2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
14 #include "afs/param.h"
16 #include <afs/param.h>
23 #include "afs/sysincludes.h"
24 #include "afsincludes.h"
30 #include <net/net_globals.h>
31 #endif /* AFS_OSF_ENV */
32 #ifdef AFS_LINUX20_ENV
35 #include "netinet/in.h"
36 #include "afs/afs_args.h"
37 #include "afs/afs_osi.h"
38 #ifdef RX_KERNEL_TRACE
39 #include "rx_kcommon.h"
41 #if (defined(AFS_AUX_ENV) || defined(AFS_AIX_ENV))
45 #undef RXDEBUG /* turn off debugging */
47 #if defined(AFS_SGI_ENV)
48 #include "sys/debug.h"
57 #endif /* AFS_ALPHA_ENV */
59 #include "afs/sysincludes.h"
60 #include "afsincludes.h"
63 #include "rx_kmutex.h"
64 #include "rx_kernel.h"
68 #include "rx_globals.h"
70 #define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
71 #define AFSOP_STOP_AFS 211 /* Stop AFS process */
72 #define AFSOP_STOP_BKG 212 /* Stop BKG process */
74 extern afs_int32 afs_termState;
76 #include "sys/lockl.h"
77 #include "sys/lock_def.h"
78 #endif /* AFS_AIX41_ENV */
79 # include "rxgen_consts.h"
81 # include <sys/types.h>
86 # include <afs/afsutil.h>
88 # include <sys/socket.h>
89 # include <sys/file.h>
91 # include <sys/stat.h>
92 # include <netinet/in.h>
93 # include <sys/time.h>
103 # include "rx_user.h"
104 # include "rx_clock.h"
105 # include "rx_queue.h"
106 # include "rx_globals.h"
107 # include "rx_trace.h"
108 # include <afs/rxgen_consts.h>
111 int (*registerProgram) () = 0;
112 int (*swapNameProgram) () = 0;
114 /* Local static routines */
115 static void rxi_DestroyConnectionNoLock(register struct rx_connection *conn);
116 #ifdef RX_ENABLE_LOCKS
117 static void rxi_SetAcksInTransmitQueue(register struct rx_call *call);
120 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
122 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
123 afs_int32 rxi_start_in_error;
125 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
128 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
129 * currently allocated within rx. This number is used to allocate the
130 * memory required to return the statistics when queried.
133 static unsigned int rxi_rpc_peer_stat_cnt;
136 * rxi_rpc_process_stat_cnt counts the total number of local process stat
137 * structures currently allocated within rx. The number is used to allocate
138 * the memory required to return the statistics when queried.
141 static unsigned int rxi_rpc_process_stat_cnt;
143 #if !defined(offsetof)
144 #include <stddef.h> /* for definition of offsetof() */
147 #ifdef AFS_PTHREAD_ENV
151 * Use procedural initialization of mutexes/condition variables
155 extern pthread_mutex_t rx_stats_mutex;
156 extern pthread_mutex_t rxkad_stats_mutex;
157 extern pthread_mutex_t des_init_mutex;
158 extern pthread_mutex_t des_random_mutex;
159 extern pthread_mutex_t rx_clock_mutex;
160 extern pthread_mutex_t rxi_connCacheMutex;
161 extern pthread_mutex_t rx_event_mutex;
162 extern pthread_mutex_t osi_malloc_mutex;
163 extern pthread_mutex_t event_handler_mutex;
164 extern pthread_mutex_t listener_mutex;
165 extern pthread_mutex_t rx_if_init_mutex;
166 extern pthread_mutex_t rx_if_mutex;
167 extern pthread_mutex_t rxkad_client_uid_mutex;
168 extern pthread_mutex_t rxkad_random_mutex;
170 extern pthread_cond_t rx_event_handler_cond;
171 extern pthread_cond_t rx_listener_cond;
173 static pthread_mutex_t epoch_mutex;
174 static pthread_mutex_t rx_init_mutex;
175 static pthread_mutex_t rx_debug_mutex;
178 rxi_InitPthread(void)
180 assert(pthread_mutex_init(&rx_clock_mutex, (const pthread_mutexattr_t *)0)
182 assert(pthread_mutex_init(&rx_stats_mutex, (const pthread_mutexattr_t *)0)
184 assert(pthread_mutex_init
185 (&rxi_connCacheMutex, (const pthread_mutexattr_t *)0) == 0);
186 assert(pthread_mutex_init(&rx_init_mutex, (const pthread_mutexattr_t *)0)
188 assert(pthread_mutex_init(&epoch_mutex, (const pthread_mutexattr_t *)0) ==
190 assert(pthread_mutex_init(&rx_event_mutex, (const pthread_mutexattr_t *)0)
192 assert(pthread_mutex_init(&des_init_mutex, (const pthread_mutexattr_t *)0)
194 assert(pthread_mutex_init
195 (&des_random_mutex, (const pthread_mutexattr_t *)0) == 0);
196 assert(pthread_mutex_init
197 (&osi_malloc_mutex, (const pthread_mutexattr_t *)0) == 0);
198 assert(pthread_mutex_init
199 (&event_handler_mutex, (const pthread_mutexattr_t *)0) == 0);
200 assert(pthread_mutex_init(&listener_mutex, (const pthread_mutexattr_t *)0)
202 assert(pthread_mutex_init
203 (&rx_if_init_mutex, (const pthread_mutexattr_t *)0) == 0);
204 assert(pthread_mutex_init(&rx_if_mutex, (const pthread_mutexattr_t *)0) ==
206 assert(pthread_mutex_init
207 (&rxkad_client_uid_mutex, (const pthread_mutexattr_t *)0) == 0);
208 assert(pthread_mutex_init
209 (&rxkad_random_mutex, (const pthread_mutexattr_t *)0) == 0);
210 assert(pthread_mutex_init
211 (&rxkad_stats_mutex, (const pthread_mutexattr_t *)0) == 0);
212 assert(pthread_mutex_init(&rx_debug_mutex, (const pthread_mutexattr_t *)0)
215 assert(pthread_cond_init
216 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
217 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
219 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
222 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
223 #define INIT_PTHREAD_LOCKS \
224 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0);
226 * The rx_stats_mutex mutex protects the following global variables:
231 * rxi_lowConnRefCount
232 * rxi_lowPeerRefCount
241 #define INIT_PTHREAD_LOCKS
245 /* Variables for handling the minProcs implementation. availProcs gives the
246 * number of threads available in the pool at this moment (not counting dudes
247 * executing right now). totalMin gives the total number of procs required
248 * for handling all minProcs requests. minDeficit is a dynamic variable
249 * tracking the # of procs required to satisfy all of the remaining minProcs
251 * For fine grain locking to work, the quota check and the reservation of
252 * a server thread has to come while rxi_availProcs and rxi_minDeficit
253 * are locked. To this end, the code has been modified under #ifdef
254 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
255 * same time. A new function, ReturnToServerPool() returns the allocation.
257 * A call can be on several queue's (but only one at a time). When
258 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
259 * that no one else is touching the queue. To this end, we store the address
260 * of the queue lock in the call structure (under the call lock) when we
261 * put the call on a queue, and we clear the call_queue_lock when the
262 * call is removed from a queue (once the call lock has been obtained).
263 * This allows rxi_ResetCall to safely synchronize with others wishing
264 * to manipulate the queue.
267 #ifdef RX_ENABLE_LOCKS
268 static afs_kmutex_t rx_rpc_stats;
269 void rxi_StartUnlocked();
272 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
273 ** pretty good that the next packet coming in is from the same connection
274 ** as the last packet, since we're send multiple packets in a transmit window.
276 struct rx_connection *rxLastConn = 0;
278 #ifdef RX_ENABLE_LOCKS
279 /* The locking hierarchy for rx fine grain locking is composed of these
282 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
283 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
284 * call->lock - locks call data fields.
285 * These are independent of each other:
286 * rx_freeCallQueue_lock
291 * serverQueueEntry->lock
293 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
294 * peer->lock - locks peer data fields.
295 * conn_data_lock - that more than one thread is not updating a conn data
296 * field at the same time.
304 * Do we need a lock to protect the peer field in the conn structure?
305 * conn->peer was previously a constant for all intents and so has no
306 * lock protecting this field. The multihomed client delta introduced
307 * a RX code change : change the peer field in the connection structure
308 * to that remote inetrface from which the last packet for this
309 * connection was sent out. This may become an issue if further changes
312 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
313 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
315 /* rxdb_fileID is used to identify the lock location, along with line#. */
316 static int rxdb_fileID = RXDB_FILE_RX;
317 #endif /* RX_LOCKS_DB */
318 #else /* RX_ENABLE_LOCKS */
319 #define SET_CALL_QUEUE_LOCK(C, L)
320 #define CLEAR_CALL_QUEUE_LOCK(C)
321 #endif /* RX_ENABLE_LOCKS */
322 struct rx_serverQueueEntry *rx_waitForPacket = 0;
323 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
325 /* ------------Exported Interfaces------------- */
327 /* This function allows rxkad to set the epoch to a suitably random number
328 * which rx_NewConnection will use in the future. The principle purpose is to
329 * get rxnull connections to use the same epoch as the rxkad connections do, at
330 * least once the first rxkad connection is established. This is important now
331 * that the host/port addresses aren't used in FindConnection: the uniqueness
332 * of epoch/cid matters and the start time won't do. */
334 #ifdef AFS_PTHREAD_ENV
336 * This mutex protects the following global variables:
340 #define LOCK_EPOCH assert(pthread_mutex_lock(&epoch_mutex)==0);
341 #define UNLOCK_EPOCH assert(pthread_mutex_unlock(&epoch_mutex)==0);
345 #endif /* AFS_PTHREAD_ENV */
348 rx_SetEpoch(afs_uint32 epoch)
350 LOCK_EPOCH rx_epoch = epoch;
353 /* Initialize rx. A port number may be mentioned, in which case this
354 * becomes the default port number for any service installed later.
355 * If 0 is provided for the port number, a random port will be chosen
356 * by the kernel. Whether this will ever overlap anything in
357 * /etc/services is anybody's guess... Returns 0 on success, -1 on
359 static int rxinit_status = 1;
360 #ifdef AFS_PTHREAD_ENV
362 * This mutex protects the following global variables:
366 #define LOCK_RX_INIT assert(pthread_mutex_lock(&rx_init_mutex)==0);
367 #define UNLOCK_RX_INIT assert(pthread_mutex_unlock(&rx_init_mutex)==0);
370 #define UNLOCK_RX_INIT
374 rx_InitHost(u_int host, u_int port)
381 char *htable, *ptable;
384 #if defined(AFS_DJGPP_ENV) && !defined(DEBUG)
385 __djgpp_set_quiet_socket(1);
390 INIT_PTHREAD_LOCKS LOCK_RX_INIT if (rxinit_status == 0) {
391 tmp_status = rxinit_status;
392 UNLOCK_RX_INIT return tmp_status; /* Already started; return previous error code. */
395 if (afs_winsockInit() < 0)
401 * Initialize anything necessary to provide a non-premptive threading
404 rxi_InitializeThreadSupport();
407 /* Allocate and initialize a socket for client and perhaps server
410 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
411 if (rx_socket == OSI_NULLSOCKET) {
412 UNLOCK_RX_INIT return RX_ADDRINUSE;
414 #ifdef RX_ENABLE_LOCKS
417 #endif /* RX_LOCKS_DB */
418 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
419 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
420 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
421 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
422 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
424 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
426 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
428 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
430 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
432 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
434 #if defined(KERNEL) && defined(AFS_HPUX110_ENV)
436 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
437 #endif /* KERNEL && AFS_HPUX110_ENV */
438 #else /* RX_ENABLE_LOCKS */
439 #if defined(KERNEL) && defined(AFS_GLOBAL_SUNLOCK) && !defined(AFS_HPUX_ENV) && !defined(AFS_OBSD_ENV)
440 mutex_init(&afs_rxglobal_lock, "afs_rxglobal_lock", MUTEX_DEFAULT, NULL);
441 #endif /* AFS_GLOBAL_SUNLOCK */
442 #endif /* RX_ENABLE_LOCKS */
445 rx_connDeadTime = 12;
446 rx_tranquil = 0; /* reset flag */
447 memset((char *)&rx_stats, 0, sizeof(struct rx_stats));
449 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
450 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
451 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
452 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
453 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
454 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
456 /* Malloc up a bunch of packets & buffers */
458 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
459 queue_Init(&rx_freePacketQueue);
460 rxi_NeedMorePackets = FALSE;
461 rxi_MorePackets(rx_nPackets);
469 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
470 tv.tv_sec = clock_now.sec;
471 tv.tv_usec = clock_now.usec;
472 srand((unsigned int)tv.tv_usec);
479 #if defined(KERNEL) && !defined(UKERNEL)
480 /* Really, this should never happen in a real kernel */
483 struct sockaddr_in addr;
484 int addrlen = sizeof(addr);
485 if (getsockname((int)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
489 rx_port = addr.sin_port;
492 rx_stats.minRtt.sec = 9999999;
494 rx_SetEpoch(tv.tv_sec | 0x80000000);
496 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
497 * will provide a randomer value. */
499 MUTEX_ENTER(&rx_stats_mutex);
500 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
501 MUTEX_EXIT(&rx_stats_mutex);
502 /* *Slightly* random start time for the cid. This is just to help
503 * out with the hashing function at the peer */
504 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
505 rx_connHashTable = (struct rx_connection **)htable;
506 rx_peerHashTable = (struct rx_peer **)ptable;
508 rx_lastAckDelay.sec = 0;
509 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
510 rx_hardAckDelay.sec = 0;
511 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
512 rx_softAckDelay.sec = 0;
513 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
515 rxevent_Init(20, rxi_ReScheduleEvents);
517 /* Initialize various global queues */
518 queue_Init(&rx_idleServerQueue);
519 queue_Init(&rx_incomingCallQueue);
520 queue_Init(&rx_freeCallQueue);
522 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
523 /* Initialize our list of usable IP addresses. */
527 /* Start listener process (exact function is dependent on the
528 * implementation environment--kernel or user space) */
533 tmp_status = rxinit_status = 0;
534 UNLOCK_RX_INIT return tmp_status;
537 int rx_Init(u_int port)
539 return rx_InitHost(htonl(INADDR_ANY), port);
542 /* called with unincremented nRequestsRunning to see if it is OK to start
543 * a new thread in this service. Could be "no" for two reasons: over the
544 * max quota, or would prevent others from reaching their min quota.
546 #ifdef RX_ENABLE_LOCKS
547 /* This verion of QuotaOK reserves quota if it's ok while the
548 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
551 QuotaOK(register struct rx_service *aservice)
553 /* check if over max quota */
554 if (aservice->nRequestsRunning >= aservice->maxProcs) {
558 /* under min quota, we're OK */
559 /* otherwise, can use only if there are enough to allow everyone
560 * to go to their min quota after this guy starts.
562 MUTEX_ENTER(&rx_stats_mutex);
563 if ((aservice->nRequestsRunning < aservice->minProcs)
564 || (rxi_availProcs > rxi_minDeficit)) {
565 aservice->nRequestsRunning++;
566 /* just started call in minProcs pool, need fewer to maintain
568 if (aservice->nRequestsRunning <= aservice->minProcs)
571 MUTEX_EXIT(&rx_stats_mutex);
574 MUTEX_EXIT(&rx_stats_mutex);
580 ReturnToServerPool(register struct rx_service *aservice)
582 aservice->nRequestsRunning--;
583 MUTEX_ENTER(&rx_stats_mutex);
584 if (aservice->nRequestsRunning < aservice->minProcs)
587 MUTEX_EXIT(&rx_stats_mutex);
590 #else /* RX_ENABLE_LOCKS */
592 QuotaOK(register struct rx_service *aservice)
595 /* under min quota, we're OK */
596 if (aservice->nRequestsRunning < aservice->minProcs)
599 /* check if over max quota */
600 if (aservice->nRequestsRunning >= aservice->maxProcs)
603 /* otherwise, can use only if there are enough to allow everyone
604 * to go to their min quota after this guy starts.
606 if (rxi_availProcs > rxi_minDeficit)
610 #endif /* RX_ENABLE_LOCKS */
613 /* Called by rx_StartServer to start up lwp's to service calls.
614 NExistingProcs gives the number of procs already existing, and which
615 therefore needn't be created. */
617 rxi_StartServerProcs(int nExistingProcs)
619 register struct rx_service *service;
624 /* For each service, reserve N processes, where N is the "minimum"
625 * number of processes that MUST be able to execute a request in parallel,
626 * at any time, for that process. Also compute the maximum difference
627 * between any service's maximum number of processes that can run
628 * (i.e. the maximum number that ever will be run, and a guarantee
629 * that this number will run if other services aren't running), and its
630 * minimum number. The result is the extra number of processes that
631 * we need in order to provide the latter guarantee */
632 for (i = 0; i < RX_MAX_SERVICES; i++) {
634 service = rx_services[i];
635 if (service == (struct rx_service *)0)
637 nProcs += service->minProcs;
638 diff = service->maxProcs - service->minProcs;
642 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
643 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
644 for (i = 0; i < nProcs; i++) {
645 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
650 /* This routine must be called if any services are exported. If the
651 * donateMe flag is set, the calling process is donated to the server
654 rx_StartServer(int donateMe)
656 register struct rx_service *service;
657 register int i, nProcs = 0;
663 /* Start server processes, if necessary (exact function is dependent
664 * on the implementation environment--kernel or user space). DonateMe
665 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
666 * case, one less new proc will be created rx_StartServerProcs.
668 rxi_StartServerProcs(donateMe);
670 /* count up the # of threads in minProcs, and add set the min deficit to
671 * be that value, too.
673 for (i = 0; i < RX_MAX_SERVICES; i++) {
674 service = rx_services[i];
675 if (service == (struct rx_service *)0)
677 MUTEX_ENTER(&rx_stats_mutex);
678 rxi_totalMin += service->minProcs;
679 /* below works even if a thread is running, since minDeficit would
680 * still have been decremented and later re-incremented.
682 rxi_minDeficit += service->minProcs;
683 MUTEX_EXIT(&rx_stats_mutex);
686 /* Turn on reaping of idle server connections */
687 rxi_ReapConnections();
696 #ifdef AFS_PTHREAD_ENV
698 pid = (pid_t) pthread_self();
699 #else /* AFS_PTHREAD_ENV */
701 LWP_CurrentProcess(&pid);
702 #endif /* AFS_PTHREAD_ENV */
704 sprintf(name, "srv_%d", ++nProcs);
706 (*registerProgram) (pid, name);
708 #endif /* AFS_NT40_ENV */
709 rx_ServerProc(); /* Never returns */
714 /* Create a new client connection to the specified service, using the
715 * specified security object to implement the security model for this
717 struct rx_connection *
718 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
719 register struct rx_securityClass *securityObject,
720 int serviceSecurityIndex)
724 register struct rx_connection *conn;
729 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", shost, sport, sservice, securityObject, serviceSecurityIndex));
731 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
732 * the case of kmem_alloc? */
733 conn = rxi_AllocConnection();
734 #ifdef RX_ENABLE_LOCKS
735 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
736 MUTEX_INIT(&conn->conn_data_lock, "conn call lock", MUTEX_DEFAULT, 0);
737 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
741 MUTEX_ENTER(&rx_connHashTable_lock);
742 cid = (rx_nextCid += RX_MAXCALLS);
743 conn->type = RX_CLIENT_CONNECTION;
745 conn->epoch = rx_epoch;
746 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
747 conn->serviceId = sservice;
748 conn->securityObject = securityObject;
749 /* This doesn't work in all compilers with void (they're buggy), so fake it
751 conn->securityData = (VOID *) 0;
752 conn->securityIndex = serviceSecurityIndex;
753 rx_SetConnDeadTime(conn, rx_connDeadTime);
754 conn->ackRate = RX_FAST_ACK_RATE;
756 conn->specific = NULL;
757 conn->challengeEvent = NULL;
758 conn->delayedAbortEvent = NULL;
759 conn->abortCount = 0;
762 RXS_NewConnection(securityObject, conn);
764 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
766 conn->refCount++; /* no lock required since only this thread knows... */
767 conn->next = rx_connHashTable[hashindex];
768 rx_connHashTable[hashindex] = conn;
769 MUTEX_ENTER(&rx_stats_mutex);
770 rx_stats.nClientConns++;
771 MUTEX_EXIT(&rx_stats_mutex);
773 MUTEX_EXIT(&rx_connHashTable_lock);
780 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
782 /* The idea is to set the dead time to a value that allows several
783 * keepalives to be dropped without timing out the connection. */
784 conn->secondsUntilDead = MAX(seconds, 6);
785 conn->secondsUntilPing = conn->secondsUntilDead / 6;
788 int rxi_lowPeerRefCount = 0;
789 int rxi_lowConnRefCount = 0;
792 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
793 * NOTE: must not be called with rx_connHashTable_lock held.
796 rxi_CleanupConnection(struct rx_connection *conn)
798 /* Notify the service exporter, if requested, that this connection
799 * is being destroyed */
800 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
801 (*conn->service->destroyConnProc) (conn);
803 /* Notify the security module that this connection is being destroyed */
804 RXS_DestroyConnection(conn->securityObject, conn);
806 /* If this is the last connection using the rx_peer struct, set its
807 * idle time to now. rxi_ReapConnections will reap it if it's still
808 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
810 MUTEX_ENTER(&rx_peerHashTable_lock);
811 if (--conn->peer->refCount <= 0) {
812 conn->peer->idleWhen = clock_Sec();
813 if (conn->peer->refCount < 0) {
814 conn->peer->refCount = 0;
815 MUTEX_ENTER(&rx_stats_mutex);
816 rxi_lowPeerRefCount++;
817 MUTEX_EXIT(&rx_stats_mutex);
820 MUTEX_EXIT(&rx_peerHashTable_lock);
822 MUTEX_ENTER(&rx_stats_mutex);
823 if (conn->type == RX_SERVER_CONNECTION)
824 rx_stats.nServerConns--;
826 rx_stats.nClientConns--;
827 MUTEX_EXIT(&rx_stats_mutex);
830 if (conn->specific) {
832 for (i = 0; i < conn->nSpecific; i++) {
833 if (conn->specific[i] && rxi_keyCreate_destructor[i])
834 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
835 conn->specific[i] = NULL;
837 free(conn->specific);
839 conn->specific = NULL;
843 MUTEX_DESTROY(&conn->conn_call_lock);
844 MUTEX_DESTROY(&conn->conn_data_lock);
845 CV_DESTROY(&conn->conn_call_cv);
847 rxi_FreeConnection(conn);
850 /* Destroy the specified connection */
852 rxi_DestroyConnection(register struct rx_connection *conn)
854 MUTEX_ENTER(&rx_connHashTable_lock);
855 rxi_DestroyConnectionNoLock(conn);
856 /* conn should be at the head of the cleanup list */
857 if (conn == rx_connCleanup_list) {
858 rx_connCleanup_list = rx_connCleanup_list->next;
859 MUTEX_EXIT(&rx_connHashTable_lock);
860 rxi_CleanupConnection(conn);
862 #ifdef RX_ENABLE_LOCKS
864 MUTEX_EXIT(&rx_connHashTable_lock);
866 #endif /* RX_ENABLE_LOCKS */
870 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
872 register struct rx_connection **conn_ptr;
873 register int havecalls = 0;
874 struct rx_packet *packet;
881 MUTEX_ENTER(&conn->conn_data_lock);
882 if (conn->refCount > 0)
885 MUTEX_ENTER(&rx_stats_mutex);
886 rxi_lowConnRefCount++;
887 MUTEX_EXIT(&rx_stats_mutex);
890 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
891 /* Busy; wait till the last guy before proceeding */
892 MUTEX_EXIT(&conn->conn_data_lock);
897 /* If the client previously called rx_NewCall, but it is still
898 * waiting, treat this as a running call, and wait to destroy the
899 * connection later when the call completes. */
900 if ((conn->type == RX_CLIENT_CONNECTION)
901 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
902 conn->flags |= RX_CONN_DESTROY_ME;
903 MUTEX_EXIT(&conn->conn_data_lock);
907 MUTEX_EXIT(&conn->conn_data_lock);
909 /* Check for extant references to this connection */
910 for (i = 0; i < RX_MAXCALLS; i++) {
911 register struct rx_call *call = conn->call[i];
914 if (conn->type == RX_CLIENT_CONNECTION) {
915 MUTEX_ENTER(&call->lock);
916 if (call->delayedAckEvent) {
917 /* Push the final acknowledgment out now--there
918 * won't be a subsequent call to acknowledge the
919 * last reply packets */
920 rxevent_Cancel(call->delayedAckEvent, call,
921 RX_CALL_REFCOUNT_DELAY);
922 if (call->state == RX_STATE_PRECALL
923 || call->state == RX_STATE_ACTIVE) {
924 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
926 rxi_AckAll(NULL, call, 0);
929 MUTEX_EXIT(&call->lock);
933 #ifdef RX_ENABLE_LOCKS
935 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
936 MUTEX_EXIT(&conn->conn_data_lock);
938 /* Someone is accessing a packet right now. */
942 #endif /* RX_ENABLE_LOCKS */
945 /* Don't destroy the connection if there are any call
946 * structures still in use */
947 MUTEX_ENTER(&conn->conn_data_lock);
948 conn->flags |= RX_CONN_DESTROY_ME;
949 MUTEX_EXIT(&conn->conn_data_lock);
954 if (conn->delayedAbortEvent) {
955 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
956 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
958 MUTEX_ENTER(&conn->conn_data_lock);
959 rxi_SendConnectionAbort(conn, packet, 0, 1);
960 MUTEX_EXIT(&conn->conn_data_lock);
961 rxi_FreePacket(packet);
965 /* Remove from connection hash table before proceeding */
967 &rx_connHashTable[CONN_HASH
968 (peer->host, peer->port, conn->cid, conn->epoch,
970 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
971 if (*conn_ptr == conn) {
972 *conn_ptr = conn->next;
976 /* if the conn that we are destroying was the last connection, then we
977 * clear rxLastConn as well */
978 if (rxLastConn == conn)
981 /* Make sure the connection is completely reset before deleting it. */
982 /* get rid of pending events that could zap us later */
983 if (conn->challengeEvent)
984 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
985 if (conn->checkReachEvent)
986 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
988 /* Add the connection to the list of destroyed connections that
989 * need to be cleaned up. This is necessary to avoid deadlocks
990 * in the routines we call to inform others that this connection is
991 * being destroyed. */
992 conn->next = rx_connCleanup_list;
993 rx_connCleanup_list = conn;
996 /* Externally available version */
998 rx_DestroyConnection(register struct rx_connection *conn)
1004 rxi_DestroyConnection(conn);
1009 /* Start a new rx remote procedure call, on the specified connection.
1010 * If wait is set to 1, wait for a free call channel; otherwise return
1011 * 0. Maxtime gives the maximum number of seconds this call may take,
1012 * after rx_MakeCall returns. After this time interval, a call to any
1013 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1014 * For fine grain locking, we hold the conn_call_lock in order to
1015 * to ensure that we don't get signalle after we found a call in an active
1016 * state and before we go to sleep.
1019 rx_NewCall(register struct rx_connection *conn)
1022 register struct rx_call *call;
1023 struct clock queueTime;
1027 dpf(("rx_MakeCall(conn %x)\n", conn));
1030 clock_GetTime(&queueTime);
1032 MUTEX_ENTER(&conn->conn_call_lock);
1035 * Check if there are others waiting for a new call.
1036 * If so, let them go first to avoid starving them.
1037 * This is a fairly simple scheme, and might not be
1038 * a complete solution for large numbers of waiters.
1040 if (conn->makeCallWaiters) {
1041 #ifdef RX_ENABLE_LOCKS
1042 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1049 for (i = 0; i < RX_MAXCALLS; i++) {
1050 call = conn->call[i];
1052 MUTEX_ENTER(&call->lock);
1053 if (call->state == RX_STATE_DALLY) {
1054 rxi_ResetCall(call, 0);
1055 (*call->callNumber)++;
1058 MUTEX_EXIT(&call->lock);
1060 call = rxi_NewCall(conn, i);
1064 if (i < RX_MAXCALLS) {
1067 MUTEX_ENTER(&conn->conn_data_lock);
1068 conn->flags |= RX_CONN_MAKECALL_WAITING;
1069 MUTEX_EXIT(&conn->conn_data_lock);
1071 conn->makeCallWaiters++;
1072 #ifdef RX_ENABLE_LOCKS
1073 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1077 conn->makeCallWaiters--;
1080 * Wake up anyone else who might be giving us a chance to
1081 * run (see code above that avoids resource starvation).
1083 #ifdef RX_ENABLE_LOCKS
1084 CV_BROADCAST(&conn->conn_call_cv);
1089 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1091 /* Client is initially in send mode */
1092 call->state = RX_STATE_ACTIVE;
1093 call->mode = RX_MODE_SENDING;
1095 /* remember start time for call in case we have hard dead time limit */
1096 call->queueTime = queueTime;
1097 clock_GetTime(&call->startTime);
1098 hzero(call->bytesSent);
1099 hzero(call->bytesRcvd);
1101 /* Turn on busy protocol. */
1102 rxi_KeepAliveOn(call);
1104 MUTEX_EXIT(&call->lock);
1105 MUTEX_EXIT(&conn->conn_call_lock);
1109 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1110 /* Now, if TQ wasn't cleared earlier, do it now. */
1112 MUTEX_ENTER(&call->lock);
1113 while (call->flags & RX_CALL_TQ_BUSY) {
1114 call->flags |= RX_CALL_TQ_WAIT;
1115 #ifdef RX_ENABLE_LOCKS
1116 CV_WAIT(&call->cv_tq, &call->lock);
1117 #else /* RX_ENABLE_LOCKS */
1118 osi_rxSleep(&call->tq);
1119 #endif /* RX_ENABLE_LOCKS */
1121 if (call->flags & RX_CALL_TQ_CLEARME) {
1122 rxi_ClearTransmitQueue(call, 0);
1123 queue_Init(&call->tq);
1125 MUTEX_EXIT(&call->lock);
1127 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1133 rxi_HasActiveCalls(register struct rx_connection *aconn)
1136 register struct rx_call *tcall;
1140 for (i = 0; i < RX_MAXCALLS; i++) {
1141 if ((tcall = aconn->call[i])) {
1142 if ((tcall->state == RX_STATE_ACTIVE)
1143 || (tcall->state == RX_STATE_PRECALL)) {
1154 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1155 register afs_int32 * aint32s)
1158 register struct rx_call *tcall;
1162 for (i = 0; i < RX_MAXCALLS; i++) {
1163 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1164 aint32s[i] = aconn->callNumber[i] + 1;
1166 aint32s[i] = aconn->callNumber[i];
1173 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1174 register afs_int32 * aint32s)
1177 register struct rx_call *tcall;
1181 for (i = 0; i < RX_MAXCALLS; i++) {
1182 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1183 aconn->callNumber[i] = aint32s[i] - 1;
1185 aconn->callNumber[i] = aint32s[i];
1191 /* Advertise a new service. A service is named locally by a UDP port
1192 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1195 char *serviceName; Name for identification purposes (e.g. the
1196 service name might be used for probing for
1199 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1200 struct rx_securityClass **securityObjects, int nSecurityObjects,
1201 afs_int32(*serviceProc) (struct rx_call * acall))
1203 osi_socket socket = OSI_NULLSOCKET;
1204 register struct rx_service *tservice;
1210 if (serviceId == 0) {
1212 "rx_NewService: service id for service %s is not non-zero.\n",
1219 "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",
1227 tservice = rxi_AllocService();
1230 for (i = 0; i < RX_MAX_SERVICES; i++) {
1231 register struct rx_service *service = rx_services[i];
1233 if (port == service->servicePort) {
1234 if (service->serviceId == serviceId) {
1235 /* The identical service has already been
1236 * installed; if the caller was intending to
1237 * change the security classes used by this
1238 * service, he/she loses. */
1240 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1241 serviceName, serviceId, service->serviceName);
1244 rxi_FreeService(tservice);
1247 /* Different service, same port: re-use the socket
1248 * which is bound to the same port */
1249 socket = service->socket;
1252 if (socket == OSI_NULLSOCKET) {
1253 /* If we don't already have a socket (from another
1254 * service on same port) get a new one */
1255 socket = rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
1256 if (socket == OSI_NULLSOCKET) {
1259 rxi_FreeService(tservice);
1264 service->socket = socket;
1265 service->servicePort = port;
1266 service->serviceId = serviceId;
1267 service->serviceName = serviceName;
1268 service->nSecurityObjects = nSecurityObjects;
1269 service->securityObjects = securityObjects;
1270 service->minProcs = 0;
1271 service->maxProcs = 1;
1272 service->idleDeadTime = 60;
1273 service->connDeadTime = rx_connDeadTime;
1274 service->executeRequestProc = serviceProc;
1275 service->checkReach = 0;
1276 rx_services[i] = service; /* not visible until now */
1284 rxi_FreeService(tservice);
1285 (osi_Msg "rx_NewService: cannot support > %d services\n",
1290 /* Generic request processing loop. This routine should be called
1291 * by the implementation dependent rx_ServerProc. If socketp is
1292 * non-null, it will be set to the file descriptor that this thread
1293 * is now listening on. If socketp is null, this routine will never
1296 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1298 register struct rx_call *call;
1299 register afs_int32 code;
1300 register struct rx_service *tservice = NULL;
1307 call = rx_GetCall(threadID, tservice, socketp);
1308 if (socketp && *socketp != OSI_NULLSOCKET) {
1309 /* We are now a listener thread */
1314 /* if server is restarting( typically smooth shutdown) then do not
1315 * allow any new calls.
1318 if (rx_tranquil && (call != NULL)) {
1323 MUTEX_ENTER(&call->lock);
1325 rxi_CallError(call, RX_RESTARTING);
1326 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1328 MUTEX_EXIT(&call->lock);
1333 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1334 #ifdef RX_ENABLE_LOCKS
1336 #endif /* RX_ENABLE_LOCKS */
1337 afs_termState = AFSOP_STOP_AFS;
1338 afs_osi_Wakeup(&afs_termState);
1339 #ifdef RX_ENABLE_LOCKS
1341 #endif /* RX_ENABLE_LOCKS */
1346 tservice = call->conn->service;
1348 if (tservice->beforeProc)
1349 (*tservice->beforeProc) (call);
1351 code = call->conn->service->executeRequestProc(call);
1353 if (tservice->afterProc)
1354 (*tservice->afterProc) (call, code);
1356 rx_EndCall(call, code);
1357 MUTEX_ENTER(&rx_stats_mutex);
1359 MUTEX_EXIT(&rx_stats_mutex);
1365 rx_WakeupServerProcs(void)
1367 struct rx_serverQueueEntry *np, *tqp;
1372 MUTEX_ENTER(&rx_serverPool_lock);
1374 #ifdef RX_ENABLE_LOCKS
1375 if (rx_waitForPacket)
1376 CV_BROADCAST(&rx_waitForPacket->cv);
1377 #else /* RX_ENABLE_LOCKS */
1378 if (rx_waitForPacket)
1379 osi_rxWakeup(rx_waitForPacket);
1380 #endif /* RX_ENABLE_LOCKS */
1381 MUTEX_ENTER(&freeSQEList_lock);
1382 for (np = rx_FreeSQEList; np; np = tqp) {
1383 tqp = *(struct rx_serverQueueEntry **)np;
1384 #ifdef RX_ENABLE_LOCKS
1385 CV_BROADCAST(&np->cv);
1386 #else /* RX_ENABLE_LOCKS */
1388 #endif /* RX_ENABLE_LOCKS */
1390 MUTEX_EXIT(&freeSQEList_lock);
1391 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1392 #ifdef RX_ENABLE_LOCKS
1393 CV_BROADCAST(&np->cv);
1394 #else /* RX_ENABLE_LOCKS */
1396 #endif /* RX_ENABLE_LOCKS */
1398 MUTEX_EXIT(&rx_serverPool_lock);
1404 * One thing that seems to happen is that all the server threads get
1405 * tied up on some empty or slow call, and then a whole bunch of calls
1406 * arrive at once, using up the packet pool, so now there are more
1407 * empty calls. The most critical resources here are server threads
1408 * and the free packet pool. The "doreclaim" code seems to help in
1409 * general. I think that eventually we arrive in this state: there
1410 * are lots of pending calls which do have all their packets present,
1411 * so they won't be reclaimed, are multi-packet calls, so they won't
1412 * be scheduled until later, and thus are tying up most of the free
1413 * packet pool for a very long time.
1415 * 1. schedule multi-packet calls if all the packets are present.
1416 * Probably CPU-bound operation, useful to return packets to pool.
1417 * Do what if there is a full window, but the last packet isn't here?
1418 * 3. preserve one thread which *only* runs "best" calls, otherwise
1419 * it sleeps and waits for that type of call.
1420 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1421 * the current dataquota business is badly broken. The quota isn't adjusted
1422 * to reflect how many packets are presently queued for a running call.
1423 * So, when we schedule a queued call with a full window of packets queued
1424 * up for it, that *should* free up a window full of packets for other 2d-class
1425 * calls to be able to use from the packet pool. But it doesn't.
1427 * NB. Most of the time, this code doesn't run -- since idle server threads
1428 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1429 * as a new call arrives.
1431 /* Sleep until a call arrives. Returns a pointer to the call, ready
1432 * for an rx_Read. */
1433 #ifdef RX_ENABLE_LOCKS
1435 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1437 struct rx_serverQueueEntry *sq;
1438 register struct rx_call *call = (struct rx_call *)0;
1439 struct rx_service *service = NULL;
1442 MUTEX_ENTER(&freeSQEList_lock);
1444 if ((sq = rx_FreeSQEList)) {
1445 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1446 MUTEX_EXIT(&freeSQEList_lock);
1447 } else { /* otherwise allocate a new one and return that */
1448 MUTEX_EXIT(&freeSQEList_lock);
1449 sq = (struct rx_serverQueueEntry *)
1450 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1451 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1452 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1455 MUTEX_ENTER(&rx_serverPool_lock);
1456 if (cur_service != NULL) {
1457 ReturnToServerPool(cur_service);
1460 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1461 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1463 /* Scan for eligible incoming calls. A call is not eligible
1464 * if the maximum number of calls for its service type are
1465 * already executing */
1466 /* One thread will process calls FCFS (to prevent starvation),
1467 * while the other threads may run ahead looking for calls which
1468 * have all their input data available immediately. This helps
1469 * keep threads from blocking, waiting for data from the client. */
1470 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1471 service = tcall->conn->service;
1472 if (!QuotaOK(service)) {
1475 if (tno == rxi_fcfs_thread_num
1476 || !tcall->queue_item_header.next) {
1477 /* If we're the fcfs thread , then we'll just use
1478 * this call. If we haven't been able to find an optimal
1479 * choice, and we're at the end of the list, then use a
1480 * 2d choice if one has been identified. Otherwise... */
1481 call = (choice2 ? choice2 : tcall);
1482 service = call->conn->service;
1483 } else if (!queue_IsEmpty(&tcall->rq)) {
1484 struct rx_packet *rp;
1485 rp = queue_First(&tcall->rq, rx_packet);
1486 if (rp->header.seq == 1) {
1488 || (rp->header.flags & RX_LAST_PACKET)) {
1490 } else if (rxi_2dchoice && !choice2
1491 && !(tcall->flags & RX_CALL_CLEARED)
1492 && (tcall->rprev > rxi_HardAckRate)) {
1501 ReturnToServerPool(service);
1508 MUTEX_EXIT(&rx_serverPool_lock);
1509 MUTEX_ENTER(&call->lock);
1511 if (call->flags & RX_CALL_WAIT_PROC) {
1512 call->flags &= ~RX_CALL_WAIT_PROC;
1513 MUTEX_ENTER(&rx_stats_mutex);
1515 MUTEX_EXIT(&rx_stats_mutex);
1518 if (call->state != RX_STATE_PRECALL || call->error) {
1519 MUTEX_EXIT(&call->lock);
1520 MUTEX_ENTER(&rx_serverPool_lock);
1521 ReturnToServerPool(service);
1526 if (queue_IsEmpty(&call->rq)
1527 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1528 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1530 CLEAR_CALL_QUEUE_LOCK(call);
1533 /* If there are no eligible incoming calls, add this process
1534 * to the idle server queue, to wait for one */
1538 *socketp = OSI_NULLSOCKET;
1540 sq->socketp = socketp;
1541 queue_Append(&rx_idleServerQueue, sq);
1542 #ifndef AFS_AIX41_ENV
1543 rx_waitForPacket = sq;
1545 rx_waitingForPacket = sq;
1546 #endif /* AFS_AIX41_ENV */
1548 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1550 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1551 MUTEX_EXIT(&rx_serverPool_lock);
1552 return (struct rx_call *)0;
1555 } while (!(call = sq->newcall)
1556 && !(socketp && *socketp != OSI_NULLSOCKET));
1557 MUTEX_EXIT(&rx_serverPool_lock);
1559 MUTEX_ENTER(&call->lock);
1565 MUTEX_ENTER(&freeSQEList_lock);
1566 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1567 rx_FreeSQEList = sq;
1568 MUTEX_EXIT(&freeSQEList_lock);
1571 clock_GetTime(&call->startTime);
1572 call->state = RX_STATE_ACTIVE;
1573 call->mode = RX_MODE_RECEIVING;
1574 #ifdef RX_KERNEL_TRACE
1575 if (ICL_SETACTIVE(afs_iclSetp)) {
1576 int glockOwner = ISAFS_GLOCK();
1579 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1580 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1587 rxi_calltrace(RX_CALL_START, call);
1588 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1589 call->conn->service->servicePort, call->conn->service->serviceId,
1592 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1593 MUTEX_EXIT(&call->lock);
1595 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1600 #else /* RX_ENABLE_LOCKS */
1602 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1604 struct rx_serverQueueEntry *sq;
1605 register struct rx_call *call = (struct rx_call *)0, *choice2;
1606 struct rx_service *service = NULL;
1611 MUTEX_ENTER(&freeSQEList_lock);
1613 if ((sq = rx_FreeSQEList)) {
1614 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1615 MUTEX_EXIT(&freeSQEList_lock);
1616 } else { /* otherwise allocate a new one and return that */
1617 MUTEX_EXIT(&freeSQEList_lock);
1618 sq = (struct rx_serverQueueEntry *)
1619 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1620 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1621 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1623 MUTEX_ENTER(&sq->lock);
1625 if (cur_service != NULL) {
1626 cur_service->nRequestsRunning--;
1627 if (cur_service->nRequestsRunning < cur_service->minProcs)
1631 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1632 register struct rx_call *tcall, *ncall;
1633 /* Scan for eligible incoming calls. A call is not eligible
1634 * if the maximum number of calls for its service type are
1635 * already executing */
1636 /* One thread will process calls FCFS (to prevent starvation),
1637 * while the other threads may run ahead looking for calls which
1638 * have all their input data available immediately. This helps
1639 * keep threads from blocking, waiting for data from the client. */
1640 choice2 = (struct rx_call *)0;
1641 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1642 service = tcall->conn->service;
1643 if (QuotaOK(service)) {
1644 if (tno == rxi_fcfs_thread_num
1645 || !tcall->queue_item_header.next) {
1646 /* If we're the fcfs thread, then we'll just use
1647 * this call. If we haven't been able to find an optimal
1648 * choice, and we're at the end of the list, then use a
1649 * 2d choice if one has been identified. Otherwise... */
1650 call = (choice2 ? choice2 : tcall);
1651 service = call->conn->service;
1652 } else if (!queue_IsEmpty(&tcall->rq)) {
1653 struct rx_packet *rp;
1654 rp = queue_First(&tcall->rq, rx_packet);
1655 if (rp->header.seq == 1
1657 || (rp->header.flags & RX_LAST_PACKET))) {
1659 } else if (rxi_2dchoice && !choice2
1660 && !(tcall->flags & RX_CALL_CLEARED)
1661 && (tcall->rprev > rxi_HardAckRate)) {
1674 /* we can't schedule a call if there's no data!!! */
1675 /* send an ack if there's no data, if we're missing the
1676 * first packet, or we're missing something between first
1677 * and last -- there's a "hole" in the incoming data. */
1678 if (queue_IsEmpty(&call->rq)
1679 || queue_First(&call->rq, rx_packet)->header.seq != 1
1680 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1681 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1683 call->flags &= (~RX_CALL_WAIT_PROC);
1684 service->nRequestsRunning++;
1685 /* just started call in minProcs pool, need fewer to maintain
1687 if (service->nRequestsRunning <= service->minProcs)
1691 /* MUTEX_EXIT(&call->lock); */
1693 /* If there are no eligible incoming calls, add this process
1694 * to the idle server queue, to wait for one */
1697 *socketp = OSI_NULLSOCKET;
1699 sq->socketp = socketp;
1700 queue_Append(&rx_idleServerQueue, sq);
1704 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1707 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1708 return (struct rx_call *)0;
1711 } while (!(call = sq->newcall)
1712 && !(socketp && *socketp != OSI_NULLSOCKET));
1714 MUTEX_EXIT(&sq->lock);
1716 MUTEX_ENTER(&freeSQEList_lock);
1717 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1718 rx_FreeSQEList = sq;
1719 MUTEX_EXIT(&freeSQEList_lock);
1722 clock_GetTime(&call->startTime);
1723 call->state = RX_STATE_ACTIVE;
1724 call->mode = RX_MODE_RECEIVING;
1725 #ifdef RX_KERNEL_TRACE
1726 if (ICL_SETACTIVE(afs_iclSetp)) {
1727 int glockOwner = ISAFS_GLOCK();
1730 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1731 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1738 rxi_calltrace(RX_CALL_START, call);
1739 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1740 call->conn->service->servicePort, call->conn->service->serviceId,
1743 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1751 #endif /* RX_ENABLE_LOCKS */
1755 /* Establish a procedure to be called when a packet arrives for a
1756 * call. This routine will be called at most once after each call,
1757 * and will also be called if there is an error condition on the or
1758 * the call is complete. Used by multi rx to build a selection
1759 * function which determines which of several calls is likely to be a
1760 * good one to read from.
1761 * NOTE: the way this is currently implemented it is probably only a
1762 * good idea to (1) use it immediately after a newcall (clients only)
1763 * and (2) only use it once. Other uses currently void your warranty
1766 rx_SetArrivalProc(register struct rx_call *call,
1767 register VOID(*proc) (register struct rx_call * call,
1768 register struct multi_handle * mh,
1769 register int index),
1770 register VOID * handle, register VOID * arg)
1772 call->arrivalProc = proc;
1773 call->arrivalProcHandle = handle;
1774 call->arrivalProcArg = arg;
1777 /* Call is finished (possibly prematurely). Return rc to the peer, if
1778 * appropriate, and return the final error code from the conversation
1782 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1784 register struct rx_connection *conn = call->conn;
1785 register struct rx_service *service;
1786 register struct rx_packet *tp; /* Temporary packet pointer */
1787 register struct rx_packet *nxp; /* Next packet pointer, for queue_Scan */
1791 dpf(("rx_EndCall(call %x)\n", call));
1795 MUTEX_ENTER(&call->lock);
1797 if (rc == 0 && call->error == 0) {
1798 call->abortCode = 0;
1799 call->abortCount = 0;
1802 call->arrivalProc = (VOID(*)())0;
1803 if (rc && call->error == 0) {
1804 rxi_CallError(call, rc);
1805 /* Send an abort message to the peer if this error code has
1806 * only just been set. If it was set previously, assume the
1807 * peer has already been sent the error code or will request it
1809 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1811 if (conn->type == RX_SERVER_CONNECTION) {
1812 /* Make sure reply or at least dummy reply is sent */
1813 if (call->mode == RX_MODE_RECEIVING) {
1814 rxi_WriteProc(call, 0, 0);
1816 if (call->mode == RX_MODE_SENDING) {
1817 rxi_FlushWrite(call);
1819 service = conn->service;
1820 rxi_calltrace(RX_CALL_END, call);
1821 /* Call goes to hold state until reply packets are acknowledged */
1822 if (call->tfirst + call->nSoftAcked < call->tnext) {
1823 call->state = RX_STATE_HOLD;
1825 call->state = RX_STATE_DALLY;
1826 rxi_ClearTransmitQueue(call, 0);
1827 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1828 rxevent_Cancel(call->keepAliveEvent, call,
1829 RX_CALL_REFCOUNT_ALIVE);
1831 } else { /* Client connection */
1833 /* Make sure server receives input packets, in the case where
1834 * no reply arguments are expected */
1835 if ((call->mode == RX_MODE_SENDING)
1836 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1837 (void)rxi_ReadProc(call, &dummy, 1);
1840 /* If we had an outstanding delayed ack, be nice to the server
1841 * and force-send it now.
1843 if (call->delayedAckEvent) {
1844 rxevent_Cancel(call->delayedAckEvent, call,
1845 RX_CALL_REFCOUNT_DELAY);
1846 call->delayedAckEvent = NULL;
1847 rxi_SendDelayedAck(NULL, call, NULL);
1850 /* We need to release the call lock since it's lower than the
1851 * conn_call_lock and we don't want to hold the conn_call_lock
1852 * over the rx_ReadProc call. The conn_call_lock needs to be held
1853 * here for the case where rx_NewCall is perusing the calls on
1854 * the connection structure. We don't want to signal until
1855 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1856 * have checked this call, found it active and by the time it
1857 * goes to sleep, will have missed the signal.
1859 MUTEX_EXIT(&call->lock);
1860 MUTEX_ENTER(&conn->conn_call_lock);
1861 MUTEX_ENTER(&call->lock);
1862 MUTEX_ENTER(&conn->conn_data_lock);
1863 conn->flags |= RX_CONN_BUSY;
1864 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1865 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1866 MUTEX_EXIT(&conn->conn_data_lock);
1867 #ifdef RX_ENABLE_LOCKS
1868 CV_BROADCAST(&conn->conn_call_cv);
1873 #ifdef RX_ENABLE_LOCKS
1875 MUTEX_EXIT(&conn->conn_data_lock);
1877 #endif /* RX_ENABLE_LOCKS */
1878 call->state = RX_STATE_DALLY;
1880 error = call->error;
1882 /* currentPacket, nLeft, and NFree must be zeroed here, because
1883 * ResetCall cannot: ResetCall may be called at splnet(), in the
1884 * kernel version, and may interrupt the macros rx_Read or
1885 * rx_Write, which run at normal priority for efficiency. */
1886 if (call->currentPacket) {
1887 rxi_FreePacket(call->currentPacket);
1888 call->currentPacket = (struct rx_packet *)0;
1889 call->nLeft = call->nFree = call->curlen = 0;
1891 call->nLeft = call->nFree = call->curlen = 0;
1893 /* Free any packets from the last call to ReadvProc/WritevProc */
1894 for (queue_Scan(&call->iovq, tp, nxp, rx_packet)) {
1899 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1900 MUTEX_EXIT(&call->lock);
1901 if (conn->type == RX_CLIENT_CONNECTION) {
1902 MUTEX_EXIT(&conn->conn_call_lock);
1903 conn->flags &= ~RX_CONN_BUSY;
1908 * Map errors to the local host's errno.h format.
1910 error = ntoh_syserr_conv(error);
1914 #if !defined(KERNEL)
1916 /* Call this routine when shutting down a server or client (especially
1917 * clients). This will allow Rx to gracefully garbage collect server
1918 * connections, and reduce the number of retries that a server might
1919 * make to a dead client.
1920 * This is not quite right, since some calls may still be ongoing and
1921 * we can't lock them to destroy them. */
1925 register struct rx_connection **conn_ptr, **conn_end;
1927 INIT_PTHREAD_LOCKS LOCK_RX_INIT if (rxinit_status == 1) {
1928 UNLOCK_RX_INIT return; /* Already shutdown. */
1930 rxi_DeleteCachedConnections();
1931 if (rx_connHashTable) {
1932 MUTEX_ENTER(&rx_connHashTable_lock);
1933 for (conn_ptr = &rx_connHashTable[0], conn_end =
1934 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
1936 struct rx_connection *conn, *next;
1937 for (conn = *conn_ptr; conn; conn = next) {
1939 if (conn->type == RX_CLIENT_CONNECTION) {
1940 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
1942 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
1943 #ifdef RX_ENABLE_LOCKS
1944 rxi_DestroyConnectionNoLock(conn);
1945 #else /* RX_ENABLE_LOCKS */
1946 rxi_DestroyConnection(conn);
1947 #endif /* RX_ENABLE_LOCKS */
1951 #ifdef RX_ENABLE_LOCKS
1952 while (rx_connCleanup_list) {
1953 struct rx_connection *conn;
1954 conn = rx_connCleanup_list;
1955 rx_connCleanup_list = rx_connCleanup_list->next;
1956 MUTEX_EXIT(&rx_connHashTable_lock);
1957 rxi_CleanupConnection(conn);
1958 MUTEX_ENTER(&rx_connHashTable_lock);
1960 MUTEX_EXIT(&rx_connHashTable_lock);
1961 #endif /* RX_ENABLE_LOCKS */
1969 /* if we wakeup packet waiter too often, can get in loop with two
1970 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
1972 rxi_PacketsUnWait(void)
1974 if (!rx_waitingForPackets) {
1978 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
1979 return; /* still over quota */
1982 rx_waitingForPackets = 0;
1983 #ifdef RX_ENABLE_LOCKS
1984 CV_BROADCAST(&rx_waitingForPackets_cv);
1986 osi_rxWakeup(&rx_waitingForPackets);
1992 /* ------------------Internal interfaces------------------------- */
1994 /* Return this process's service structure for the
1995 * specified socket and service */
1997 rxi_FindService(register osi_socket socket, register u_short serviceId)
1999 register struct rx_service **sp;
2000 for (sp = &rx_services[0]; *sp; sp++) {
2001 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2007 /* Allocate a call structure, for the indicated channel of the
2008 * supplied connection. The mode and state of the call must be set by
2009 * the caller. Returns the call with mutex locked. */
2011 rxi_NewCall(register struct rx_connection *conn, register int channel)
2013 register struct rx_call *call;
2014 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2015 register struct rx_call *cp; /* Call pointer temp */
2016 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2017 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2019 /* Grab an existing call structure, or allocate a new one.
2020 * Existing call structures are assumed to have been left reset by
2022 MUTEX_ENTER(&rx_freeCallQueue_lock);
2024 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2026 * EXCEPT that the TQ might not yet be cleared out.
2027 * Skip over those with in-use TQs.
2030 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2031 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2037 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2038 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2039 call = queue_First(&rx_freeCallQueue, rx_call);
2040 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2042 MUTEX_ENTER(&rx_stats_mutex);
2043 rx_stats.nFreeCallStructs--;
2044 MUTEX_EXIT(&rx_stats_mutex);
2045 MUTEX_EXIT(&rx_freeCallQueue_lock);
2046 MUTEX_ENTER(&call->lock);
2047 CLEAR_CALL_QUEUE_LOCK(call);
2048 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2049 /* Now, if TQ wasn't cleared earlier, do it now. */
2050 if (call->flags & RX_CALL_TQ_CLEARME) {
2051 rxi_ClearTransmitQueue(call, 0);
2052 queue_Init(&call->tq);
2054 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2055 /* Bind the call to its connection structure */
2057 rxi_ResetCall(call, 1);
2059 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2061 MUTEX_EXIT(&rx_freeCallQueue_lock);
2062 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2063 MUTEX_ENTER(&call->lock);
2064 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2065 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2066 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2068 MUTEX_ENTER(&rx_stats_mutex);
2069 rx_stats.nCallStructs++;
2070 MUTEX_EXIT(&rx_stats_mutex);
2071 /* Initialize once-only items */
2072 queue_Init(&call->tq);
2073 queue_Init(&call->rq);
2074 queue_Init(&call->iovq);
2075 /* Bind the call to its connection structure (prereq for reset) */
2077 rxi_ResetCall(call, 1);
2079 call->channel = channel;
2080 call->callNumber = &conn->callNumber[channel];
2081 /* Note that the next expected call number is retained (in
2082 * conn->callNumber[i]), even if we reallocate the call structure
2084 conn->call[channel] = call;
2085 /* if the channel's never been used (== 0), we should start at 1, otherwise
2086 * the call number is valid from the last time this channel was used */
2087 if (*call->callNumber == 0)
2088 *call->callNumber = 1;
2093 /* A call has been inactive long enough that so we can throw away
2094 * state, including the call structure, which is placed on the call
2096 * Call is locked upon entry.
2097 * haveCTLock set if called from rxi_ReapConnections
2099 #ifdef RX_ENABLE_LOCKS
2101 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2102 #else /* RX_ENABLE_LOCKS */
2104 rxi_FreeCall(register struct rx_call *call)
2105 #endif /* RX_ENABLE_LOCKS */
2107 register int channel = call->channel;
2108 register struct rx_connection *conn = call->conn;
2111 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2112 (*call->callNumber)++;
2113 rxi_ResetCall(call, 0);
2114 call->conn->call[channel] = (struct rx_call *)0;
2116 MUTEX_ENTER(&rx_freeCallQueue_lock);
2117 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2118 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2119 /* A call may be free even though its transmit queue is still in use.
2120 * Since we search the call list from head to tail, put busy calls at
2121 * the head of the list, and idle calls at the tail.
2123 if (call->flags & RX_CALL_TQ_BUSY)
2124 queue_Prepend(&rx_freeCallQueue, call);
2126 queue_Append(&rx_freeCallQueue, call);
2127 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2128 queue_Append(&rx_freeCallQueue, call);
2129 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2130 MUTEX_ENTER(&rx_stats_mutex);
2131 rx_stats.nFreeCallStructs++;
2132 MUTEX_EXIT(&rx_stats_mutex);
2134 MUTEX_EXIT(&rx_freeCallQueue_lock);
2136 /* Destroy the connection if it was previously slated for
2137 * destruction, i.e. the Rx client code previously called
2138 * rx_DestroyConnection (client connections), or
2139 * rxi_ReapConnections called the same routine (server
2140 * connections). Only do this, however, if there are no
2141 * outstanding calls. Note that for fine grain locking, there appears
2142 * to be a deadlock in that rxi_FreeCall has a call locked and
2143 * DestroyConnectionNoLock locks each call in the conn. But note a
2144 * few lines up where we have removed this call from the conn.
2145 * If someone else destroys a connection, they either have no
2146 * call lock held or are going through this section of code.
2148 if (conn->flags & RX_CONN_DESTROY_ME) {
2149 MUTEX_ENTER(&conn->conn_data_lock);
2151 MUTEX_EXIT(&conn->conn_data_lock);
2152 #ifdef RX_ENABLE_LOCKS
2154 rxi_DestroyConnectionNoLock(conn);
2156 rxi_DestroyConnection(conn);
2157 #else /* RX_ENABLE_LOCKS */
2158 rxi_DestroyConnection(conn);
2159 #endif /* RX_ENABLE_LOCKS */
2163 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2165 rxi_Alloc(register size_t size)
2169 #if defined(AFS_AIX41_ENV) && defined(KERNEL)
2170 /* Grab the AFS filesystem lock. See afs/osi.h for the lock
2173 int glockOwner = ISAFS_GLOCK();
2177 MUTEX_ENTER(&rx_stats_mutex);
2179 rxi_Allocsize += size;
2180 MUTEX_EXIT(&rx_stats_mutex);
2181 #if (defined(AFS_AIX32_ENV) || defined(AFS_HPUX_ENV)) && !defined(AFS_HPUX100_ENV) && defined(KERNEL)
2182 if (size > AFS_SMALLOCSIZ) {
2183 p = (char *)osi_AllocMediumSpace(size);
2185 p = (char *)osi_AllocSmall(size, 1);
2186 #if defined(AFS_AIX41_ENV) && defined(KERNEL)
2191 p = (char *)osi_Alloc(size);
2194 osi_Panic("rxi_Alloc error");
2200 rxi_Free(void *addr, register size_t size)
2202 #if defined(AFS_AIX41_ENV) && defined(KERNEL)
2203 /* Grab the AFS filesystem lock. See afs/osi.h for the lock
2206 int glockOwner = ISAFS_GLOCK();
2210 MUTEX_ENTER(&rx_stats_mutex);
2212 rxi_Allocsize -= size;
2213 MUTEX_EXIT(&rx_stats_mutex);
2214 #if (defined(AFS_AIX32_ENV) || defined(AFS_HPUX_ENV)) && !defined(AFS_HPUX100_ENV) && defined(KERNEL)
2215 if (size > AFS_SMALLOCSIZ)
2216 osi_FreeMediumSpace(addr);
2218 osi_FreeSmall(addr);
2219 #if defined(AFS_AIX41_ENV) && defined(KERNEL)
2224 osi_Free(addr, size);
2228 /* Find the peer process represented by the supplied (host,port)
2229 * combination. If there is no appropriate active peer structure, a
2230 * new one will be allocated and initialized
2231 * The origPeer, if set, is a pointer to a peer structure on which the
2232 * refcount will be be decremented. This is used to replace the peer
2233 * structure hanging off a connection structure */
2235 rxi_FindPeer(register afs_uint32 host, register u_short port,
2236 struct rx_peer *origPeer, int create)
2238 register struct rx_peer *pp;
2240 hashIndex = PEER_HASH(host, port);
2241 MUTEX_ENTER(&rx_peerHashTable_lock);
2242 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2243 if ((pp->host == host) && (pp->port == port))
2248 pp = rxi_AllocPeer(); /* This bzero's *pp */
2249 pp->host = host; /* set here or in InitPeerParams is zero */
2251 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2252 queue_Init(&pp->congestionQueue);
2253 queue_Init(&pp->rpcStats);
2254 pp->next = rx_peerHashTable[hashIndex];
2255 rx_peerHashTable[hashIndex] = pp;
2256 rxi_InitPeerParams(pp);
2257 MUTEX_ENTER(&rx_stats_mutex);
2258 rx_stats.nPeerStructs++;
2259 MUTEX_EXIT(&rx_stats_mutex);
2266 origPeer->refCount--;
2267 MUTEX_EXIT(&rx_peerHashTable_lock);
2272 /* Find the connection at (host, port) started at epoch, and with the
2273 * given connection id. Creates the server connection if necessary.
2274 * The type specifies whether a client connection or a server
2275 * connection is desired. In both cases, (host, port) specify the
2276 * peer's (host, pair) pair. Client connections are not made
2277 * automatically by this routine. The parameter socket gives the
2278 * socket descriptor on which the packet was received. This is used,
2279 * in the case of server connections, to check that *new* connections
2280 * come via a valid (port, serviceId). Finally, the securityIndex
2281 * parameter must match the existing index for the connection. If a
2282 * server connection is created, it will be created using the supplied
2283 * index, if the index is valid for this service */
2284 struct rx_connection *
2285 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2286 register u_short port, u_short serviceId, afs_uint32 cid,
2287 afs_uint32 epoch, int type, u_int securityIndex)
2289 int hashindex, flag;
2290 register struct rx_connection *conn;
2291 hashindex = CONN_HASH(host, port, cid, epoch, type);
2292 MUTEX_ENTER(&rx_connHashTable_lock);
2293 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2294 rx_connHashTable[hashindex],
2297 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2298 && (epoch == conn->epoch)) {
2299 register struct rx_peer *pp = conn->peer;
2300 if (securityIndex != conn->securityIndex) {
2301 /* this isn't supposed to happen, but someone could forge a packet
2302 * like this, and there seems to be some CM bug that makes this
2303 * happen from time to time -- in which case, the fileserver
2305 MUTEX_EXIT(&rx_connHashTable_lock);
2306 return (struct rx_connection *)0;
2308 if (pp->host == host && pp->port == port)
2310 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2312 if (type == RX_CLIENT_CONNECTION && (conn->epoch & 0x80000000))
2316 /* the connection rxLastConn that was used the last time is not the
2317 ** one we are looking for now. Hence, start searching in the hash */
2319 conn = rx_connHashTable[hashindex];
2324 struct rx_service *service;
2325 if (type == RX_CLIENT_CONNECTION) {
2326 MUTEX_EXIT(&rx_connHashTable_lock);
2327 return (struct rx_connection *)0;
2329 service = rxi_FindService(socket, serviceId);
2330 if (!service || (securityIndex >= service->nSecurityObjects)
2331 || (service->securityObjects[securityIndex] == 0)) {
2332 MUTEX_EXIT(&rx_connHashTable_lock);
2333 return (struct rx_connection *)0;
2335 conn = rxi_AllocConnection(); /* This bzero's the connection */
2336 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2337 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2338 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2339 conn->next = rx_connHashTable[hashindex];
2340 rx_connHashTable[hashindex] = conn;
2341 conn->peer = rxi_FindPeer(host, port, 0, 1);
2342 conn->type = RX_SERVER_CONNECTION;
2343 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2344 conn->epoch = epoch;
2345 conn->cid = cid & RX_CIDMASK;
2346 /* conn->serial = conn->lastSerial = 0; */
2347 /* conn->timeout = 0; */
2348 conn->ackRate = RX_FAST_ACK_RATE;
2349 conn->service = service;
2350 conn->serviceId = serviceId;
2351 conn->securityIndex = securityIndex;
2352 conn->securityObject = service->securityObjects[securityIndex];
2353 conn->nSpecific = 0;
2354 conn->specific = NULL;
2355 rx_SetConnDeadTime(conn, service->connDeadTime);
2356 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2357 /* Notify security object of the new connection */
2358 RXS_NewConnection(conn->securityObject, conn);
2359 /* XXXX Connection timeout? */
2360 if (service->newConnProc)
2361 (*service->newConnProc) (conn);
2362 MUTEX_ENTER(&rx_stats_mutex);
2363 rx_stats.nServerConns++;
2364 MUTEX_EXIT(&rx_stats_mutex);
2367 MUTEX_ENTER(&conn->conn_data_lock);
2369 MUTEX_EXIT(&conn->conn_data_lock);
2371 rxLastConn = conn; /* store this connection as the last conn used */
2372 MUTEX_EXIT(&rx_connHashTable_lock);
2376 /* There are two packet tracing routines available for testing and monitoring
2377 * Rx. One is called just after every packet is received and the other is
2378 * called just before every packet is sent. Received packets, have had their
2379 * headers decoded, and packets to be sent have not yet had their headers
2380 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2381 * containing the network address. Both can be modified. The return value, if
2382 * non-zero, indicates that the packet should be dropped. */
2384 int (*rx_justReceived) () = 0;
2385 int (*rx_almostSent) () = 0;
2387 /* A packet has been received off the interface. Np is the packet, socket is
2388 * the socket number it was received from (useful in determining which service
2389 * this packet corresponds to), and (host, port) reflect the host,port of the
2390 * sender. This call returns the packet to the caller if it is finished with
2391 * it, rather than de-allocating it, just as a small performance hack */
2394 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2395 afs_uint32 host, u_short port, int *tnop,
2396 struct rx_call **newcallp)
2398 register struct rx_call *call;
2399 register struct rx_connection *conn;
2401 afs_uint32 currentCallNumber;
2407 struct rx_packet *tnp;
2410 /* We don't print out the packet until now because (1) the time may not be
2411 * accurate enough until now in the lwp implementation (rx_Listener only gets
2412 * the time after the packet is read) and (2) from a protocol point of view,
2413 * this is the first time the packet has been seen */
2414 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2415 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2416 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2417 np->header.serial, packetType, host, port, np->header.serviceId,
2418 np->header.epoch, np->header.cid, np->header.callNumber,
2419 np->header.seq, np->header.flags, np));
2422 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2423 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2426 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2427 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2430 /* If an input tracer function is defined, call it with the packet and
2431 * network address. Note this function may modify its arguments. */
2432 if (rx_justReceived) {
2433 struct sockaddr_in addr;
2435 addr.sin_family = AF_INET;
2436 addr.sin_port = port;
2437 addr.sin_addr.s_addr = host;
2438 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2439 addr.sin_len = sizeof(addr);
2440 #endif /* AFS_OSF_ENV */
2441 drop = (*rx_justReceived) (np, &addr);
2442 /* drop packet if return value is non-zero */
2445 port = addr.sin_port; /* in case fcn changed addr */
2446 host = addr.sin_addr.s_addr;
2450 /* If packet was not sent by the client, then *we* must be the client */
2451 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2452 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2454 /* Find the connection (or fabricate one, if we're the server & if
2455 * necessary) associated with this packet */
2457 rxi_FindConnection(socket, host, port, np->header.serviceId,
2458 np->header.cid, np->header.epoch, type,
2459 np->header.securityIndex);
2462 /* If no connection found or fabricated, just ignore the packet.
2463 * (An argument could be made for sending an abort packet for
2468 MUTEX_ENTER(&conn->conn_data_lock);
2469 if (conn->maxSerial < np->header.serial)
2470 conn->maxSerial = np->header.serial;
2471 MUTEX_EXIT(&conn->conn_data_lock);
2473 /* If the connection is in an error state, send an abort packet and ignore
2474 * the incoming packet */
2476 /* Don't respond to an abort packet--we don't want loops! */
2477 MUTEX_ENTER(&conn->conn_data_lock);
2478 if (np->header.type != RX_PACKET_TYPE_ABORT)
2479 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2481 MUTEX_EXIT(&conn->conn_data_lock);
2485 /* Check for connection-only requests (i.e. not call specific). */
2486 if (np->header.callNumber == 0) {
2487 switch (np->header.type) {
2488 case RX_PACKET_TYPE_ABORT:
2489 /* What if the supplied error is zero? */
2490 rxi_ConnectionError(conn, ntohl(rx_GetInt32(np, 0)));
2491 MUTEX_ENTER(&conn->conn_data_lock);
2493 MUTEX_EXIT(&conn->conn_data_lock);
2495 case RX_PACKET_TYPE_CHALLENGE:
2496 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2497 MUTEX_ENTER(&conn->conn_data_lock);
2499 MUTEX_EXIT(&conn->conn_data_lock);
2501 case RX_PACKET_TYPE_RESPONSE:
2502 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2503 MUTEX_ENTER(&conn->conn_data_lock);
2505 MUTEX_EXIT(&conn->conn_data_lock);
2507 case RX_PACKET_TYPE_PARAMS:
2508 case RX_PACKET_TYPE_PARAMS + 1:
2509 case RX_PACKET_TYPE_PARAMS + 2:
2510 /* ignore these packet types for now */
2511 MUTEX_ENTER(&conn->conn_data_lock);
2513 MUTEX_EXIT(&conn->conn_data_lock);
2518 /* Should not reach here, unless the peer is broken: send an
2520 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2521 MUTEX_ENTER(&conn->conn_data_lock);
2522 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2524 MUTEX_EXIT(&conn->conn_data_lock);
2529 channel = np->header.cid & RX_CHANNELMASK;
2530 call = conn->call[channel];
2531 #ifdef RX_ENABLE_LOCKS
2533 MUTEX_ENTER(&call->lock);
2534 /* Test to see if call struct is still attached to conn. */
2535 if (call != conn->call[channel]) {
2537 MUTEX_EXIT(&call->lock);
2538 if (type == RX_SERVER_CONNECTION) {
2539 call = conn->call[channel];
2540 /* If we started with no call attached and there is one now,
2541 * another thread is also running this routine and has gotten
2542 * the connection channel. We should drop this packet in the tests
2543 * below. If there was a call on this connection and it's now
2544 * gone, then we'll be making a new call below.
2545 * If there was previously a call and it's now different then
2546 * the old call was freed and another thread running this routine
2547 * has created a call on this channel. One of these two threads
2548 * has a packet for the old call and the code below handles those
2552 MUTEX_ENTER(&call->lock);
2554 /* This packet can't be for this call. If the new call address is
2555 * 0 then no call is running on this channel. If there is a call
2556 * then, since this is a client connection we're getting data for
2557 * it must be for the previous call.
2559 MUTEX_ENTER(&rx_stats_mutex);
2560 rx_stats.spuriousPacketsRead++;
2561 MUTEX_EXIT(&rx_stats_mutex);
2562 MUTEX_ENTER(&conn->conn_data_lock);
2564 MUTEX_EXIT(&conn->conn_data_lock);
2569 currentCallNumber = conn->callNumber[channel];
2571 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2572 if (np->header.callNumber < currentCallNumber) {
2573 MUTEX_ENTER(&rx_stats_mutex);
2574 rx_stats.spuriousPacketsRead++;
2575 MUTEX_EXIT(&rx_stats_mutex);
2576 #ifdef RX_ENABLE_LOCKS
2578 MUTEX_EXIT(&call->lock);
2580 MUTEX_ENTER(&conn->conn_data_lock);
2582 MUTEX_EXIT(&conn->conn_data_lock);
2586 MUTEX_ENTER(&conn->conn_call_lock);
2587 call = rxi_NewCall(conn, channel);
2588 MUTEX_EXIT(&conn->conn_call_lock);
2589 *call->callNumber = np->header.callNumber;
2590 call->state = RX_STATE_PRECALL;
2591 clock_GetTime(&call->queueTime);
2592 hzero(call->bytesSent);
2593 hzero(call->bytesRcvd);
2594 rxi_KeepAliveOn(call);
2595 } else if (np->header.callNumber != currentCallNumber) {
2596 /* Wait until the transmit queue is idle before deciding
2597 * whether to reset the current call. Chances are that the
2598 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2601 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2602 while ((call->state == RX_STATE_ACTIVE)
2603 && (call->flags & RX_CALL_TQ_BUSY)) {
2604 call->flags |= RX_CALL_TQ_WAIT;
2605 #ifdef RX_ENABLE_LOCKS
2606 CV_WAIT(&call->cv_tq, &call->lock);
2607 #else /* RX_ENABLE_LOCKS */
2608 osi_rxSleep(&call->tq);
2609 #endif /* RX_ENABLE_LOCKS */
2611 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2612 /* If the new call cannot be taken right now send a busy and set
2613 * the error condition in this call, so that it terminates as
2614 * quickly as possible */
2615 if (call->state == RX_STATE_ACTIVE) {
2616 struct rx_packet *tp;
2618 rxi_CallError(call, RX_CALL_DEAD);
2619 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2621 MUTEX_EXIT(&call->lock);
2622 MUTEX_ENTER(&conn->conn_data_lock);
2624 MUTEX_EXIT(&conn->conn_data_lock);
2627 rxi_ResetCall(call, 0);
2628 *call->callNumber = np->header.callNumber;
2629 call->state = RX_STATE_PRECALL;
2630 clock_GetTime(&call->queueTime);
2631 hzero(call->bytesSent);
2632 hzero(call->bytesRcvd);
2634 * If the number of queued calls exceeds the overload
2635 * threshold then abort this call.
2637 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2638 struct rx_packet *tp;
2640 rxi_CallError(call, rx_BusyError);
2641 tp = rxi_SendCallAbort(call, np, 1, 0);
2642 MUTEX_EXIT(&call->lock);
2643 MUTEX_ENTER(&conn->conn_data_lock);
2645 MUTEX_EXIT(&conn->conn_data_lock);
2648 rxi_KeepAliveOn(call);
2650 /* Continuing call; do nothing here. */
2652 } else { /* we're the client */
2653 /* Ignore all incoming acknowledgements for calls in DALLY state */
2654 if (call && (call->state == RX_STATE_DALLY)
2655 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2656 MUTEX_ENTER(&rx_stats_mutex);
2657 rx_stats.ignorePacketDally++;
2658 MUTEX_EXIT(&rx_stats_mutex);
2659 #ifdef RX_ENABLE_LOCKS
2661 MUTEX_EXIT(&call->lock);
2664 MUTEX_ENTER(&conn->conn_data_lock);
2666 MUTEX_EXIT(&conn->conn_data_lock);
2670 /* Ignore anything that's not relevant to the current call. If there
2671 * isn't a current call, then no packet is relevant. */
2672 if (!call || (np->header.callNumber != currentCallNumber)) {
2673 MUTEX_ENTER(&rx_stats_mutex);
2674 rx_stats.spuriousPacketsRead++;
2675 MUTEX_EXIT(&rx_stats_mutex);
2676 #ifdef RX_ENABLE_LOCKS
2678 MUTEX_EXIT(&call->lock);
2681 MUTEX_ENTER(&conn->conn_data_lock);
2683 MUTEX_EXIT(&conn->conn_data_lock);
2686 /* If the service security object index stamped in the packet does not
2687 * match the connection's security index, ignore the packet */
2688 if (np->header.securityIndex != conn->securityIndex) {
2689 #ifdef RX_ENABLE_LOCKS
2690 MUTEX_EXIT(&call->lock);
2692 MUTEX_ENTER(&conn->conn_data_lock);
2694 MUTEX_EXIT(&conn->conn_data_lock);
2698 /* If we're receiving the response, then all transmit packets are
2699 * implicitly acknowledged. Get rid of them. */
2700 if (np->header.type == RX_PACKET_TYPE_DATA) {
2701 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2702 /* XXX Hack. Because we must release the global rx lock when
2703 * sending packets (osi_NetSend) we drop all acks while we're
2704 * traversing the tq in rxi_Start sending packets out because
2705 * packets may move to the freePacketQueue as result of being here!
2706 * So we drop these packets until we're safely out of the
2707 * traversing. Really ugly!
2708 * For fine grain RX locking, we set the acked field in the
2709 * packets and let rxi_Start remove them from the transmit queue.
2711 if (call->flags & RX_CALL_TQ_BUSY) {
2712 #ifdef RX_ENABLE_LOCKS
2713 rxi_SetAcksInTransmitQueue(call);
2716 return np; /* xmitting; drop packet */
2719 rxi_ClearTransmitQueue(call, 0);
2721 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2722 rxi_ClearTransmitQueue(call, 0);
2723 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2725 if (np->header.type == RX_PACKET_TYPE_ACK) {
2726 /* now check to see if this is an ack packet acknowledging that the
2727 * server actually *lost* some hard-acked data. If this happens we
2728 * ignore this packet, as it may indicate that the server restarted in
2729 * the middle of a call. It is also possible that this is an old ack
2730 * packet. We don't abort the connection in this case, because this
2731 * *might* just be an old ack packet. The right way to detect a server
2732 * restart in the midst of a call is to notice that the server epoch
2734 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2735 * XXX unacknowledged. I think that this is off-by-one, but
2736 * XXX I don't dare change it just yet, since it will
2737 * XXX interact badly with the server-restart detection
2738 * XXX code in receiveackpacket. */
2739 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2740 MUTEX_ENTER(&rx_stats_mutex);
2741 rx_stats.spuriousPacketsRead++;
2742 MUTEX_EXIT(&rx_stats_mutex);
2743 MUTEX_EXIT(&call->lock);
2744 MUTEX_ENTER(&conn->conn_data_lock);
2746 MUTEX_EXIT(&conn->conn_data_lock);
2750 } /* else not a data packet */
2753 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2754 /* Set remote user defined status from packet */
2755 call->remoteStatus = np->header.userStatus;
2757 /* Note the gap between the expected next packet and the actual
2758 * packet that arrived, when the new packet has a smaller serial number
2759 * than expected. Rioses frequently reorder packets all by themselves,
2760 * so this will be quite important with very large window sizes.
2761 * Skew is checked against 0 here to avoid any dependence on the type of
2762 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2764 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2765 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2766 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2768 MUTEX_ENTER(&conn->conn_data_lock);
2769 skew = conn->lastSerial - np->header.serial;
2770 conn->lastSerial = np->header.serial;
2771 MUTEX_EXIT(&conn->conn_data_lock);
2773 register struct rx_peer *peer;
2775 if (skew > peer->inPacketSkew) {
2776 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2778 peer->inPacketSkew = skew;
2782 /* Now do packet type-specific processing */
2783 switch (np->header.type) {
2784 case RX_PACKET_TYPE_DATA:
2785 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2788 case RX_PACKET_TYPE_ACK:
2789 /* Respond immediately to ack packets requesting acknowledgement
2791 if (np->header.flags & RX_REQUEST_ACK) {
2793 (void)rxi_SendCallAbort(call, 0, 1, 0);
2795 (void)rxi_SendAck(call, 0, np->header.serial,
2796 RX_ACK_PING_RESPONSE, 1);
2798 np = rxi_ReceiveAckPacket(call, np, 1);
2800 case RX_PACKET_TYPE_ABORT:
2801 /* An abort packet: reset the connection, passing the error up to
2803 /* What if error is zero? */
2804 rxi_CallError(call, ntohl(*(afs_int32 *) rx_DataOf(np)));
2806 case RX_PACKET_TYPE_BUSY:
2809 case RX_PACKET_TYPE_ACKALL:
2810 /* All packets acknowledged, so we can drop all packets previously
2811 * readied for sending */
2812 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2813 /* XXX Hack. We because we can't release the global rx lock when
2814 * sending packets (osi_NetSend) we drop all ack pkts while we're
2815 * traversing the tq in rxi_Start sending packets out because
2816 * packets may move to the freePacketQueue as result of being
2817 * here! So we drop these packets until we're safely out of the
2818 * traversing. Really ugly!
2819 * For fine grain RX locking, we set the acked field in the packets
2820 * and let rxi_Start remove the packets from the transmit queue.
2822 if (call->flags & RX_CALL_TQ_BUSY) {
2823 #ifdef RX_ENABLE_LOCKS
2824 rxi_SetAcksInTransmitQueue(call);
2826 #else /* RX_ENABLE_LOCKS */
2828 return np; /* xmitting; drop packet */
2829 #endif /* RX_ENABLE_LOCKS */
2831 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2832 rxi_ClearTransmitQueue(call, 0);
2835 /* Should not reach here, unless the peer is broken: send an abort
2837 rxi_CallError(call, RX_PROTOCOL_ERROR);
2838 np = rxi_SendCallAbort(call, np, 1, 0);
2841 /* Note when this last legitimate packet was received, for keep-alive
2842 * processing. Note, we delay getting the time until now in the hope that
2843 * the packet will be delivered to the user before any get time is required
2844 * (if not, then the time won't actually be re-evaluated here). */
2845 call->lastReceiveTime = clock_Sec();
2846 MUTEX_EXIT(&call->lock);
2847 MUTEX_ENTER(&conn->conn_data_lock);
2849 MUTEX_EXIT(&conn->conn_data_lock);
2853 /* return true if this is an "interesting" connection from the point of view
2854 of someone trying to debug the system */
2856 rxi_IsConnInteresting(struct rx_connection *aconn)
2859 register struct rx_call *tcall;
2861 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2863 for (i = 0; i < RX_MAXCALLS; i++) {
2864 tcall = aconn->call[i];
2866 if ((tcall->state == RX_STATE_PRECALL)
2867 || (tcall->state == RX_STATE_ACTIVE))
2869 if ((tcall->mode == RX_MODE_SENDING)
2870 || (tcall->mode == RX_MODE_RECEIVING))
2878 /* if this is one of the last few packets AND it wouldn't be used by the
2879 receiving call to immediately satisfy a read request, then drop it on
2880 the floor, since accepting it might prevent a lock-holding thread from
2881 making progress in its reading. If a call has been cleared while in
2882 the precall state then ignore all subsequent packets until the call
2883 is assigned to a thread. */
2886 TooLow(struct rx_packet *ap, struct rx_call *acall)
2889 MUTEX_ENTER(&rx_stats_mutex);
2890 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
2891 && (acall->state == RX_STATE_PRECALL))
2892 || ((rx_nFreePackets < rxi_dataQuota + 2)
2893 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
2894 && (acall->flags & RX_CALL_READER_WAIT)))) {
2897 MUTEX_EXIT(&rx_stats_mutex);
2903 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
2904 struct rx_call *acall)
2906 struct rx_call *call = acall;
2910 MUTEX_ENTER(&conn->conn_data_lock);
2911 conn->checkReachEvent = NULL;
2912 waiting = conn->flags & RX_CONN_ATTACHWAIT;
2915 MUTEX_EXIT(&conn->conn_data_lock);
2919 MUTEX_ENTER(&conn->conn_call_lock);
2920 MUTEX_ENTER(&conn->conn_data_lock);
2921 for (i = 0; i < RX_MAXCALLS; i++) {
2922 struct rx_call *tc = conn->call[i];
2923 if (tc && tc->state == RX_STATE_PRECALL) {
2929 /* Indicate that rxi_CheckReachEvent is no longer running by
2930 * clearing the flag. Must be atomic under conn_data_lock to
2931 * avoid a new call slipping by: rxi_CheckConnReach holds
2932 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
2934 conn->flags &= ~RX_CONN_ATTACHWAIT;
2935 MUTEX_EXIT(&conn->conn_data_lock);
2936 MUTEX_EXIT(&conn->conn_call_lock);
2941 MUTEX_ENTER(&call->lock);
2942 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
2944 MUTEX_EXIT(&call->lock);
2946 clock_GetTime(&when);
2947 when.sec += RX_CHECKREACH_TIMEOUT;
2948 MUTEX_ENTER(&conn->conn_data_lock);
2949 if (!conn->checkReachEvent) {
2951 conn->checkReachEvent =
2952 rxevent_Post(&when, rxi_CheckReachEvent, conn, NULL);
2954 MUTEX_EXIT(&conn->conn_data_lock);
2960 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
2962 struct rx_service *service = conn->service;
2963 struct rx_peer *peer = conn->peer;
2964 afs_uint32 now, lastReach;
2966 if (service->checkReach == 0)
2970 MUTEX_ENTER(&peer->peer_lock);
2971 lastReach = peer->lastReachTime;
2972 MUTEX_EXIT(&peer->peer_lock);
2973 if (now - lastReach < RX_CHECKREACH_TTL)
2976 MUTEX_ENTER(&conn->conn_data_lock);
2977 if (conn->flags & RX_CONN_ATTACHWAIT) {
2978 MUTEX_EXIT(&conn->conn_data_lock);
2981 conn->flags |= RX_CONN_ATTACHWAIT;
2982 MUTEX_EXIT(&conn->conn_data_lock);
2983 if (!conn->checkReachEvent)
2984 rxi_CheckReachEvent(NULL, conn, call);
2989 /* try to attach call, if authentication is complete */
2991 TryAttach(register struct rx_call *acall, register osi_socket socket,
2992 register int *tnop, register struct rx_call **newcallp,
2995 struct rx_connection *conn = acall->conn;
2997 if (conn->type == RX_SERVER_CONNECTION
2998 && acall->state == RX_STATE_PRECALL) {
2999 /* Don't attach until we have any req'd. authentication. */
3000 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3001 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3002 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3003 /* Note: this does not necessarily succeed; there
3004 * may not any proc available
3007 rxi_ChallengeOn(acall->conn);
3012 /* A data packet has been received off the interface. This packet is
3013 * appropriate to the call (the call is in the right state, etc.). This
3014 * routine can return a packet to the caller, for re-use */
3017 rxi_ReceiveDataPacket(register struct rx_call *call,
3018 register struct rx_packet *np, int istack,
3019 osi_socket socket, afs_uint32 host, u_short port,
3020 int *tnop, struct rx_call **newcallp)
3022 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3026 afs_uint32 seq, serial, flags;
3028 struct rx_packet *tnp;
3030 MUTEX_ENTER(&rx_stats_mutex);
3031 rx_stats.dataPacketsRead++;
3032 MUTEX_EXIT(&rx_stats_mutex);
3035 /* If there are no packet buffers, drop this new packet, unless we can find
3036 * packet buffers from inactive calls */
3038 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3039 MUTEX_ENTER(&rx_freePktQ_lock);
3040 rxi_NeedMorePackets = TRUE;
3041 MUTEX_EXIT(&rx_freePktQ_lock);
3042 MUTEX_ENTER(&rx_stats_mutex);
3043 rx_stats.noPacketBuffersOnRead++;
3044 MUTEX_EXIT(&rx_stats_mutex);
3045 call->rprev = np->header.serial;
3046 rxi_calltrace(RX_TRACE_DROP, call);
3047 dpf(("packet %x dropped on receipt - quota problems", np));
3049 rxi_ClearReceiveQueue(call);
3050 clock_GetTime(&when);
3051 clock_Add(&when, &rx_softAckDelay);
3052 if (!call->delayedAckEvent
3053 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3054 rxevent_Cancel(call->delayedAckEvent, call,
3055 RX_CALL_REFCOUNT_DELAY);
3056 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3057 call->delayedAckEvent =
3058 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3060 /* we've damaged this call already, might as well do it in. */
3066 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3067 * packet is one of several packets transmitted as a single
3068 * datagram. Do not send any soft or hard acks until all packets
3069 * in a jumbogram have been processed. Send negative acks right away.
3071 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3072 /* tnp is non-null when there are more packets in the
3073 * current jumbo gram */
3080 seq = np->header.seq;
3081 serial = np->header.serial;
3082 flags = np->header.flags;
3084 /* If the call is in an error state, send an abort message */
3086 return rxi_SendCallAbort(call, np, istack, 0);
3088 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3089 * AFS 3.5 jumbogram. */
3090 if (flags & RX_JUMBO_PACKET) {
3091 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3096 if (np->header.spare != 0) {
3097 MUTEX_ENTER(&call->conn->conn_data_lock);
3098 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3099 MUTEX_EXIT(&call->conn->conn_data_lock);
3102 /* The usual case is that this is the expected next packet */
3103 if (seq == call->rnext) {
3105 /* Check to make sure it is not a duplicate of one already queued */
3106 if (queue_IsNotEmpty(&call->rq)
3107 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3108 MUTEX_ENTER(&rx_stats_mutex);
3109 rx_stats.dupPacketsRead++;
3110 MUTEX_EXIT(&rx_stats_mutex);
3111 dpf(("packet %x dropped on receipt - duplicate", np));
3112 rxevent_Cancel(call->delayedAckEvent, call,
3113 RX_CALL_REFCOUNT_DELAY);
3114 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3120 /* It's the next packet. Stick it on the receive queue
3121 * for this call. Set newPackets to make sure we wake
3122 * the reader once all packets have been processed */
3123 queue_Prepend(&call->rq, np);
3125 np = NULL; /* We can't use this anymore */
3128 /* If an ack is requested then set a flag to make sure we
3129 * send an acknowledgement for this packet */
3130 if (flags & RX_REQUEST_ACK) {
3131 ackNeeded = RX_ACK_REQUESTED;
3134 /* Keep track of whether we have received the last packet */
3135 if (flags & RX_LAST_PACKET) {
3136 call->flags |= RX_CALL_HAVE_LAST;
3140 /* Check whether we have all of the packets for this call */
3141 if (call->flags & RX_CALL_HAVE_LAST) {
3142 afs_uint32 tseq; /* temporary sequence number */
3143 struct rx_packet *tp; /* Temporary packet pointer */
3144 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3146 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3147 if (tseq != tp->header.seq)
3149 if (tp->header.flags & RX_LAST_PACKET) {
3150 call->flags |= RX_CALL_RECEIVE_DONE;
3157 /* Provide asynchronous notification for those who want it
3158 * (e.g. multi rx) */
3159 if (call->arrivalProc) {
3160 (*call->arrivalProc) (call, call->arrivalProcHandle,
3161 (int)call->arrivalProcArg);
3162 call->arrivalProc = (VOID(*)())0;
3165 /* Update last packet received */
3168 /* If there is no server process serving this call, grab
3169 * one, if available. We only need to do this once. If a
3170 * server thread is available, this thread becomes a server
3171 * thread and the server thread becomes a listener thread. */
3173 TryAttach(call, socket, tnop, newcallp, 0);
3176 /* This is not the expected next packet. */
3178 /* Determine whether this is a new or old packet, and if it's
3179 * a new one, whether it fits into the current receive window.
3180 * Also figure out whether the packet was delivered in sequence.
3181 * We use the prev variable to determine whether the new packet
3182 * is the successor of its immediate predecessor in the
3183 * receive queue, and the missing flag to determine whether
3184 * any of this packets predecessors are missing. */
3186 afs_uint32 prev; /* "Previous packet" sequence number */
3187 struct rx_packet *tp; /* Temporary packet pointer */
3188 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3189 int missing; /* Are any predecessors missing? */
3191 /* If the new packet's sequence number has been sent to the
3192 * application already, then this is a duplicate */
3193 if (seq < call->rnext) {
3194 MUTEX_ENTER(&rx_stats_mutex);
3195 rx_stats.dupPacketsRead++;
3196 MUTEX_EXIT(&rx_stats_mutex);
3197 rxevent_Cancel(call->delayedAckEvent, call,
3198 RX_CALL_REFCOUNT_DELAY);
3199 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3205 /* If the sequence number is greater than what can be
3206 * accomodated by the current window, then send a negative
3207 * acknowledge and drop the packet */
3208 if ((call->rnext + call->rwind) <= seq) {
3209 rxevent_Cancel(call->delayedAckEvent, call,
3210 RX_CALL_REFCOUNT_DELAY);
3211 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3218 /* Look for the packet in the queue of old received packets */
3219 for (prev = call->rnext - 1, missing =
3220 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3221 /*Check for duplicate packet */
3222 if (seq == tp->header.seq) {
3223 MUTEX_ENTER(&rx_stats_mutex);
3224 rx_stats.dupPacketsRead++;
3225 MUTEX_EXIT(&rx_stats_mutex);
3226 rxevent_Cancel(call->delayedAckEvent, call,
3227 RX_CALL_REFCOUNT_DELAY);
3228 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3234 /* If we find a higher sequence packet, break out and
3235 * insert the new packet here. */
3236 if (seq < tp->header.seq)
3238 /* Check for missing packet */
3239 if (tp->header.seq != prev + 1) {
3243 prev = tp->header.seq;
3246 /* Keep track of whether we have received the last packet. */
3247 if (flags & RX_LAST_PACKET) {
3248 call->flags |= RX_CALL_HAVE_LAST;
3251 /* It's within the window: add it to the the receive queue.
3252 * tp is left by the previous loop either pointing at the
3253 * packet before which to insert the new packet, or at the
3254 * queue head if the queue is empty or the packet should be
3256 queue_InsertBefore(tp, np);
3260 /* Check whether we have all of the packets for this call */
3261 if ((call->flags & RX_CALL_HAVE_LAST)
3262 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3263 afs_uint32 tseq; /* temporary sequence number */
3266 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3267 if (tseq != tp->header.seq)
3269 if (tp->header.flags & RX_LAST_PACKET) {
3270 call->flags |= RX_CALL_RECEIVE_DONE;
3277 /* We need to send an ack of the packet is out of sequence,
3278 * or if an ack was requested by the peer. */
3279 if (seq != prev + 1 || missing || (flags & RX_REQUEST_ACK)) {
3280 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3283 /* Acknowledge the last packet for each call */
3284 if (flags & RX_LAST_PACKET) {
3295 * If the receiver is waiting for an iovec, fill the iovec
3296 * using the data from the receive queue */
3297 if (call->flags & RX_CALL_IOVEC_WAIT) {
3298 didHardAck = rxi_FillReadVec(call, serial);
3299 /* the call may have been aborted */
3308 /* Wakeup the reader if any */
3309 if ((call->flags & RX_CALL_READER_WAIT)
3310 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3311 || (call->iovNext >= call->iovMax)
3312 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3313 call->flags &= ~RX_CALL_READER_WAIT;
3314 #ifdef RX_ENABLE_LOCKS
3315 CV_BROADCAST(&call->cv_rq);
3317 osi_rxWakeup(&call->rq);
3323 * Send an ack when requested by the peer, or once every
3324 * rxi_SoftAckRate packets until the last packet has been
3325 * received. Always send a soft ack for the last packet in
3326 * the server's reply. */
3328 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3329 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3330 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3331 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3332 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3333 } else if (call->nSoftAcks) {
3334 clock_GetTime(&when);
3335 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3336 clock_Add(&when, &rx_lastAckDelay);
3338 clock_Add(&when, &rx_softAckDelay);
3340 if (!call->delayedAckEvent
3341 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3342 rxevent_Cancel(call->delayedAckEvent, call,
3343 RX_CALL_REFCOUNT_DELAY);
3344 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3345 call->delayedAckEvent =
3346 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3348 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3349 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3356 static void rxi_ComputeRate();
3360 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3362 struct rx_peer *peer = conn->peer;
3364 MUTEX_ENTER(&peer->peer_lock);
3365 peer->lastReachTime = clock_Sec();
3366 MUTEX_EXIT(&peer->peer_lock);
3368 MUTEX_ENTER(&conn->conn_data_lock);
3369 if (conn->flags & RX_CONN_ATTACHWAIT) {
3372 conn->flags &= ~RX_CONN_ATTACHWAIT;
3373 MUTEX_EXIT(&conn->conn_data_lock);
3375 for (i = 0; i < RX_MAXCALLS; i++) {
3376 struct rx_call *call = conn->call[i];
3379 MUTEX_ENTER(&call->lock);
3380 /* tnop can be null if newcallp is null */
3381 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3383 MUTEX_EXIT(&call->lock);
3387 MUTEX_EXIT(&conn->conn_data_lock);
3390 /* rxi_ComputePeerNetStats
3392 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3393 * estimates (like RTT and throughput) based on ack packets. Caller
3394 * must ensure that the packet in question is the right one (i.e.
3395 * serial number matches).
3398 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3399 struct rx_ackPacket *ap, struct rx_packet *np)
3401 struct rx_peer *peer = call->conn->peer;
3403 /* Use RTT if not delayed by client. */
3404 if (ap->reason != RX_ACK_DELAY)
3405 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3407 rxi_ComputeRate(peer, call, p, np, ap->reason);
3411 /* The real smarts of the whole thing. */
3413 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3416 struct rx_ackPacket *ap;
3418 register struct rx_packet *tp;
3419 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3420 register struct rx_connection *conn = call->conn;
3421 struct rx_peer *peer = conn->peer;
3424 /* because there are CM's that are bogus, sending weird values for this. */
3425 afs_uint32 skew = 0;
3430 int newAckCount = 0;
3431 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3432 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3434 MUTEX_ENTER(&rx_stats_mutex);
3435 rx_stats.ackPacketsRead++;
3436 MUTEX_EXIT(&rx_stats_mutex);
3437 ap = (struct rx_ackPacket *)rx_DataOf(np);
3438 nbytes = rx_Contiguous(np) - ((ap->acks) - (u_char *) ap);
3440 return np; /* truncated ack packet */
3442 /* depends on ack packet struct */
3443 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3444 first = ntohl(ap->firstPacket);
3445 serial = ntohl(ap->serial);
3446 /* temporarily disabled -- needs to degrade over time
3447 * skew = ntohs(ap->maxSkew); */
3449 /* Ignore ack packets received out of order */
3450 if (first < call->tfirst) {
3454 if (np->header.flags & RX_SLOW_START_OK) {
3455 call->flags |= RX_CALL_SLOW_START_OK;
3458 if (ap->reason == RX_ACK_PING_RESPONSE)
3459 rxi_UpdatePeerReach(conn, call);
3464 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3465 ap->reason, ntohl(ap->previousPacket),
3466 (unsigned int)np->header.seq, (unsigned int)serial,
3467 (unsigned int)skew, ntohl(ap->firstPacket));
3470 for (offset = 0; offset < nAcks; offset++)
3471 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3478 /* Update the outgoing packet skew value to the latest value of
3479 * the peer's incoming packet skew value. The ack packet, of
3480 * course, could arrive out of order, but that won't affect things
3482 MUTEX_ENTER(&peer->peer_lock);
3483 peer->outPacketSkew = skew;
3485 /* Check for packets that no longer need to be transmitted, and
3486 * discard them. This only applies to packets positively
3487 * acknowledged as having been sent to the peer's upper level.
3488 * All other packets must be retained. So only packets with
3489 * sequence numbers < ap->firstPacket are candidates. */
3490 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3491 if (tp->header.seq >= first)
3493 call->tfirst = tp->header.seq + 1;
3495 && (tp->header.serial == serial || tp->firstSerial == serial))
3496 rxi_ComputePeerNetStats(call, tp, ap, np);
3497 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3498 /* XXX Hack. Because we have to release the global rx lock when sending
3499 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3500 * in rxi_Start sending packets out because packets may move to the
3501 * freePacketQueue as result of being here! So we drop these packets until
3502 * we're safely out of the traversing. Really ugly!
3503 * To make it even uglier, if we're using fine grain locking, we can
3504 * set the ack bits in the packets and have rxi_Start remove the packets
3505 * when it's done transmitting.
3507 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3510 if (call->flags & RX_CALL_TQ_BUSY) {
3511 #ifdef RX_ENABLE_LOCKS
3512 tp->flags |= RX_PKTFLAG_ACKED;
3513 call->flags |= RX_CALL_TQ_SOME_ACKED;
3514 #else /* RX_ENABLE_LOCKS */
3516 #endif /* RX_ENABLE_LOCKS */
3518 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3521 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3526 /* Give rate detector a chance to respond to ping requests */
3527 if (ap->reason == RX_ACK_PING_RESPONSE) {
3528 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3532 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3534 /* Now go through explicit acks/nacks and record the results in
3535 * the waiting packets. These are packets that can't be released
3536 * yet, even with a positive acknowledge. This positive
3537 * acknowledge only means the packet has been received by the
3538 * peer, not that it will be retained long enough to be sent to
3539 * the peer's upper level. In addition, reset the transmit timers
3540 * of any missing packets (those packets that must be missing
3541 * because this packet was out of sequence) */
3543 call->nSoftAcked = 0;
3544 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3545 /* Update round trip time if the ack was stimulated on receipt
3547 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3548 #ifdef RX_ENABLE_LOCKS
3549 if (tp->header.seq >= first)
3550 #endif /* RX_ENABLE_LOCKS */
3551 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3553 && (tp->header.serial == serial || tp->firstSerial == serial))
3554 rxi_ComputePeerNetStats(call, tp, ap, np);
3556 /* Set the acknowledge flag per packet based on the
3557 * information in the ack packet. An acknowlegded packet can
3558 * be downgraded when the server has discarded a packet it
3559 * soacked previously, or when an ack packet is received
3560 * out of sequence. */
3561 if (tp->header.seq < first) {
3562 /* Implicit ack information */
3563 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3566 tp->flags |= RX_PKTFLAG_ACKED;
3567 } else if (tp->header.seq < first + nAcks) {
3568 /* Explicit ack information: set it in the packet appropriately */
3569 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3570 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3572 tp->flags |= RX_PKTFLAG_ACKED;
3580 tp->flags &= ~RX_PKTFLAG_ACKED;
3584 tp->flags &= ~RX_PKTFLAG_ACKED;
3588 /* If packet isn't yet acked, and it has been transmitted at least
3589 * once, reset retransmit time using latest timeout
3590 * ie, this should readjust the retransmit timer for all outstanding
3591 * packets... So we don't just retransmit when we should know better*/
3593 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3594 tp->retryTime = tp->timeSent;
3595 clock_Add(&tp->retryTime, &peer->timeout);
3596 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3597 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3601 /* If the window has been extended by this acknowledge packet,
3602 * then wakeup a sender waiting in alloc for window space, or try
3603 * sending packets now, if he's been sitting on packets due to
3604 * lack of window space */
3605 if (call->tnext < (call->tfirst + call->twind)) {
3606 #ifdef RX_ENABLE_LOCKS
3607 CV_SIGNAL(&call->cv_twind);
3609 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3610 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3611 osi_rxWakeup(&call->twind);
3614 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3615 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3619 /* if the ack packet has a receivelen field hanging off it,
3620 * update our state */
3621 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3624 /* If the ack packet has a "recommended" size that is less than
3625 * what I am using now, reduce my size to match */
3626 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3627 sizeof(afs_int32), &tSize);
3628 tSize = (afs_uint32) ntohl(tSize);
3629 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3631 /* Get the maximum packet size to send to this peer */
3632 rx_packetread(np, rx_AckDataSize(ap->nAcks), sizeof(afs_int32),
3634 tSize = (afs_uint32) ntohl(tSize);
3635 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3636 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3638 /* sanity check - peer might have restarted with different params.
3639 * If peer says "send less", dammit, send less... Peer should never
3640 * be unable to accept packets of the size that prior AFS versions would
3641 * send without asking. */
3642 if (peer->maxMTU != tSize) {
3643 peer->maxMTU = tSize;
3644 peer->MTU = MIN(tSize, peer->MTU);
3645 call->MTU = MIN(call->MTU, tSize);
3649 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3652 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3653 sizeof(afs_int32), &tSize);
3654 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3655 if (tSize < call->twind) { /* smaller than our send */
3656 call->twind = tSize; /* window, we must send less... */
3657 call->ssthresh = MIN(call->twind, call->ssthresh);
3660 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3661 * network MTU confused with the loopback MTU. Calculate the
3662 * maximum MTU here for use in the slow start code below.
3664 maxMTU = peer->maxMTU;
3665 /* Did peer restart with older RX version? */
3666 if (peer->maxDgramPackets > 1) {
3667 peer->maxDgramPackets = 1;
3669 } else if (np->length >=
3670 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3673 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3674 sizeof(afs_int32), &tSize);
3675 tSize = (afs_uint32) ntohl(tSize);
3677 * As of AFS 3.5 we set the send window to match the receive window.
3679 if (tSize < call->twind) {
3680 call->twind = tSize;
3681 call->ssthresh = MIN(call->twind, call->ssthresh);
3682 } else if (tSize > call->twind) {
3683 call->twind = tSize;
3687 * As of AFS 3.5, a jumbogram is more than one fixed size
3688 * packet transmitted in a single UDP datagram. If the remote
3689 * MTU is smaller than our local MTU then never send a datagram
3690 * larger than the natural MTU.
3693 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3694 sizeof(afs_int32), &tSize);
3695 maxDgramPackets = (afs_uint32) ntohl(tSize);
3696 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3698 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
3699 maxDgramPackets = MIN(maxDgramPackets, tSize);
3700 if (maxDgramPackets > 1) {
3701 peer->maxDgramPackets = maxDgramPackets;
3702 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3704 peer->maxDgramPackets = 1;
3705 call->MTU = peer->natMTU;
3707 } else if (peer->maxDgramPackets > 1) {
3708 /* Restarted with lower version of RX */
3709 peer->maxDgramPackets = 1;
3711 } else if (peer->maxDgramPackets > 1
3712 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3713 /* Restarted with lower version of RX */
3714 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3715 peer->natMTU = OLD_MAX_PACKET_SIZE;
3716 peer->MTU = OLD_MAX_PACKET_SIZE;
3717 peer->maxDgramPackets = 1;
3718 peer->nDgramPackets = 1;
3720 call->MTU = OLD_MAX_PACKET_SIZE;
3725 * Calculate how many datagrams were successfully received after
3726 * the first missing packet and adjust the negative ack counter
3731 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3732 if (call->nNacks < nNacked) {
3733 call->nNacks = nNacked;
3742 if (call->flags & RX_CALL_FAST_RECOVER) {
3744 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3746 call->flags &= ~RX_CALL_FAST_RECOVER;
3747 call->cwind = call->nextCwind;
3748 call->nextCwind = 0;
3751 call->nCwindAcks = 0;
3752 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3753 /* Three negative acks in a row trigger congestion recovery */
3754 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3755 MUTEX_EXIT(&peer->peer_lock);
3756 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3757 /* someone else is waiting to start recovery */
3760 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3761 while (call->flags & RX_CALL_TQ_BUSY) {
3762 call->flags |= RX_CALL_TQ_WAIT;
3763 #ifdef RX_ENABLE_LOCKS
3764 CV_WAIT(&call->cv_tq, &call->lock);
3765 #else /* RX_ENABLE_LOCKS */
3766 osi_rxSleep(&call->tq);
3767 #endif /* RX_ENABLE_LOCKS */
3769 MUTEX_ENTER(&peer->peer_lock);
3770 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3771 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3772 call->flags |= RX_CALL_FAST_RECOVER;
3773 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3775 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3776 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3777 call->nextCwind = call->ssthresh;
3780 peer->MTU = call->MTU;
3781 peer->cwind = call->nextCwind;
3782 peer->nDgramPackets = call->nDgramPackets;
3784 call->congestSeq = peer->congestSeq;
3785 /* Reset the resend times on the packets that were nacked
3786 * so we will retransmit as soon as the window permits*/
3787 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3789 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3790 clock_Zero(&tp->retryTime);
3792 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3797 /* If cwind is smaller than ssthresh, then increase
3798 * the window one packet for each ack we receive (exponential
3800 * If cwind is greater than or equal to ssthresh then increase
3801 * the congestion window by one packet for each cwind acks we
3802 * receive (linear growth). */
3803 if (call->cwind < call->ssthresh) {
3805 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3806 call->nCwindAcks = 0;
3808 call->nCwindAcks += newAckCount;
3809 if (call->nCwindAcks >= call->cwind) {
3810 call->nCwindAcks = 0;
3811 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3815 * If we have received several acknowledgements in a row then
3816 * it is time to increase the size of our datagrams
3818 if ((int)call->nAcks > rx_nDgramThreshold) {
3819 if (peer->maxDgramPackets > 1) {
3820 if (call->nDgramPackets < peer->maxDgramPackets) {
3821 call->nDgramPackets++;
3823 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3824 } else if (call->MTU < peer->maxMTU) {
3825 call->MTU += peer->natMTU;
3826 call->MTU = MIN(call->MTU, peer->maxMTU);
3832 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3834 /* Servers need to hold the call until all response packets have
3835 * been acknowledged. Soft acks are good enough since clients
3836 * are not allowed to clear their receive queues. */
3837 if (call->state == RX_STATE_HOLD
3838 && call->tfirst + call->nSoftAcked >= call->tnext) {
3839 call->state = RX_STATE_DALLY;
3840 rxi_ClearTransmitQueue(call, 0);
3841 } else if (!queue_IsEmpty(&call->tq)) {
3842 rxi_Start(0, call, istack);
3847 /* Received a response to a challenge packet */
3849 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
3850 register struct rx_packet *np, int istack)
3854 /* Ignore the packet if we're the client */
3855 if (conn->type == RX_CLIENT_CONNECTION)
3858 /* If already authenticated, ignore the packet (it's probably a retry) */
3859 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
3862 /* Otherwise, have the security object evaluate the response packet */
3863 error = RXS_CheckResponse(conn->securityObject, conn, np);
3865 /* If the response is invalid, reset the connection, sending
3866 * an abort to the peer */
3870 rxi_ConnectionError(conn, error);
3871 MUTEX_ENTER(&conn->conn_data_lock);
3872 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3873 MUTEX_EXIT(&conn->conn_data_lock);
3876 /* If the response is valid, any calls waiting to attach
3877 * servers can now do so */
3880 for (i = 0; i < RX_MAXCALLS; i++) {
3881 struct rx_call *call = conn->call[i];
3883 MUTEX_ENTER(&call->lock);
3884 if (call->state == RX_STATE_PRECALL)
3885 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
3886 /* tnop can be null if newcallp is null */
3887 MUTEX_EXIT(&call->lock);
3891 /* Update the peer reachability information, just in case
3892 * some calls went into attach-wait while we were waiting
3893 * for authentication..
3895 rxi_UpdatePeerReach(conn, NULL);
3900 /* A client has received an authentication challenge: the security
3901 * object is asked to cough up a respectable response packet to send
3902 * back to the server. The server is responsible for retrying the
3903 * challenge if it fails to get a response. */
3906 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
3907 register struct rx_packet *np, int istack)
3911 /* Ignore the challenge if we're the server */
3912 if (conn->type == RX_SERVER_CONNECTION)
3915 /* Ignore the challenge if the connection is otherwise idle; someone's
3916 * trying to use us as an oracle. */
3917 if (!rxi_HasActiveCalls(conn))
3920 /* Send the security object the challenge packet. It is expected to fill
3921 * in the response. */
3922 error = RXS_GetResponse(conn->securityObject, conn, np);
3924 /* If the security object is unable to return a valid response, reset the
3925 * connection and send an abort to the peer. Otherwise send the response
3926 * packet to the peer connection. */
3928 rxi_ConnectionError(conn, error);
3929 MUTEX_ENTER(&conn->conn_data_lock);
3930 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3931 MUTEX_EXIT(&conn->conn_data_lock);
3933 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
3934 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
3940 /* Find an available server process to service the current request in
3941 * the given call structure. If one isn't available, queue up this
3942 * call so it eventually gets one */
3944 rxi_AttachServerProc(register struct rx_call *call,
3945 register osi_socket socket, register int *tnop,
3946 register struct rx_call **newcallp)
3948 register struct rx_serverQueueEntry *sq;
3949 register struct rx_service *service = call->conn->service;
3950 register int haveQuota = 0;
3952 /* May already be attached */
3953 if (call->state == RX_STATE_ACTIVE)
3956 MUTEX_ENTER(&rx_serverPool_lock);
3958 haveQuota = QuotaOK(service);
3959 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
3960 /* If there are no processes available to service this call,
3961 * put the call on the incoming call queue (unless it's
3962 * already on the queue).
3964 #ifdef RX_ENABLE_LOCKS
3966 ReturnToServerPool(service);
3967 #endif /* RX_ENABLE_LOCKS */
3969 if (!(call->flags & RX_CALL_WAIT_PROC)) {
3970 call->flags |= RX_CALL_WAIT_PROC;
3971 MUTEX_ENTER(&rx_stats_mutex);
3973 MUTEX_EXIT(&rx_stats_mutex);
3974 rxi_calltrace(RX_CALL_ARRIVAL, call);
3975 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
3976 queue_Append(&rx_incomingCallQueue, call);
3979 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
3981 /* If hot threads are enabled, and both newcallp and sq->socketp
3982 * are non-null, then this thread will process the call, and the
3983 * idle server thread will start listening on this threads socket.
3986 if (rx_enable_hot_thread && newcallp && sq->socketp) {
3989 *sq->socketp = socket;
3990 clock_GetTime(&call->startTime);
3991 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
3995 if (call->flags & RX_CALL_WAIT_PROC) {
3996 /* Conservative: I don't think this should happen */
3997 call->flags &= ~RX_CALL_WAIT_PROC;
3998 MUTEX_ENTER(&rx_stats_mutex);
4000 MUTEX_EXIT(&rx_stats_mutex);
4001 if (queue_IsOnQueue(call))
4004 call->state = RX_STATE_ACTIVE;
4005 call->mode = RX_MODE_RECEIVING;
4006 #ifdef RX_KERNEL_TRACE
4008 int glockOwner = ISAFS_GLOCK();
4011 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4012 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4018 if (call->flags & RX_CALL_CLEARED) {
4019 /* send an ack now to start the packet flow up again */
4020 call->flags &= ~RX_CALL_CLEARED;
4021 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4023 #ifdef RX_ENABLE_LOCKS
4026 service->nRequestsRunning++;
4027 if (service->nRequestsRunning <= service->minProcs)
4033 MUTEX_EXIT(&rx_serverPool_lock);
4036 /* Delay the sending of an acknowledge event for a short while, while
4037 * a new call is being prepared (in the case of a client) or a reply
4038 * is being prepared (in the case of a server). Rather than sending
4039 * an ack packet, an ACKALL packet is sent. */
4041 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4043 #ifdef RX_ENABLE_LOCKS
4045 MUTEX_ENTER(&call->lock);
4046 call->delayedAckEvent = NULL;
4047 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4049 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4050 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4052 MUTEX_EXIT(&call->lock);
4053 #else /* RX_ENABLE_LOCKS */
4055 call->delayedAckEvent = NULL;
4056 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4057 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4058 #endif /* RX_ENABLE_LOCKS */
4062 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4065 #ifdef RX_ENABLE_LOCKS
4067 MUTEX_ENTER(&call->lock);
4068 if (event == call->delayedAckEvent)
4069 call->delayedAckEvent = NULL;
4070 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4072 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4074 MUTEX_EXIT(&call->lock);
4075 #else /* RX_ENABLE_LOCKS */
4077 call->delayedAckEvent = NULL;
4078 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4079 #endif /* RX_ENABLE_LOCKS */
4083 #ifdef RX_ENABLE_LOCKS
4084 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4085 * clearing them out.
4088 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4090 register struct rx_packet *p, *tp;
4093 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4096 p->flags |= RX_PKTFLAG_ACKED;
4100 call->flags |= RX_CALL_TQ_CLEARME;
4101 call->flags |= RX_CALL_TQ_SOME_ACKED;
4104 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4105 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4106 call->tfirst = call->tnext;
4107 call->nSoftAcked = 0;
4109 if (call->flags & RX_CALL_FAST_RECOVER) {
4110 call->flags &= ~RX_CALL_FAST_RECOVER;
4111 call->cwind = call->nextCwind;
4112 call->nextCwind = 0;
4115 CV_SIGNAL(&call->cv_twind);
4117 #endif /* RX_ENABLE_LOCKS */
4119 /* Clear out the transmit queue for the current call (all packets have
4120 * been received by peer) */
4122 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4124 register struct rx_packet *p, *tp;
4126 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4127 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4129 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4132 p->flags |= RX_PKTFLAG_ACKED;
4136 call->flags |= RX_CALL_TQ_CLEARME;
4137 call->flags |= RX_CALL_TQ_SOME_ACKED;
4140 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4141 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4147 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4148 call->flags &= ~RX_CALL_TQ_CLEARME;
4150 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4152 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4153 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4154 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4155 call->nSoftAcked = 0;
4157 if (call->flags & RX_CALL_FAST_RECOVER) {
4158 call->flags &= ~RX_CALL_FAST_RECOVER;
4159 call->cwind = call->nextCwind;
4161 #ifdef RX_ENABLE_LOCKS
4162 CV_SIGNAL(&call->cv_twind);
4164 osi_rxWakeup(&call->twind);
4169 rxi_ClearReceiveQueue(register struct rx_call *call)
4171 register struct rx_packet *p, *tp;
4172 if (queue_IsNotEmpty(&call->rq)) {
4173 for (queue_Scan(&call->rq, p, tp, rx_packet)) {
4178 rx_packetReclaims++;
4180 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4182 if (call->state == RX_STATE_PRECALL) {
4183 call->flags |= RX_CALL_CLEARED;
4187 /* Send an abort packet for the specified call */
4189 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4190 int istack, int force)
4198 /* Clients should never delay abort messages */
4199 if (rx_IsClientConn(call->conn))
4202 if (call->abortCode != call->error) {
4203 call->abortCode = call->error;
4204 call->abortCount = 0;
4207 if (force || rxi_callAbortThreshhold == 0
4208 || call->abortCount < rxi_callAbortThreshhold) {
4209 if (call->delayedAbortEvent) {
4210 rxevent_Cancel(call->delayedAbortEvent, call,
4211 RX_CALL_REFCOUNT_ABORT);
4213 error = htonl(call->error);
4216 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4217 (char *)&error, sizeof(error), istack);
4218 } else if (!call->delayedAbortEvent) {
4219 clock_GetTime(&when);
4220 clock_Addmsec(&when, rxi_callAbortDelay);
4221 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4222 call->delayedAbortEvent =
4223 rxevent_Post(&when, rxi_SendDelayedCallAbort, call, 0);
4228 /* Send an abort packet for the specified connection. Packet is an
4229 * optional pointer to a packet that can be used to send the abort.
4230 * Once the number of abort messages reaches the threshhold, an
4231 * event is scheduled to send the abort. Setting the force flag
4232 * overrides sending delayed abort messages.
4234 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4235 * to send the abort packet.
4238 rxi_SendConnectionAbort(register struct rx_connection *conn,
4239 struct rx_packet *packet, int istack, int force)
4247 /* Clients should never delay abort messages */
4248 if (rx_IsClientConn(conn))
4251 if (force || rxi_connAbortThreshhold == 0
4252 || conn->abortCount < rxi_connAbortThreshhold) {
4253 if (conn->delayedAbortEvent) {
4254 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4256 error = htonl(conn->error);
4258 MUTEX_EXIT(&conn->conn_data_lock);
4260 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4261 RX_PACKET_TYPE_ABORT, (char *)&error,
4262 sizeof(error), istack);
4263 MUTEX_ENTER(&conn->conn_data_lock);
4264 } else if (!conn->delayedAbortEvent) {
4265 clock_GetTime(&when);
4266 clock_Addmsec(&when, rxi_connAbortDelay);
4267 conn->delayedAbortEvent =
4268 rxevent_Post(&when, rxi_SendDelayedConnAbort, conn, 0);
4273 /* Associate an error all of the calls owned by a connection. Called
4274 * with error non-zero. This is only for really fatal things, like
4275 * bad authentication responses. The connection itself is set in
4276 * error at this point, so that future packets received will be
4279 rxi_ConnectionError(register struct rx_connection *conn,
4280 register afs_int32 error)
4284 MUTEX_ENTER(&conn->conn_data_lock);
4285 if (conn->challengeEvent)
4286 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4287 if (conn->checkReachEvent) {
4288 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4289 conn->checkReachEvent = 0;
4290 conn->flags &= ~RX_CONN_ATTACHWAIT;
4293 MUTEX_EXIT(&conn->conn_data_lock);
4294 for (i = 0; i < RX_MAXCALLS; i++) {
4295 struct rx_call *call = conn->call[i];
4297 MUTEX_ENTER(&call->lock);
4298 rxi_CallError(call, error);
4299 MUTEX_EXIT(&call->lock);
4302 conn->error = error;
4303 MUTEX_ENTER(&rx_stats_mutex);
4304 rx_stats.fatalErrors++;
4305 MUTEX_EXIT(&rx_stats_mutex);
4310 rxi_CallError(register struct rx_call *call, afs_int32 error)
4313 error = call->error;
4314 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4315 if (!(call->flags & RX_CALL_TQ_BUSY)) {
4316 rxi_ResetCall(call, 0);
4319 rxi_ResetCall(call, 0);
4321 call->error = error;
4322 call->mode = RX_MODE_ERROR;
4325 /* Reset various fields in a call structure, and wakeup waiting
4326 * processes. Some fields aren't changed: state & mode are not
4327 * touched (these must be set by the caller), and bufptr, nLeft, and
4328 * nFree are not reset, since these fields are manipulated by
4329 * unprotected macros, and may only be reset by non-interrupting code.
4332 /* this code requires that call->conn be set properly as a pre-condition. */
4333 #endif /* ADAPT_WINDOW */
4336 rxi_ResetCall(register struct rx_call *call, register int newcall)
4339 register struct rx_peer *peer;
4340 struct rx_packet *packet;
4342 /* Notify anyone who is waiting for asynchronous packet arrival */
4343 if (call->arrivalProc) {
4344 (*call->arrivalProc) (call, call->arrivalProcHandle,
4345 (int)call->arrivalProcArg);
4346 call->arrivalProc = (VOID(*)())0;
4349 if (call->delayedAbortEvent) {
4350 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4351 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4353 rxi_SendCallAbort(call, packet, 0, 1);
4354 rxi_FreePacket(packet);
4359 * Update the peer with the congestion information in this call
4360 * so other calls on this connection can pick up where this call
4361 * left off. If the congestion sequence numbers don't match then
4362 * another call experienced a retransmission.
4364 peer = call->conn->peer;
4365 MUTEX_ENTER(&peer->peer_lock);
4367 if (call->congestSeq == peer->congestSeq) {
4368 peer->cwind = MAX(peer->cwind, call->cwind);
4369 peer->MTU = MAX(peer->MTU, call->MTU);
4370 peer->nDgramPackets =
4371 MAX(peer->nDgramPackets, call->nDgramPackets);
4374 call->abortCode = 0;
4375 call->abortCount = 0;
4377 if (peer->maxDgramPackets > 1) {
4378 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4380 call->MTU = peer->MTU;
4382 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4383 call->ssthresh = rx_maxSendWindow;
4384 call->nDgramPackets = peer->nDgramPackets;
4385 call->congestSeq = peer->congestSeq;
4386 MUTEX_EXIT(&peer->peer_lock);
4388 flags = call->flags;
4389 rxi_ClearReceiveQueue(call);
4390 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4391 if (call->flags & RX_CALL_TQ_BUSY) {
4392 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4393 call->flags |= (flags & RX_CALL_TQ_WAIT);
4395 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4397 rxi_ClearTransmitQueue(call, 0);
4398 queue_Init(&call->tq);
4401 queue_Init(&call->rq);
4403 call->rwind = rx_initReceiveWindow;
4404 call->twind = rx_initSendWindow;
4405 call->nSoftAcked = 0;
4406 call->nextCwind = 0;
4409 call->nCwindAcks = 0;
4410 call->nSoftAcks = 0;
4411 call->nHardAcks = 0;
4413 call->tfirst = call->rnext = call->tnext = 1;
4415 call->lastAcked = 0;
4416 call->localStatus = call->remoteStatus = 0;
4418 if (flags & RX_CALL_READER_WAIT) {
4419 #ifdef RX_ENABLE_LOCKS
4420 CV_BROADCAST(&call->cv_rq);
4422 osi_rxWakeup(&call->rq);
4425 if (flags & RX_CALL_WAIT_PACKETS) {
4426 MUTEX_ENTER(&rx_freePktQ_lock);
4427 rxi_PacketsUnWait(); /* XXX */
4428 MUTEX_EXIT(&rx_freePktQ_lock);
4430 #ifdef RX_ENABLE_LOCKS
4431 CV_SIGNAL(&call->cv_twind);
4433 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4434 osi_rxWakeup(&call->twind);
4437 #ifdef RX_ENABLE_LOCKS
4438 /* The following ensures that we don't mess with any queue while some
4439 * other thread might also be doing so. The call_queue_lock field is
4440 * is only modified under the call lock. If the call is in the process
4441 * of being removed from a queue, the call is not locked until the
4442 * the queue lock is dropped and only then is the call_queue_lock field
4443 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4444 * Note that any other routine which removes a call from a queue has to
4445 * obtain the queue lock before examing the queue and removing the call.
4447 if (call->call_queue_lock) {
4448 MUTEX_ENTER(call->call_queue_lock);
4449 if (queue_IsOnQueue(call)) {
4451 if (flags & RX_CALL_WAIT_PROC) {
4452 MUTEX_ENTER(&rx_stats_mutex);
4454 MUTEX_EXIT(&rx_stats_mutex);
4457 MUTEX_EXIT(call->call_queue_lock);
4458 CLEAR_CALL_QUEUE_LOCK(call);
4460 #else /* RX_ENABLE_LOCKS */
4461 if (queue_IsOnQueue(call)) {
4463 if (flags & RX_CALL_WAIT_PROC)
4466 #endif /* RX_ENABLE_LOCKS */
4468 rxi_KeepAliveOff(call);
4469 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4472 /* Send an acknowledge for the indicated packet (seq,serial) of the
4473 * indicated call, for the indicated reason (reason). This
4474 * acknowledge will specifically acknowledge receiving the packet, and
4475 * will also specify which other packets for this call have been
4476 * received. This routine returns the packet that was used to the
4477 * caller. The caller is responsible for freeing it or re-using it.
4478 * This acknowledgement also returns the highest sequence number
4479 * actually read out by the higher level to the sender; the sender
4480 * promises to keep around packets that have not been read by the
4481 * higher level yet (unless, of course, the sender decides to abort
4482 * the call altogether). Any of p, seq, serial, pflags, or reason may
4483 * be set to zero without ill effect. That is, if they are zero, they
4484 * will not convey any information.
4485 * NOW there is a trailer field, after the ack where it will safely be
4486 * ignored by mundanes, which indicates the maximum size packet this
4487 * host can swallow. */
4489 register struct rx_packet *optionalPacket; use to send ack (or null)
4490 int seq; Sequence number of the packet we are acking
4491 int serial; Serial number of the packet
4492 int pflags; Flags field from packet header
4493 int reason; Reason an acknowledge was prompted
4497 rxi_SendAck(register struct rx_call *call,
4498 register struct rx_packet *optionalPacket, int serial, int reason,
4501 struct rx_ackPacket *ap;
4502 register struct rx_packet *rqp;
4503 register struct rx_packet *nxp; /* For queue_Scan */
4504 register struct rx_packet *p;
4509 * Open the receive window once a thread starts reading packets
4511 if (call->rnext > 1) {
4512 call->rwind = rx_maxReceiveWindow;
4515 call->nHardAcks = 0;
4516 call->nSoftAcks = 0;
4517 if (call->rnext > call->lastAcked)
4518 call->lastAcked = call->rnext;
4522 rx_computelen(p, p->length); /* reset length, you never know */
4523 } /* where that's been... */
4524 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4525 /* We won't send the ack, but don't panic. */
4526 return optionalPacket;
4530 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4533 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL)) {
4534 if (!optionalPacket)
4536 return optionalPacket;
4538 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4539 if (rx_Contiguous(p) < templ) {
4540 if (!optionalPacket)
4542 return optionalPacket;
4547 /* MTUXXX failing to send an ack is very serious. We should */
4548 /* try as hard as possible to send even a partial ack; it's */
4549 /* better than nothing. */
4550 ap = (struct rx_ackPacket *)rx_DataOf(p);
4551 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4552 ap->reason = reason;
4554 /* The skew computation used to be bogus, I think it's better now. */
4555 /* We should start paying attention to skew. XXX */
4556 ap->serial = htonl(serial);
4557 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4559 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4560 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4562 /* No fear of running out of ack packet here because there can only be at most
4563 * one window full of unacknowledged packets. The window size must be constrained
4564 * to be less than the maximum ack size, of course. Also, an ack should always
4565 * fit into a single packet -- it should not ever be fragmented. */
4566 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4567 if (!rqp || !call->rq.next
4568 || (rqp->header.seq > (call->rnext + call->rwind))) {
4569 if (!optionalPacket)
4571 rxi_CallError(call, RX_CALL_DEAD);
4572 return optionalPacket;
4575 while (rqp->header.seq > call->rnext + offset)
4576 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4577 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4579 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4580 if (!optionalPacket)
4582 rxi_CallError(call, RX_CALL_DEAD);
4583 return optionalPacket;
4588 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4590 /* these are new for AFS 3.3 */
4591 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4592 templ = htonl(templ);
4593 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4594 templ = htonl(call->conn->peer->ifMTU);
4595 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4596 sizeof(afs_int32), &templ);
4598 /* new for AFS 3.4 */
4599 templ = htonl(call->rwind);
4600 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4601 sizeof(afs_int32), &templ);
4603 /* new for AFS 3.5 */
4604 templ = htonl(call->conn->peer->ifDgramPackets);
4605 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4606 sizeof(afs_int32), &templ);
4608 p->header.serviceId = call->conn->serviceId;
4609 p->header.cid = (call->conn->cid | call->channel);
4610 p->header.callNumber = *call->callNumber;
4612 p->header.securityIndex = call->conn->securityIndex;
4613 p->header.epoch = call->conn->epoch;
4614 p->header.type = RX_PACKET_TYPE_ACK;
4615 p->header.flags = RX_SLOW_START_OK;
4616 if (reason == RX_ACK_PING) {
4617 p->header.flags |= RX_REQUEST_ACK;
4619 clock_GetTime(&call->pingRequestTime);
4622 if (call->conn->type == RX_CLIENT_CONNECTION)
4623 p->header.flags |= RX_CLIENT_INITIATED;
4627 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u",
4628 ap->reason, ntohl(ap->previousPacket),
4629 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4631 for (offset = 0; offset < ap->nAcks; offset++)
4632 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4640 register int i, nbytes = p->length;
4642 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4643 if (nbytes <= p->wirevec[i].iov_len) {
4644 register int savelen, saven;
4646 savelen = p->wirevec[i].iov_len;
4648 p->wirevec[i].iov_len = nbytes;
4650 rxi_Send(call, p, istack);
4651 p->wirevec[i].iov_len = savelen;
4655 nbytes -= p->wirevec[i].iov_len;
4658 MUTEX_ENTER(&rx_stats_mutex);
4659 rx_stats.ackPacketsSent++;
4660 MUTEX_EXIT(&rx_stats_mutex);
4661 if (!optionalPacket)
4663 return optionalPacket; /* Return packet for re-use by caller */
4666 /* Send all of the packets in the list in single datagram */
4668 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4669 int istack, int moreFlag, struct clock *now,
4670 struct clock *retryTime, int resending)
4675 struct rx_connection *conn = call->conn;
4676 struct rx_peer *peer = conn->peer;
4678 MUTEX_ENTER(&peer->peer_lock);
4681 peer->reSends += len;
4682 MUTEX_ENTER(&rx_stats_mutex);
4683 rx_stats.dataPacketsSent += len;
4684 MUTEX_EXIT(&rx_stats_mutex);
4685 MUTEX_EXIT(&peer->peer_lock);
4687 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4691 /* Set the packet flags and schedule the resend events */
4692 /* Only request an ack for the last packet in the list */
4693 for (i = 0; i < len; i++) {
4694 list[i]->retryTime = *retryTime;
4695 if (list[i]->header.serial) {
4696 /* Exponentially backoff retry times */
4697 if (list[i]->backoff < MAXBACKOFF) {
4698 /* so it can't stay == 0 */
4699 list[i]->backoff = (list[i]->backoff << 1) + 1;
4702 clock_Addmsec(&(list[i]->retryTime),
4703 ((afs_uint32) list[i]->backoff) << 8);
4706 /* Wait a little extra for the ack on the last packet */
4707 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4708 clock_Addmsec(&(list[i]->retryTime), 400);
4711 /* Record the time sent */
4712 list[i]->timeSent = *now;
4714 /* Ask for an ack on retransmitted packets, on every other packet
4715 * if the peer doesn't support slow start. Ask for an ack on every
4716 * packet until the congestion window reaches the ack rate. */
4717 if (list[i]->header.serial) {
4719 MUTEX_ENTER(&rx_stats_mutex);
4720 rx_stats.dataPacketsReSent++;
4721 MUTEX_EXIT(&rx_stats_mutex);
4723 /* improved RTO calculation- not Karn */
4724 list[i]->firstSent = *now;
4725 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4726 || (!(call->flags & RX_CALL_SLOW_START_OK)
4727 && (list[i]->header.seq & 1)))) {
4732 MUTEX_ENTER(&peer->peer_lock);
4736 MUTEX_ENTER(&rx_stats_mutex);
4737 rx_stats.dataPacketsSent++;
4738 MUTEX_EXIT(&rx_stats_mutex);
4739 MUTEX_EXIT(&peer->peer_lock);
4741 /* Tag this packet as not being the last in this group,
4742 * for the receiver's benefit */
4743 if (i < len - 1 || moreFlag) {
4744 list[i]->header.flags |= RX_MORE_PACKETS;
4747 /* Install the new retransmit time for the packet, and
4748 * record the time sent */
4749 list[i]->timeSent = *now;
4753 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4756 /* Since we're about to send a data packet to the peer, it's
4757 * safe to nuke any scheduled end-of-packets ack */
4758 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4760 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4761 MUTEX_EXIT(&call->lock);
4763 rxi_SendPacketList(call, conn, list, len, istack);
4765 rxi_SendPacket(call, conn, list[0], istack);
4767 MUTEX_ENTER(&call->lock);
4768 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4770 /* Update last send time for this call (for keep-alive
4771 * processing), and for the connection (so that we can discover
4772 * idle connections) */
4773 conn->lastSendTime = call->lastSendTime = clock_Sec();
4776 /* When sending packets we need to follow these rules:
4777 * 1. Never send more than maxDgramPackets in a jumbogram.
4778 * 2. Never send a packet with more than two iovecs in a jumbogram.
4779 * 3. Never send a retransmitted packet in a jumbogram.
4780 * 4. Never send more than cwind/4 packets in a jumbogram
4781 * We always keep the last list we should have sent so we
4782 * can set the RX_MORE_PACKETS flags correctly.
4785 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4786 int istack, struct clock *now, struct clock *retryTime,
4789 int i, cnt, lastCnt = 0;
4790 struct rx_packet **listP, **lastP = 0;
4791 struct rx_peer *peer = call->conn->peer;
4792 int morePackets = 0;
4794 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4795 /* Does the current packet force us to flush the current list? */
4797 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4798 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4800 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
4802 /* If the call enters an error state stop sending, or if
4803 * we entered congestion recovery mode, stop sending */
4804 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4812 /* Add the current packet to the list if it hasn't been acked.
4813 * Otherwise adjust the list pointer to skip the current packet. */
4814 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
4816 /* Do we need to flush the list? */
4817 if (cnt >= (int)peer->maxDgramPackets
4818 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
4819 || list[i]->header.serial
4820 || list[i]->length != RX_JUMBOBUFFERSIZE) {
4822 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
4823 retryTime, resending);
4824 /* If the call enters an error state stop sending, or if
4825 * we entered congestion recovery mode, stop sending */
4827 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4832 listP = &list[i + 1];
4837 osi_Panic("rxi_SendList error");
4839 listP = &list[i + 1];
4843 /* Send the whole list when the call is in receive mode, when
4844 * the call is in eof mode, when we are in fast recovery mode,
4845 * and when we have the last packet */
4846 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
4847 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
4848 || (call->flags & RX_CALL_FAST_RECOVER)) {
4849 /* Check for the case where the current list contains
4850 * an acked packet. Since we always send retransmissions
4851 * in a separate packet, we only need to check the first
4852 * packet in the list */
4853 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
4857 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
4858 retryTime, resending);
4859 /* If the call enters an error state stop sending, or if
4860 * we entered congestion recovery mode, stop sending */
4861 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4865 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
4868 } else if (lastCnt > 0) {
4869 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
4874 #ifdef RX_ENABLE_LOCKS
4875 /* Call rxi_Start, below, but with the call lock held. */
4877 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
4880 MUTEX_ENTER(&call->lock);
4881 rxi_Start(event, call, istack);
4882 MUTEX_EXIT(&call->lock);
4884 #endif /* RX_ENABLE_LOCKS */
4886 /* This routine is called when new packets are readied for
4887 * transmission and when retransmission may be necessary, or when the
4888 * transmission window or burst count are favourable. This should be
4889 * better optimized for new packets, the usual case, now that we've
4890 * got rid of queues of send packets. XXXXXXXXXXX */
4892 rxi_Start(struct rxevent *event, register struct rx_call *call, int istack)
4894 struct rx_packet *p;
4895 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
4896 struct rx_peer *peer = call->conn->peer;
4897 struct clock now, retryTime;
4901 struct rx_packet **xmitList;
4904 /* If rxi_Start is being called as a result of a resend event,
4905 * then make sure that the event pointer is removed from the call
4906 * structure, since there is no longer a per-call retransmission
4908 if (event && event == call->resendEvent) {
4909 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
4910 call->resendEvent = NULL;
4912 if (queue_IsEmpty(&call->tq)) {
4916 /* Timeouts trigger congestion recovery */
4917 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4918 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4919 /* someone else is waiting to start recovery */
4922 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4923 while (call->flags & RX_CALL_TQ_BUSY) {
4924 call->flags |= RX_CALL_TQ_WAIT;
4925 #ifdef RX_ENABLE_LOCKS
4926 CV_WAIT(&call->cv_tq, &call->lock);
4927 #else /* RX_ENABLE_LOCKS */
4928 osi_rxSleep(&call->tq);
4929 #endif /* RX_ENABLE_LOCKS */
4931 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4932 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4933 call->flags |= RX_CALL_FAST_RECOVER;
4934 if (peer->maxDgramPackets > 1) {
4935 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4937 call->MTU = MIN(peer->natMTU, peer->maxMTU);
4939 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4940 call->nDgramPackets = 1;
4942 call->nextCwind = 1;
4945 MUTEX_ENTER(&peer->peer_lock);
4946 peer->MTU = call->MTU;
4947 peer->cwind = call->cwind;
4948 peer->nDgramPackets = 1;
4950 call->congestSeq = peer->congestSeq;
4951 MUTEX_EXIT(&peer->peer_lock);
4952 /* Clear retry times on packets. Otherwise, it's possible for
4953 * some packets in the queue to force resends at rates faster
4954 * than recovery rates.
4956 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
4957 if (!(p->flags & RX_PKTFLAG_ACKED)) {
4958 clock_Zero(&p->retryTime);
4963 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4964 MUTEX_ENTER(&rx_stats_mutex);
4965 rx_tq_debug.rxi_start_in_error++;
4966 MUTEX_EXIT(&rx_stats_mutex);
4971 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
4972 /* Get clock to compute the re-transmit time for any packets
4973 * in this burst. Note, if we back off, it's reasonable to
4974 * back off all of the packets in the same manner, even if
4975 * some of them have been retransmitted more times than more
4976 * recent additions */
4977 clock_GetTime(&now);
4978 retryTime = now; /* initialize before use */
4979 MUTEX_ENTER(&peer->peer_lock);
4980 clock_Add(&retryTime, &peer->timeout);
4981 MUTEX_EXIT(&peer->peer_lock);
4983 /* Send (or resend) any packets that need it, subject to
4984 * window restrictions and congestion burst control
4985 * restrictions. Ask for an ack on the last packet sent in
4986 * this burst. For now, we're relying upon the window being
4987 * considerably bigger than the largest number of packets that
4988 * are typically sent at once by one initial call to
4989 * rxi_Start. This is probably bogus (perhaps we should ask
4990 * for an ack when we're half way through the current
4991 * window?). Also, for non file transfer applications, this
4992 * may end up asking for an ack for every packet. Bogus. XXXX
4995 * But check whether we're here recursively, and let the other guy
4998 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4999 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5000 call->flags |= RX_CALL_TQ_BUSY;
5002 call->flags &= ~RX_CALL_NEED_START;
5003 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5005 maxXmitPackets = MIN(call->twind, call->cwind);
5006 xmitList = (struct rx_packet **)
5007 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5008 if (xmitList == NULL)
5009 osi_Panic("rxi_Start, failed to allocate xmit list");
5010 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5011 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5012 /* We shouldn't be sending packets if a thread is waiting
5013 * to initiate congestion recovery */
5017 && (call->flags & RX_CALL_FAST_RECOVER)) {
5018 /* Only send one packet during fast recovery */
5021 if ((p->flags & RX_PKTFLAG_FREE)
5022 || (!queue_IsEnd(&call->tq, nxp)
5023 && (nxp->flags & RX_PKTFLAG_FREE))
5024 || (p == (struct rx_packet *)&rx_freePacketQueue)
5025 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5026 osi_Panic("rxi_Start: xmit queue clobbered");
5028 if (p->flags & RX_PKTFLAG_ACKED) {
5029 MUTEX_ENTER(&rx_stats_mutex);
5030 rx_stats.ignoreAckedPacket++;
5031 MUTEX_EXIT(&rx_stats_mutex);
5032 continue; /* Ignore this packet if it has been acknowledged */
5035 /* Turn off all flags except these ones, which are the same
5036 * on each transmission */
5037 p->header.flags &= RX_PRESET_FLAGS;
5039 if (p->header.seq >=
5040 call->tfirst + MIN((int)call->twind,
5041 (int)(call->nSoftAcked +
5043 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5044 /* Note: if we're waiting for more window space, we can
5045 * still send retransmits; hence we don't return here, but
5046 * break out to schedule a retransmit event */
5047 dpf(("call %d waiting for window",
5048 *(call->callNumber)));
5052 /* Transmit the packet if it needs to be sent. */
5053 if (!clock_Lt(&now, &p->retryTime)) {
5054 if (nXmitPackets == maxXmitPackets) {
5055 osi_Panic("rxi_Start: xmit list overflowed");
5057 xmitList[nXmitPackets++] = p;
5061 /* xmitList now hold pointers to all of the packets that are
5062 * ready to send. Now we loop to send the packets */
5063 if (nXmitPackets > 0) {
5064 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5065 &now, &retryTime, resending);
5068 maxXmitPackets * sizeof(struct rx_packet *));
5070 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5072 * TQ references no longer protected by this flag; they must remain
5073 * protected by the global lock.
5075 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5076 call->flags &= ~RX_CALL_TQ_BUSY;
5077 if (call->flags & RX_CALL_TQ_WAIT) {
5078 call->flags &= ~RX_CALL_TQ_WAIT;
5079 #ifdef RX_ENABLE_LOCKS
5080 CV_BROADCAST(&call->cv_tq);
5081 #else /* RX_ENABLE_LOCKS */
5082 osi_rxWakeup(&call->tq);
5083 #endif /* RX_ENABLE_LOCKS */
5088 /* We went into the error state while sending packets. Now is
5089 * the time to reset the call. This will also inform the using
5090 * process that the call is in an error state.
5092 MUTEX_ENTER(&rx_stats_mutex);
5093 rx_tq_debug.rxi_start_aborted++;
5094 MUTEX_EXIT(&rx_stats_mutex);
5095 call->flags &= ~RX_CALL_TQ_BUSY;
5096 if (call->flags & RX_CALL_TQ_WAIT) {
5097 call->flags &= ~RX_CALL_TQ_WAIT;
5098 #ifdef RX_ENABLE_LOCKS
5099 CV_BROADCAST(&call->cv_tq);
5100 #else /* RX_ENABLE_LOCKS */
5101 osi_rxWakeup(&call->tq);
5102 #endif /* RX_ENABLE_LOCKS */
5104 rxi_CallError(call, call->error);
5107 #ifdef RX_ENABLE_LOCKS
5108 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5109 register int missing;
5110 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5111 /* Some packets have received acks. If they all have, we can clear
5112 * the transmit queue.
5115 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5116 if (p->header.seq < call->tfirst
5117 && (p->flags & RX_PKTFLAG_ACKED)) {
5124 call->flags |= RX_CALL_TQ_CLEARME;
5126 #endif /* RX_ENABLE_LOCKS */
5127 /* Don't bother doing retransmits if the TQ is cleared. */
5128 if (call->flags & RX_CALL_TQ_CLEARME) {
5129 rxi_ClearTransmitQueue(call, 1);
5131 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5134 /* Always post a resend event, if there is anything in the
5135 * queue, and resend is possible. There should be at least
5136 * one unacknowledged packet in the queue ... otherwise none
5137 * of these packets should be on the queue in the first place.
5139 if (call->resendEvent) {
5140 /* Cancel the existing event and post a new one */
5141 rxevent_Cancel(call->resendEvent, call,
5142 RX_CALL_REFCOUNT_RESEND);
5145 /* The retry time is the retry time on the first unacknowledged
5146 * packet inside the current window */
5148 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5149 /* Don't set timers for packets outside the window */
5150 if (p->header.seq >= call->tfirst + call->twind) {
5154 if (!(p->flags & RX_PKTFLAG_ACKED)
5155 && !clock_IsZero(&p->retryTime)) {
5157 retryTime = p->retryTime;
5162 /* Post a new event to re-run rxi_Start when retries may be needed */
5163 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5164 #ifdef RX_ENABLE_LOCKS
5165 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5167 rxevent_Post(&retryTime, rxi_StartUnlocked,
5168 (void *)call, (void *)istack);
5169 #else /* RX_ENABLE_LOCKS */
5171 rxevent_Post(&retryTime, rxi_Start, (void *)call,
5173 #endif /* RX_ENABLE_LOCKS */
5176 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5177 } while (call->flags & RX_CALL_NEED_START);
5179 * TQ references no longer protected by this flag; they must remain
5180 * protected by the global lock.
5182 call->flags &= ~RX_CALL_TQ_BUSY;
5183 if (call->flags & RX_CALL_TQ_WAIT) {
5184 call->flags &= ~RX_CALL_TQ_WAIT;
5185 #ifdef RX_ENABLE_LOCKS
5186 CV_BROADCAST(&call->cv_tq);
5187 #else /* RX_ENABLE_LOCKS */
5188 osi_rxWakeup(&call->tq);
5189 #endif /* RX_ENABLE_LOCKS */
5192 call->flags |= RX_CALL_NEED_START;
5194 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5196 if (call->resendEvent) {
5197 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5202 /* Also adjusts the keep alive parameters for the call, to reflect
5203 * that we have just sent a packet (so keep alives aren't sent
5206 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5209 register struct rx_connection *conn = call->conn;
5211 /* Stamp each packet with the user supplied status */
5212 p->header.userStatus = call->localStatus;
5214 /* Allow the security object controlling this call's security to
5215 * make any last-minute changes to the packet */
5216 RXS_SendPacket(conn->securityObject, call, p);
5218 /* Since we're about to send SOME sort of packet to the peer, it's
5219 * safe to nuke any scheduled end-of-packets ack */
5220 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5222 /* Actually send the packet, filling in more connection-specific fields */
5223 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5224 MUTEX_EXIT(&call->lock);
5225 rxi_SendPacket(call, conn, p, istack);
5226 MUTEX_ENTER(&call->lock);
5227 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5229 /* Update last send time for this call (for keep-alive
5230 * processing), and for the connection (so that we can discover
5231 * idle connections) */
5232 conn->lastSendTime = call->lastSendTime = clock_Sec();
5236 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5237 * that things are fine. Also called periodically to guarantee that nothing
5238 * falls through the cracks (e.g. (error + dally) connections have keepalive
5239 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5241 * haveCTLock Set if calling from rxi_ReapConnections
5243 #ifdef RX_ENABLE_LOCKS
5245 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5246 #else /* RX_ENABLE_LOCKS */
5248 rxi_CheckCall(register struct rx_call *call)
5249 #endif /* RX_ENABLE_LOCKS */
5251 register struct rx_connection *conn = call->conn;
5253 afs_uint32 deadTime;
5255 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5256 if (call->flags & RX_CALL_TQ_BUSY) {
5257 /* Call is active and will be reset by rxi_Start if it's
5258 * in an error state.
5263 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5265 (((afs_uint32) conn->secondsUntilDead << 10) +
5266 ((afs_uint32) conn->peer->rtt >> 3) +
5267 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5269 /* These are computed to the second (+- 1 second). But that's
5270 * good enough for these values, which should be a significant
5271 * number of seconds. */
5272 if (now > (call->lastReceiveTime + deadTime)) {
5273 if (call->state == RX_STATE_ACTIVE) {
5274 rxi_CallError(call, RX_CALL_DEAD);
5277 #ifdef RX_ENABLE_LOCKS
5278 /* Cancel pending events */
5279 rxevent_Cancel(call->delayedAckEvent, call,
5280 RX_CALL_REFCOUNT_DELAY);
5281 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5282 rxevent_Cancel(call->keepAliveEvent, call,
5283 RX_CALL_REFCOUNT_ALIVE);
5284 if (call->refCount == 0) {
5285 rxi_FreeCall(call, haveCTLock);
5289 #else /* RX_ENABLE_LOCKS */
5292 #endif /* RX_ENABLE_LOCKS */
5294 /* Non-active calls are destroyed if they are not responding
5295 * to pings; active calls are simply flagged in error, so the
5296 * attached process can die reasonably gracefully. */
5298 /* see if we have a non-activity timeout */
5299 if (call->startWait && conn->idleDeadTime
5300 && ((call->startWait + conn->idleDeadTime) < now)) {
5301 if (call->state == RX_STATE_ACTIVE) {
5302 rxi_CallError(call, RX_CALL_TIMEOUT);
5306 /* see if we have a hard timeout */
5307 if (conn->hardDeadTime
5308 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5309 if (call->state == RX_STATE_ACTIVE)
5310 rxi_CallError(call, RX_CALL_TIMEOUT);
5317 /* When a call is in progress, this routine is called occasionally to
5318 * make sure that some traffic has arrived (or been sent to) the peer.
5319 * If nothing has arrived in a reasonable amount of time, the call is
5320 * declared dead; if nothing has been sent for a while, we send a
5321 * keep-alive packet (if we're actually trying to keep the call alive)
5324 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5327 struct rx_connection *conn;
5330 MUTEX_ENTER(&call->lock);
5331 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5332 if (event == call->keepAliveEvent)
5333 call->keepAliveEvent = NULL;
5336 #ifdef RX_ENABLE_LOCKS
5337 if (rxi_CheckCall(call, 0)) {
5338 MUTEX_EXIT(&call->lock);
5341 #else /* RX_ENABLE_LOCKS */
5342 if (rxi_CheckCall(call))
5344 #endif /* RX_ENABLE_LOCKS */
5346 /* Don't try to keep alive dallying calls */
5347 if (call->state == RX_STATE_DALLY) {
5348 MUTEX_EXIT(&call->lock);
5353 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5354 /* Don't try to send keepalives if there is unacknowledged data */
5355 /* the rexmit code should be good enough, this little hack
5356 * doesn't quite work XXX */
5357 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5359 rxi_ScheduleKeepAliveEvent(call);
5360 MUTEX_EXIT(&call->lock);
5365 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5367 if (!call->keepAliveEvent) {
5369 clock_GetTime(&when);
5370 when.sec += call->conn->secondsUntilPing;
5371 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5372 call->keepAliveEvent =
5373 rxevent_Post(&when, rxi_KeepAliveEvent, call, 0);
5377 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5379 rxi_KeepAliveOn(register struct rx_call *call)
5381 /* Pretend last packet received was received now--i.e. if another
5382 * packet isn't received within the keep alive time, then the call
5383 * will die; Initialize last send time to the current time--even
5384 * if a packet hasn't been sent yet. This will guarantee that a
5385 * keep-alive is sent within the ping time */
5386 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5387 rxi_ScheduleKeepAliveEvent(call);
5390 /* This routine is called to send connection abort messages
5391 * that have been delayed to throttle looping clients. */
5393 rxi_SendDelayedConnAbort(struct rxevent *event,
5394 register struct rx_connection *conn, char *dummy)
5397 struct rx_packet *packet;
5399 MUTEX_ENTER(&conn->conn_data_lock);
5400 conn->delayedAbortEvent = NULL;
5401 error = htonl(conn->error);
5403 MUTEX_EXIT(&conn->conn_data_lock);
5404 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5407 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5408 RX_PACKET_TYPE_ABORT, (char *)&error,
5410 rxi_FreePacket(packet);
5414 /* This routine is called to send call abort messages
5415 * that have been delayed to throttle looping clients. */
5417 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5421 struct rx_packet *packet;
5423 MUTEX_ENTER(&call->lock);
5424 call->delayedAbortEvent = NULL;
5425 error = htonl(call->error);
5427 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5430 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5431 (char *)&error, sizeof(error), 0);
5432 rxi_FreePacket(packet);
5434 MUTEX_EXIT(&call->lock);
5437 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5438 * seconds) to ask the client to authenticate itself. The routine
5439 * issues a challenge to the client, which is obtained from the
5440 * security object associated with the connection */
5442 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5445 int tries = (int)atries;
5446 conn->challengeEvent = NULL;
5447 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5448 register struct rx_packet *packet;
5452 /* We've failed to authenticate for too long.
5453 * Reset any calls waiting for authentication;
5454 * they are all in RX_STATE_PRECALL.
5458 MUTEX_ENTER(&conn->conn_call_lock);
5459 for (i = 0; i < RX_MAXCALLS; i++) {
5460 struct rx_call *call = conn->call[i];
5462 MUTEX_ENTER(&call->lock);
5463 if (call->state == RX_STATE_PRECALL) {
5464 rxi_CallError(call, RX_CALL_DEAD);
5465 rxi_SendCallAbort(call, NULL, 0, 0);
5467 MUTEX_EXIT(&call->lock);
5470 MUTEX_EXIT(&conn->conn_call_lock);
5474 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5476 /* If there's no packet available, do this later. */
5477 RXS_GetChallenge(conn->securityObject, conn, packet);
5478 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5479 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5480 rxi_FreePacket(packet);
5482 clock_GetTime(&when);
5483 when.sec += RX_CHALLENGE_TIMEOUT;
5484 conn->challengeEvent =
5485 rxevent_Post(&when, rxi_ChallengeEvent, conn,
5486 (void *)(tries - 1));
5490 /* Call this routine to start requesting the client to authenticate
5491 * itself. This will continue until authentication is established,
5492 * the call times out, or an invalid response is returned. The
5493 * security object associated with the connection is asked to create
5494 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5495 * defined earlier. */
5497 rxi_ChallengeOn(register struct rx_connection *conn)
5499 if (!conn->challengeEvent) {
5500 RXS_CreateChallenge(conn->securityObject, conn);
5501 rxi_ChallengeEvent(NULL, conn, (void *)RX_CHALLENGE_MAXTRIES);
5506 /* Compute round trip time of the packet provided, in *rttp.
5509 /* rxi_ComputeRoundTripTime is called with peer locked. */
5510 /* sentp and/or peer may be null */
5512 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5513 register struct clock *sentp,
5514 register struct rx_peer *peer)
5516 struct clock thisRtt, *rttp = &thisRtt;
5518 #if defined(AFS_ALPHA_LINUX22_ENV) && defined(AFS_PTHREAD_ENV) && !defined(KERNEL)
5519 /* making year 2038 bugs to get this running now - stroucki */
5520 struct timeval temptime;
5522 register int rtt_timeout;
5524 #if defined(AFS_ALPHA_LINUX20_ENV) && defined(AFS_PTHREAD_ENV) && !defined(KERNEL)
5525 /* yet again. This was the worst Heisenbug of the port - stroucki */
5526 clock_GetTime(&temptime);
5527 rttp->sec = (afs_int32) temptime.tv_sec;
5528 rttp->usec = (afs_int32) temptime.tv_usec;
5530 clock_GetTime(rttp);
5532 if (clock_Lt(rttp, sentp)) {
5534 return; /* somebody set the clock back, don't count this time. */
5536 clock_Sub(rttp, sentp);
5537 MUTEX_ENTER(&rx_stats_mutex);
5538 if (clock_Lt(rttp, &rx_stats.minRtt))
5539 rx_stats.minRtt = *rttp;
5540 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5541 if (rttp->sec > 60) {
5542 MUTEX_EXIT(&rx_stats_mutex);
5543 return; /* somebody set the clock ahead */
5545 rx_stats.maxRtt = *rttp;
5547 clock_Add(&rx_stats.totalRtt, rttp);
5548 rx_stats.nRttSamples++;
5549 MUTEX_EXIT(&rx_stats_mutex);
5551 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5553 /* Apply VanJacobson round-trip estimations */
5558 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5559 * srtt is stored as fixed point with 3 bits after the binary
5560 * point (i.e., scaled by 8). The following magic is
5561 * equivalent to the smoothing algorithm in rfc793 with an
5562 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5563 * srtt*8 = srtt*8 + rtt - srtt
5564 * srtt = srtt + rtt/8 - srtt/8
5567 delta = MSEC(rttp) - (peer->rtt >> 3);
5571 * We accumulate a smoothed rtt variance (actually, a smoothed
5572 * mean difference), then set the retransmit timer to smoothed
5573 * rtt + 4 times the smoothed variance (was 2x in van's original
5574 * paper, but 4x works better for me, and apparently for him as
5576 * rttvar is stored as
5577 * fixed point with 2 bits after the binary point (scaled by
5578 * 4). The following is equivalent to rfc793 smoothing with
5579 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5580 * replaces rfc793's wired-in beta.
5581 * dev*4 = dev*4 + (|actual - expected| - dev)
5587 delta -= (peer->rtt_dev >> 2);
5588 peer->rtt_dev += delta;
5590 /* I don't have a stored RTT so I start with this value. Since I'm
5591 * probably just starting a call, and will be pushing more data down
5592 * this, I expect congestion to increase rapidly. So I fudge a
5593 * little, and I set deviance to half the rtt. In practice,
5594 * deviance tends to approach something a little less than
5595 * half the smoothed rtt. */
5596 peer->rtt = (MSEC(rttp) << 3) + 8;
5597 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5599 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5600 * the other of these connections is usually in a user process, and can
5601 * be switched and/or swapped out. So on fast, reliable networks, the
5602 * timeout would otherwise be too short.
5604 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5605 clock_Zero(&(peer->timeout));
5606 clock_Addmsec(&(peer->timeout), rtt_timeout);
5608 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)));
5612 /* Find all server connections that have not been active for a long time, and
5615 rxi_ReapConnections(void)
5618 clock_GetTime(&now);
5620 /* Find server connection structures that haven't been used for
5621 * greater than rx_idleConnectionTime */
5623 struct rx_connection **conn_ptr, **conn_end;
5624 int i, havecalls = 0;
5625 MUTEX_ENTER(&rx_connHashTable_lock);
5626 for (conn_ptr = &rx_connHashTable[0], conn_end =
5627 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5629 struct rx_connection *conn, *next;
5630 struct rx_call *call;
5634 for (conn = *conn_ptr; conn; conn = next) {
5635 /* XXX -- Shouldn't the connection be locked? */
5638 for (i = 0; i < RX_MAXCALLS; i++) {
5639 call = conn->call[i];
5642 MUTEX_ENTER(&call->lock);
5643 #ifdef RX_ENABLE_LOCKS
5644 result = rxi_CheckCall(call, 1);
5645 #else /* RX_ENABLE_LOCKS */
5646 result = rxi_CheckCall(call);
5647 #endif /* RX_ENABLE_LOCKS */
5648 MUTEX_EXIT(&call->lock);
5650 /* If CheckCall freed the call, it might
5651 * have destroyed the connection as well,
5652 * which screws up the linked lists.
5658 if (conn->type == RX_SERVER_CONNECTION) {
5659 /* This only actually destroys the connection if
5660 * there are no outstanding calls */
5661 MUTEX_ENTER(&conn->conn_data_lock);
5662 if (!havecalls && !conn->refCount
5663 && ((conn->lastSendTime + rx_idleConnectionTime) <
5665 conn->refCount++; /* it will be decr in rx_DestroyConn */
5666 MUTEX_EXIT(&conn->conn_data_lock);
5667 #ifdef RX_ENABLE_LOCKS
5668 rxi_DestroyConnectionNoLock(conn);
5669 #else /* RX_ENABLE_LOCKS */
5670 rxi_DestroyConnection(conn);
5671 #endif /* RX_ENABLE_LOCKS */
5673 #ifdef RX_ENABLE_LOCKS
5675 MUTEX_EXIT(&conn->conn_data_lock);
5677 #endif /* RX_ENABLE_LOCKS */
5681 #ifdef RX_ENABLE_LOCKS
5682 while (rx_connCleanup_list) {
5683 struct rx_connection *conn;
5684 conn = rx_connCleanup_list;
5685 rx_connCleanup_list = rx_connCleanup_list->next;
5686 MUTEX_EXIT(&rx_connHashTable_lock);
5687 rxi_CleanupConnection(conn);
5688 MUTEX_ENTER(&rx_connHashTable_lock);
5690 MUTEX_EXIT(&rx_connHashTable_lock);
5691 #endif /* RX_ENABLE_LOCKS */
5694 /* Find any peer structures that haven't been used (haven't had an
5695 * associated connection) for greater than rx_idlePeerTime */
5697 struct rx_peer **peer_ptr, **peer_end;
5699 MUTEX_ENTER(&rx_rpc_stats);
5700 MUTEX_ENTER(&rx_peerHashTable_lock);
5701 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5702 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5704 struct rx_peer *peer, *next, *prev;
5705 for (prev = peer = *peer_ptr; peer; peer = next) {
5707 code = MUTEX_TRYENTER(&peer->peer_lock);
5708 if ((code) && (peer->refCount == 0)
5709 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5710 rx_interface_stat_p rpc_stat, nrpc_stat;
5712 MUTEX_EXIT(&peer->peer_lock);
5713 MUTEX_DESTROY(&peer->peer_lock);
5715 (&peer->rpcStats, rpc_stat, nrpc_stat,
5716 rx_interface_stat)) {
5717 unsigned int num_funcs;
5720 queue_Remove(&rpc_stat->queue_header);
5721 queue_Remove(&rpc_stat->all_peers);
5722 num_funcs = rpc_stat->stats[0].func_total;
5724 sizeof(rx_interface_stat_t) +
5725 rpc_stat->stats[0].func_total *
5726 sizeof(rx_function_entry_v1_t);
5728 rxi_Free(rpc_stat, space);
5729 rxi_rpc_peer_stat_cnt -= num_funcs;
5732 MUTEX_ENTER(&rx_stats_mutex);
5733 rx_stats.nPeerStructs--;
5734 MUTEX_EXIT(&rx_stats_mutex);
5735 if (prev == *peer_ptr) {
5742 MUTEX_EXIT(&peer->peer_lock);
5748 MUTEX_EXIT(&rx_peerHashTable_lock);
5749 MUTEX_EXIT(&rx_rpc_stats);
5752 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5753 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5754 * GC, just below. Really, we shouldn't have to keep moving packets from
5755 * one place to another, but instead ought to always know if we can
5756 * afford to hold onto a packet in its particular use. */
5757 MUTEX_ENTER(&rx_freePktQ_lock);
5758 if (rx_waitingForPackets) {
5759 rx_waitingForPackets = 0;
5760 #ifdef RX_ENABLE_LOCKS
5761 CV_BROADCAST(&rx_waitingForPackets_cv);
5763 osi_rxWakeup(&rx_waitingForPackets);
5766 MUTEX_EXIT(&rx_freePktQ_lock);
5768 now.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5769 rxevent_Post(&now, rxi_ReapConnections, 0, 0);
5773 /* rxs_Release - This isn't strictly necessary but, since the macro name from
5774 * rx.h is sort of strange this is better. This is called with a security
5775 * object before it is discarded. Each connection using a security object has
5776 * its own refcount to the object so it won't actually be freed until the last
5777 * connection is destroyed.
5779 * This is the only rxs module call. A hold could also be written but no one
5783 rxs_Release(struct rx_securityClass *aobj)
5785 return RXS_Close(aobj);
5789 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
5790 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
5791 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
5792 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
5794 /* Adjust our estimate of the transmission rate to this peer, given
5795 * that the packet p was just acked. We can adjust peer->timeout and
5796 * call->twind. Pragmatically, this is called
5797 * only with packets of maximal length.
5798 * Called with peer and call locked.
5802 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
5803 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
5805 afs_int32 xferSize, xferMs;
5806 register afs_int32 minTime;
5809 /* Count down packets */
5810 if (peer->rateFlag > 0)
5812 /* Do nothing until we're enabled */
5813 if (peer->rateFlag != 0)
5818 /* Count only when the ack seems legitimate */
5819 switch (ackReason) {
5820 case RX_ACK_REQUESTED:
5822 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
5826 case RX_ACK_PING_RESPONSE:
5827 if (p) /* want the response to ping-request, not data send */
5829 clock_GetTime(&newTO);
5830 if (clock_Gt(&newTO, &call->pingRequestTime)) {
5831 clock_Sub(&newTO, &call->pingRequestTime);
5832 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
5836 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
5843 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));
5845 /* Track only packets that are big enough. */
5846 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
5850 /* absorb RTT data (in milliseconds) for these big packets */
5851 if (peer->smRtt == 0) {
5852 peer->smRtt = xferMs;
5854 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
5859 if (peer->countDown) {
5863 peer->countDown = 10; /* recalculate only every so often */
5865 /* In practice, we can measure only the RTT for full packets,
5866 * because of the way Rx acks the data that it receives. (If it's
5867 * smaller than a full packet, it often gets implicitly acked
5868 * either by the call response (from a server) or by the next call
5869 * (from a client), and either case confuses transmission times
5870 * with processing times.) Therefore, replace the above
5871 * more-sophisticated processing with a simpler version, where the
5872 * smoothed RTT is kept for full-size packets, and the time to
5873 * transmit a windowful of full-size packets is simply RTT *
5874 * windowSize. Again, we take two steps:
5875 - ensure the timeout is large enough for a single packet's RTT;
5876 - ensure that the window is small enough to fit in the desired timeout.*/
5878 /* First, the timeout check. */
5879 minTime = peer->smRtt;
5880 /* Get a reasonable estimate for a timeout period */
5882 newTO.sec = minTime / 1000;
5883 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
5885 /* Increase the timeout period so that we can always do at least
5886 * one packet exchange */
5887 if (clock_Gt(&newTO, &peer->timeout)) {
5889 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));
5891 peer->timeout = newTO;
5894 /* Now, get an estimate for the transmit window size. */
5895 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
5896 /* Now, convert to the number of full packets that could fit in a
5897 * reasonable fraction of that interval */
5898 minTime /= (peer->smRtt << 1);
5899 xferSize = minTime; /* (make a copy) */
5901 /* Now clamp the size to reasonable bounds. */
5904 else if (minTime > rx_Window)
5905 minTime = rx_Window;
5906 /* if (minTime != peer->maxWindow) {
5907 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
5908 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
5909 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
5911 peer->maxWindow = minTime;
5912 elide... call->twind = minTime;
5916 /* Cut back on the peer timeout if it had earlier grown unreasonably.
5917 * Discern this by calculating the timeout necessary for rx_Window
5919 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
5920 /* calculate estimate for transmission interval in milliseconds */
5921 minTime = rx_Window * peer->smRtt;
5922 if (minTime < 1000) {
5923 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
5924 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
5925 peer->timeout.usec, peer->smRtt, peer->packetSize));
5927 newTO.sec = 0; /* cut back on timeout by half a second */
5928 newTO.usec = 500000;
5929 clock_Sub(&peer->timeout, &newTO);
5934 } /* end of rxi_ComputeRate */
5935 #endif /* ADAPT_WINDOW */
5943 /* Don't call this debugging routine directly; use dpf */
5945 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
5946 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
5950 clock_GetTime(&now);
5951 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
5952 (unsigned int)now.usec / 1000);
5953 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
5961 * This function is used to process the rx_stats structure that is local
5962 * to a process as well as an rx_stats structure received from a remote
5963 * process (via rxdebug). Therefore, it needs to do minimal version
5967 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
5968 afs_int32 freePackets, char version)
5972 if (size != sizeof(struct rx_stats)) {
5974 "Unexpected size of stats structure: was %d, expected %d\n",
5975 size, sizeof(struct rx_stats));
5978 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
5981 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
5982 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
5983 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
5984 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
5985 s->specialPktAllocFailures);
5987 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
5988 s->receivePktAllocFailures, s->sendPktAllocFailures,
5989 s->specialPktAllocFailures);
5993 " greedy %d, " "bogusReads %d (last from host %x), "
5994 "noPackets %d, " "noBuffers %d, " "selects %d, "
5995 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
5996 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
5997 s->selects, s->sendSelects);
5999 fprintf(file, " packets read: ");
6000 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6001 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6003 fprintf(file, "\n");
6006 " other read counters: data %d, " "ack %d, " "dup %d "
6007 "spurious %d " "dally %d\n", s->dataPacketsRead,
6008 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6009 s->ignorePacketDally);
6011 fprintf(file, " packets sent: ");
6012 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6013 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6015 fprintf(file, "\n");
6018 " other send counters: ack %d, " "data %d (not resends), "
6019 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6020 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6021 s->dataPacketsPushed, s->ignoreAckedPacket);
6024 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6025 s->netSendFailures, (int)s->fatalErrors);
6027 if (s->nRttSamples) {
6028 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6029 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6031 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6032 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6036 " %d server connections, " "%d client connections, "
6037 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6038 s->nServerConns, s->nClientConns, s->nPeerStructs,
6039 s->nCallStructs, s->nFreeCallStructs);
6041 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6042 fprintf(file, " %d clock updates\n", clock_nUpdates);
6047 /* for backward compatibility */
6049 rx_PrintStats(FILE * file)
6051 MUTEX_ENTER(&rx_stats_mutex);
6052 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6054 MUTEX_EXIT(&rx_stats_mutex);
6058 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6060 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6061 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6062 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6065 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6066 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6067 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6070 " Packet size %d, " "max in packet skew %d, "
6071 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6072 (int)peer->outPacketSkew);
6075 #ifdef AFS_PTHREAD_ENV
6077 * This mutex protects the following static variables:
6081 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0);
6082 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0);
6084 #define LOCK_RX_DEBUG
6085 #define UNLOCK_RX_DEBUG
6086 #endif /* AFS_PTHREAD_ENV */
6089 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6090 u_char type, void *inputData, size_t inputLength,
6091 void *outputData, size_t outputLength)
6093 static afs_int32 counter = 100;
6095 struct rx_header theader;
6097 register afs_int32 code;
6099 struct sockaddr_in taddr, faddr;
6104 endTime = time(0) + 20; /* try for 20 seconds */
6105 LOCK_RX_DEBUG counter++;
6106 UNLOCK_RX_DEBUG tp = &tbuffer[sizeof(struct rx_header)];
6107 taddr.sin_family = AF_INET;
6108 taddr.sin_port = remotePort;
6109 taddr.sin_addr.s_addr = remoteAddr;
6110 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6111 taddr.sin_len = sizeof(struct sockaddr_in);
6114 memset(&theader, 0, sizeof(theader));
6115 theader.epoch = htonl(999);
6117 theader.callNumber = htonl(counter);
6120 theader.type = type;
6121 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6122 theader.serviceId = 0;
6124 memcpy(tbuffer, &theader, sizeof(theader));
6125 memcpy(tp, inputData, inputLength);
6127 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6128 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6130 /* see if there's a packet available */
6132 FD_SET(socket, &imask);
6135 code = select(socket + 1, &imask, 0, 0, &tv);
6136 if (code == 1 && FD_ISSET(socket,&imask)) {
6137 /* now receive a packet */
6138 faddrLen = sizeof(struct sockaddr_in);
6140 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6141 (struct sockaddr *)&faddr, &faddrLen);
6144 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6145 if (counter == ntohl(theader.callNumber))
6150 /* see if we've timed out */
6151 if (endTime < time(0))
6154 code -= sizeof(struct rx_header);
6155 if (code > outputLength)
6156 code = outputLength;
6157 memcpy(outputData, tp, code);
6162 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6163 afs_uint16 remotePort, struct rx_debugStats * stat,
6164 afs_uint32 * supportedValues)
6166 struct rx_debugIn in;
6169 *supportedValues = 0;
6170 in.type = htonl(RX_DEBUGI_GETSTATS);
6173 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6174 &in, sizeof(in), stat, sizeof(*stat));
6177 * If the call was successful, fixup the version and indicate
6178 * what contents of the stat structure are valid.
6179 * Also do net to host conversion of fields here.
6183 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6184 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6186 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6187 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6189 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6190 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6192 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6193 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6195 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6196 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6198 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6199 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6201 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6202 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6205 stat->nFreePackets = ntohl(stat->nFreePackets);
6206 stat->packetReclaims = ntohl(stat->packetReclaims);
6207 stat->callsExecuted = ntohl(stat->callsExecuted);
6208 stat->nWaiting = ntohl(stat->nWaiting);
6209 stat->idleThreads = ntohl(stat->idleThreads);
6216 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6217 afs_uint16 remotePort, struct rx_stats * stat,
6218 afs_uint32 * supportedValues)
6220 struct rx_debugIn in;
6221 afs_int32 *lp = (afs_int32 *) stat;
6226 * supportedValues is currently unused, but added to allow future
6227 * versioning of this function.
6230 *supportedValues = 0;
6231 in.type = htonl(RX_DEBUGI_RXSTATS);
6233 memset(stat, 0, sizeof(*stat));
6235 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6236 &in, sizeof(in), stat, sizeof(*stat));
6241 * Do net to host conversion here
6244 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6253 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6254 afs_uint16 remotePort, size_t version_length,
6258 return MakeDebugCall(socket, remoteAddr, remotePort,
6259 RX_PACKET_TYPE_VERSION, a, 1, version,
6264 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6265 afs_uint16 remotePort, afs_int32 * nextConnection,
6266 int allConnections, afs_uint32 debugSupportedValues,
6267 struct rx_debugConn * conn,
6268 afs_uint32 * supportedValues)
6270 struct rx_debugIn in;
6275 * supportedValues is currently unused, but added to allow future
6276 * versioning of this function.
6279 *supportedValues = 0;
6280 if (allConnections) {
6281 in.type = htonl(RX_DEBUGI_GETALLCONN);
6283 in.type = htonl(RX_DEBUGI_GETCONN);
6285 in.index = htonl(*nextConnection);
6286 memset(conn, 0, sizeof(*conn));
6288 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6289 &in, sizeof(in), conn, sizeof(*conn));
6292 *nextConnection += 1;
6295 * Convert old connection format to new structure.
6298 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6299 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6300 #define MOVEvL(a) (conn->a = vL->a)
6302 /* any old or unrecognized version... */
6303 for (i = 0; i < RX_MAXCALLS; i++) {
6304 MOVEvL(callState[i]);
6305 MOVEvL(callMode[i]);
6306 MOVEvL(callFlags[i]);
6307 MOVEvL(callOther[i]);
6309 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6310 MOVEvL(secStats.type);
6311 MOVEvL(secStats.level);
6312 MOVEvL(secStats.flags);
6313 MOVEvL(secStats.expires);
6314 MOVEvL(secStats.packetsReceived);
6315 MOVEvL(secStats.packetsSent);
6316 MOVEvL(secStats.bytesReceived);
6317 MOVEvL(secStats.bytesSent);
6322 * Do net to host conversion here
6324 * I don't convert host or port since we are most likely
6325 * going to want these in NBO.
6327 conn->cid = ntohl(conn->cid);
6328 conn->serial = ntohl(conn->serial);
6329 for (i = 0; i < RX_MAXCALLS; i++) {
6330 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6332 conn->error = ntohl(conn->error);
6333 conn->secStats.flags = ntohl(conn->secStats.flags);
6334 conn->secStats.expires = ntohl(conn->secStats.expires);
6335 conn->secStats.packetsReceived =
6336 ntohl(conn->secStats.packetsReceived);
6337 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6338 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6339 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6340 conn->epoch = ntohl(conn->epoch);
6341 conn->natMTU = ntohl(conn->natMTU);
6348 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6349 afs_uint16 remotePort, afs_int32 * nextPeer,
6350 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6351 afs_uint32 * supportedValues)
6353 struct rx_debugIn in;
6357 * supportedValues is currently unused, but added to allow future
6358 * versioning of this function.
6361 *supportedValues = 0;
6362 in.type = htonl(RX_DEBUGI_GETPEER);
6363 in.index = htonl(*nextPeer);
6364 memset(peer, 0, sizeof(*peer));
6366 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6367 &in, sizeof(in), peer, sizeof(*peer));
6373 * Do net to host conversion here
6375 * I don't convert host or port since we are most likely
6376 * going to want these in NBO.
6378 peer->ifMTU = ntohs(peer->ifMTU);
6379 peer->idleWhen = ntohl(peer->idleWhen);
6380 peer->refCount = ntohs(peer->refCount);
6381 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6382 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6383 peer->rtt = ntohl(peer->rtt);
6384 peer->rtt_dev = ntohl(peer->rtt_dev);
6385 peer->timeout.sec = ntohl(peer->timeout.sec);
6386 peer->timeout.usec = ntohl(peer->timeout.usec);
6387 peer->nSent = ntohl(peer->nSent);
6388 peer->reSends = ntohl(peer->reSends);
6389 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6390 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6391 peer->rateFlag = ntohl(peer->rateFlag);
6392 peer->natMTU = ntohs(peer->natMTU);
6393 peer->maxMTU = ntohs(peer->maxMTU);
6394 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6395 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6396 peer->MTU = ntohs(peer->MTU);
6397 peer->cwind = ntohs(peer->cwind);
6398 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6399 peer->congestSeq = ntohs(peer->congestSeq);
6400 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6401 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6402 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6403 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6408 #endif /* RXDEBUG */
6413 struct rx_serverQueueEntry *np;
6416 register struct rx_call *call;
6417 register struct rx_serverQueueEntry *sq;
6420 LOCK_RX_INIT if (rxinit_status == 1) {
6421 UNLOCK_RX_INIT return; /* Already shutdown. */
6425 #ifndef AFS_PTHREAD_ENV
6426 FD_ZERO(&rx_selectMask);
6427 #endif /* AFS_PTHREAD_ENV */
6428 rxi_dataQuota = RX_MAX_QUOTA;
6429 #ifndef AFS_PTHREAD_ENV
6431 #endif /* AFS_PTHREAD_ENV */
6434 #ifndef AFS_PTHREAD_ENV
6435 #ifndef AFS_USE_GETTIMEOFDAY
6437 #endif /* AFS_USE_GETTIMEOFDAY */
6438 #endif /* AFS_PTHREAD_ENV */
6440 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6441 call = queue_First(&rx_freeCallQueue, rx_call);
6443 rxi_Free(call, sizeof(struct rx_call));
6446 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6447 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6453 struct rx_peer **peer_ptr, **peer_end;
6454 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6455 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6457 struct rx_peer *peer, *next;
6458 for (peer = *peer_ptr; peer; peer = next) {
6459 rx_interface_stat_p rpc_stat, nrpc_stat;
6462 (&peer->rpcStats, rpc_stat, nrpc_stat,
6463 rx_interface_stat)) {
6464 unsigned int num_funcs;
6467 queue_Remove(&rpc_stat->queue_header);
6468 queue_Remove(&rpc_stat->all_peers);
6469 num_funcs = rpc_stat->stats[0].func_total;
6471 sizeof(rx_interface_stat_t) +
6472 rpc_stat->stats[0].func_total *
6473 sizeof(rx_function_entry_v1_t);
6475 rxi_Free(rpc_stat, space);
6476 MUTEX_ENTER(&rx_rpc_stats);
6477 rxi_rpc_peer_stat_cnt -= num_funcs;
6478 MUTEX_EXIT(&rx_rpc_stats);
6482 MUTEX_ENTER(&rx_stats_mutex);
6483 rx_stats.nPeerStructs--;
6484 MUTEX_EXIT(&rx_stats_mutex);
6488 for (i = 0; i < RX_MAX_SERVICES; i++) {
6490 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6492 for (i = 0; i < rx_hashTableSize; i++) {
6493 register struct rx_connection *tc, *ntc;
6494 MUTEX_ENTER(&rx_connHashTable_lock);
6495 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6497 for (j = 0; j < RX_MAXCALLS; j++) {
6499 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6502 rxi_Free(tc, sizeof(*tc));
6504 MUTEX_EXIT(&rx_connHashTable_lock);
6507 MUTEX_ENTER(&freeSQEList_lock);
6509 while ((np = rx_FreeSQEList)) {
6510 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6511 MUTEX_DESTROY(&np->lock);
6512 rxi_Free(np, sizeof(*np));
6515 MUTEX_EXIT(&freeSQEList_lock);
6516 MUTEX_DESTROY(&freeSQEList_lock);
6517 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6518 MUTEX_DESTROY(&rx_connHashTable_lock);
6519 MUTEX_DESTROY(&rx_peerHashTable_lock);
6520 MUTEX_DESTROY(&rx_serverPool_lock);
6522 osi_Free(rx_connHashTable,
6523 rx_hashTableSize * sizeof(struct rx_connection *));
6524 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6526 UNPIN(rx_connHashTable,
6527 rx_hashTableSize * sizeof(struct rx_connection *));
6528 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6530 rxi_FreeAllPackets();
6532 MUTEX_ENTER(&rx_stats_mutex);
6533 rxi_dataQuota = RX_MAX_QUOTA;
6534 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6535 MUTEX_EXIT(&rx_stats_mutex);
6540 #ifdef RX_ENABLE_LOCKS
6542 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6544 if (!MUTEX_ISMINE(lockaddr))
6545 osi_Panic("Lock not held: %s", msg);
6547 #endif /* RX_ENABLE_LOCKS */
6552 * Routines to implement connection specific data.
6556 rx_KeyCreate(rx_destructor_t rtn)
6559 MUTEX_ENTER(&rxi_keyCreate_lock);
6560 key = rxi_keyCreate_counter++;
6561 rxi_keyCreate_destructor = (rx_destructor_t *)
6562 realloc((void *)rxi_keyCreate_destructor,
6563 (key + 1) * sizeof(rx_destructor_t));
6564 rxi_keyCreate_destructor[key] = rtn;
6565 MUTEX_EXIT(&rxi_keyCreate_lock);
6570 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6573 MUTEX_ENTER(&conn->conn_data_lock);
6574 if (!conn->specific) {
6575 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6576 for (i = 0; i < key; i++)
6577 conn->specific[i] = NULL;
6578 conn->nSpecific = key + 1;
6579 conn->specific[key] = ptr;
6580 } else if (key >= conn->nSpecific) {
6581 conn->specific = (void **)
6582 realloc(conn->specific, (key + 1) * sizeof(void *));
6583 for (i = conn->nSpecific; i < key; i++)
6584 conn->specific[i] = NULL;
6585 conn->nSpecific = key + 1;
6586 conn->specific[key] = ptr;
6588 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6589 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6590 conn->specific[key] = ptr;
6592 MUTEX_EXIT(&conn->conn_data_lock);
6596 rx_GetSpecific(struct rx_connection *conn, int key)
6599 MUTEX_ENTER(&conn->conn_data_lock);
6600 if (key >= conn->nSpecific)
6603 ptr = conn->specific[key];
6604 MUTEX_EXIT(&conn->conn_data_lock);
6608 #endif /* !KERNEL */
6611 * processStats is a queue used to store the statistics for the local
6612 * process. Its contents are similar to the contents of the rpcStats
6613 * queue on a rx_peer structure, but the actual data stored within
6614 * this queue contains totals across the lifetime of the process (assuming
6615 * the stats have not been reset) - unlike the per peer structures
6616 * which can come and go based upon the peer lifetime.
6619 static struct rx_queue processStats = { &processStats, &processStats };
6622 * peerStats is a queue used to store the statistics for all peer structs.
6623 * Its contents are the union of all the peer rpcStats queues.
6626 static struct rx_queue peerStats = { &peerStats, &peerStats };
6629 * rxi_monitor_processStats is used to turn process wide stat collection
6633 static int rxi_monitor_processStats = 0;
6636 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6639 static int rxi_monitor_peerStats = 0;
6642 * rxi_AddRpcStat - given all of the information for a particular rpc
6643 * call, create (if needed) and update the stat totals for the rpc.
6647 * IN stats - the queue of stats that will be updated with the new value
6649 * IN rxInterface - a unique number that identifies the rpc interface
6651 * IN currentFunc - the index of the function being invoked
6653 * IN totalFunc - the total number of functions in this interface
6655 * IN queueTime - the amount of time this function waited for a thread
6657 * IN execTime - the amount of time this function invocation took to execute
6659 * IN bytesSent - the number bytes sent by this invocation
6661 * IN bytesRcvd - the number bytes received by this invocation
6663 * IN isServer - if true, this invocation was made to a server
6665 * IN remoteHost - the ip address of the remote host
6667 * IN remotePort - the port of the remote host
6669 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6671 * INOUT counter - if a new stats structure is allocated, the counter will
6672 * be updated with the new number of allocated stat structures
6680 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6681 afs_uint32 currentFunc, afs_uint32 totalFunc,
6682 struct clock *queueTime, struct clock *execTime,
6683 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6684 afs_uint32 remoteHost, afs_uint32 remotePort,
6685 int addToPeerList, unsigned int *counter)
6688 rx_interface_stat_p rpc_stat, nrpc_stat;
6691 * See if there's already a structure for this interface
6694 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6695 if ((rpc_stat->stats[0].interfaceId == rxInterface)
6696 && (rpc_stat->stats[0].remote_is_server == isServer))
6701 * Didn't find a match so allocate a new structure and add it to the
6705 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
6706 || (rpc_stat->stats[0].interfaceId != rxInterface)
6707 || (rpc_stat->stats[0].remote_is_server != isServer)) {
6712 sizeof(rx_interface_stat_t) +
6713 totalFunc * sizeof(rx_function_entry_v1_t);
6715 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
6716 if (rpc_stat == NULL) {
6720 *counter += totalFunc;
6721 for (i = 0; i < totalFunc; i++) {
6722 rpc_stat->stats[i].remote_peer = remoteHost;
6723 rpc_stat->stats[i].remote_port = remotePort;
6724 rpc_stat->stats[i].remote_is_server = isServer;
6725 rpc_stat->stats[i].interfaceId = rxInterface;
6726 rpc_stat->stats[i].func_total = totalFunc;
6727 rpc_stat->stats[i].func_index = i;
6728 hzero(rpc_stat->stats[i].invocations);
6729 hzero(rpc_stat->stats[i].bytes_sent);
6730 hzero(rpc_stat->stats[i].bytes_rcvd);
6731 rpc_stat->stats[i].queue_time_sum.sec = 0;
6732 rpc_stat->stats[i].queue_time_sum.usec = 0;
6733 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
6734 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
6735 rpc_stat->stats[i].queue_time_min.sec = 9999999;
6736 rpc_stat->stats[i].queue_time_min.usec = 9999999;
6737 rpc_stat->stats[i].queue_time_max.sec = 0;
6738 rpc_stat->stats[i].queue_time_max.usec = 0;
6739 rpc_stat->stats[i].execution_time_sum.sec = 0;
6740 rpc_stat->stats[i].execution_time_sum.usec = 0;
6741 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
6742 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
6743 rpc_stat->stats[i].execution_time_min.sec = 9999999;
6744 rpc_stat->stats[i].execution_time_min.usec = 9999999;
6745 rpc_stat->stats[i].execution_time_max.sec = 0;
6746 rpc_stat->stats[i].execution_time_max.usec = 0;
6748 queue_Prepend(stats, rpc_stat);
6749 if (addToPeerList) {
6750 queue_Prepend(&peerStats, &rpc_stat->all_peers);
6755 * Increment the stats for this function
6758 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
6759 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
6760 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
6761 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
6762 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
6763 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
6764 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
6766 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
6767 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
6769 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
6770 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
6772 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
6773 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
6775 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
6776 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
6784 * rx_IncrementTimeAndCount - increment the times and count for a particular
6789 * IN peer - the peer who invoked the rpc
6791 * IN rxInterface - a unique number that identifies the rpc interface
6793 * IN currentFunc - the index of the function being invoked
6795 * IN totalFunc - the total number of functions in this interface
6797 * IN queueTime - the amount of time this function waited for a thread
6799 * IN execTime - the amount of time this function invocation took to execute
6801 * IN bytesSent - the number bytes sent by this invocation
6803 * IN bytesRcvd - the number bytes received by this invocation
6805 * IN isServer - if true, this invocation was made to a server
6813 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
6814 afs_uint32 currentFunc, afs_uint32 totalFunc,
6815 struct clock *queueTime, struct clock *execTime,
6816 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
6820 MUTEX_ENTER(&rx_rpc_stats);
6821 MUTEX_ENTER(&peer->peer_lock);
6823 if (rxi_monitor_peerStats) {
6824 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
6825 queueTime, execTime, bytesSent, bytesRcvd, isServer,
6826 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
6829 if (rxi_monitor_processStats) {
6830 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
6831 queueTime, execTime, bytesSent, bytesRcvd, isServer,
6832 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
6835 MUTEX_EXIT(&peer->peer_lock);
6836 MUTEX_EXIT(&rx_rpc_stats);
6841 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
6845 * IN callerVersion - the rpc stat version of the caller.
6847 * IN count - the number of entries to marshall.
6849 * IN stats - pointer to stats to be marshalled.
6851 * OUT ptr - Where to store the marshalled data.
6858 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
6859 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
6865 * We only support the first version
6867 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
6868 *(ptr++) = stats->remote_peer;
6869 *(ptr++) = stats->remote_port;
6870 *(ptr++) = stats->remote_is_server;
6871 *(ptr++) = stats->interfaceId;
6872 *(ptr++) = stats->func_total;
6873 *(ptr++) = stats->func_index;
6874 *(ptr++) = hgethi(stats->invocations);
6875 *(ptr++) = hgetlo(stats->invocations);
6876 *(ptr++) = hgethi(stats->bytes_sent);
6877 *(ptr++) = hgetlo(stats->bytes_sent);
6878 *(ptr++) = hgethi(stats->bytes_rcvd);
6879 *(ptr++) = hgetlo(stats->bytes_rcvd);
6880 *(ptr++) = stats->queue_time_sum.sec;
6881 *(ptr++) = stats->queue_time_sum.usec;
6882 *(ptr++) = stats->queue_time_sum_sqr.sec;
6883 *(ptr++) = stats->queue_time_sum_sqr.usec;
6884 *(ptr++) = stats->queue_time_min.sec;
6885 *(ptr++) = stats->queue_time_min.usec;
6886 *(ptr++) = stats->queue_time_max.sec;
6887 *(ptr++) = stats->queue_time_max.usec;
6888 *(ptr++) = stats->execution_time_sum.sec;
6889 *(ptr++) = stats->execution_time_sum.usec;
6890 *(ptr++) = stats->execution_time_sum_sqr.sec;
6891 *(ptr++) = stats->execution_time_sum_sqr.usec;
6892 *(ptr++) = stats->execution_time_min.sec;
6893 *(ptr++) = stats->execution_time_min.usec;
6894 *(ptr++) = stats->execution_time_max.sec;
6895 *(ptr++) = stats->execution_time_max.usec;
6901 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
6906 * IN callerVersion - the rpc stat version of the caller
6908 * OUT myVersion - the rpc stat version of this function
6910 * OUT clock_sec - local time seconds
6912 * OUT clock_usec - local time microseconds
6914 * OUT allocSize - the number of bytes allocated to contain stats
6916 * OUT statCount - the number stats retrieved from this process.
6918 * OUT stats - the actual stats retrieved from this process.
6922 * Returns void. If successful, stats will != NULL.
6926 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
6927 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
6928 size_t * allocSize, afs_uint32 * statCount,
6929 afs_uint32 ** stats)
6939 *myVersion = RX_STATS_RETRIEVAL_VERSION;
6942 * Check to see if stats are enabled
6945 MUTEX_ENTER(&rx_rpc_stats);
6946 if (!rxi_monitor_processStats) {
6947 MUTEX_EXIT(&rx_rpc_stats);
6951 clock_GetTime(&now);
6952 *clock_sec = now.sec;
6953 *clock_usec = now.usec;
6956 * Allocate the space based upon the caller version
6958 * If the client is at an older version than we are,
6959 * we return the statistic data in the older data format, but
6960 * we still return our version number so the client knows we
6961 * are maintaining more data than it can retrieve.
6964 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
6965 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
6966 *statCount = rxi_rpc_process_stat_cnt;
6969 * This can't happen yet, but in the future version changes
6970 * can be handled by adding additional code here
6974 if (space > (size_t) 0) {
6976 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
6979 rx_interface_stat_p rpc_stat, nrpc_stat;
6983 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6985 * Copy the data based upon the caller version
6987 rx_MarshallProcessRPCStats(callerVersion,
6988 rpc_stat->stats[0].func_total,
6989 rpc_stat->stats, &ptr);
6995 MUTEX_EXIT(&rx_rpc_stats);
7000 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7004 * IN callerVersion - the rpc stat version of the caller
7006 * OUT myVersion - the rpc stat version of this function
7008 * OUT clock_sec - local time seconds
7010 * OUT clock_usec - local time microseconds
7012 * OUT allocSize - the number of bytes allocated to contain stats
7014 * OUT statCount - the number of stats retrieved from the individual
7017 * OUT stats - the actual stats retrieved from the individual peer structures.
7021 * Returns void. If successful, stats will != NULL.
7025 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7026 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7027 size_t * allocSize, afs_uint32 * statCount,
7028 afs_uint32 ** stats)
7038 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7041 * Check to see if stats are enabled
7044 MUTEX_ENTER(&rx_rpc_stats);
7045 if (!rxi_monitor_peerStats) {
7046 MUTEX_EXIT(&rx_rpc_stats);
7050 clock_GetTime(&now);
7051 *clock_sec = now.sec;
7052 *clock_usec = now.usec;
7055 * Allocate the space based upon the caller version
7057 * If the client is at an older version than we are,
7058 * we return the statistic data in the older data format, but
7059 * we still return our version number so the client knows we
7060 * are maintaining more data than it can retrieve.
7063 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7064 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7065 *statCount = rxi_rpc_peer_stat_cnt;
7068 * This can't happen yet, but in the future version changes
7069 * can be handled by adding additional code here
7073 if (space > (size_t) 0) {
7075 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7078 rx_interface_stat_p rpc_stat, nrpc_stat;
7082 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7084 * We have to fix the offset of rpc_stat since we are
7085 * keeping this structure on two rx_queues. The rx_queue
7086 * package assumes that the rx_queue member is the first
7087 * member of the structure. That is, rx_queue assumes that
7088 * any one item is only on one queue at a time. We are
7089 * breaking that assumption and so we have to do a little
7090 * math to fix our pointers.
7093 fix_offset = (char *)rpc_stat;
7094 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7095 rpc_stat = (rx_interface_stat_p) fix_offset;
7098 * Copy the data based upon the caller version
7100 rx_MarshallProcessRPCStats(callerVersion,
7101 rpc_stat->stats[0].func_total,
7102 rpc_stat->stats, &ptr);
7108 MUTEX_EXIT(&rx_rpc_stats);
7113 * rx_FreeRPCStats - free memory allocated by
7114 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7118 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7119 * rx_RetrievePeerRPCStats
7121 * IN allocSize - the number of bytes in stats.
7129 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7131 rxi_Free(stats, allocSize);
7135 * rx_queryProcessRPCStats - see if process rpc stat collection is
7136 * currently enabled.
7142 * Returns 0 if stats are not enabled != 0 otherwise
7146 rx_queryProcessRPCStats(void)
7149 MUTEX_ENTER(&rx_rpc_stats);
7150 rc = rxi_monitor_processStats;
7151 MUTEX_EXIT(&rx_rpc_stats);
7156 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7162 * Returns 0 if stats are not enabled != 0 otherwise
7166 rx_queryPeerRPCStats(void)
7169 MUTEX_ENTER(&rx_rpc_stats);
7170 rc = rxi_monitor_peerStats;
7171 MUTEX_EXIT(&rx_rpc_stats);
7176 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7186 rx_enableProcessRPCStats(void)
7188 MUTEX_ENTER(&rx_rpc_stats);
7189 rx_enable_stats = 1;
7190 rxi_monitor_processStats = 1;
7191 MUTEX_EXIT(&rx_rpc_stats);
7195 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7205 rx_enablePeerRPCStats(void)
7207 MUTEX_ENTER(&rx_rpc_stats);
7208 rx_enable_stats = 1;
7209 rxi_monitor_peerStats = 1;
7210 MUTEX_EXIT(&rx_rpc_stats);
7214 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7224 rx_disableProcessRPCStats(void)
7226 rx_interface_stat_p rpc_stat, nrpc_stat;
7229 MUTEX_ENTER(&rx_rpc_stats);
7232 * Turn off process statistics and if peer stats is also off, turn
7236 rxi_monitor_processStats = 0;
7237 if (rxi_monitor_peerStats == 0) {
7238 rx_enable_stats = 0;
7241 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7242 unsigned int num_funcs = 0;
7245 queue_Remove(rpc_stat);
7246 num_funcs = rpc_stat->stats[0].func_total;
7248 sizeof(rx_interface_stat_t) +
7249 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7251 rxi_Free(rpc_stat, space);
7252 rxi_rpc_process_stat_cnt -= num_funcs;
7254 MUTEX_EXIT(&rx_rpc_stats);
7258 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7268 rx_disablePeerRPCStats(void)
7270 struct rx_peer **peer_ptr, **peer_end;
7273 MUTEX_ENTER(&rx_rpc_stats);
7276 * Turn off peer statistics and if process stats is also off, turn
7280 rxi_monitor_peerStats = 0;
7281 if (rxi_monitor_processStats == 0) {
7282 rx_enable_stats = 0;
7285 MUTEX_ENTER(&rx_peerHashTable_lock);
7286 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7287 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7289 struct rx_peer *peer, *next, *prev;
7290 for (prev = peer = *peer_ptr; peer; peer = next) {
7292 code = MUTEX_TRYENTER(&peer->peer_lock);
7294 rx_interface_stat_p rpc_stat, nrpc_stat;
7297 (&peer->rpcStats, rpc_stat, nrpc_stat,
7298 rx_interface_stat)) {
7299 unsigned int num_funcs = 0;
7302 queue_Remove(&rpc_stat->queue_header);
7303 queue_Remove(&rpc_stat->all_peers);
7304 num_funcs = rpc_stat->stats[0].func_total;
7306 sizeof(rx_interface_stat_t) +
7307 rpc_stat->stats[0].func_total *
7308 sizeof(rx_function_entry_v1_t);
7310 rxi_Free(rpc_stat, space);
7311 rxi_rpc_peer_stat_cnt -= num_funcs;
7313 MUTEX_EXIT(&peer->peer_lock);
7314 if (prev == *peer_ptr) {
7324 MUTEX_EXIT(&rx_peerHashTable_lock);
7325 MUTEX_EXIT(&rx_rpc_stats);
7329 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7334 * IN clearFlag - flag indicating which stats to clear
7342 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7344 rx_interface_stat_p rpc_stat, nrpc_stat;
7346 MUTEX_ENTER(&rx_rpc_stats);
7348 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7349 unsigned int num_funcs = 0, i;
7350 num_funcs = rpc_stat->stats[0].func_total;
7351 for (i = 0; i < num_funcs; i++) {
7352 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7353 hzero(rpc_stat->stats[i].invocations);
7355 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7356 hzero(rpc_stat->stats[i].bytes_sent);
7358 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7359 hzero(rpc_stat->stats[i].bytes_rcvd);
7361 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7362 rpc_stat->stats[i].queue_time_sum.sec = 0;
7363 rpc_stat->stats[i].queue_time_sum.usec = 0;
7365 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7366 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7367 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7369 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7370 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7371 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7373 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7374 rpc_stat->stats[i].queue_time_max.sec = 0;
7375 rpc_stat->stats[i].queue_time_max.usec = 0;
7377 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7378 rpc_stat->stats[i].execution_time_sum.sec = 0;
7379 rpc_stat->stats[i].execution_time_sum.usec = 0;
7381 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7382 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7383 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7385 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7386 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7387 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7389 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7390 rpc_stat->stats[i].execution_time_max.sec = 0;
7391 rpc_stat->stats[i].execution_time_max.usec = 0;
7396 MUTEX_EXIT(&rx_rpc_stats);
7400 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7405 * IN clearFlag - flag indicating which stats to clear
7413 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7415 rx_interface_stat_p rpc_stat, nrpc_stat;
7417 MUTEX_ENTER(&rx_rpc_stats);
7419 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7420 unsigned int num_funcs = 0, i;
7423 * We have to fix the offset of rpc_stat since we are
7424 * keeping this structure on two rx_queues. The rx_queue
7425 * package assumes that the rx_queue member is the first
7426 * member of the structure. That is, rx_queue assumes that
7427 * any one item is only on one queue at a time. We are
7428 * breaking that assumption and so we have to do a little
7429 * math to fix our pointers.
7432 fix_offset = (char *)rpc_stat;
7433 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7434 rpc_stat = (rx_interface_stat_p) fix_offset;
7436 num_funcs = rpc_stat->stats[0].func_total;
7437 for (i = 0; i < num_funcs; i++) {
7438 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7439 hzero(rpc_stat->stats[i].invocations);
7441 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7442 hzero(rpc_stat->stats[i].bytes_sent);
7444 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7445 hzero(rpc_stat->stats[i].bytes_rcvd);
7447 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7448 rpc_stat->stats[i].queue_time_sum.sec = 0;
7449 rpc_stat->stats[i].queue_time_sum.usec = 0;
7451 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7452 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7453 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7455 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7456 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7457 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7459 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7460 rpc_stat->stats[i].queue_time_max.sec = 0;
7461 rpc_stat->stats[i].queue_time_max.usec = 0;
7463 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7464 rpc_stat->stats[i].execution_time_sum.sec = 0;
7465 rpc_stat->stats[i].execution_time_sum.usec = 0;
7467 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7468 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7469 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7471 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7472 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7473 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7475 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7476 rpc_stat->stats[i].execution_time_max.sec = 0;
7477 rpc_stat->stats[i].execution_time_max.usec = 0;
7482 MUTEX_EXIT(&rx_rpc_stats);
7486 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7487 * is authorized to enable/disable/clear RX statistics.
7489 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7492 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7494 rxi_rxstat_userok = proc;
7498 rx_RxStatUserOk(struct rx_call *call)
7500 if (!rxi_rxstat_userok)
7502 return rxi_rxstat_userok(call);