5238 * idle connections) */
5239 conn->lastSendTime = call->lastSendTime = clock_Sec();
5243 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5244 * that things are fine. Also called periodically to guarantee that nothing
5245 * falls through the cracks (e.g. (error + dally) connections have keepalive
5246 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5248 * haveCTLock Set if calling from rxi_ReapConnections
5250 #ifdef RX_ENABLE_LOCKS
5252 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5253 #else /* RX_ENABLE_LOCKS */
5255 rxi_CheckCall(register struct rx_call *call)
5256 #endif /* RX_ENABLE_LOCKS */
5258 register struct rx_connection *conn = call->conn;
5260 afs_uint32 deadTime;
5262 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5263 if (call->flags & RX_CALL_TQ_BUSY) {
5264 /* Call is active and will be reset by rxi_Start if it's
5265 * in an error state.
5270 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5272 (((afs_uint32) conn->secondsUntilDead << 10) +
5273 ((afs_uint32) conn->peer->rtt >> 3) +
5274 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5276 /* These are computed to the second (+- 1 second). But that's
5277 * good enough for these values, which should be a significant
5278 * number of seconds. */
5279 if (now > (call->lastReceiveTime + deadTime)) {
5280 if (call->state == RX_STATE_ACTIVE) {
5281 rxi_CallError(call, RX_CALL_DEAD);
5284 #ifdef RX_ENABLE_LOCKS
5285 /* Cancel pending events */
5286 rxevent_Cancel(call->delayedAckEvent, call,
5287 RX_CALL_REFCOUNT_DELAY);
5288 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5289 rxevent_Cancel(call->keepAliveEvent, call,
5290 RX_CALL_REFCOUNT_ALIVE);
5291 if (call->refCount == 0) {
5292 rxi_FreeCall(call, haveCTLock);
5296 #else /* RX_ENABLE_LOCKS */
5299 #endif /* RX_ENABLE_LOCKS */
5301 /* Non-active calls are destroyed if they are not responding
5302 * to pings; active calls are simply flagged in error, so the
5303 * attached process can die reasonably gracefully. */
5305 /* see if we have a non-activity timeout */
5306 if (call->startWait && conn->idleDeadTime
5307 && ((call->startWait + conn->idleDeadTime) < now)) {
5308 if (call->state == RX_STATE_ACTIVE) {
5309 rxi_CallError(call, RX_CALL_TIMEOUT);
5313 /* see if we have a hard timeout */
5314 if (conn->hardDeadTime
5315 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5316 if (call->state == RX_STATE_ACTIVE)
5317 rxi_CallError(call, RX_CALL_TIMEOUT);
5324 /* When a call is in progress, this routine is called occasionally to
5325 * make sure that some traffic has arrived (or been sent to) the peer.
5326 * If nothing has arrived in a reasonable amount of time, the call is
5327 * declared dead; if nothing has been sent for a while, we send a
5328 * keep-alive packet (if we're actually trying to keep the call alive)
5331 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5334 struct rx_connection *conn;
5337 MUTEX_ENTER(&call->lock);
5338 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5339 if (event == call->keepAliveEvent)
5340 call->keepAliveEvent = NULL;
5343 #ifdef RX_ENABLE_LOCKS
5344 if (rxi_CheckCall(call, 0)) {
5345 MUTEX_EXIT(&call->lock);
5348 #else /* RX_ENABLE_LOCKS */
5349 if (rxi_CheckCall(call))
5351 #endif /* RX_ENABLE_LOCKS */
5353 /* Don't try to keep alive dallying calls */
5354 if (call->state == RX_STATE_DALLY) {
5355 MUTEX_EXIT(&call->lock);
5360 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5361 /* Don't try to send keepalives if there is unacknowledged data */
5362 /* the rexmit code should be good enough, this little hack
5363 * doesn't quite work XXX */
5364 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5366 rxi_ScheduleKeepAliveEvent(call);
5367 MUTEX_EXIT(&call->lock);
5372 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5374 if (!call->keepAliveEvent) {
5376 clock_GetTime(&when);
5377 when.sec += call->conn->secondsUntilPing;
5378 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5379 call->keepAliveEvent =
5380 rxevent_Post(&when, rxi_KeepAliveEvent, call, 0);
5384 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5386 rxi_KeepAliveOn(register struct rx_call *call)
5388 /* Pretend last packet received was received now--i.e. if another
5389 * packet isn't received within the keep alive time, then the call
5390 * will die; Initialize last send time to the current time--even
5391 * if a packet hasn't been sent yet. This will guarantee that a
5392 * keep-alive is sent within the ping time */
5393 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5394 rxi_ScheduleKeepAliveEvent(call);
5397 /* This routine is called to send connection abort messages
5398 * that have been delayed to throttle looping clients. */
5400 rxi_SendDelayedConnAbort(struct rxevent *event,
5401 register struct rx_connection *conn, char *dummy)
5404 struct rx_packet *packet;
5406 MUTEX_ENTER(&conn->conn_data_lock);
5407 conn->delayedAbortEvent = NULL;
5408 error = htonl(conn->error);
5410 MUTEX_EXIT(&conn->conn_data_lock);
5411 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5414 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5415 RX_PACKET_TYPE_ABORT, (char *)&error,
5417 rxi_FreePacket(packet);
5421 /* This routine is called to send call abort messages
5422 * that have been delayed to throttle looping clients. */
5424 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5428 struct rx_packet *packet;
5430 MUTEX_ENTER(&call->lock);
5431 call->delayedAbortEvent = NULL;
5432 error = htonl(call->error);
5434 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5437 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5438 (char *)&error, sizeof(error), 0);
5439 rxi_FreePacket(packet);
5441 MUTEX_EXIT(&call->lock);
5444 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5445 * seconds) to ask the client to authenticate itself. The routine
5446 * issues a challenge to the client, which is obtained from the
5447 * security object associated with the connection */
5449 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5450 void *arg1, int tries)
5452 conn->challengeEvent = NULL;
5453 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5454 register struct rx_packet *packet;
5458 /* We've failed to authenticate for too long.
5459 * Reset any calls waiting for authentication;
5460 * they are all in RX_STATE_PRECALL.
5464 MUTEX_ENTER(&conn->conn_call_lock);
5465 for (i = 0; i < RX_MAXCALLS; i++) {
5466 struct rx_call *call = conn->call[i];
5468 MUTEX_ENTER(&call->lock);
5469 if (call->state == RX_STATE_PRECALL) {
5470 rxi_CallError(call, RX_CALL_DEAD);
5471 rxi_SendCallAbort(call, NULL, 0, 0);
5473 MUTEX_EXIT(&call->lock);
5476 MUTEX_EXIT(&conn->conn_call_lock);
5480 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5482 /* If there's no packet available, do this later. */
5483 RXS_GetChallenge(conn->securityObject, conn, packet);
5484 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5485 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5486 rxi_FreePacket(packet);
5488 clock_GetTime(&when);
5489 when.sec += RX_CHALLENGE_TIMEOUT;
5490 conn->challengeEvent =
5491 rxevent_Post2(&when, rxi_ChallengeEvent, conn, 0,
5496 /* Call this routine to start requesting the client to authenticate
5497 * itself. This will continue until authentication is established,
5498 * the call times out, or an invalid response is returned. The
5499 * security object associated with the connection is asked to create
5500 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5501 * defined earlier. */
5503 rxi_ChallengeOn(register struct rx_connection *conn)
5505 if (!conn->challengeEvent) {
5506 RXS_CreateChallenge(conn->securityObject, conn);
5507 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5512 /* Compute round trip time of the packet provided, in *rttp.
5515 /* rxi_ComputeRoundTripTime is called with peer locked. */
5516 /* sentp and/or peer may be null */
5518 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5519 register struct clock *sentp,
5520 register struct rx_peer *peer)
5522 struct clock thisRtt, *rttp = &thisRtt;
5524 register int rtt_timeout;
5526 clock_GetTime(rttp);
5528 if (clock_Lt(rttp, sentp)) {
5530 return; /* somebody set the clock back, don't count this time. */
5532 clock_Sub(rttp, sentp);
5533 MUTEX_ENTER(&rx_stats_mutex);
5534 if (clock_Lt(rttp, &rx_stats.minRtt))
5535 rx_stats.minRtt = *rttp;
5536 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5537 if (rttp->sec > 60) {
5538 MUTEX_EXIT(&rx_stats_mutex);
5539 return; /* somebody set the clock ahead */
5541 rx_stats.maxRtt = *rttp;
5543 clock_Add(&rx_stats.totalRtt, rttp);
5544 rx_stats.nRttSamples++;
5545 MUTEX_EXIT(&rx_stats_mutex);
5547 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5549 /* Apply VanJacobson round-trip estimations */
5554 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5555 * srtt is stored as fixed point with 3 bits after the binary
5556 * point (i.e., scaled by 8). The following magic is
5557 * equivalent to the smoothing algorithm in rfc793 with an
5558 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5559 * srtt*8 = srtt*8 + rtt - srtt
5560 * srtt = srtt + rtt/8 - srtt/8
5563 delta = MSEC(rttp) - (peer->rtt >> 3);
5567 * We accumulate a smoothed rtt variance (actually, a smoothed
5568 * mean difference), then set the retransmit timer to smoothed
5569 * rtt + 4 times the smoothed variance (was 2x in van's original
5570 * paper, but 4x works better for me, and apparently for him as
5572 * rttvar is stored as
5573 * fixed point with 2 bits after the binary point (scaled by
5574 * 4). The following is equivalent to rfc793 smoothing with
5575 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5576 * replaces rfc793's wired-in beta.
5577 * dev*4 = dev*4 + (|actual - expected| - dev)
5583 delta -= (peer->rtt_dev >> 2);
5584 peer->rtt_dev += delta;
5586 /* I don't have a stored RTT so I start with this value. Since I'm
5587 * probably just starting a call, and will be pushing more data down
5588 * this, I expect congestion to increase rapidly. So I fudge a
5589 * little, and I set deviance to half the rtt. In practice,
5590 * deviance tends to approach something a little less than
5591 * half the smoothed rtt. */
5592 peer->rtt = (MSEC(rttp) << 3) + 8;
5593 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5595 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5596 * the other of these connections is usually in a user process, and can
5597 * be switched and/or swapped out. So on fast, reliable networks, the
5598 * timeout would otherwise be too short.
5600 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5601 clock_Zero(&(peer->timeout));
5602 clock_Addmsec(&(peer->timeout), rtt_timeout);
5604 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)));
5608 /* Find all server connections that have not been active for a long time, and
5611 rxi_ReapConnections(void)
5614 clock_GetTime(&now);
5616 /* Find server connection structures that haven't been used for
5617 * greater than rx_idleConnectionTime */
5619 struct rx_connection **conn_ptr, **conn_end;
5620 int i, havecalls = 0;
5621 MUTEX_ENTER(&rx_connHashTable_lock);
5622 for (conn_ptr = &rx_connHashTable[0], conn_end =
5623 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5625 struct rx_connection *conn, *next;
5626 struct rx_call *call;
5630 for (conn = *conn_ptr; conn; conn = next) {
5631 /* XXX -- Shouldn't the connection be locked? */
5634 for (i = 0; i < RX_MAXCALLS; i++) {
5635 call = conn->call[i];
5638 MUTEX_ENTER(&call->lock);
5639 #ifdef RX_ENABLE_LOCKS
5640 result = rxi_CheckCall(call, 1);
5641 #else /* RX_ENABLE_LOCKS */
5642 result = rxi_CheckCall(call);
5643 #endif /* RX_ENABLE_LOCKS */
5644 MUTEX_EXIT(&call->lock);
5646 /* If CheckCall freed the call, it might
5647 * have destroyed the connection as well,
5648 * which screws up the linked lists.
5654 if (conn->type == RX_SERVER_CONNECTION) {
5655 /* This only actually destroys the connection if
5656 * there are no outstanding calls */
5657 MUTEX_ENTER(&conn->conn_data_lock);
5658 if (!havecalls && !conn->refCount
5659 && ((conn->lastSendTime + rx_idleConnectionTime) <
5661 conn->refCount++; /* it will be decr in rx_DestroyConn */
5662 MUTEX_EXIT(&conn->conn_data_lock);
5663 #ifdef RX_ENABLE_LOCKS
5664 rxi_DestroyConnectionNoLock(conn);
5665 #else /* RX_ENABLE_LOCKS */
5666 rxi_DestroyConnection(conn);
5667 #endif /* RX_ENABLE_LOCKS */
5669 #ifdef RX_ENABLE_LOCKS
5671 MUTEX_EXIT(&conn->conn_data_lock);
5673 #endif /* RX_ENABLE_LOCKS */
5677 #ifdef RX_ENABLE_LOCKS
5678 while (rx_connCleanup_list) {
5679 struct rx_connection *conn;
5680 conn = rx_connCleanup_list;
5681 rx_connCleanup_list = rx_connCleanup_list->next;
5682 MUTEX_EXIT(&rx_connHashTable_lock);
5683 rxi_CleanupConnection(conn);
5684 MUTEX_ENTER(&rx_connHashTable_lock);
5686 MUTEX_EXIT(&rx_connHashTable_lock);
5687 #endif /* RX_ENABLE_LOCKS */
5690 /* Find any peer structures that haven't been used (haven't had an
5691 * associated connection) for greater than rx_idlePeerTime */
5693 struct rx_peer **peer_ptr, **peer_end;
5695 MUTEX_ENTER(&rx_rpc_stats);
5696 MUTEX_ENTER(&rx_peerHashTable_lock);
5697 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5698 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5700 struct rx_peer *peer, *next, *prev;
5701 for (prev = peer = *peer_ptr; peer; peer = next) {
5703 code = MUTEX_TRYENTER(&peer->peer_lock);
5704 if ((code) && (peer->refCount == 0)
5705 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5706 rx_interface_stat_p rpc_stat, nrpc_stat;
5708 MUTEX_EXIT(&peer->peer_lock);
5709 MUTEX_DESTROY(&peer->peer_lock);
5711 (&peer->rpcStats, rpc_stat, nrpc_stat,
5712 rx_interface_stat)) {
5713 unsigned int num_funcs;
5716 queue_Remove(&rpc_stat->queue_header);
5717 queue_Remove(&rpc_stat->all_peers);
5718 num_funcs = rpc_stat->stats[0].func_total;
5720 sizeof(rx_interface_stat_t) +
5721 rpc_stat->stats[0].func_total *
5722 sizeof(rx_function_entry_v1_t);
5724 rxi_Free(rpc_stat, space);
5725 rxi_rpc_peer_stat_cnt -= num_funcs;
5728 MUTEX_ENTER(&rx_stats_mutex);
5729 rx_stats.nPeerStructs--;
5730 MUTEX_EXIT(&rx_stats_mutex);
5731 if (peer == *peer_ptr) {
5738 MUTEX_EXIT(&peer->peer_lock);
5744 MUTEX_EXIT(&rx_peerHashTable_lock);
5745 MUTEX_EXIT(&rx_rpc_stats);
5748 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5749 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5750 * GC, just below. Really, we shouldn't have to keep moving packets from
5751 * one place to another, but instead ought to always know if we can
5752 * afford to hold onto a packet in its particular use. */
5753 MUTEX_ENTER(&rx_freePktQ_lock);
5754 if (rx_waitingForPackets) {
5755 rx_waitingForPackets = 0;
5756 #ifdef RX_ENABLE_LOCKS
5757 CV_BROADCAST(&rx_waitingForPackets_cv);
5759 osi_rxWakeup(&rx_waitingForPackets);
5762 MUTEX_EXIT(&rx_freePktQ_lock);
5764 now.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5765 rxevent_Post(&now, rxi_ReapConnections, 0, 0);
5769 /* rxs_Release - This isn't strictly necessary but, since the macro name from
5770 * rx.h is sort of strange this is better. This is called with a security
5771 * object before it is discarded. Each connection using a security object has
5772 * its own refcount to the object so it won't actually be freed until the last
5773 * connection is destroyed.
5775 * This is the only rxs module call. A hold could also be written but no one
5779 rxs_Release(struct rx_securityClass *aobj)
5781 return RXS_Close(aobj);
5785 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
5786 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
5787 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
5788 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
5790 /* Adjust our estimate of the transmission rate to this peer, given
5791 * that the packet p was just acked. We can adjust peer->timeout and
5792 * call->twind. Pragmatically, this is called
5793 * only with packets of maximal length.
5794 * Called with peer and call locked.
5798 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
5799 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
5801 afs_int32 xferSize, xferMs;
5802 register afs_int32 minTime;
5805 /* Count down packets */
5806 if (peer->rateFlag > 0)
5808 /* Do nothing until we're enabled */
5809 if (peer->rateFlag != 0)
5814 /* Count only when the ack seems legitimate */
5815 switch (ackReason) {
5816 case RX_ACK_REQUESTED:
5818 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
5822 case RX_ACK_PING_RESPONSE:
5823 if (p) /* want the response to ping-request, not data send */
5825 clock_GetTime(&newTO);
5826 if (clock_Gt(&newTO, &call->pingRequestTime)) {
5827 clock_Sub(&newTO, &call->pingRequestTime);
5828 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
5832 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
5839 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));
5841 /* Track only packets that are big enough. */
5842 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
5846 /* absorb RTT data (in milliseconds) for these big packets */
5847 if (peer->smRtt == 0) {
5848 peer->smRtt = xferMs;
5850 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
5855 if (peer->countDown) {
5859 peer->countDown = 10; /* recalculate only every so often */
5861 /* In practice, we can measure only the RTT for full packets,
5862 * because of the way Rx acks the data that it receives. (If it's
5863 * smaller than a full packet, it often gets implicitly acked
5864 * either by the call response (from a server) or by the next call
5865 * (from a client), and either case confuses transmission times
5866 * with processing times.) Therefore, replace the above
5867 * more-sophisticated processing with a simpler version, where the
5868 * smoothed RTT is kept for full-size packets, and the time to
5869 * transmit a windowful of full-size packets is simply RTT *
5870 * windowSize. Again, we take two steps:
5871 - ensure the timeout is large enough for a single packet's RTT;
5872 - ensure that the window is small enough to fit in the desired timeout.*/
5874 /* First, the timeout check. */
5875 minTime = peer->smRtt;
5876 /* Get a reasonable estimate for a timeout period */
5878 newTO.sec = minTime / 1000;
5879 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
5881 /* Increase the timeout period so that we can always do at least
5882 * one packet exchange */
5883 if (clock_Gt(&newTO, &peer->timeout)) {
5885 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));
5887 peer->timeout = newTO;
5890 /* Now, get an estimate for the transmit window size. */
5891 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
5892 /* Now, convert to the number of full packets that could fit in a
5893 * reasonable fraction of that interval */
5894 minTime /= (peer->smRtt << 1);
5895 xferSize = minTime; /* (make a copy) */
5897 /* Now clamp the size to reasonable bounds. */
5900 else if (minTime > rx_Window)
5901 minTime = rx_Window;
5902 /* if (minTime != peer->maxWindow) {
5903 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
5904 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
5905 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
5907 peer->maxWindow = minTime;
5908 elide... call->twind = minTime;
5912 /* Cut back on the peer timeout if it had earlier grown unreasonably.
5913 * Discern this by calculating the timeout necessary for rx_Window
5915 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
5916 /* calculate estimate for transmission interval in milliseconds */
5917 minTime = rx_Window * peer->smRtt;
5918 if (minTime < 1000) {
5919 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
5920 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
5921 peer->timeout.usec, peer->smRtt, peer->packetSize));
5923 newTO.sec = 0; /* cut back on timeout by half a second */
5924 newTO.usec = 500000;
5925 clock_Sub(&peer->timeout, &newTO);
5930 } /* end of rxi_ComputeRate */
5931 #endif /* ADAPT_WINDOW */
5939 /* Don't call this debugging routine directly; use dpf */
5941 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
5942 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
5946 clock_GetTime(&now);
5947 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
5948 (unsigned int)now.usec / 1000);
5949 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
5957 * This function is used to process the rx_stats structure that is local
5958 * to a process as well as an rx_stats structure received from a remote
5959 * process (via rxdebug). Therefore, it needs to do minimal version
5963 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
5964 afs_int32 freePackets, char version)
5968 if (size != sizeof(struct rx_stats)) {
5970 "Unexpected size of stats structure: was %d, expected %d\n",
5971 size, sizeof(struct rx_stats));
5974 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
5977 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
5978 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
5979 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
5980 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
5981 s->specialPktAllocFailures);
5983 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
5984 s->receivePktAllocFailures, s->sendPktAllocFailures,
5985 s->specialPktAllocFailures);
5989 " greedy %d, " "bogusReads %d (last from host %x), "
5990 "noPackets %d, " "noBuffers %d, " "selects %d, "
5991 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
5992 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
5993 s->selects, s->sendSelects);
5995 fprintf(file, " packets read: ");
5996 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
5997 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
5999 fprintf(file, "\n");
6002 " other read counters: data %d, " "ack %d, " "dup %d "
6003 "spurious %d " "dally %d\n", s->dataPacketsRead,
6004 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6005 s->ignorePacketDally);
6007 fprintf(file, " packets sent: ");
6008 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6009 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6011 fprintf(file, "\n");
6014 " other send counters: ack %d, " "data %d (not resends), "
6015 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6016 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6017 s->dataPacketsPushed, s->ignoreAckedPacket);
6020 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6021 s->netSendFailures, (int)s->fatalErrors);
6023 if (s->nRttSamples) {
6024 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6025 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6027 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6028 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6032 " %d server connections, " "%d client connections, "
6033 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6034 s->nServerConns, s->nClientConns, s->nPeerStructs,
6035 s->nCallStructs, s->nFreeCallStructs);
6037 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6038 fprintf(file, " %d clock updates\n", clock_nUpdates);
6043 /* for backward compatibility */
6045 rx_PrintStats(FILE * file)
6047 MUTEX_ENTER(&rx_stats_mutex);
6048 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6050 MUTEX_EXIT(&rx_stats_mutex);
6054 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6056 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6057 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6058 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6061 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6062 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6063 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6066 " Packet size %d, " "max in packet skew %d, "
6067 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6068 (int)peer->outPacketSkew);
6071 #ifdef AFS_PTHREAD_ENV
6073 * This mutex protects the following static variables:
6077 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6078 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6080 #define LOCK_RX_DEBUG
6081 #define UNLOCK_RX_DEBUG
6082 #endif /* AFS_PTHREAD_ENV */
6085 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6086 u_char type, void *inputData, size_t inputLength,
6087 void *outputData, size_t outputLength)
6089 static afs_int32 counter = 100;
6091 struct rx_header theader;
6093 register afs_int32 code;
6095 struct sockaddr_in taddr, faddr;
6100 endTime = time(0) + 20; /* try for 20 seconds */
6104 tp = &tbuffer[sizeof(struct rx_header)];
6105 taddr.sin_family = AF_INET;
6106 taddr.sin_port = remotePort;
6107 taddr.sin_addr.s_addr = remoteAddr;
6108 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6109 taddr.sin_len = sizeof(struct sockaddr_in);
6112 memset(&theader, 0, sizeof(theader));
6113 theader.epoch = htonl(999);
6115 theader.callNumber = htonl(counter);
6118 theader.type = type;
6119 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6120 theader.serviceId = 0;
6122 memcpy(tbuffer, &theader, sizeof(theader));
6123 memcpy(tp, inputData, inputLength);
6125 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6126 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6128 /* see if there's a packet available */
6130 FD_SET(socket, &imask);
6133 code = select(socket + 1, &imask, 0, 0, &tv);
6134 if (code == 1 && FD_ISSET(socket, &imask)) {
6135 /* now receive a packet */
6136 faddrLen = sizeof(struct sockaddr_in);
6138 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6139 (struct sockaddr *)&faddr, &faddrLen);
6142 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6143 if (counter == ntohl(theader.callNumber))
6148 /* see if we've timed out */
6149 if (endTime < time(0))
6152 code -= sizeof(struct rx_header);
6153 if (code > outputLength)
6154 code = outputLength;
6155 memcpy(outputData, tp, code);
6160 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6161 afs_uint16 remotePort, struct rx_debugStats * stat,
6162 afs_uint32 * supportedValues)
6164 struct rx_debugIn in;
6167 *supportedValues = 0;
6168 in.type = htonl(RX_DEBUGI_GETSTATS);
6171 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6172 &in, sizeof(in), stat, sizeof(*stat));
6175 * If the call was successful, fixup the version and indicate
6176 * what contents of the stat structure are valid.
6177 * Also do net to host conversion of fields here.
6181 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6182 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6184 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6185 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6187 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6188 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6190 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6191 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6193 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6194 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6196 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6197 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6199 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6200 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6202 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6203 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6206 stat->nFreePackets = ntohl(stat->nFreePackets);
6207 stat->packetReclaims = ntohl(stat->packetReclaims);
6208 stat->callsExecuted = ntohl(stat->callsExecuted);
6209 stat->nWaiting = ntohl(stat->nWaiting);
6210 stat->idleThreads = ntohl(stat->idleThreads);
6217 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6218 afs_uint16 remotePort, struct rx_stats * stat,
6219 afs_uint32 * supportedValues)
6221 struct rx_debugIn in;
6222 afs_int32 *lp = (afs_int32 *) stat;
6227 * supportedValues is currently unused, but added to allow future
6228 * versioning of this function.
6231 *supportedValues = 0;
6232 in.type = htonl(RX_DEBUGI_RXSTATS);
6234 memset(stat, 0, sizeof(*stat));
6236 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6237 &in, sizeof(in), stat, sizeof(*stat));
6242 * Do net to host conversion here
6245 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6254 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6255 afs_uint16 remotePort, size_t version_length,
6259 return MakeDebugCall(socket, remoteAddr, remotePort,
6260 RX_PACKET_TYPE_VERSION, a, 1, version,
6265 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6266 afs_uint16 remotePort, afs_int32 * nextConnection,
6267 int allConnections, afs_uint32 debugSupportedValues,
6268 struct rx_debugConn * conn,
6269 afs_uint32 * supportedValues)
6271 struct rx_debugIn in;
6276 * supportedValues is currently unused, but added to allow future
6277 * versioning of this function.
6280 *supportedValues = 0;
6281 if (allConnections) {
6282 in.type = htonl(RX_DEBUGI_GETALLCONN);
6284 in.type = htonl(RX_DEBUGI_GETCONN);
6286 in.index = htonl(*nextConnection);
6287 memset(conn, 0, sizeof(*conn));
6289 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6290 &in, sizeof(in), conn, sizeof(*conn));
6293 *nextConnection += 1;
6296 * Convert old connection format to new structure.
6299 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6300 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6301 #define MOVEvL(a) (conn->a = vL->a)
6303 /* any old or unrecognized version... */
6304 for (i = 0; i < RX_MAXCALLS; i++) {
6305 MOVEvL(callState[i]);
6306 MOVEvL(callMode[i]);
6307 MOVEvL(callFlags[i]);
6308 MOVEvL(callOther[i]);
6310 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6311 MOVEvL(secStats.type);
6312 MOVEvL(secStats.level);
6313 MOVEvL(secStats.flags);
6314 MOVEvL(secStats.expires);
6315 MOVEvL(secStats.packetsReceived);
6316 MOVEvL(secStats.packetsSent);
6317 MOVEvL(secStats.bytesReceived);
6318 MOVEvL(secStats.bytesSent);
6323 * Do net to host conversion here
6325 * I don't convert host or port since we are most likely
6326 * going to want these in NBO.
6328 conn->cid = ntohl(conn->cid);
6329 conn->serial = ntohl(conn->serial);
6330 for (i = 0; i < RX_MAXCALLS; i++) {
6331 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6333 conn->error = ntohl(conn->error);
6334 conn->secStats.flags = ntohl(conn->secStats.flags);
6335 conn->secStats.expires = ntohl(conn->secStats.expires);
6336 conn->secStats.packetsReceived =
6337 ntohl(conn->secStats.packetsReceived);
6338 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6339 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6340 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6341 conn->epoch = ntohl(conn->epoch);
6342 conn->natMTU = ntohl(conn->natMTU);
6349 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6350 afs_uint16 remotePort, afs_int32 * nextPeer,
6351 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6352 afs_uint32 * supportedValues)
6354 struct rx_debugIn in;
6358 * supportedValues is currently unused, but added to allow future
6359 * versioning of this function.
6362 *supportedValues = 0;
6363 in.type = htonl(RX_DEBUGI_GETPEER);
6364 in.index = htonl(*nextPeer);
6365 memset(peer, 0, sizeof(*peer));
6367 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6368 &in, sizeof(in), peer, sizeof(*peer));
6374 * Do net to host conversion here
6376 * I don't convert host or port since we are most likely
6377 * going to want these in NBO.
6379 peer->ifMTU = ntohs(peer->ifMTU);
6380 peer->idleWhen = ntohl(peer->idleWhen);
6381 peer->refCount = ntohs(peer->refCount);
6382 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6383 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6384 peer->rtt = ntohl(peer->rtt);
6385 peer->rtt_dev = ntohl(peer->rtt_dev);
6386 peer->timeout.sec = ntohl(peer->timeout.sec);
6387 peer->timeout.usec = ntohl(peer->timeout.usec);
6388 peer->nSent = ntohl(peer->nSent);
6389 peer->reSends = ntohl(peer->reSends);
6390 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6391 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6392 peer->rateFlag = ntohl(peer->rateFlag);
6393 peer->natMTU = ntohs(peer->natMTU);
6394 peer->maxMTU = ntohs(peer->maxMTU);
6395 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6396 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6397 peer->MTU = ntohs(peer->MTU);
6398 peer->cwind = ntohs(peer->cwind);
6399 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6400 peer->congestSeq = ntohs(peer->congestSeq);
6401 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6402 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6403 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6404 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6409 #endif /* RXDEBUG */
6414 struct rx_serverQueueEntry *np;
6417 register struct rx_call *call;
6418 register struct rx_serverQueueEntry *sq;
6422 if (rxinit_status == 1) {
6424 return; /* Already shutdown. */
6428 #ifndef AFS_PTHREAD_ENV
6429 FD_ZERO(&rx_selectMask);
6430 #endif /* AFS_PTHREAD_ENV */
6431 rxi_dataQuota = RX_MAX_QUOTA;
6432 #ifndef AFS_PTHREAD_ENV
6434 #endif /* AFS_PTHREAD_ENV */
6437 #ifndef AFS_PTHREAD_ENV
6438 #ifndef AFS_USE_GETTIMEOFDAY
6440 #endif /* AFS_USE_GETTIMEOFDAY */
6441 #endif /* AFS_PTHREAD_ENV */
6443 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6444 call = queue_First(&rx_freeCallQueue, rx_call);
6446 rxi_Free(call, sizeof(struct rx_call));
6449 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6450 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6456 struct rx_peer **peer_ptr, **peer_end;
6457 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6458 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6460 struct rx_peer *peer, *next;
6461 for (peer = *peer_ptr; peer; peer = next) {
6462 rx_interface_stat_p rpc_stat, nrpc_stat;
6465 (&peer->rpcStats, rpc_stat, nrpc_stat,
6466 rx_interface_stat)) {
6467 unsigned int num_funcs;
6470 queue_Remove(&rpc_stat->queue_header);
6471 queue_Remove(&rpc_stat->all_peers);
6472 num_funcs = rpc_stat->stats[0].func_total;
6474 sizeof(rx_interface_stat_t) +
6475 rpc_stat->stats[0].func_total *
6476 sizeof(rx_function_entry_v1_t);
6478 rxi_Free(rpc_stat, space);
6479 MUTEX_ENTER(&rx_rpc_stats);
6480 rxi_rpc_peer_stat_cnt -= num_funcs;
6481 MUTEX_EXIT(&rx_rpc_stats);
6485 MUTEX_ENTER(&rx_stats_mutex);
6486 rx_stats.nPeerStructs--;
6487 MUTEX_EXIT(&rx_stats_mutex);
6491 for (i = 0; i < RX_MAX_SERVICES; i++) {
6493 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6495 for (i = 0; i < rx_hashTableSize; i++) {
6496 register struct rx_connection *tc, *ntc;
6497 MUTEX_ENTER(&rx_connHashTable_lock);
6498 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6500 for (j = 0; j < RX_MAXCALLS; j++) {
6502 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6505 rxi_Free(tc, sizeof(*tc));
6507 MUTEX_EXIT(&rx_connHashTable_lock);
6510 MUTEX_ENTER(&freeSQEList_lock);
6512 while ((np = rx_FreeSQEList)) {
6513 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6514 MUTEX_DESTROY(&np->lock);
6515 rxi_Free(np, sizeof(*np));
6518 MUTEX_EXIT(&freeSQEList_lock);
6519 MUTEX_DESTROY(&freeSQEList_lock);
6520 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6521 MUTEX_DESTROY(&rx_connHashTable_lock);
6522 MUTEX_DESTROY(&rx_peerHashTable_lock);
6523 MUTEX_DESTROY(&rx_serverPool_lock);
6525 osi_Free(rx_connHashTable,
6526 rx_hashTableSize * sizeof(struct rx_connection *));
6527 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6529 UNPIN(rx_connHashTable,
6530 rx_hashTableSize * sizeof(struct rx_connection *));
6531 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6533 rxi_FreeAllPackets();
6535 MUTEX_ENTER(&rx_stats_mutex);
6536 rxi_dataQuota = RX_MAX_QUOTA;
6537 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6538 MUTEX_EXIT(&rx_stats_mutex);
6544 #ifdef RX_ENABLE_LOCKS
6546 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6548 if (!MUTEX_ISMINE(lockaddr))
6549 osi_Panic("Lock not held: %s", msg);
6551 #endif /* RX_ENABLE_LOCKS */
6556 * Routines to implement connection specific data.
6560 rx_KeyCreate(rx_destructor_t rtn)
6563 MUTEX_ENTER(&rxi_keyCreate_lock);
6564 key = rxi_keyCreate_counter++;
6565 rxi_keyCreate_destructor = (rx_destructor_t *)
6566 realloc((void *)rxi_keyCreate_destructor,
6567 (key + 1) * sizeof(rx_destructor_t));
6568 rxi_keyCreate_destructor[key] = rtn;
6569 MUTEX_EXIT(&rxi_keyCreate_lock);
6574 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6577 MUTEX_ENTER(&conn->conn_data_lock);
6578 if (!conn->specific) {
6579 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6580 for (i = 0; i < key; i++)
6581 conn->specific[i] = NULL;
6582 conn->nSpecific = key + 1;
6583 conn->specific[key] = ptr;
6584 } else if (key >= conn->nSpecific) {
6585 conn->specific = (void **)
6586 realloc(conn->specific, (key + 1) * sizeof(void *));
6587 for (i = conn->nSpecific; i < key; i++)
6588 conn->specific[i] = NULL;
6589 conn->nSpecific = key + 1;
6590 conn->specific[key] = ptr;
6592 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6593 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6594 conn->specific[key] = ptr;
6596 MUTEX_EXIT(&conn->conn_data_lock);
6600 rx_GetSpecific(struct rx_connection *conn, int key)
6603 MUTEX_ENTER(&conn->conn_data_lock);
6604 if (key >= conn->nSpecific)
6607 ptr = conn->specific[key];
6608 MUTEX_EXIT(&conn->conn_data_lock);
6612 #endif /* !KERNEL */
6615 * processStats is a queue used to store the statistics for the local
6616 * process. Its contents are similar to the contents of the rpcStats
6617 * queue on a rx_peer structure, but the actual data stored within
6618 * this queue contains totals across the lifetime of the process (assuming
6619 * the stats have not been reset) - unlike the per peer structures
6620 * which can come and go based upon the peer lifetime.
6623 static struct rx_queue processStats = { &processStats, &processStats };
6626 * peerStats is a queue used to store the statistics for all peer structs.
6627 * Its contents are the union of all the peer rpcStats queues.
6630 static struct rx_queue peerStats = { &peerStats, &peerStats };
6633 * rxi_monitor_processStats is used to turn process wide stat collection
6637 static int rxi_monitor_processStats = 0;
6640 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6643 static int rxi_monitor_peerStats = 0;
6646 * rxi_AddRpcStat - given all of the information for a particular rpc
6647 * call, create (if needed) and update the stat totals for the rpc.
6651 * IN stats - the queue of stats that will be updated with the new value
6653 * IN rxInterface - a unique number that identifies the rpc interface
6655 * IN currentFunc - the index of the function being invoked
6657 * IN totalFunc - the total number of functions in this interface
6659 * IN queueTime - the amount of time this function waited for a thread
6661 * IN execTime - the amount of time this function invocation took to execute
6663 * IN bytesSent - the number bytes sent by this invocation
6665 * IN bytesRcvd - the number bytes received by this invocation
6667 * IN isServer - if true, this invocation was made to a server
6669 * IN remoteHost - the ip address of the remote host
6671 * IN remotePort - the port of the remote host
6673 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6675 * INOUT counter - if a new stats structure is allocated, the counter will
6676 * be updated with the new number of allocated stat structures
6684 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6685 afs_uint32 currentFunc, afs_uint32 totalFunc,
6686 struct clock *queueTime, struct clock *execTime,
6687 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6688 afs_uint32 remoteHost, afs_uint32 remotePort,
6689 int addToPeerList, unsigned int *counter)
6692 rx_interface_stat_p rpc_stat, nrpc_stat;
6695 * See if there's already a structure for this interface
6698 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6699 if ((rpc_stat->stats[0].interfaceId == rxInterface)
6700 && (rpc_stat->stats[0].remote_is_server == isServer))
6705 * Didn't find a match so allocate a new structure and add it to the
6709 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
6710 || (rpc_stat->stats[0].interfaceId != rxInterface)
6711 || (rpc_stat->stats[0].remote_is_server != isServer)) {
6716 sizeof(rx_interface_stat_t) +
6717 totalFunc * sizeof(rx_function_entry_v1_t);
6719 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
6720 if (rpc_stat == NULL) {
6724 *counter += totalFunc;
6725 for (i = 0; i < totalFunc; i++) {
6726 rpc_stat->stats[i].remote_peer = remoteHost;
6727 rpc_stat->stats[i].remote_port = remotePort;
6728 rpc_stat->stats[i].remote_is_server = isServer;
6729 rpc_stat->stats[i].interfaceId = rxInterface;
6730 rpc_stat->stats[i].func_total = totalFunc;
6731 rpc_stat->stats[i].func_index = i;
6732 hzero(rpc_stat->stats[i].invocations);
6733 hzero(rpc_stat->stats[i].bytes_sent);
6734 hzero(rpc_stat->stats[i].bytes_rcvd);
6735 rpc_stat->stats[i].queue_time_sum.sec = 0;
6736 rpc_stat->stats[i].queue_time_sum.usec = 0;
6737 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
6738 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
6739 rpc_stat->stats[i].queue_time_min.sec = 9999999;
6740 rpc_stat->stats[i].queue_time_min.usec = 9999999;
6741 rpc_stat->stats[i].queue_time_max.sec = 0;
6742 rpc_stat->stats[i].queue_time_max.usec = 0;
6743 rpc_stat->stats[i].execution_time_sum.sec = 0;
6744 rpc_stat->stats[i].execution_time_sum.usec = 0;
6745 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
6746 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
6747 rpc_stat->stats[i].execution_time_min.sec = 9999999;
6748 rpc_stat->stats[i].execution_time_min.usec = 9999999;
6749 rpc_stat->stats[i].execution_time_max.sec = 0;
6750 rpc_stat->stats[i].execution_time_max.usec = 0;
6752 queue_Prepend(stats, rpc_stat);
6753 if (addToPeerList) {
6754 queue_Prepend(&peerStats, &rpc_stat->all_peers);
6759 * Increment the stats for this function
6762 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
6763 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
6764 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
6765 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
6766 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
6767 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
6768 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
6770 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
6771 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
6773 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
6774 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
6776 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
6777 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
6779 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
6780 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
6788 * rx_IncrementTimeAndCount - increment the times and count for a particular
6793 * IN peer - the peer who invoked the rpc
6795 * IN rxInterface - a unique number that identifies the rpc interface
6797 * IN currentFunc - the index of the function being invoked
6799 * IN totalFunc - the total number of functions in this interface
6801 * IN queueTime - the amount of time this function waited for a thread
6803 * IN execTime - the amount of time this function invocation took to execute
6805 * IN bytesSent - the number bytes sent by this invocation
6807 * IN bytesRcvd - the number bytes received by this invocation
6809 * IN isServer - if true, this invocation was made to a server
6817 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
6818 afs_uint32 currentFunc, afs_uint32 totalFunc,
6819 struct clock *queueTime, struct clock *execTime,
6820 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
6824 MUTEX_ENTER(&rx_rpc_stats);
6825 MUTEX_ENTER(&peer->peer_lock);
6827 if (rxi_monitor_peerStats) {
6828 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
6829 queueTime, execTime, bytesSent, bytesRcvd, isServer,
6830 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
6833 if (rxi_monitor_processStats) {
6834 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
6835 queueTime, execTime, bytesSent, bytesRcvd, isServer,
6836 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
6839 MUTEX_EXIT(&peer->peer_lock);
6840 MUTEX_EXIT(&rx_rpc_stats);
6845 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
6849 * IN callerVersion - the rpc stat version of the caller.
6851 * IN count - the number of entries to marshall.
6853 * IN stats - pointer to stats to be marshalled.
6855 * OUT ptr - Where to store the marshalled data.
6862 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
6863 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
6869 * We only support the first version
6871 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
6872 *(ptr++) = stats->remote_peer;
6873 *(ptr++) = stats->remote_port;
6874 *(ptr++) = stats->remote_is_server;
6875 *(ptr++) = stats->interfaceId;
6876 *(ptr++) = stats->func_total;
6877 *(ptr++) = stats->func_index;
6878 *(ptr++) = hgethi(stats->invocations);
6879 *(ptr++) = hgetlo(stats->invocations);
6880 *(ptr++) = hgethi(stats->bytes_sent);
6881 *(ptr++) = hgetlo(stats->bytes_sent);
6882 *(ptr++) = hgethi(stats->bytes_rcvd);
6883 *(ptr++) = hgetlo(stats->bytes_rcvd);
6884 *(ptr++) = stats->queue_time_sum.sec;
6885 *(ptr++) = stats->queue_time_sum.usec;
6886 *(ptr++) = stats->queue_time_sum_sqr.sec;
6887 *(ptr++) = stats->queue_time_sum_sqr.usec;
6888 *(ptr++) = stats->queue_time_min.sec;
6889 *(ptr++) = stats->queue_time_min.usec;
6890 *(ptr++) = stats->queue_time_max.sec;
6891 *(ptr++) = stats->queue_time_max.usec;
6892 *(ptr++) = stats->execution_time_sum.sec;
6893 *(ptr++) = stats->execution_time_sum.usec;
6894 *(ptr++) = stats->execution_time_sum_sqr.sec;
6895 *(ptr++) = stats->execution_time_sum_sqr.usec;
6896 *(ptr++) = stats->execution_time_min.sec;
6897 *(ptr++) = stats->execution_time_min.usec;
6898 *(ptr++) = stats->execution_time_max.sec;
6899 *(ptr++) = stats->execution_time_max.usec;
6905 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
6910 * IN callerVersion - the rpc stat version of the caller
6912 * OUT myVersion - the rpc stat version of this function
6914 * OUT clock_sec - local time seconds
6916 * OUT clock_usec - local time microseconds
6918 * OUT allocSize - the number of bytes allocated to contain stats
6920 * OUT statCount - the number stats retrieved from this process.
6922 * OUT stats - the actual stats retrieved from this process.
6926 * Returns void. If successful, stats will != NULL.
6930 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
6931 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
6932 size_t * allocSize, afs_uint32 * statCount,
6933 afs_uint32 ** stats)
6943 *myVersion = RX_STATS_RETRIEVAL_VERSION;
6946 * Check to see if stats are enabled
6949 MUTEX_ENTER(&rx_rpc_stats);
6950 if (!rxi_monitor_processStats) {
6951 MUTEX_EXIT(&rx_rpc_stats);
6955 clock_GetTime(&now);
6956 *clock_sec = now.sec;
6957 *clock_usec = now.usec;
6960 * Allocate the space based upon the caller version
6962 * If the client is at an older version than we are,
6963 * we return the statistic data in the older data format, but
6964 * we still return our version number so the client knows we
6965 * are maintaining more data than it can retrieve.
6968 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
6969 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
6970 *statCount = rxi_rpc_process_stat_cnt;
6973 * This can't happen yet, but in the future version changes
6974 * can be handled by adding additional code here
6978 if (space > (size_t) 0) {
6980 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
6983 rx_interface_stat_p rpc_stat, nrpc_stat;
6987 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6989 * Copy the data based upon the caller version
6991 rx_MarshallProcessRPCStats(callerVersion,
6992 rpc_stat->stats[0].func_total,
6993 rpc_stat->stats, &ptr);
6999 MUTEX_EXIT(&rx_rpc_stats);
7004 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7008 * IN callerVersion - the rpc stat version of the caller
7010 * OUT myVersion - the rpc stat version of this function
7012 * OUT clock_sec - local time seconds
7014 * OUT clock_usec - local time microseconds
7016 * OUT allocSize - the number of bytes allocated to contain stats
7018 * OUT statCount - the number of stats retrieved from the individual
7021 * OUT stats - the actual stats retrieved from the individual peer structures.
7025 * Returns void. If successful, stats will != NULL.
7029 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7030 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7031 size_t * allocSize, afs_uint32 * statCount,
7032 afs_uint32 ** stats)
7042 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7045 * Check to see if stats are enabled
7048 MUTEX_ENTER(&rx_rpc_stats);
7049 if (!rxi_monitor_peerStats) {
7050 MUTEX_EXIT(&rx_rpc_stats);
7054 clock_GetTime(&now);
7055 *clock_sec = now.sec;
7056 *clock_usec = now.usec;
7059 * Allocate the space based upon the caller version
7061 * If the client is at an older version than we are,
7062 * we return the statistic data in the older data format, but
7063 * we still return our version number so the client knows we
7064 * are maintaining more data than it can retrieve.
7067 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7068 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7069 *statCount = rxi_rpc_peer_stat_cnt;
7072 * This can't happen yet, but in the future version changes
7073 * can be handled by adding additional code here
7077 if (space > (size_t) 0) {
7079 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7082 rx_interface_stat_p rpc_stat, nrpc_stat;
7086 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7088 * We have to fix the offset of rpc_stat since we are
7089 * keeping this structure on two rx_queues. The rx_queue
7090 * package assumes that the rx_queue member is the first
7091 * member of the structure. That is, rx_queue assumes that
7092 * any one item is only on one queue at a time. We are
7093 * breaking that assumption and so we have to do a little
7094 * math to fix our pointers.
7097 fix_offset = (char *)rpc_stat;
7098 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7099 rpc_stat = (rx_interface_stat_p) fix_offset;
7102 * Copy the data based upon the caller version
7104 rx_MarshallProcessRPCStats(callerVersion,
7105 rpc_stat->stats[0].func_total,
7106 rpc_stat->stats, &ptr);
7112 MUTEX_EXIT(&rx_rpc_stats);
7117 * rx_FreeRPCStats - free memory allocated by
7118 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7122 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7123 * rx_RetrievePeerRPCStats
7125 * IN allocSize - the number of bytes in stats.
7133 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7135 rxi_Free(stats, allocSize);
7139 * rx_queryProcessRPCStats - see if process rpc stat collection is
7140 * currently enabled.
7146 * Returns 0 if stats are not enabled != 0 otherwise
7150 rx_queryProcessRPCStats(void)
7153 MUTEX_ENTER(&rx_rpc_stats);
7154 rc = rxi_monitor_processStats;
7155 MUTEX_EXIT(&rx_rpc_stats);
7160 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7166 * Returns 0 if stats are not enabled != 0 otherwise
7170 rx_queryPeerRPCStats(void)
7173 MUTEX_ENTER(&rx_rpc_stats);
7174 rc = rxi_monitor_peerStats;
7175 MUTEX_EXIT(&rx_rpc_stats);
7180 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7190 rx_enableProcessRPCStats(void)
7192 MUTEX_ENTER(&rx_rpc_stats);
7193 rx_enable_stats = 1;
7194 rxi_monitor_processStats = 1;
7195 MUTEX_EXIT(&rx_rpc_stats);
7199 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7209 rx_enablePeerRPCStats(void)
7211 MUTEX_ENTER(&rx_rpc_stats);
7212 rx_enable_stats = 1;
7213 rxi_monitor_peerStats = 1;
7214 MUTEX_EXIT(&rx_rpc_stats);
7218 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7228 rx_disableProcessRPCStats(void)
7230 rx_interface_stat_p rpc_stat, nrpc_stat;
7233 MUTEX_ENTER(&rx_rpc_stats);
7236 * Turn off process statistics and if peer stats is also off, turn
7240 rxi_monitor_processStats = 0;
7241 if (rxi_monitor_peerStats == 0) {
7242 rx_enable_stats = 0;
7245 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7246 unsigned int num_funcs = 0;
7249 queue_Remove(rpc_stat);
7250 num_funcs = rpc_stat->stats[0].func_total;
7252 sizeof(rx_interface_stat_t) +
7253 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7255 rxi_Free(rpc_stat, space);
7256 rxi_rpc_process_stat_cnt -= num_funcs;
7258 MUTEX_EXIT(&rx_rpc_stats);
7262 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7272 rx_disablePeerRPCStats(void)
7274 struct rx_peer **peer_ptr, **peer_end;
7277 MUTEX_ENTER(&rx_rpc_stats);
7280 * Turn off peer statistics and if process stats is also off, turn
7284 rxi_monitor_peerStats = 0;
7285 if (rxi_monitor_processStats == 0) {
7286 rx_enable_stats = 0;
7289 MUTEX_ENTER(&rx_peerHashTable_lock);
7290 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7291 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7293 struct rx_peer *peer, *next, *prev;
7294 for (prev = peer = *peer_ptr; peer; peer = next) {
7296 code = MUTEX_TRYENTER(&peer->peer_lock);
7298 rx_interface_stat_p rpc_stat, nrpc_stat;
7301 (&peer->rpcStats, rpc_stat, nrpc_stat,
7302 rx_interface_stat)) {
7303 unsigned int num_funcs = 0;
7306 queue_Remove(&rpc_stat->queue_header);
7307 queue_Remove(&rpc_stat->all_peers);
7308 num_funcs = rpc_stat->stats[0].func_total;
7310 sizeof(rx_interface_stat_t) +
7311 rpc_stat->stats[0].func_total *
7312 sizeof(rx_function_entry_v1_t);
7314 rxi_Free(rpc_stat, space);
7315 rxi_rpc_peer_stat_cnt -= num_funcs;
7317 MUTEX_EXIT(&peer->peer_lock);
7318 if (prev == *peer_ptr) {
7328 MUTEX_EXIT(&rx_peerHashTable_lock);
7329 MUTEX_EXIT(&rx_rpc_stats);
7333 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7338 * IN clearFlag - flag indicating which stats to clear
7346 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7348 rx_interface_stat_p rpc_stat, nrpc_stat;
7350 MUTEX_ENTER(&rx_rpc_stats);
7352 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7353 unsigned int num_funcs = 0, i;
7354 num_funcs = rpc_stat->stats[0].func_total;
7355 for (i = 0; i < num_funcs; i++) {
7356 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7357 hzero(rpc_stat->stats[i].invocations);
7359 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7360 hzero(rpc_stat->stats[i].bytes_sent);
7362 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7363 hzero(rpc_stat->stats[i].bytes_rcvd);
7365 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7366 rpc_stat->stats[i].queue_time_sum.sec = 0;
7367 rpc_stat->stats[i].queue_time_sum.usec = 0;
7369 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7370 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7371 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7373 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7374 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7375 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7377 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7378 rpc_stat->stats[i].queue_time_max.sec = 0;
7379 rpc_stat->stats[i].queue_time_max.usec = 0;
7381 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7382 rpc_stat->stats[i].execution_time_sum.sec = 0;
7383 rpc_stat->stats[i].execution_time_sum.usec = 0;
7385 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7386 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7387 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7389 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7390 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7391 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7393 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7394 rpc_stat->stats[i].execution_time_max.sec = 0;
7395 rpc_stat->stats[i].execution_time_max.usec = 0;
7400 MUTEX_EXIT(&rx_rpc_stats);
7404 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7409 * IN clearFlag - flag indicating which stats to clear
7417 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7419 rx_interface_stat_p rpc_stat, nrpc_stat;
7421 MUTEX_ENTER(&rx_rpc_stats);
7423 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7424 unsigned int num_funcs = 0, i;
7427 * We have to fix the offset of rpc_stat since we are
7428 * keeping this structure on two rx_queues. The rx_queue
7429 * package assumes that the rx_queue member is the first
7430 * member of the structure. That is, rx_queue assumes that
7431 * any one item is only on one queue at a time. We are
7432 * breaking that assumption and so we have to do a little
7433 * math to fix our pointers.
7436 fix_offset = (char *)rpc_stat;
7437 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7438 rpc_stat = (rx_interface_stat_p) fix_offset;
7440 num_funcs = rpc_stat->stats[0].func_total;
7441 for (i = 0; i < num_funcs; i++) {
7442 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7443 hzero(rpc_stat->stats[i].invocations);
7445 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7446 hzero(rpc_stat->stats[i].bytes_sent);
7448 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7449 hzero(rpc_stat->stats[i].bytes_rcvd);
7451 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7452 rpc_stat->stats[i].queue_time_sum.sec = 0;
7453 rpc_stat->stats[i].queue_time_sum.usec = 0;
7455 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7456 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7457 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7459 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7460 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7461 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7463 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7464 rpc_stat->stats[i].queue_time_max.sec = 0;
7465 rpc_stat->stats[i].queue_time_max.usec = 0;
7467 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7468 rpc_stat->stats[i].execution_time_sum.sec = 0;
7469 rpc_stat->stats[i].execution_time_sum.usec = 0;
7471 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7472 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7473 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7475 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7476 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7477 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7479 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7480 rpc_stat->stats[i].execution_time_max.sec = 0;
7481 rpc_stat->stats[i].execution_time_max.usec = 0;
7486 MUTEX_EXIT(&rx_rpc_stats);
7490 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7491 * is authorized to enable/disable/clear RX statistics.
7493 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7496 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7498 rxi_rxstat_userok = proc;
7502 rx_RxStatUserOk(struct rx_call *call)
7504 if (!rxi_rxstat_userok)
7506 return rxi_rxstat_userok(call);