2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
14 #include "afs/param.h"
16 #include <afs/param.h>
23 #include "afs/sysincludes.h"
24 #include "afsincludes.h"
30 #include <net/net_globals.h>
31 #endif /* AFS_OSF_ENV */
32 #ifdef AFS_LINUX20_ENV
35 #include "netinet/in.h"
36 #include "afs/afs_args.h"
37 #include "afs/afs_osi.h"
38 #ifdef RX_KERNEL_TRACE
39 #include "rx_kcommon.h"
41 #if (defined(AFS_AUX_ENV) || defined(AFS_AIX_ENV))
45 #undef RXDEBUG /* turn off debugging */
47 #if defined(AFS_SGI_ENV)
48 #include "sys/debug.h"
57 #endif /* AFS_OSF_ENV */
59 #include "afs/sysincludes.h"
60 #include "afsincludes.h"
63 #include "rx_kmutex.h"
64 #include "rx_kernel.h"
68 #include "rx_globals.h"
70 #define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
71 #define AFSOP_STOP_AFS 211 /* Stop AFS process */
72 #define AFSOP_STOP_BKG 212 /* Stop BKG process */
74 extern afs_int32 afs_termState;
76 #include "sys/lockl.h"
77 #include "sys/lock_def.h"
78 #endif /* AFS_AIX41_ENV */
79 # include "rxgen_consts.h"
81 # include <sys/types.h>
86 # include <afs/afsutil.h>
88 # include <sys/socket.h>
89 # include <sys/file.h>
91 # include <sys/stat.h>
92 # include <netinet/in.h>
93 # include <sys/time.h>
103 # include "rx_user.h"
104 # include "rx_clock.h"
105 # include "rx_queue.h"
106 # include "rx_globals.h"
107 # include "rx_trace.h"
108 # include <afs/rxgen_consts.h>
111 int (*registerProgram) () = 0;
112 int (*swapNameProgram) () = 0;
114 /* Local static routines */
115 static void rxi_DestroyConnectionNoLock(register struct rx_connection *conn);
116 #ifdef RX_ENABLE_LOCKS
117 static void rxi_SetAcksInTransmitQueue(register struct rx_call *call);
120 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
122 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
123 afs_int32 rxi_start_in_error;
125 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
128 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
129 * currently allocated within rx. This number is used to allocate the
130 * memory required to return the statistics when queried.
133 static unsigned int rxi_rpc_peer_stat_cnt;
136 * rxi_rpc_process_stat_cnt counts the total number of local process stat
137 * structures currently allocated within rx. The number is used to allocate
138 * the memory required to return the statistics when queried.
141 static unsigned int rxi_rpc_process_stat_cnt;
143 #if !defined(offsetof)
144 #include <stddef.h> /* for definition of offsetof() */
147 #ifdef AFS_PTHREAD_ENV
151 * Use procedural initialization of mutexes/condition variables
155 extern pthread_mutex_t rx_stats_mutex;
156 extern pthread_mutex_t rxkad_stats_mutex;
157 extern pthread_mutex_t des_init_mutex;
158 extern pthread_mutex_t des_random_mutex;
159 extern pthread_mutex_t rx_clock_mutex;
160 extern pthread_mutex_t rxi_connCacheMutex;
161 extern pthread_mutex_t rx_event_mutex;
162 extern pthread_mutex_t osi_malloc_mutex;
163 extern pthread_mutex_t event_handler_mutex;
164 extern pthread_mutex_t listener_mutex;
165 extern pthread_mutex_t rx_if_init_mutex;
166 extern pthread_mutex_t rx_if_mutex;
167 extern pthread_mutex_t rxkad_client_uid_mutex;
168 extern pthread_mutex_t rxkad_random_mutex;
170 extern pthread_cond_t rx_event_handler_cond;
171 extern pthread_cond_t rx_listener_cond;
173 static pthread_mutex_t epoch_mutex;
174 static pthread_mutex_t rx_init_mutex;
175 static pthread_mutex_t rx_debug_mutex;
178 rxi_InitPthread(void)
180 assert(pthread_mutex_init(&rx_clock_mutex, (const pthread_mutexattr_t *)0)
182 assert(pthread_mutex_init(&rx_stats_mutex, (const pthread_mutexattr_t *)0)
184 assert(pthread_mutex_init
185 (&rxi_connCacheMutex, (const pthread_mutexattr_t *)0) == 0);
186 assert(pthread_mutex_init(&rx_init_mutex, (const pthread_mutexattr_t *)0)
188 assert(pthread_mutex_init(&epoch_mutex, (const pthread_mutexattr_t *)0) ==
190 assert(pthread_mutex_init(&rx_event_mutex, (const pthread_mutexattr_t *)0)
192 assert(pthread_mutex_init(&des_init_mutex, (const pthread_mutexattr_t *)0)
194 assert(pthread_mutex_init
195 (&des_random_mutex, (const pthread_mutexattr_t *)0) == 0);
196 assert(pthread_mutex_init
197 (&osi_malloc_mutex, (const pthread_mutexattr_t *)0) == 0);
198 assert(pthread_mutex_init
199 (&event_handler_mutex, (const pthread_mutexattr_t *)0) == 0);
200 assert(pthread_mutex_init(&listener_mutex, (const pthread_mutexattr_t *)0)
202 assert(pthread_mutex_init
203 (&rx_if_init_mutex, (const pthread_mutexattr_t *)0) == 0);
204 assert(pthread_mutex_init(&rx_if_mutex, (const pthread_mutexattr_t *)0) ==
206 assert(pthread_mutex_init
207 (&rxkad_client_uid_mutex, (const pthread_mutexattr_t *)0) == 0);
208 assert(pthread_mutex_init
209 (&rxkad_random_mutex, (const pthread_mutexattr_t *)0) == 0);
210 assert(pthread_mutex_init
211 (&rxkad_stats_mutex, (const pthread_mutexattr_t *)0) == 0);
212 assert(pthread_mutex_init(&rx_debug_mutex, (const pthread_mutexattr_t *)0)
215 assert(pthread_cond_init
216 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
217 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
219 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
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)
355 /* Initialize rx. A port number may be mentioned, in which case this
356 * becomes the default port number for any service installed later.
357 * If 0 is provided for the port number, a random port will be chosen
358 * by the kernel. Whether this will ever overlap anything in
359 * /etc/services is anybody's guess... Returns 0 on success, -1 on
361 static int rxinit_status = 1;
362 #ifdef AFS_PTHREAD_ENV
364 * This mutex protects the following global variables:
368 #define LOCK_RX_INIT assert(pthread_mutex_lock(&rx_init_mutex)==0)
369 #define UNLOCK_RX_INIT assert(pthread_mutex_unlock(&rx_init_mutex)==0)
372 #define UNLOCK_RX_INIT
376 rx_InitHost(u_int host, u_int port)
383 char *htable, *ptable;
386 #if defined(AFS_DJGPP_ENV) && !defined(DEBUG)
387 __djgpp_set_quiet_socket(1);
394 if (rxinit_status == 0) {
395 tmp_status = rxinit_status;
397 return tmp_status; /* Already started; return previous error code. */
400 if (afs_winsockInit() < 0)
406 * Initialize anything necessary to provide a non-premptive threading
409 rxi_InitializeThreadSupport();
412 /* Allocate and initialize a socket for client and perhaps server
415 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
416 if (rx_socket == OSI_NULLSOCKET) {
420 #ifdef RX_ENABLE_LOCKS
423 #endif /* RX_LOCKS_DB */
424 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
425 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
426 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
427 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
428 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
430 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
432 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
434 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
436 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
438 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
440 #if defined(KERNEL) && defined(AFS_HPUX110_ENV)
442 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
443 #endif /* KERNEL && AFS_HPUX110_ENV */
444 #else /* RX_ENABLE_LOCKS */
445 #if defined(KERNEL) && defined(AFS_GLOBAL_SUNLOCK) && !defined(AFS_HPUX_ENV) && !defined(AFS_OBSD_ENV)
446 mutex_init(&afs_rxglobal_lock, "afs_rxglobal_lock", MUTEX_DEFAULT, NULL);
447 #endif /* AFS_GLOBAL_SUNLOCK */
448 #endif /* RX_ENABLE_LOCKS */
451 rx_connDeadTime = 12;
452 rx_tranquil = 0; /* reset flag */
453 memset((char *)&rx_stats, 0, sizeof(struct rx_stats));
455 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
456 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
457 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
458 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
459 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
460 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
462 /* Malloc up a bunch of packets & buffers */
464 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
465 queue_Init(&rx_freePacketQueue);
466 rxi_NeedMorePackets = FALSE;
467 rxi_MorePackets(rx_nPackets);
474 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
475 tv.tv_sec = clock_now.sec;
476 tv.tv_usec = clock_now.usec;
477 srand((unsigned int)tv.tv_usec);
484 #if defined(KERNEL) && !defined(UKERNEL)
485 /* Really, this should never happen in a real kernel */
488 struct sockaddr_in addr;
489 int addrlen = sizeof(addr);
490 if (getsockname((int)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
494 rx_port = addr.sin_port;
497 rx_stats.minRtt.sec = 9999999;
499 rx_SetEpoch(tv.tv_sec | 0x80000000);
501 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
502 * will provide a randomer value. */
504 MUTEX_ENTER(&rx_stats_mutex);
505 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
506 MUTEX_EXIT(&rx_stats_mutex);
507 /* *Slightly* random start time for the cid. This is just to help
508 * out with the hashing function at the peer */
509 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
510 rx_connHashTable = (struct rx_connection **)htable;
511 rx_peerHashTable = (struct rx_peer **)ptable;
513 rx_lastAckDelay.sec = 0;
514 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
515 rx_hardAckDelay.sec = 0;
516 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
517 rx_softAckDelay.sec = 0;
518 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
520 rxevent_Init(20, rxi_ReScheduleEvents);
522 /* Initialize various global queues */
523 queue_Init(&rx_idleServerQueue);
524 queue_Init(&rx_incomingCallQueue);
525 queue_Init(&rx_freeCallQueue);
527 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
528 /* Initialize our list of usable IP addresses. */
532 /* Start listener process (exact function is dependent on the
533 * implementation environment--kernel or user space) */
537 tmp_status = rxinit_status = 0;
545 return rx_InitHost(htonl(INADDR_ANY), port);
548 /* called with unincremented nRequestsRunning to see if it is OK to start
549 * a new thread in this service. Could be "no" for two reasons: over the
550 * max quota, or would prevent others from reaching their min quota.
552 #ifdef RX_ENABLE_LOCKS
553 /* This verion of QuotaOK reserves quota if it's ok while the
554 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
557 QuotaOK(register struct rx_service *aservice)
559 /* check if over max quota */
560 if (aservice->nRequestsRunning >= aservice->maxProcs) {
564 /* under min quota, we're OK */
565 /* otherwise, can use only if there are enough to allow everyone
566 * to go to their min quota after this guy starts.
568 MUTEX_ENTER(&rx_stats_mutex);
569 if ((aservice->nRequestsRunning < aservice->minProcs)
570 || (rxi_availProcs > rxi_minDeficit)) {
571 aservice->nRequestsRunning++;
572 /* just started call in minProcs pool, need fewer to maintain
574 if (aservice->nRequestsRunning <= aservice->minProcs)
577 MUTEX_EXIT(&rx_stats_mutex);
580 MUTEX_EXIT(&rx_stats_mutex);
586 ReturnToServerPool(register struct rx_service *aservice)
588 aservice->nRequestsRunning--;
589 MUTEX_ENTER(&rx_stats_mutex);
590 if (aservice->nRequestsRunning < aservice->minProcs)
593 MUTEX_EXIT(&rx_stats_mutex);
596 #else /* RX_ENABLE_LOCKS */
598 QuotaOK(register struct rx_service *aservice)
601 /* under min quota, we're OK */
602 if (aservice->nRequestsRunning < aservice->minProcs)
605 /* check if over max quota */
606 if (aservice->nRequestsRunning >= aservice->maxProcs)
609 /* otherwise, can use only if there are enough to allow everyone
610 * to go to their min quota after this guy starts.
612 if (rxi_availProcs > rxi_minDeficit)
616 #endif /* RX_ENABLE_LOCKS */
619 /* Called by rx_StartServer to start up lwp's to service calls.
620 NExistingProcs gives the number of procs already existing, and which
621 therefore needn't be created. */
623 rxi_StartServerProcs(int nExistingProcs)
625 register struct rx_service *service;
630 /* For each service, reserve N processes, where N is the "minimum"
631 * number of processes that MUST be able to execute a request in parallel,
632 * at any time, for that process. Also compute the maximum difference
633 * between any service's maximum number of processes that can run
634 * (i.e. the maximum number that ever will be run, and a guarantee
635 * that this number will run if other services aren't running), and its
636 * minimum number. The result is the extra number of processes that
637 * we need in order to provide the latter guarantee */
638 for (i = 0; i < RX_MAX_SERVICES; i++) {
640 service = rx_services[i];
641 if (service == (struct rx_service *)0)
643 nProcs += service->minProcs;
644 diff = service->maxProcs - service->minProcs;
648 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
649 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
650 for (i = 0; i < nProcs; i++) {
651 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
656 /* This routine must be called if any services are exported. If the
657 * donateMe flag is set, the calling process is donated to the server
660 rx_StartServer(int donateMe)
662 register struct rx_service *service;
668 /* Start server processes, if necessary (exact function is dependent
669 * on the implementation environment--kernel or user space). DonateMe
670 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
671 * case, one less new proc will be created rx_StartServerProcs.
673 rxi_StartServerProcs(donateMe);
675 /* count up the # of threads in minProcs, and add set the min deficit to
676 * be that value, too.
678 for (i = 0; i < RX_MAX_SERVICES; i++) {
679 service = rx_services[i];
680 if (service == (struct rx_service *)0)
682 MUTEX_ENTER(&rx_stats_mutex);
683 rxi_totalMin += service->minProcs;
684 /* below works even if a thread is running, since minDeficit would
685 * still have been decremented and later re-incremented.
687 rxi_minDeficit += service->minProcs;
688 MUTEX_EXIT(&rx_stats_mutex);
691 /* Turn on reaping of idle server connections */
692 rxi_ReapConnections();
701 #ifdef AFS_PTHREAD_ENV
703 pid = (pid_t) pthread_self();
704 #else /* AFS_PTHREAD_ENV */
706 LWP_CurrentProcess(&pid);
707 #endif /* AFS_PTHREAD_ENV */
709 sprintf(name, "srv_%d", ++nProcs);
711 (*registerProgram) (pid, name);
713 #endif /* AFS_NT40_ENV */
714 rx_ServerProc(); /* Never returns */
719 /* Create a new client connection to the specified service, using the
720 * specified security object to implement the security model for this
722 struct rx_connection *
723 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
724 register struct rx_securityClass *securityObject,
725 int serviceSecurityIndex)
729 register struct rx_connection *conn;
734 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", shost, sport, sservice, securityObject, serviceSecurityIndex));
736 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
737 * the case of kmem_alloc? */
738 conn = rxi_AllocConnection();
739 #ifdef RX_ENABLE_LOCKS
740 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
741 MUTEX_INIT(&conn->conn_data_lock, "conn call lock", MUTEX_DEFAULT, 0);
742 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
745 MUTEX_ENTER(&rx_connHashTable_lock);
746 cid = (rx_nextCid += RX_MAXCALLS);
747 conn->type = RX_CLIENT_CONNECTION;
749 conn->epoch = rx_epoch;
750 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
751 conn->serviceId = sservice;
752 conn->securityObject = securityObject;
753 /* This doesn't work in all compilers with void (they're buggy), so fake it
755 conn->securityData = (VOID *) 0;
756 conn->securityIndex = serviceSecurityIndex;
757 rx_SetConnDeadTime(conn, rx_connDeadTime);
758 conn->ackRate = RX_FAST_ACK_RATE;
760 conn->specific = NULL;
761 conn->challengeEvent = NULL;
762 conn->delayedAbortEvent = NULL;
763 conn->abortCount = 0;
766 RXS_NewConnection(securityObject, conn);
768 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
770 conn->refCount++; /* no lock required since only this thread knows... */
771 conn->next = rx_connHashTable[hashindex];
772 rx_connHashTable[hashindex] = conn;
773 MUTEX_ENTER(&rx_stats_mutex);
774 rx_stats.nClientConns++;
775 MUTEX_EXIT(&rx_stats_mutex);
777 MUTEX_EXIT(&rx_connHashTable_lock);
783 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
785 /* The idea is to set the dead time to a value that allows several
786 * keepalives to be dropped without timing out the connection. */
787 conn->secondsUntilDead = MAX(seconds, 6);
788 conn->secondsUntilPing = conn->secondsUntilDead / 6;
791 int rxi_lowPeerRefCount = 0;
792 int rxi_lowConnRefCount = 0;
795 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
796 * NOTE: must not be called with rx_connHashTable_lock held.
799 rxi_CleanupConnection(struct rx_connection *conn)
801 /* Notify the service exporter, if requested, that this connection
802 * is being destroyed */
803 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
804 (*conn->service->destroyConnProc) (conn);
806 /* Notify the security module that this connection is being destroyed */
807 RXS_DestroyConnection(conn->securityObject, conn);
809 /* If this is the last connection using the rx_peer struct, set its
810 * idle time to now. rxi_ReapConnections will reap it if it's still
811 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
813 MUTEX_ENTER(&rx_peerHashTable_lock);
814 if (conn->peer->refCount < 2) {
815 conn->peer->idleWhen = clock_Sec();
816 if (conn->peer->refCount < 1) {
817 conn->peer->refCount = 1;
818 MUTEX_ENTER(&rx_stats_mutex);
819 rxi_lowPeerRefCount++;
820 MUTEX_EXIT(&rx_stats_mutex);
823 conn->peer->refCount--;
824 MUTEX_EXIT(&rx_peerHashTable_lock);
826 MUTEX_ENTER(&rx_stats_mutex);
827 if (conn->type == RX_SERVER_CONNECTION)
828 rx_stats.nServerConns--;
830 rx_stats.nClientConns--;
831 MUTEX_EXIT(&rx_stats_mutex);
834 if (conn->specific) {
836 for (i = 0; i < conn->nSpecific; i++) {
837 if (conn->specific[i] && rxi_keyCreate_destructor[i])
838 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
839 conn->specific[i] = NULL;
841 free(conn->specific);
843 conn->specific = NULL;
847 MUTEX_DESTROY(&conn->conn_call_lock);
848 MUTEX_DESTROY(&conn->conn_data_lock);
849 CV_DESTROY(&conn->conn_call_cv);
851 rxi_FreeConnection(conn);
854 /* Destroy the specified connection */
856 rxi_DestroyConnection(register struct rx_connection *conn)
858 MUTEX_ENTER(&rx_connHashTable_lock);
859 rxi_DestroyConnectionNoLock(conn);
860 /* conn should be at the head of the cleanup list */
861 if (conn == rx_connCleanup_list) {
862 rx_connCleanup_list = rx_connCleanup_list->next;
863 MUTEX_EXIT(&rx_connHashTable_lock);
864 rxi_CleanupConnection(conn);
866 #ifdef RX_ENABLE_LOCKS
868 MUTEX_EXIT(&rx_connHashTable_lock);
870 #endif /* RX_ENABLE_LOCKS */
874 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
876 register struct rx_connection **conn_ptr;
877 register int havecalls = 0;
878 struct rx_packet *packet;
885 MUTEX_ENTER(&conn->conn_data_lock);
886 if (conn->refCount > 0)
889 MUTEX_ENTER(&rx_stats_mutex);
890 rxi_lowConnRefCount++;
891 MUTEX_EXIT(&rx_stats_mutex);
894 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
895 /* Busy; wait till the last guy before proceeding */
896 MUTEX_EXIT(&conn->conn_data_lock);
901 /* If the client previously called rx_NewCall, but it is still
902 * waiting, treat this as a running call, and wait to destroy the
903 * connection later when the call completes. */
904 if ((conn->type == RX_CLIENT_CONNECTION)
905 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
906 conn->flags |= RX_CONN_DESTROY_ME;
907 MUTEX_EXIT(&conn->conn_data_lock);
911 MUTEX_EXIT(&conn->conn_data_lock);
913 /* Check for extant references to this connection */
914 for (i = 0; i < RX_MAXCALLS; i++) {
915 register struct rx_call *call = conn->call[i];
918 if (conn->type == RX_CLIENT_CONNECTION) {
919 MUTEX_ENTER(&call->lock);
920 if (call->delayedAckEvent) {
921 /* Push the final acknowledgment out now--there
922 * won't be a subsequent call to acknowledge the
923 * last reply packets */
924 rxevent_Cancel(call->delayedAckEvent, call,
925 RX_CALL_REFCOUNT_DELAY);
926 if (call->state == RX_STATE_PRECALL
927 || call->state == RX_STATE_ACTIVE) {
928 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
930 rxi_AckAll(NULL, call, 0);
933 MUTEX_EXIT(&call->lock);
937 #ifdef RX_ENABLE_LOCKS
939 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
940 MUTEX_EXIT(&conn->conn_data_lock);
942 /* Someone is accessing a packet right now. */
946 #endif /* RX_ENABLE_LOCKS */
949 /* Don't destroy the connection if there are any call
950 * structures still in use */
951 MUTEX_ENTER(&conn->conn_data_lock);
952 conn->flags |= RX_CONN_DESTROY_ME;
953 MUTEX_EXIT(&conn->conn_data_lock);
958 if (conn->delayedAbortEvent) {
959 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
960 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
962 MUTEX_ENTER(&conn->conn_data_lock);
963 rxi_SendConnectionAbort(conn, packet, 0, 1);
964 MUTEX_EXIT(&conn->conn_data_lock);
965 rxi_FreePacket(packet);
969 /* Remove from connection hash table before proceeding */
971 &rx_connHashTable[CONN_HASH
972 (peer->host, peer->port, conn->cid, conn->epoch,
974 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
975 if (*conn_ptr == conn) {
976 *conn_ptr = conn->next;
980 /* if the conn that we are destroying was the last connection, then we
981 * clear rxLastConn as well */
982 if (rxLastConn == conn)
985 /* Make sure the connection is completely reset before deleting it. */
986 /* get rid of pending events that could zap us later */
987 if (conn->challengeEvent)
988 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
989 if (conn->checkReachEvent)
990 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
992 /* Add the connection to the list of destroyed connections that
993 * need to be cleaned up. This is necessary to avoid deadlocks
994 * in the routines we call to inform others that this connection is
995 * being destroyed. */
996 conn->next = rx_connCleanup_list;
997 rx_connCleanup_list = conn;
1000 /* Externally available version */
1002 rx_DestroyConnection(register struct rx_connection *conn)
1007 rxi_DestroyConnection(conn);
1012 rx_GetConnection(register struct rx_connection *conn)
1017 MUTEX_ENTER(&conn->conn_data_lock);
1019 MUTEX_EXIT(&conn->conn_data_lock);
1023 /* Start a new rx remote procedure call, on the specified connection.
1024 * If wait is set to 1, wait for a free call channel; otherwise return
1025 * 0. Maxtime gives the maximum number of seconds this call may take,
1026 * after rx_MakeCall returns. After this time interval, a call to any
1027 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1028 * For fine grain locking, we hold the conn_call_lock in order to
1029 * to ensure that we don't get signalle after we found a call in an active
1030 * state and before we go to sleep.
1033 rx_NewCall(register struct rx_connection *conn)
1036 register struct rx_call *call;
1037 struct clock queueTime;
1041 dpf(("rx_MakeCall(conn %x)\n", conn));
1044 clock_GetTime(&queueTime);
1045 MUTEX_ENTER(&conn->conn_call_lock);
1048 * Check if there are others waiting for a new call.
1049 * If so, let them go first to avoid starving them.
1050 * This is a fairly simple scheme, and might not be
1051 * a complete solution for large numbers of waiters.
1053 if (conn->makeCallWaiters) {
1054 #ifdef RX_ENABLE_LOCKS
1055 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1062 for (i = 0; i < RX_MAXCALLS; i++) {
1063 call = conn->call[i];
1065 MUTEX_ENTER(&call->lock);
1066 if (call->state == RX_STATE_DALLY) {
1067 rxi_ResetCall(call, 0);
1068 (*call->callNumber)++;
1071 MUTEX_EXIT(&call->lock);
1073 call = rxi_NewCall(conn, i);
1077 if (i < RX_MAXCALLS) {
1080 MUTEX_ENTER(&conn->conn_data_lock);
1081 conn->flags |= RX_CONN_MAKECALL_WAITING;
1082 MUTEX_EXIT(&conn->conn_data_lock);
1084 conn->makeCallWaiters++;
1085 #ifdef RX_ENABLE_LOCKS
1086 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1090 conn->makeCallWaiters--;
1093 * Wake up anyone else who might be giving us a chance to
1094 * run (see code above that avoids resource starvation).
1096 #ifdef RX_ENABLE_LOCKS
1097 CV_BROADCAST(&conn->conn_call_cv);
1102 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1104 /* Client is initially in send mode */
1105 call->state = RX_STATE_ACTIVE;
1106 call->mode = RX_MODE_SENDING;
1108 /* remember start time for call in case we have hard dead time limit */
1109 call->queueTime = queueTime;
1110 clock_GetTime(&call->startTime);
1111 hzero(call->bytesSent);
1112 hzero(call->bytesRcvd);
1114 /* Turn on busy protocol. */
1115 rxi_KeepAliveOn(call);
1117 MUTEX_EXIT(&call->lock);
1118 MUTEX_EXIT(&conn->conn_call_lock);
1121 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1122 /* Now, if TQ wasn't cleared earlier, do it now. */
1123 MUTEX_ENTER(&call->lock);
1124 while (call->flags & RX_CALL_TQ_BUSY) {
1125 call->flags |= RX_CALL_TQ_WAIT;
1126 #ifdef RX_ENABLE_LOCKS
1127 CV_WAIT(&call->cv_tq, &call->lock);
1128 #else /* RX_ENABLE_LOCKS */
1129 osi_rxSleep(&call->tq);
1130 #endif /* RX_ENABLE_LOCKS */
1132 if (call->flags & RX_CALL_TQ_CLEARME) {
1133 rxi_ClearTransmitQueue(call, 0);
1134 queue_Init(&call->tq);
1136 MUTEX_EXIT(&call->lock);
1137 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1143 rxi_HasActiveCalls(register struct rx_connection *aconn)
1146 register struct rx_call *tcall;
1150 for (i = 0; i < RX_MAXCALLS; i++) {
1151 if ((tcall = aconn->call[i])) {
1152 if ((tcall->state == RX_STATE_ACTIVE)
1153 || (tcall->state == RX_STATE_PRECALL)) {
1164 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1165 register afs_int32 * aint32s)
1168 register struct rx_call *tcall;
1172 for (i = 0; i < RX_MAXCALLS; i++) {
1173 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1174 aint32s[i] = aconn->callNumber[i] + 1;
1176 aint32s[i] = aconn->callNumber[i];
1183 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1184 register afs_int32 * aint32s)
1187 register struct rx_call *tcall;
1191 for (i = 0; i < RX_MAXCALLS; i++) {
1192 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1193 aconn->callNumber[i] = aint32s[i] - 1;
1195 aconn->callNumber[i] = aint32s[i];
1201 /* Advertise a new service. A service is named locally by a UDP port
1202 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1205 char *serviceName; Name for identification purposes (e.g. the
1206 service name might be used for probing for
1209 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1210 struct rx_securityClass **securityObjects, int nSecurityObjects,
1211 afs_int32(*serviceProc) (struct rx_call * acall))
1213 osi_socket socket = OSI_NULLSOCKET;
1214 register struct rx_service *tservice;
1220 if (serviceId == 0) {
1222 "rx_NewService: service id for service %s is not non-zero.\n",
1229 "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",
1237 tservice = rxi_AllocService();
1239 for (i = 0; i < RX_MAX_SERVICES; i++) {
1240 register struct rx_service *service = rx_services[i];
1242 if (port == service->servicePort) {
1243 if (service->serviceId == serviceId) {
1244 /* The identical service has already been
1245 * installed; if the caller was intending to
1246 * change the security classes used by this
1247 * service, he/she loses. */
1249 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1250 serviceName, serviceId, service->serviceName);
1252 rxi_FreeService(tservice);
1255 /* Different service, same port: re-use the socket
1256 * which is bound to the same port */
1257 socket = service->socket;
1260 if (socket == OSI_NULLSOCKET) {
1261 /* If we don't already have a socket (from another
1262 * service on same port) get a new one */
1263 socket = rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
1264 if (socket == OSI_NULLSOCKET) {
1266 rxi_FreeService(tservice);
1271 service->socket = socket;
1272 service->servicePort = port;
1273 service->serviceId = serviceId;
1274 service->serviceName = serviceName;
1275 service->nSecurityObjects = nSecurityObjects;
1276 service->securityObjects = securityObjects;
1277 service->minProcs = 0;
1278 service->maxProcs = 1;
1279 service->idleDeadTime = 60;
1280 service->connDeadTime = rx_connDeadTime;
1281 service->executeRequestProc = serviceProc;
1282 service->checkReach = 0;
1283 rx_services[i] = service; /* not visible until now */
1289 rxi_FreeService(tservice);
1290 (osi_Msg "rx_NewService: cannot support > %d services\n",
1295 /* Generic request processing loop. This routine should be called
1296 * by the implementation dependent rx_ServerProc. If socketp is
1297 * non-null, it will be set to the file descriptor that this thread
1298 * is now listening on. If socketp is null, this routine will never
1301 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1303 register struct rx_call *call;
1304 register afs_int32 code;
1305 register struct rx_service *tservice = NULL;
1312 call = rx_GetCall(threadID, tservice, socketp);
1313 if (socketp && *socketp != OSI_NULLSOCKET) {
1314 /* We are now a listener thread */
1319 /* if server is restarting( typically smooth shutdown) then do not
1320 * allow any new calls.
1323 if (rx_tranquil && (call != NULL)) {
1327 MUTEX_ENTER(&call->lock);
1329 rxi_CallError(call, RX_RESTARTING);
1330 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1332 MUTEX_EXIT(&call->lock);
1336 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1337 #ifdef RX_ENABLE_LOCKS
1339 #endif /* RX_ENABLE_LOCKS */
1340 afs_termState = AFSOP_STOP_AFS;
1341 afs_osi_Wakeup(&afs_termState);
1342 #ifdef RX_ENABLE_LOCKS
1344 #endif /* RX_ENABLE_LOCKS */
1349 tservice = call->conn->service;
1351 if (tservice->beforeProc)
1352 (*tservice->beforeProc) (call);
1354 code = call->conn->service->executeRequestProc(call);
1356 if (tservice->afterProc)
1357 (*tservice->afterProc) (call, code);
1359 rx_EndCall(call, code);
1360 MUTEX_ENTER(&rx_stats_mutex);
1362 MUTEX_EXIT(&rx_stats_mutex);
1368 rx_WakeupServerProcs(void)
1370 struct rx_serverQueueEntry *np, *tqp;
1374 MUTEX_ENTER(&rx_serverPool_lock);
1376 #ifdef RX_ENABLE_LOCKS
1377 if (rx_waitForPacket)
1378 CV_BROADCAST(&rx_waitForPacket->cv);
1379 #else /* RX_ENABLE_LOCKS */
1380 if (rx_waitForPacket)
1381 osi_rxWakeup(rx_waitForPacket);
1382 #endif /* RX_ENABLE_LOCKS */
1383 MUTEX_ENTER(&freeSQEList_lock);
1384 for (np = rx_FreeSQEList; np; np = tqp) {
1385 tqp = *(struct rx_serverQueueEntry **)np;
1386 #ifdef RX_ENABLE_LOCKS
1387 CV_BROADCAST(&np->cv);
1388 #else /* RX_ENABLE_LOCKS */
1390 #endif /* RX_ENABLE_LOCKS */
1392 MUTEX_EXIT(&freeSQEList_lock);
1393 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1394 #ifdef RX_ENABLE_LOCKS
1395 CV_BROADCAST(&np->cv);
1396 #else /* RX_ENABLE_LOCKS */
1398 #endif /* RX_ENABLE_LOCKS */
1400 MUTEX_EXIT(&rx_serverPool_lock);
1405 * One thing that seems to happen is that all the server threads get
1406 * tied up on some empty or slow call, and then a whole bunch of calls
1407 * arrive at once, using up the packet pool, so now there are more
1408 * empty calls. The most critical resources here are server threads
1409 * and the free packet pool. The "doreclaim" code seems to help in
1410 * general. I think that eventually we arrive in this state: there
1411 * are lots of pending calls which do have all their packets present,
1412 * so they won't be reclaimed, are multi-packet calls, so they won't
1413 * be scheduled until later, and thus are tying up most of the free
1414 * packet pool for a very long time.
1416 * 1. schedule multi-packet calls if all the packets are present.
1417 * Probably CPU-bound operation, useful to return packets to pool.
1418 * Do what if there is a full window, but the last packet isn't here?
1419 * 3. preserve one thread which *only* runs "best" calls, otherwise
1420 * it sleeps and waits for that type of call.
1421 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1422 * the current dataquota business is badly broken. The quota isn't adjusted
1423 * to reflect how many packets are presently queued for a running call.
1424 * So, when we schedule a queued call with a full window of packets queued
1425 * up for it, that *should* free up a window full of packets for other 2d-class
1426 * calls to be able to use from the packet pool. But it doesn't.
1428 * NB. Most of the time, this code doesn't run -- since idle server threads
1429 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1430 * as a new call arrives.
1432 /* Sleep until a call arrives. Returns a pointer to the call, ready
1433 * for an rx_Read. */
1434 #ifdef RX_ENABLE_LOCKS
1436 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1438 struct rx_serverQueueEntry *sq;
1439 register struct rx_call *call = (struct rx_call *)0;
1440 struct rx_service *service = NULL;
1443 MUTEX_ENTER(&freeSQEList_lock);
1445 if ((sq = rx_FreeSQEList)) {
1446 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1447 MUTEX_EXIT(&freeSQEList_lock);
1448 } else { /* otherwise allocate a new one and return that */
1449 MUTEX_EXIT(&freeSQEList_lock);
1450 sq = (struct rx_serverQueueEntry *)
1451 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1452 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1453 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1456 MUTEX_ENTER(&rx_serverPool_lock);
1457 if (cur_service != NULL) {
1458 ReturnToServerPool(cur_service);
1461 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1462 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1464 /* Scan for eligible incoming calls. A call is not eligible
1465 * if the maximum number of calls for its service type are
1466 * already executing */
1467 /* One thread will process calls FCFS (to prevent starvation),
1468 * while the other threads may run ahead looking for calls which
1469 * have all their input data available immediately. This helps
1470 * keep threads from blocking, waiting for data from the client. */
1471 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1472 service = tcall->conn->service;
1473 if (!QuotaOK(service)) {
1476 if (tno == rxi_fcfs_thread_num
1477 || !tcall->queue_item_header.next) {
1478 /* If we're the fcfs thread , then we'll just use
1479 * this call. If we haven't been able to find an optimal
1480 * choice, and we're at the end of the list, then use a
1481 * 2d choice if one has been identified. Otherwise... */
1482 call = (choice2 ? choice2 : tcall);
1483 service = call->conn->service;
1484 } else if (!queue_IsEmpty(&tcall->rq)) {
1485 struct rx_packet *rp;
1486 rp = queue_First(&tcall->rq, rx_packet);
1487 if (rp->header.seq == 1) {
1489 || (rp->header.flags & RX_LAST_PACKET)) {
1491 } else if (rxi_2dchoice && !choice2
1492 && !(tcall->flags & RX_CALL_CLEARED)
1493 && (tcall->rprev > rxi_HardAckRate)) {
1502 ReturnToServerPool(service);
1509 MUTEX_EXIT(&rx_serverPool_lock);
1510 MUTEX_ENTER(&call->lock);
1512 if (call->flags & RX_CALL_WAIT_PROC) {
1513 call->flags &= ~RX_CALL_WAIT_PROC;
1514 MUTEX_ENTER(&rx_stats_mutex);
1516 MUTEX_EXIT(&rx_stats_mutex);
1519 if (call->state != RX_STATE_PRECALL || call->error) {
1520 MUTEX_EXIT(&call->lock);
1521 MUTEX_ENTER(&rx_serverPool_lock);
1522 ReturnToServerPool(service);
1527 if (queue_IsEmpty(&call->rq)
1528 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1529 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1531 CLEAR_CALL_QUEUE_LOCK(call);
1534 /* If there are no eligible incoming calls, add this process
1535 * to the idle server queue, to wait for one */
1539 *socketp = OSI_NULLSOCKET;
1541 sq->socketp = socketp;
1542 queue_Append(&rx_idleServerQueue, sq);
1543 #ifndef AFS_AIX41_ENV
1544 rx_waitForPacket = sq;
1546 rx_waitingForPacket = sq;
1547 #endif /* AFS_AIX41_ENV */
1549 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1551 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1552 MUTEX_EXIT(&rx_serverPool_lock);
1553 return (struct rx_call *)0;
1556 } while (!(call = sq->newcall)
1557 && !(socketp && *socketp != OSI_NULLSOCKET));
1558 MUTEX_EXIT(&rx_serverPool_lock);
1560 MUTEX_ENTER(&call->lock);
1566 MUTEX_ENTER(&freeSQEList_lock);
1567 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1568 rx_FreeSQEList = sq;
1569 MUTEX_EXIT(&freeSQEList_lock);
1572 clock_GetTime(&call->startTime);
1573 call->state = RX_STATE_ACTIVE;
1574 call->mode = RX_MODE_RECEIVING;
1575 #ifdef RX_KERNEL_TRACE
1576 if (ICL_SETACTIVE(afs_iclSetp)) {
1577 int glockOwner = ISAFS_GLOCK();
1580 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1581 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1588 rxi_calltrace(RX_CALL_START, call);
1589 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1590 call->conn->service->servicePort, call->conn->service->serviceId,
1593 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1594 MUTEX_EXIT(&call->lock);
1596 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1601 #else /* RX_ENABLE_LOCKS */
1603 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1605 struct rx_serverQueueEntry *sq;
1606 register struct rx_call *call = (struct rx_call *)0, *choice2;
1607 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) {
1706 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1707 return (struct rx_call *)0;
1710 } while (!(call = sq->newcall)
1711 && !(socketp && *socketp != OSI_NULLSOCKET));
1713 MUTEX_EXIT(&sq->lock);
1715 MUTEX_ENTER(&freeSQEList_lock);
1716 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1717 rx_FreeSQEList = sq;
1718 MUTEX_EXIT(&freeSQEList_lock);
1721 clock_GetTime(&call->startTime);
1722 call->state = RX_STATE_ACTIVE;
1723 call->mode = RX_MODE_RECEIVING;
1724 #ifdef RX_KERNEL_TRACE
1725 if (ICL_SETACTIVE(afs_iclSetp)) {
1726 int glockOwner = ISAFS_GLOCK();
1729 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1730 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1737 rxi_calltrace(RX_CALL_START, call);
1738 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1739 call->conn->service->servicePort, call->conn->service->serviceId,
1742 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1749 #endif /* RX_ENABLE_LOCKS */
1753 /* Establish a procedure to be called when a packet arrives for a
1754 * call. This routine will be called at most once after each call,
1755 * and will also be called if there is an error condition on the or
1756 * the call is complete. Used by multi rx to build a selection
1757 * function which determines which of several calls is likely to be a
1758 * good one to read from.
1759 * NOTE: the way this is currently implemented it is probably only a
1760 * good idea to (1) use it immediately after a newcall (clients only)
1761 * and (2) only use it once. Other uses currently void your warranty
1764 rx_SetArrivalProc(register struct rx_call *call,
1765 register void (*proc) (register struct rx_call * call,
1767 register int index),
1768 register VOID * handle, register int arg)
1770 call->arrivalProc = proc;
1771 call->arrivalProcHandle = handle;
1772 call->arrivalProcArg = arg;
1775 /* Call is finished (possibly prematurely). Return rc to the peer, if
1776 * appropriate, and return the final error code from the conversation
1780 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1782 register struct rx_connection *conn = call->conn;
1783 register struct rx_service *service;
1784 register struct rx_packet *tp; /* Temporary packet pointer */
1785 register struct rx_packet *nxp; /* Next packet pointer, for queue_Scan */
1789 dpf(("rx_EndCall(call %x)\n", call));
1792 MUTEX_ENTER(&call->lock);
1794 if (rc == 0 && call->error == 0) {
1795 call->abortCode = 0;
1796 call->abortCount = 0;
1799 call->arrivalProc = (void (*)())0;
1800 if (rc && call->error == 0) {
1801 rxi_CallError(call, rc);
1802 /* Send an abort message to the peer if this error code has
1803 * only just been set. If it was set previously, assume the
1804 * peer has already been sent the error code or will request it
1806 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1808 if (conn->type == RX_SERVER_CONNECTION) {
1809 /* Make sure reply or at least dummy reply is sent */
1810 if (call->mode == RX_MODE_RECEIVING) {
1811 rxi_WriteProc(call, 0, 0);
1813 if (call->mode == RX_MODE_SENDING) {
1814 rxi_FlushWrite(call);
1816 service = conn->service;
1817 rxi_calltrace(RX_CALL_END, call);
1818 /* Call goes to hold state until reply packets are acknowledged */
1819 if (call->tfirst + call->nSoftAcked < call->tnext) {
1820 call->state = RX_STATE_HOLD;
1822 call->state = RX_STATE_DALLY;
1823 rxi_ClearTransmitQueue(call, 0);
1824 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1825 rxevent_Cancel(call->keepAliveEvent, call,
1826 RX_CALL_REFCOUNT_ALIVE);
1828 } else { /* Client connection */
1830 /* Make sure server receives input packets, in the case where
1831 * no reply arguments are expected */
1832 if ((call->mode == RX_MODE_SENDING)
1833 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1834 (void)rxi_ReadProc(call, &dummy, 1);
1837 /* If we had an outstanding delayed ack, be nice to the server
1838 * and force-send it now.
1840 if (call->delayedAckEvent) {
1841 rxevent_Cancel(call->delayedAckEvent, call,
1842 RX_CALL_REFCOUNT_DELAY);
1843 call->delayedAckEvent = NULL;
1844 rxi_SendDelayedAck(NULL, call, NULL);
1847 /* We need to release the call lock since it's lower than the
1848 * conn_call_lock and we don't want to hold the conn_call_lock
1849 * over the rx_ReadProc call. The conn_call_lock needs to be held
1850 * here for the case where rx_NewCall is perusing the calls on
1851 * the connection structure. We don't want to signal until
1852 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1853 * have checked this call, found it active and by the time it
1854 * goes to sleep, will have missed the signal.
1856 MUTEX_EXIT(&call->lock);
1857 MUTEX_ENTER(&conn->conn_call_lock);
1858 MUTEX_ENTER(&call->lock);
1859 MUTEX_ENTER(&conn->conn_data_lock);
1860 conn->flags |= RX_CONN_BUSY;
1861 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1862 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1863 MUTEX_EXIT(&conn->conn_data_lock);
1864 #ifdef RX_ENABLE_LOCKS
1865 CV_BROADCAST(&conn->conn_call_cv);
1870 #ifdef RX_ENABLE_LOCKS
1872 MUTEX_EXIT(&conn->conn_data_lock);
1874 #endif /* RX_ENABLE_LOCKS */
1875 call->state = RX_STATE_DALLY;
1877 error = call->error;
1879 /* currentPacket, nLeft, and NFree must be zeroed here, because
1880 * ResetCall cannot: ResetCall may be called at splnet(), in the
1881 * kernel version, and may interrupt the macros rx_Read or
1882 * rx_Write, which run at normal priority for efficiency. */
1883 if (call->currentPacket) {
1884 rxi_FreePacket(call->currentPacket);
1885 call->currentPacket = (struct rx_packet *)0;
1886 call->nLeft = call->nFree = call->curlen = 0;
1888 call->nLeft = call->nFree = call->curlen = 0;
1890 /* Free any packets from the last call to ReadvProc/WritevProc */
1891 for (queue_Scan(&call->iovq, tp, nxp, rx_packet)) {
1896 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1897 MUTEX_EXIT(&call->lock);
1898 if (conn->type == RX_CLIENT_CONNECTION) {
1899 MUTEX_EXIT(&conn->conn_call_lock);
1900 conn->flags &= ~RX_CONN_BUSY;
1904 * Map errors to the local host's errno.h format.
1906 error = ntoh_syserr_conv(error);
1910 #if !defined(KERNEL)
1912 /* Call this routine when shutting down a server or client (especially
1913 * clients). This will allow Rx to gracefully garbage collect server
1914 * connections, and reduce the number of retries that a server might
1915 * make to a dead client.
1916 * This is not quite right, since some calls may still be ongoing and
1917 * we can't lock them to destroy them. */
1921 register struct rx_connection **conn_ptr, **conn_end;
1925 if (rxinit_status == 1) {
1927 return; /* Already shutdown. */
1929 rxi_DeleteCachedConnections();
1930 if (rx_connHashTable) {
1931 MUTEX_ENTER(&rx_connHashTable_lock);
1932 for (conn_ptr = &rx_connHashTable[0], conn_end =
1933 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
1935 struct rx_connection *conn, *next;
1936 for (conn = *conn_ptr; conn; conn = next) {
1938 if (conn->type == RX_CLIENT_CONNECTION) {
1939 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
1941 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
1942 #ifdef RX_ENABLE_LOCKS
1943 rxi_DestroyConnectionNoLock(conn);
1944 #else /* RX_ENABLE_LOCKS */
1945 rxi_DestroyConnection(conn);
1946 #endif /* RX_ENABLE_LOCKS */
1950 #ifdef RX_ENABLE_LOCKS
1951 while (rx_connCleanup_list) {
1952 struct rx_connection *conn;
1953 conn = rx_connCleanup_list;
1954 rx_connCleanup_list = rx_connCleanup_list->next;
1955 MUTEX_EXIT(&rx_connHashTable_lock);
1956 rxi_CleanupConnection(conn);
1957 MUTEX_ENTER(&rx_connHashTable_lock);
1959 MUTEX_EXIT(&rx_connHashTable_lock);
1960 #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 MUTEX_ENTER(&rx_stats_mutex);
2171 rxi_Allocsize += size;
2172 MUTEX_EXIT(&rx_stats_mutex);
2174 p = (char *)osi_Alloc(size);
2177 osi_Panic("rxi_Alloc error");
2183 rxi_Free(void *addr, register size_t size)
2185 MUTEX_ENTER(&rx_stats_mutex);
2187 rxi_Allocsize -= size;
2188 MUTEX_EXIT(&rx_stats_mutex);
2190 osi_Free(addr, size);
2193 /* Find the peer process represented by the supplied (host,port)
2194 * combination. If there is no appropriate active peer structure, a
2195 * new one will be allocated and initialized
2196 * The origPeer, if set, is a pointer to a peer structure on which the
2197 * refcount will be be decremented. This is used to replace the peer
2198 * structure hanging off a connection structure */
2200 rxi_FindPeer(register afs_uint32 host, register u_short port,
2201 struct rx_peer *origPeer, int create)
2203 register struct rx_peer *pp;
2205 hashIndex = PEER_HASH(host, port);
2206 MUTEX_ENTER(&rx_peerHashTable_lock);
2207 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2208 if ((pp->host == host) && (pp->port == port))
2213 pp = rxi_AllocPeer(); /* This bzero's *pp */
2214 pp->host = host; /* set here or in InitPeerParams is zero */
2216 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2217 queue_Init(&pp->congestionQueue);
2218 queue_Init(&pp->rpcStats);
2219 pp->next = rx_peerHashTable[hashIndex];
2220 rx_peerHashTable[hashIndex] = pp;
2221 rxi_InitPeerParams(pp);
2222 MUTEX_ENTER(&rx_stats_mutex);
2223 rx_stats.nPeerStructs++;
2224 MUTEX_EXIT(&rx_stats_mutex);
2231 origPeer->refCount--;
2232 MUTEX_EXIT(&rx_peerHashTable_lock);
2237 /* Find the connection at (host, port) started at epoch, and with the
2238 * given connection id. Creates the server connection if necessary.
2239 * The type specifies whether a client connection or a server
2240 * connection is desired. In both cases, (host, port) specify the
2241 * peer's (host, pair) pair. Client connections are not made
2242 * automatically by this routine. The parameter socket gives the
2243 * socket descriptor on which the packet was received. This is used,
2244 * in the case of server connections, to check that *new* connections
2245 * come via a valid (port, serviceId). Finally, the securityIndex
2246 * parameter must match the existing index for the connection. If a
2247 * server connection is created, it will be created using the supplied
2248 * index, if the index is valid for this service */
2249 struct rx_connection *
2250 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2251 register u_short port, u_short serviceId, afs_uint32 cid,
2252 afs_uint32 epoch, int type, u_int securityIndex)
2254 int hashindex, flag;
2255 register struct rx_connection *conn;
2256 hashindex = CONN_HASH(host, port, cid, epoch, type);
2257 MUTEX_ENTER(&rx_connHashTable_lock);
2258 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2259 rx_connHashTable[hashindex],
2262 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2263 && (epoch == conn->epoch)) {
2264 register struct rx_peer *pp = conn->peer;
2265 if (securityIndex != conn->securityIndex) {
2266 /* this isn't supposed to happen, but someone could forge a packet
2267 * like this, and there seems to be some CM bug that makes this
2268 * happen from time to time -- in which case, the fileserver
2270 MUTEX_EXIT(&rx_connHashTable_lock);
2271 return (struct rx_connection *)0;
2273 if (pp->host == host && pp->port == port)
2275 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2277 /* So what happens when it's a callback connection? */
2278 if ( /*type == RX_CLIENT_CONNECTION && */
2279 (conn->epoch & 0x80000000))
2283 /* the connection rxLastConn that was used the last time is not the
2284 ** one we are looking for now. Hence, start searching in the hash */
2286 conn = rx_connHashTable[hashindex];
2291 struct rx_service *service;
2292 if (type == RX_CLIENT_CONNECTION) {
2293 MUTEX_EXIT(&rx_connHashTable_lock);
2294 return (struct rx_connection *)0;
2296 service = rxi_FindService(socket, serviceId);
2297 if (!service || (securityIndex >= service->nSecurityObjects)
2298 || (service->securityObjects[securityIndex] == 0)) {
2299 MUTEX_EXIT(&rx_connHashTable_lock);
2300 return (struct rx_connection *)0;
2302 conn = rxi_AllocConnection(); /* This bzero's the connection */
2303 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2304 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2305 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2306 conn->next = rx_connHashTable[hashindex];
2307 rx_connHashTable[hashindex] = conn;
2308 conn->peer = rxi_FindPeer(host, port, 0, 1);
2309 conn->type = RX_SERVER_CONNECTION;
2310 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2311 conn->epoch = epoch;
2312 conn->cid = cid & RX_CIDMASK;
2313 /* conn->serial = conn->lastSerial = 0; */
2314 /* conn->timeout = 0; */
2315 conn->ackRate = RX_FAST_ACK_RATE;
2316 conn->service = service;
2317 conn->serviceId = serviceId;
2318 conn->securityIndex = securityIndex;
2319 conn->securityObject = service->securityObjects[securityIndex];
2320 conn->nSpecific = 0;
2321 conn->specific = NULL;
2322 rx_SetConnDeadTime(conn, service->connDeadTime);
2323 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2324 /* Notify security object of the new connection */
2325 RXS_NewConnection(conn->securityObject, conn);
2326 /* XXXX Connection timeout? */
2327 if (service->newConnProc)
2328 (*service->newConnProc) (conn);
2329 MUTEX_ENTER(&rx_stats_mutex);
2330 rx_stats.nServerConns++;
2331 MUTEX_EXIT(&rx_stats_mutex);
2334 MUTEX_ENTER(&conn->conn_data_lock);
2336 MUTEX_EXIT(&conn->conn_data_lock);
2338 rxLastConn = conn; /* store this connection as the last conn used */
2339 MUTEX_EXIT(&rx_connHashTable_lock);
2343 /* There are two packet tracing routines available for testing and monitoring
2344 * Rx. One is called just after every packet is received and the other is
2345 * called just before every packet is sent. Received packets, have had their
2346 * headers decoded, and packets to be sent have not yet had their headers
2347 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2348 * containing the network address. Both can be modified. The return value, if
2349 * non-zero, indicates that the packet should be dropped. */
2351 int (*rx_justReceived) () = 0;
2352 int (*rx_almostSent) () = 0;
2354 /* A packet has been received off the interface. Np is the packet, socket is
2355 * the socket number it was received from (useful in determining which service
2356 * this packet corresponds to), and (host, port) reflect the host,port of the
2357 * sender. This call returns the packet to the caller if it is finished with
2358 * it, rather than de-allocating it, just as a small performance hack */
2361 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2362 afs_uint32 host, u_short port, int *tnop,
2363 struct rx_call **newcallp)
2365 register struct rx_call *call;
2366 register struct rx_connection *conn;
2368 afs_uint32 currentCallNumber;
2374 struct rx_packet *tnp;
2377 /* We don't print out the packet until now because (1) the time may not be
2378 * accurate enough until now in the lwp implementation (rx_Listener only gets
2379 * the time after the packet is read) and (2) from a protocol point of view,
2380 * this is the first time the packet has been seen */
2381 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2382 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2383 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2384 np->header.serial, packetType, host, port, np->header.serviceId,
2385 np->header.epoch, np->header.cid, np->header.callNumber,
2386 np->header.seq, np->header.flags, np));
2389 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2390 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2393 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2394 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2397 /* If an input tracer function is defined, call it with the packet and
2398 * network address. Note this function may modify its arguments. */
2399 if (rx_justReceived) {
2400 struct sockaddr_in addr;
2402 addr.sin_family = AF_INET;
2403 addr.sin_port = port;
2404 addr.sin_addr.s_addr = host;
2405 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2406 addr.sin_len = sizeof(addr);
2407 #endif /* AFS_OSF_ENV */
2408 drop = (*rx_justReceived) (np, &addr);
2409 /* drop packet if return value is non-zero */
2412 port = addr.sin_port; /* in case fcn changed addr */
2413 host = addr.sin_addr.s_addr;
2417 /* If packet was not sent by the client, then *we* must be the client */
2418 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2419 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2421 /* Find the connection (or fabricate one, if we're the server & if
2422 * necessary) associated with this packet */
2424 rxi_FindConnection(socket, host, port, np->header.serviceId,
2425 np->header.cid, np->header.epoch, type,
2426 np->header.securityIndex);
2429 /* If no connection found or fabricated, just ignore the packet.
2430 * (An argument could be made for sending an abort packet for
2435 MUTEX_ENTER(&conn->conn_data_lock);
2436 if (conn->maxSerial < np->header.serial)
2437 conn->maxSerial = np->header.serial;
2438 MUTEX_EXIT(&conn->conn_data_lock);
2440 /* If the connection is in an error state, send an abort packet and ignore
2441 * the incoming packet */
2443 /* Don't respond to an abort packet--we don't want loops! */
2444 MUTEX_ENTER(&conn->conn_data_lock);
2445 if (np->header.type != RX_PACKET_TYPE_ABORT)
2446 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2448 MUTEX_EXIT(&conn->conn_data_lock);
2452 /* Check for connection-only requests (i.e. not call specific). */
2453 if (np->header.callNumber == 0) {
2454 switch (np->header.type) {
2455 case RX_PACKET_TYPE_ABORT:
2456 /* What if the supplied error is zero? */
2457 rxi_ConnectionError(conn, ntohl(rx_GetInt32(np, 0)));
2458 MUTEX_ENTER(&conn->conn_data_lock);
2460 MUTEX_EXIT(&conn->conn_data_lock);
2462 case RX_PACKET_TYPE_CHALLENGE:
2463 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2464 MUTEX_ENTER(&conn->conn_data_lock);
2466 MUTEX_EXIT(&conn->conn_data_lock);
2468 case RX_PACKET_TYPE_RESPONSE:
2469 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2470 MUTEX_ENTER(&conn->conn_data_lock);
2472 MUTEX_EXIT(&conn->conn_data_lock);
2474 case RX_PACKET_TYPE_PARAMS:
2475 case RX_PACKET_TYPE_PARAMS + 1:
2476 case RX_PACKET_TYPE_PARAMS + 2:
2477 /* ignore these packet types for now */
2478 MUTEX_ENTER(&conn->conn_data_lock);
2480 MUTEX_EXIT(&conn->conn_data_lock);
2485 /* Should not reach here, unless the peer is broken: send an
2487 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2488 MUTEX_ENTER(&conn->conn_data_lock);
2489 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2491 MUTEX_EXIT(&conn->conn_data_lock);
2496 channel = np->header.cid & RX_CHANNELMASK;
2497 call = conn->call[channel];
2498 #ifdef RX_ENABLE_LOCKS
2500 MUTEX_ENTER(&call->lock);
2501 /* Test to see if call struct is still attached to conn. */
2502 if (call != conn->call[channel]) {
2504 MUTEX_EXIT(&call->lock);
2505 if (type == RX_SERVER_CONNECTION) {
2506 call = conn->call[channel];
2507 /* If we started with no call attached and there is one now,
2508 * another thread is also running this routine and has gotten
2509 * the connection channel. We should drop this packet in the tests
2510 * below. If there was a call on this connection and it's now
2511 * gone, then we'll be making a new call below.
2512 * If there was previously a call and it's now different then
2513 * the old call was freed and another thread running this routine
2514 * has created a call on this channel. One of these two threads
2515 * has a packet for the old call and the code below handles those
2519 MUTEX_ENTER(&call->lock);
2521 /* This packet can't be for this call. If the new call address is
2522 * 0 then no call is running on this channel. If there is a call
2523 * then, since this is a client connection we're getting data for
2524 * it must be for the previous call.
2526 MUTEX_ENTER(&rx_stats_mutex);
2527 rx_stats.spuriousPacketsRead++;
2528 MUTEX_EXIT(&rx_stats_mutex);
2529 MUTEX_ENTER(&conn->conn_data_lock);
2531 MUTEX_EXIT(&conn->conn_data_lock);
2536 currentCallNumber = conn->callNumber[channel];
2538 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2539 if (np->header.callNumber < currentCallNumber) {
2540 MUTEX_ENTER(&rx_stats_mutex);
2541 rx_stats.spuriousPacketsRead++;
2542 MUTEX_EXIT(&rx_stats_mutex);
2543 #ifdef RX_ENABLE_LOCKS
2545 MUTEX_EXIT(&call->lock);
2547 MUTEX_ENTER(&conn->conn_data_lock);
2549 MUTEX_EXIT(&conn->conn_data_lock);
2553 MUTEX_ENTER(&conn->conn_call_lock);
2554 call = rxi_NewCall(conn, channel);
2555 MUTEX_EXIT(&conn->conn_call_lock);
2556 *call->callNumber = np->header.callNumber;
2557 call->state = RX_STATE_PRECALL;
2558 clock_GetTime(&call->queueTime);
2559 hzero(call->bytesSent);
2560 hzero(call->bytesRcvd);
2561 rxi_KeepAliveOn(call);
2562 } else if (np->header.callNumber != currentCallNumber) {
2563 /* Wait until the transmit queue is idle before deciding
2564 * whether to reset the current call. Chances are that the
2565 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2568 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2569 while ((call->state == RX_STATE_ACTIVE)
2570 && (call->flags & RX_CALL_TQ_BUSY)) {
2571 call->flags |= RX_CALL_TQ_WAIT;
2572 #ifdef RX_ENABLE_LOCKS
2573 CV_WAIT(&call->cv_tq, &call->lock);
2574 #else /* RX_ENABLE_LOCKS */
2575 osi_rxSleep(&call->tq);
2576 #endif /* RX_ENABLE_LOCKS */
2578 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2579 /* If the new call cannot be taken right now send a busy and set
2580 * the error condition in this call, so that it terminates as
2581 * quickly as possible */
2582 if (call->state == RX_STATE_ACTIVE) {
2583 struct rx_packet *tp;
2585 rxi_CallError(call, RX_CALL_DEAD);
2586 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2588 MUTEX_EXIT(&call->lock);
2589 MUTEX_ENTER(&conn->conn_data_lock);
2591 MUTEX_EXIT(&conn->conn_data_lock);
2594 rxi_ResetCall(call, 0);
2595 *call->callNumber = np->header.callNumber;
2596 call->state = RX_STATE_PRECALL;
2597 clock_GetTime(&call->queueTime);
2598 hzero(call->bytesSent);
2599 hzero(call->bytesRcvd);
2601 * If the number of queued calls exceeds the overload
2602 * threshold then abort this call.
2604 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2605 struct rx_packet *tp;
2607 rxi_CallError(call, rx_BusyError);
2608 tp = rxi_SendCallAbort(call, np, 1, 0);
2609 MUTEX_EXIT(&call->lock);
2610 MUTEX_ENTER(&conn->conn_data_lock);
2612 MUTEX_EXIT(&conn->conn_data_lock);
2613 MUTEX_ENTER(&rx_stats_mutex);
2615 MUTEX_EXIT(&rx_stats_mutex);
2618 rxi_KeepAliveOn(call);
2620 /* Continuing call; do nothing here. */
2622 } else { /* we're the client */
2623 /* Ignore all incoming acknowledgements for calls in DALLY state */
2624 if (call && (call->state == RX_STATE_DALLY)
2625 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2626 MUTEX_ENTER(&rx_stats_mutex);
2627 rx_stats.ignorePacketDally++;
2628 MUTEX_EXIT(&rx_stats_mutex);
2629 #ifdef RX_ENABLE_LOCKS
2631 MUTEX_EXIT(&call->lock);
2634 MUTEX_ENTER(&conn->conn_data_lock);
2636 MUTEX_EXIT(&conn->conn_data_lock);
2640 /* Ignore anything that's not relevant to the current call. If there
2641 * isn't a current call, then no packet is relevant. */
2642 if (!call || (np->header.callNumber != currentCallNumber)) {
2643 MUTEX_ENTER(&rx_stats_mutex);
2644 rx_stats.spuriousPacketsRead++;
2645 MUTEX_EXIT(&rx_stats_mutex);
2646 #ifdef RX_ENABLE_LOCKS
2648 MUTEX_EXIT(&call->lock);
2651 MUTEX_ENTER(&conn->conn_data_lock);
2653 MUTEX_EXIT(&conn->conn_data_lock);
2656 /* If the service security object index stamped in the packet does not
2657 * match the connection's security index, ignore the packet */
2658 if (np->header.securityIndex != conn->securityIndex) {
2659 #ifdef RX_ENABLE_LOCKS
2660 MUTEX_EXIT(&call->lock);
2662 MUTEX_ENTER(&conn->conn_data_lock);
2664 MUTEX_EXIT(&conn->conn_data_lock);
2668 /* If we're receiving the response, then all transmit packets are
2669 * implicitly acknowledged. Get rid of them. */
2670 if (np->header.type == RX_PACKET_TYPE_DATA) {
2671 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2672 /* XXX Hack. Because we must release the global rx lock when
2673 * sending packets (osi_NetSend) we drop all acks while we're
2674 * traversing the tq in rxi_Start sending packets out because
2675 * packets may move to the freePacketQueue as result of being here!
2676 * So we drop these packets until we're safely out of the
2677 * traversing. Really ugly!
2678 * For fine grain RX locking, we set the acked field in the
2679 * packets and let rxi_Start remove them from the transmit queue.
2681 if (call->flags & RX_CALL_TQ_BUSY) {
2682 #ifdef RX_ENABLE_LOCKS
2683 rxi_SetAcksInTransmitQueue(call);
2686 return np; /* xmitting; drop packet */
2689 rxi_ClearTransmitQueue(call, 0);
2691 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2692 rxi_ClearTransmitQueue(call, 0);
2693 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2695 if (np->header.type == RX_PACKET_TYPE_ACK) {
2696 /* now check to see if this is an ack packet acknowledging that the
2697 * server actually *lost* some hard-acked data. If this happens we
2698 * ignore this packet, as it may indicate that the server restarted in
2699 * the middle of a call. It is also possible that this is an old ack
2700 * packet. We don't abort the connection in this case, because this
2701 * *might* just be an old ack packet. The right way to detect a server
2702 * restart in the midst of a call is to notice that the server epoch
2704 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2705 * XXX unacknowledged. I think that this is off-by-one, but
2706 * XXX I don't dare change it just yet, since it will
2707 * XXX interact badly with the server-restart detection
2708 * XXX code in receiveackpacket. */
2709 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2710 MUTEX_ENTER(&rx_stats_mutex);
2711 rx_stats.spuriousPacketsRead++;
2712 MUTEX_EXIT(&rx_stats_mutex);
2713 MUTEX_EXIT(&call->lock);
2714 MUTEX_ENTER(&conn->conn_data_lock);
2716 MUTEX_EXIT(&conn->conn_data_lock);
2720 } /* else not a data packet */
2723 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2724 /* Set remote user defined status from packet */
2725 call->remoteStatus = np->header.userStatus;
2727 /* Note the gap between the expected next packet and the actual
2728 * packet that arrived, when the new packet has a smaller serial number
2729 * than expected. Rioses frequently reorder packets all by themselves,
2730 * so this will be quite important with very large window sizes.
2731 * Skew is checked against 0 here to avoid any dependence on the type of
2732 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2734 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2735 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2736 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2738 MUTEX_ENTER(&conn->conn_data_lock);
2739 skew = conn->lastSerial - np->header.serial;
2740 conn->lastSerial = np->header.serial;
2741 MUTEX_EXIT(&conn->conn_data_lock);
2743 register struct rx_peer *peer;
2745 if (skew > peer->inPacketSkew) {
2746 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2748 peer->inPacketSkew = skew;
2752 /* Now do packet type-specific processing */
2753 switch (np->header.type) {
2754 case RX_PACKET_TYPE_DATA:
2755 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2758 case RX_PACKET_TYPE_ACK:
2759 /* Respond immediately to ack packets requesting acknowledgement
2761 if (np->header.flags & RX_REQUEST_ACK) {
2763 (void)rxi_SendCallAbort(call, 0, 1, 0);
2765 (void)rxi_SendAck(call, 0, np->header.serial,
2766 RX_ACK_PING_RESPONSE, 1);
2768 np = rxi_ReceiveAckPacket(call, np, 1);
2770 case RX_PACKET_TYPE_ABORT:
2771 /* An abort packet: reset the connection, passing the error up to
2773 /* What if error is zero? */
2774 rxi_CallError(call, ntohl(*(afs_int32 *) rx_DataOf(np)));
2776 case RX_PACKET_TYPE_BUSY:
2779 case RX_PACKET_TYPE_ACKALL:
2780 /* All packets acknowledged, so we can drop all packets previously
2781 * readied for sending */
2782 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2783 /* XXX Hack. We because we can't release the global rx lock when
2784 * sending packets (osi_NetSend) we drop all ack pkts while we're
2785 * traversing the tq in rxi_Start sending packets out because
2786 * packets may move to the freePacketQueue as result of being
2787 * here! So we drop these packets until we're safely out of the
2788 * traversing. Really ugly!
2789 * For fine grain RX locking, we set the acked field in the packets
2790 * and let rxi_Start remove the packets from the transmit queue.
2792 if (call->flags & RX_CALL_TQ_BUSY) {
2793 #ifdef RX_ENABLE_LOCKS
2794 rxi_SetAcksInTransmitQueue(call);
2796 #else /* RX_ENABLE_LOCKS */
2798 return np; /* xmitting; drop packet */
2799 #endif /* RX_ENABLE_LOCKS */
2801 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2802 rxi_ClearTransmitQueue(call, 0);
2805 /* Should not reach here, unless the peer is broken: send an abort
2807 rxi_CallError(call, RX_PROTOCOL_ERROR);
2808 np = rxi_SendCallAbort(call, np, 1, 0);
2811 /* Note when this last legitimate packet was received, for keep-alive
2812 * processing. Note, we delay getting the time until now in the hope that
2813 * the packet will be delivered to the user before any get time is required
2814 * (if not, then the time won't actually be re-evaluated here). */
2815 call->lastReceiveTime = clock_Sec();
2816 MUTEX_EXIT(&call->lock);
2817 MUTEX_ENTER(&conn->conn_data_lock);
2819 MUTEX_EXIT(&conn->conn_data_lock);
2823 /* return true if this is an "interesting" connection from the point of view
2824 of someone trying to debug the system */
2826 rxi_IsConnInteresting(struct rx_connection *aconn)
2829 register struct rx_call *tcall;
2831 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2833 for (i = 0; i < RX_MAXCALLS; i++) {
2834 tcall = aconn->call[i];
2836 if ((tcall->state == RX_STATE_PRECALL)
2837 || (tcall->state == RX_STATE_ACTIVE))
2839 if ((tcall->mode == RX_MODE_SENDING)
2840 || (tcall->mode == RX_MODE_RECEIVING))
2848 /* if this is one of the last few packets AND it wouldn't be used by the
2849 receiving call to immediately satisfy a read request, then drop it on
2850 the floor, since accepting it might prevent a lock-holding thread from
2851 making progress in its reading. If a call has been cleared while in
2852 the precall state then ignore all subsequent packets until the call
2853 is assigned to a thread. */
2856 TooLow(struct rx_packet *ap, struct rx_call *acall)
2859 MUTEX_ENTER(&rx_stats_mutex);
2860 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
2861 && (acall->state == RX_STATE_PRECALL))
2862 || ((rx_nFreePackets < rxi_dataQuota + 2)
2863 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
2864 && (acall->flags & RX_CALL_READER_WAIT)))) {
2867 MUTEX_EXIT(&rx_stats_mutex);
2873 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
2874 struct rx_call *acall)
2876 struct rx_call *call = acall;
2880 MUTEX_ENTER(&conn->conn_data_lock);
2881 conn->checkReachEvent = NULL;
2882 waiting = conn->flags & RX_CONN_ATTACHWAIT;
2885 MUTEX_EXIT(&conn->conn_data_lock);
2889 MUTEX_ENTER(&conn->conn_call_lock);
2890 MUTEX_ENTER(&conn->conn_data_lock);
2891 for (i = 0; i < RX_MAXCALLS; i++) {
2892 struct rx_call *tc = conn->call[i];
2893 if (tc && tc->state == RX_STATE_PRECALL) {
2899 /* Indicate that rxi_CheckReachEvent is no longer running by
2900 * clearing the flag. Must be atomic under conn_data_lock to
2901 * avoid a new call slipping by: rxi_CheckConnReach holds
2902 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
2904 conn->flags &= ~RX_CONN_ATTACHWAIT;
2905 MUTEX_EXIT(&conn->conn_data_lock);
2906 MUTEX_EXIT(&conn->conn_call_lock);
2911 MUTEX_ENTER(&call->lock);
2912 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
2914 MUTEX_EXIT(&call->lock);
2916 clock_GetTime(&when);
2917 when.sec += RX_CHECKREACH_TIMEOUT;
2918 MUTEX_ENTER(&conn->conn_data_lock);
2919 if (!conn->checkReachEvent) {
2921 conn->checkReachEvent =
2922 rxevent_Post(&when, rxi_CheckReachEvent, conn, NULL);
2924 MUTEX_EXIT(&conn->conn_data_lock);
2930 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
2932 struct rx_service *service = conn->service;
2933 struct rx_peer *peer = conn->peer;
2934 afs_uint32 now, lastReach;
2936 if (service->checkReach == 0)
2940 MUTEX_ENTER(&peer->peer_lock);
2941 lastReach = peer->lastReachTime;
2942 MUTEX_EXIT(&peer->peer_lock);
2943 if (now - lastReach < RX_CHECKREACH_TTL)
2946 MUTEX_ENTER(&conn->conn_data_lock);
2947 if (conn->flags & RX_CONN_ATTACHWAIT) {
2948 MUTEX_EXIT(&conn->conn_data_lock);
2951 conn->flags |= RX_CONN_ATTACHWAIT;
2952 MUTEX_EXIT(&conn->conn_data_lock);
2953 if (!conn->checkReachEvent)
2954 rxi_CheckReachEvent(NULL, conn, call);
2959 /* try to attach call, if authentication is complete */
2961 TryAttach(register struct rx_call *acall, register osi_socket socket,
2962 register int *tnop, register struct rx_call **newcallp,
2965 struct rx_connection *conn = acall->conn;
2967 if (conn->type == RX_SERVER_CONNECTION
2968 && acall->state == RX_STATE_PRECALL) {
2969 /* Don't attach until we have any req'd. authentication. */
2970 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
2971 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
2972 rxi_AttachServerProc(acall, socket, tnop, newcallp);
2973 /* Note: this does not necessarily succeed; there
2974 * may not any proc available
2977 rxi_ChallengeOn(acall->conn);
2982 /* A data packet has been received off the interface. This packet is
2983 * appropriate to the call (the call is in the right state, etc.). This
2984 * routine can return a packet to the caller, for re-use */
2987 rxi_ReceiveDataPacket(register struct rx_call *call,
2988 register struct rx_packet *np, int istack,
2989 osi_socket socket, afs_uint32 host, u_short port,
2990 int *tnop, struct rx_call **newcallp)
2992 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
2996 afs_uint32 seq, serial, flags;
2998 struct rx_packet *tnp;
3000 MUTEX_ENTER(&rx_stats_mutex);
3001 rx_stats.dataPacketsRead++;
3002 MUTEX_EXIT(&rx_stats_mutex);
3005 /* If there are no packet buffers, drop this new packet, unless we can find
3006 * packet buffers from inactive calls */
3008 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3009 MUTEX_ENTER(&rx_freePktQ_lock);
3010 rxi_NeedMorePackets = TRUE;
3011 MUTEX_EXIT(&rx_freePktQ_lock);
3012 MUTEX_ENTER(&rx_stats_mutex);
3013 rx_stats.noPacketBuffersOnRead++;
3014 MUTEX_EXIT(&rx_stats_mutex);
3015 call->rprev = np->header.serial;
3016 rxi_calltrace(RX_TRACE_DROP, call);
3017 dpf(("packet %x dropped on receipt - quota problems", np));
3019 rxi_ClearReceiveQueue(call);
3020 clock_GetTime(&when);
3021 clock_Add(&when, &rx_softAckDelay);
3022 if (!call->delayedAckEvent
3023 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3024 rxevent_Cancel(call->delayedAckEvent, call,
3025 RX_CALL_REFCOUNT_DELAY);
3026 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3027 call->delayedAckEvent =
3028 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3030 /* we've damaged this call already, might as well do it in. */
3036 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3037 * packet is one of several packets transmitted as a single
3038 * datagram. Do not send any soft or hard acks until all packets
3039 * in a jumbogram have been processed. Send negative acks right away.
3041 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3042 /* tnp is non-null when there are more packets in the
3043 * current jumbo gram */
3050 seq = np->header.seq;
3051 serial = np->header.serial;
3052 flags = np->header.flags;
3054 /* If the call is in an error state, send an abort message */
3056 return rxi_SendCallAbort(call, np, istack, 0);
3058 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3059 * AFS 3.5 jumbogram. */
3060 if (flags & RX_JUMBO_PACKET) {
3061 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3066 if (np->header.spare != 0) {
3067 MUTEX_ENTER(&call->conn->conn_data_lock);
3068 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3069 MUTEX_EXIT(&call->conn->conn_data_lock);
3072 /* The usual case is that this is the expected next packet */
3073 if (seq == call->rnext) {
3075 /* Check to make sure it is not a duplicate of one already queued */
3076 if (queue_IsNotEmpty(&call->rq)
3077 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3078 MUTEX_ENTER(&rx_stats_mutex);
3079 rx_stats.dupPacketsRead++;
3080 MUTEX_EXIT(&rx_stats_mutex);
3081 dpf(("packet %x dropped on receipt - duplicate", np));
3082 rxevent_Cancel(call->delayedAckEvent, call,
3083 RX_CALL_REFCOUNT_DELAY);
3084 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3090 /* It's the next packet. Stick it on the receive queue
3091 * for this call. Set newPackets to make sure we wake
3092 * the reader once all packets have been processed */
3093 queue_Prepend(&call->rq, np);
3095 np = NULL; /* We can't use this anymore */
3098 /* If an ack is requested then set a flag to make sure we
3099 * send an acknowledgement for this packet */
3100 if (flags & RX_REQUEST_ACK) {
3101 ackNeeded = RX_ACK_REQUESTED;
3104 /* Keep track of whether we have received the last packet */
3105 if (flags & RX_LAST_PACKET) {
3106 call->flags |= RX_CALL_HAVE_LAST;
3110 /* Check whether we have all of the packets for this call */
3111 if (call->flags & RX_CALL_HAVE_LAST) {
3112 afs_uint32 tseq; /* temporary sequence number */
3113 struct rx_packet *tp; /* Temporary packet pointer */
3114 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3116 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3117 if (tseq != tp->header.seq)
3119 if (tp->header.flags & RX_LAST_PACKET) {
3120 call->flags |= RX_CALL_RECEIVE_DONE;
3127 /* Provide asynchronous notification for those who want it
3128 * (e.g. multi rx) */
3129 if (call->arrivalProc) {
3130 (*call->arrivalProc) (call, call->arrivalProcHandle,
3131 call->arrivalProcArg);
3132 call->arrivalProc = (void (*)())0;
3135 /* Update last packet received */
3138 /* If there is no server process serving this call, grab
3139 * one, if available. We only need to do this once. If a
3140 * server thread is available, this thread becomes a server
3141 * thread and the server thread becomes a listener thread. */
3143 TryAttach(call, socket, tnop, newcallp, 0);
3146 /* This is not the expected next packet. */
3148 /* Determine whether this is a new or old packet, and if it's
3149 * a new one, whether it fits into the current receive window.
3150 * Also figure out whether the packet was delivered in sequence.
3151 * We use the prev variable to determine whether the new packet
3152 * is the successor of its immediate predecessor in the
3153 * receive queue, and the missing flag to determine whether
3154 * any of this packets predecessors are missing. */
3156 afs_uint32 prev; /* "Previous packet" sequence number */
3157 struct rx_packet *tp; /* Temporary packet pointer */
3158 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3159 int missing; /* Are any predecessors missing? */
3161 /* If the new packet's sequence number has been sent to the
3162 * application already, then this is a duplicate */
3163 if (seq < call->rnext) {
3164 MUTEX_ENTER(&rx_stats_mutex);
3165 rx_stats.dupPacketsRead++;
3166 MUTEX_EXIT(&rx_stats_mutex);
3167 rxevent_Cancel(call->delayedAckEvent, call,
3168 RX_CALL_REFCOUNT_DELAY);
3169 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3175 /* If the sequence number is greater than what can be
3176 * accomodated by the current window, then send a negative
3177 * acknowledge and drop the packet */
3178 if ((call->rnext + call->rwind) <= seq) {
3179 rxevent_Cancel(call->delayedAckEvent, call,
3180 RX_CALL_REFCOUNT_DELAY);
3181 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3188 /* Look for the packet in the queue of old received packets */
3189 for (prev = call->rnext - 1, missing =
3190 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3191 /*Check for duplicate packet */
3192 if (seq == tp->header.seq) {
3193 MUTEX_ENTER(&rx_stats_mutex);
3194 rx_stats.dupPacketsRead++;
3195 MUTEX_EXIT(&rx_stats_mutex);
3196 rxevent_Cancel(call->delayedAckEvent, call,
3197 RX_CALL_REFCOUNT_DELAY);
3198 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3204 /* If we find a higher sequence packet, break out and
3205 * insert the new packet here. */
3206 if (seq < tp->header.seq)
3208 /* Check for missing packet */
3209 if (tp->header.seq != prev + 1) {
3213 prev = tp->header.seq;
3216 /* Keep track of whether we have received the last packet. */
3217 if (flags & RX_LAST_PACKET) {
3218 call->flags |= RX_CALL_HAVE_LAST;
3221 /* It's within the window: add it to the the receive queue.
3222 * tp is left by the previous loop either pointing at the
3223 * packet before which to insert the new packet, or at the
3224 * queue head if the queue is empty or the packet should be
3226 queue_InsertBefore(tp, np);
3230 /* Check whether we have all of the packets for this call */
3231 if ((call->flags & RX_CALL_HAVE_LAST)
3232 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3233 afs_uint32 tseq; /* temporary sequence number */
3236 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3237 if (tseq != tp->header.seq)
3239 if (tp->header.flags & RX_LAST_PACKET) {
3240 call->flags |= RX_CALL_RECEIVE_DONE;
3247 /* We need to send an ack of the packet is out of sequence,
3248 * or if an ack was requested by the peer. */
3249 if (seq != prev + 1 || missing || (flags & RX_REQUEST_ACK)) {
3250 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3253 /* Acknowledge the last packet for each call */
3254 if (flags & RX_LAST_PACKET) {
3265 * If the receiver is waiting for an iovec, fill the iovec
3266 * using the data from the receive queue */
3267 if (call->flags & RX_CALL_IOVEC_WAIT) {
3268 didHardAck = rxi_FillReadVec(call, serial);
3269 /* the call may have been aborted */
3278 /* Wakeup the reader if any */
3279 if ((call->flags & RX_CALL_READER_WAIT)
3280 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3281 || (call->iovNext >= call->iovMax)
3282 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3283 call->flags &= ~RX_CALL_READER_WAIT;
3284 #ifdef RX_ENABLE_LOCKS
3285 CV_BROADCAST(&call->cv_rq);
3287 osi_rxWakeup(&call->rq);
3293 * Send an ack when requested by the peer, or once every
3294 * rxi_SoftAckRate packets until the last packet has been
3295 * received. Always send a soft ack for the last packet in
3296 * the server's reply. */
3298 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3299 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3300 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3301 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3302 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3303 } else if (call->nSoftAcks) {
3304 clock_GetTime(&when);
3305 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3306 clock_Add(&when, &rx_lastAckDelay);
3308 clock_Add(&when, &rx_softAckDelay);
3310 if (!call->delayedAckEvent
3311 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3312 rxevent_Cancel(call->delayedAckEvent, call,
3313 RX_CALL_REFCOUNT_DELAY);
3314 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3315 call->delayedAckEvent =
3316 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3318 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3319 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3326 static void rxi_ComputeRate();
3330 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3332 struct rx_peer *peer = conn->peer;
3334 MUTEX_ENTER(&peer->peer_lock);
3335 peer->lastReachTime = clock_Sec();
3336 MUTEX_EXIT(&peer->peer_lock);
3338 MUTEX_ENTER(&conn->conn_data_lock);
3339 if (conn->flags & RX_CONN_ATTACHWAIT) {
3342 conn->flags &= ~RX_CONN_ATTACHWAIT;
3343 MUTEX_EXIT(&conn->conn_data_lock);
3345 for (i = 0; i < RX_MAXCALLS; i++) {
3346 struct rx_call *call = conn->call[i];
3349 MUTEX_ENTER(&call->lock);
3350 /* tnop can be null if newcallp is null */
3351 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3353 MUTEX_EXIT(&call->lock);
3357 MUTEX_EXIT(&conn->conn_data_lock);
3360 /* rxi_ComputePeerNetStats
3362 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3363 * estimates (like RTT and throughput) based on ack packets. Caller
3364 * must ensure that the packet in question is the right one (i.e.
3365 * serial number matches).
3368 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3369 struct rx_ackPacket *ap, struct rx_packet *np)
3371 struct rx_peer *peer = call->conn->peer;
3373 /* Use RTT if not delayed by client. */
3374 if (ap->reason != RX_ACK_DELAY)
3375 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3377 rxi_ComputeRate(peer, call, p, np, ap->reason);
3381 /* The real smarts of the whole thing. */
3383 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3386 struct rx_ackPacket *ap;
3388 register struct rx_packet *tp;
3389 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3390 register struct rx_connection *conn = call->conn;
3391 struct rx_peer *peer = conn->peer;
3394 /* because there are CM's that are bogus, sending weird values for this. */
3395 afs_uint32 skew = 0;
3400 int newAckCount = 0;
3401 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3402 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3404 MUTEX_ENTER(&rx_stats_mutex);
3405 rx_stats.ackPacketsRead++;
3406 MUTEX_EXIT(&rx_stats_mutex);
3407 ap = (struct rx_ackPacket *)rx_DataOf(np);
3408 nbytes = rx_Contiguous(np) - ((ap->acks) - (u_char *) ap);
3410 return np; /* truncated ack packet */
3412 /* depends on ack packet struct */
3413 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3414 first = ntohl(ap->firstPacket);
3415 serial = ntohl(ap->serial);
3416 /* temporarily disabled -- needs to degrade over time
3417 * skew = ntohs(ap->maxSkew); */
3419 /* Ignore ack packets received out of order */
3420 if (first < call->tfirst) {
3424 if (np->header.flags & RX_SLOW_START_OK) {
3425 call->flags |= RX_CALL_SLOW_START_OK;
3428 if (ap->reason == RX_ACK_PING_RESPONSE)
3429 rxi_UpdatePeerReach(conn, call);
3434 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3435 ap->reason, ntohl(ap->previousPacket),
3436 (unsigned int)np->header.seq, (unsigned int)serial,
3437 (unsigned int)skew, ntohl(ap->firstPacket));
3440 for (offset = 0; offset < nAcks; offset++)
3441 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3448 /* Update the outgoing packet skew value to the latest value of
3449 * the peer's incoming packet skew value. The ack packet, of
3450 * course, could arrive out of order, but that won't affect things
3452 MUTEX_ENTER(&peer->peer_lock);
3453 peer->outPacketSkew = skew;
3455 /* Check for packets that no longer need to be transmitted, and
3456 * discard them. This only applies to packets positively
3457 * acknowledged as having been sent to the peer's upper level.
3458 * All other packets must be retained. So only packets with
3459 * sequence numbers < ap->firstPacket are candidates. */
3460 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3461 if (tp->header.seq >= first)
3463 call->tfirst = tp->header.seq + 1;
3465 && (tp->header.serial == serial || tp->firstSerial == serial))
3466 rxi_ComputePeerNetStats(call, tp, ap, np);
3467 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3468 /* XXX Hack. Because we have to release the global rx lock when sending
3469 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3470 * in rxi_Start sending packets out because packets may move to the
3471 * freePacketQueue as result of being here! So we drop these packets until
3472 * we're safely out of the traversing. Really ugly!
3473 * To make it even uglier, if we're using fine grain locking, we can
3474 * set the ack bits in the packets and have rxi_Start remove the packets
3475 * when it's done transmitting.
3477 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3480 if (call->flags & RX_CALL_TQ_BUSY) {
3481 #ifdef RX_ENABLE_LOCKS
3482 tp->flags |= RX_PKTFLAG_ACKED;
3483 call->flags |= RX_CALL_TQ_SOME_ACKED;
3484 #else /* RX_ENABLE_LOCKS */
3486 #endif /* RX_ENABLE_LOCKS */
3488 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3491 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3496 /* Give rate detector a chance to respond to ping requests */
3497 if (ap->reason == RX_ACK_PING_RESPONSE) {
3498 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3502 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3504 /* Now go through explicit acks/nacks and record the results in
3505 * the waiting packets. These are packets that can't be released
3506 * yet, even with a positive acknowledge. This positive
3507 * acknowledge only means the packet has been received by the
3508 * peer, not that it will be retained long enough to be sent to
3509 * the peer's upper level. In addition, reset the transmit timers
3510 * of any missing packets (those packets that must be missing
3511 * because this packet was out of sequence) */
3513 call->nSoftAcked = 0;
3514 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3515 /* Update round trip time if the ack was stimulated on receipt
3517 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3518 #ifdef RX_ENABLE_LOCKS
3519 if (tp->header.seq >= first)
3520 #endif /* RX_ENABLE_LOCKS */
3521 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3523 && (tp->header.serial == serial || tp->firstSerial == serial))
3524 rxi_ComputePeerNetStats(call, tp, ap, np);
3526 /* Set the acknowledge flag per packet based on the
3527 * information in the ack packet. An acknowlegded packet can
3528 * be downgraded when the server has discarded a packet it
3529 * soacked previously, or when an ack packet is received
3530 * out of sequence. */
3531 if (tp->header.seq < first) {
3532 /* Implicit ack information */
3533 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3536 tp->flags |= RX_PKTFLAG_ACKED;
3537 } else if (tp->header.seq < first + nAcks) {
3538 /* Explicit ack information: set it in the packet appropriately */
3539 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3540 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3542 tp->flags |= RX_PKTFLAG_ACKED;
3550 tp->flags &= ~RX_PKTFLAG_ACKED;
3554 tp->flags &= ~RX_PKTFLAG_ACKED;
3558 /* If packet isn't yet acked, and it has been transmitted at least
3559 * once, reset retransmit time using latest timeout
3560 * ie, this should readjust the retransmit timer for all outstanding
3561 * packets... So we don't just retransmit when we should know better*/
3563 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3564 tp->retryTime = tp->timeSent;
3565 clock_Add(&tp->retryTime, &peer->timeout);
3566 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3567 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3571 /* If the window has been extended by this acknowledge packet,
3572 * then wakeup a sender waiting in alloc for window space, or try
3573 * sending packets now, if he's been sitting on packets due to
3574 * lack of window space */
3575 if (call->tnext < (call->tfirst + call->twind)) {
3576 #ifdef RX_ENABLE_LOCKS
3577 CV_SIGNAL(&call->cv_twind);
3579 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3580 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3581 osi_rxWakeup(&call->twind);
3584 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3585 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3589 /* if the ack packet has a receivelen field hanging off it,
3590 * update our state */
3591 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3594 /* If the ack packet has a "recommended" size that is less than
3595 * what I am using now, reduce my size to match */
3596 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3597 sizeof(afs_int32), &tSize);
3598 tSize = (afs_uint32) ntohl(tSize);
3599 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3601 /* Get the maximum packet size to send to this peer */
3602 rx_packetread(np, rx_AckDataSize(ap->nAcks), sizeof(afs_int32),
3604 tSize = (afs_uint32) ntohl(tSize);
3605 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3606 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3608 /* sanity check - peer might have restarted with different params.
3609 * If peer says "send less", dammit, send less... Peer should never
3610 * be unable to accept packets of the size that prior AFS versions would
3611 * send without asking. */
3612 if (peer->maxMTU != tSize) {
3613 peer->maxMTU = tSize;
3614 peer->MTU = MIN(tSize, peer->MTU);
3615 call->MTU = MIN(call->MTU, tSize);
3619 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3622 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3623 sizeof(afs_int32), &tSize);
3624 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3625 if (tSize < call->twind) { /* smaller than our send */
3626 call->twind = tSize; /* window, we must send less... */
3627 call->ssthresh = MIN(call->twind, call->ssthresh);
3630 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3631 * network MTU confused with the loopback MTU. Calculate the
3632 * maximum MTU here for use in the slow start code below.
3634 maxMTU = peer->maxMTU;
3635 /* Did peer restart with older RX version? */
3636 if (peer->maxDgramPackets > 1) {
3637 peer->maxDgramPackets = 1;
3639 } else if (np->length >=
3640 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3643 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3644 sizeof(afs_int32), &tSize);
3645 tSize = (afs_uint32) ntohl(tSize);
3647 * As of AFS 3.5 we set the send window to match the receive window.
3649 if (tSize < call->twind) {
3650 call->twind = tSize;
3651 call->ssthresh = MIN(call->twind, call->ssthresh);
3652 } else if (tSize > call->twind) {
3653 call->twind = tSize;
3657 * As of AFS 3.5, a jumbogram is more than one fixed size
3658 * packet transmitted in a single UDP datagram. If the remote
3659 * MTU is smaller than our local MTU then never send a datagram
3660 * larger than the natural MTU.
3663 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3664 sizeof(afs_int32), &tSize);
3665 maxDgramPackets = (afs_uint32) ntohl(tSize);
3666 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3668 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
3669 maxDgramPackets = MIN(maxDgramPackets, tSize);
3670 if (maxDgramPackets > 1) {
3671 peer->maxDgramPackets = maxDgramPackets;
3672 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3674 peer->maxDgramPackets = 1;
3675 call->MTU = peer->natMTU;
3677 } else if (peer->maxDgramPackets > 1) {
3678 /* Restarted with lower version of RX */
3679 peer->maxDgramPackets = 1;
3681 } else if (peer->maxDgramPackets > 1
3682 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3683 /* Restarted with lower version of RX */
3684 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3685 peer->natMTU = OLD_MAX_PACKET_SIZE;
3686 peer->MTU = OLD_MAX_PACKET_SIZE;
3687 peer->maxDgramPackets = 1;
3688 peer->nDgramPackets = 1;
3690 call->MTU = OLD_MAX_PACKET_SIZE;
3695 * Calculate how many datagrams were successfully received after
3696 * the first missing packet and adjust the negative ack counter
3701 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3702 if (call->nNacks < nNacked) {
3703 call->nNacks = nNacked;
3712 if (call->flags & RX_CALL_FAST_RECOVER) {
3714 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3716 call->flags &= ~RX_CALL_FAST_RECOVER;
3717 call->cwind = call->nextCwind;
3718 call->nextCwind = 0;
3721 call->nCwindAcks = 0;
3722 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3723 /* Three negative acks in a row trigger congestion recovery */
3724 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3725 MUTEX_EXIT(&peer->peer_lock);
3726 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3727 /* someone else is waiting to start recovery */
3730 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3731 while (call->flags & RX_CALL_TQ_BUSY) {
3732 call->flags |= RX_CALL_TQ_WAIT;
3733 #ifdef RX_ENABLE_LOCKS
3734 CV_WAIT(&call->cv_tq, &call->lock);
3735 #else /* RX_ENABLE_LOCKS */
3736 osi_rxSleep(&call->tq);
3737 #endif /* RX_ENABLE_LOCKS */
3739 MUTEX_ENTER(&peer->peer_lock);
3740 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3741 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3742 call->flags |= RX_CALL_FAST_RECOVER;
3743 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3745 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3746 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3747 call->nextCwind = call->ssthresh;
3750 peer->MTU = call->MTU;
3751 peer->cwind = call->nextCwind;
3752 peer->nDgramPackets = call->nDgramPackets;
3754 call->congestSeq = peer->congestSeq;
3755 /* Reset the resend times on the packets that were nacked
3756 * so we will retransmit as soon as the window permits*/
3757 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3759 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3760 clock_Zero(&tp->retryTime);
3762 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3767 /* If cwind is smaller than ssthresh, then increase
3768 * the window one packet for each ack we receive (exponential
3770 * If cwind is greater than or equal to ssthresh then increase
3771 * the congestion window by one packet for each cwind acks we
3772 * receive (linear growth). */
3773 if (call->cwind < call->ssthresh) {
3775 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3776 call->nCwindAcks = 0;
3778 call->nCwindAcks += newAckCount;
3779 if (call->nCwindAcks >= call->cwind) {
3780 call->nCwindAcks = 0;
3781 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3785 * If we have received several acknowledgements in a row then
3786 * it is time to increase the size of our datagrams
3788 if ((int)call->nAcks > rx_nDgramThreshold) {
3789 if (peer->maxDgramPackets > 1) {
3790 if (call->nDgramPackets < peer->maxDgramPackets) {
3791 call->nDgramPackets++;
3793 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3794 } else if (call->MTU < peer->maxMTU) {
3795 call->MTU += peer->natMTU;
3796 call->MTU = MIN(call->MTU, peer->maxMTU);
3802 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3804 /* Servers need to hold the call until all response packets have
3805 * been acknowledged. Soft acks are good enough since clients
3806 * are not allowed to clear their receive queues. */
3807 if (call->state == RX_STATE_HOLD
3808 && call->tfirst + call->nSoftAcked >= call->tnext) {
3809 call->state = RX_STATE_DALLY;
3810 rxi_ClearTransmitQueue(call, 0);
3811 } else if (!queue_IsEmpty(&call->tq)) {
3812 rxi_Start(0, call, 0, istack);
3817 /* Received a response to a challenge packet */
3819 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
3820 register struct rx_packet *np, int istack)
3824 /* Ignore the packet if we're the client */
3825 if (conn->type == RX_CLIENT_CONNECTION)
3828 /* If already authenticated, ignore the packet (it's probably a retry) */
3829 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
3832 /* Otherwise, have the security object evaluate the response packet */
3833 error = RXS_CheckResponse(conn->securityObject, conn, np);
3835 /* If the response is invalid, reset the connection, sending
3836 * an abort to the peer */
3840 rxi_ConnectionError(conn, error);
3841 MUTEX_ENTER(&conn->conn_data_lock);
3842 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3843 MUTEX_EXIT(&conn->conn_data_lock);
3846 /* If the response is valid, any calls waiting to attach
3847 * servers can now do so */
3850 for (i = 0; i < RX_MAXCALLS; i++) {
3851 struct rx_call *call = conn->call[i];
3853 MUTEX_ENTER(&call->lock);
3854 if (call->state == RX_STATE_PRECALL)
3855 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
3856 /* tnop can be null if newcallp is null */
3857 MUTEX_EXIT(&call->lock);
3861 /* Update the peer reachability information, just in case
3862 * some calls went into attach-wait while we were waiting
3863 * for authentication..
3865 rxi_UpdatePeerReach(conn, NULL);
3870 /* A client has received an authentication challenge: the security
3871 * object is asked to cough up a respectable response packet to send
3872 * back to the server. The server is responsible for retrying the
3873 * challenge if it fails to get a response. */
3876 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
3877 register struct rx_packet *np, int istack)
3881 /* Ignore the challenge if we're the server */
3882 if (conn->type == RX_SERVER_CONNECTION)
3885 /* Ignore the challenge if the connection is otherwise idle; someone's
3886 * trying to use us as an oracle. */
3887 if (!rxi_HasActiveCalls(conn))
3890 /* Send the security object the challenge packet. It is expected to fill
3891 * in the response. */
3892 error = RXS_GetResponse(conn->securityObject, conn, np);
3894 /* If the security object is unable to return a valid response, reset the
3895 * connection and send an abort to the peer. Otherwise send the response
3896 * packet to the peer connection. */
3898 rxi_ConnectionError(conn, error);
3899 MUTEX_ENTER(&conn->conn_data_lock);
3900 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3901 MUTEX_EXIT(&conn->conn_data_lock);
3903 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
3904 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
3910 /* Find an available server process to service the current request in
3911 * the given call structure. If one isn't available, queue up this
3912 * call so it eventually gets one */
3914 rxi_AttachServerProc(register struct rx_call *call,
3915 register osi_socket socket, register int *tnop,
3916 register struct rx_call **newcallp)
3918 register struct rx_serverQueueEntry *sq;
3919 register struct rx_service *service = call->conn->service;
3920 register int haveQuota = 0;
3922 /* May already be attached */
3923 if (call->state == RX_STATE_ACTIVE)
3926 MUTEX_ENTER(&rx_serverPool_lock);
3928 haveQuota = QuotaOK(service);
3929 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
3930 /* If there are no processes available to service this call,
3931 * put the call on the incoming call queue (unless it's
3932 * already on the queue).
3934 #ifdef RX_ENABLE_LOCKS
3936 ReturnToServerPool(service);
3937 #endif /* RX_ENABLE_LOCKS */
3939 if (!(call->flags & RX_CALL_WAIT_PROC)) {
3940 call->flags |= RX_CALL_WAIT_PROC;
3941 MUTEX_ENTER(&rx_stats_mutex);
3944 MUTEX_EXIT(&rx_stats_mutex);
3945 rxi_calltrace(RX_CALL_ARRIVAL, call);
3946 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
3947 queue_Append(&rx_incomingCallQueue, call);
3950 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
3952 /* If hot threads are enabled, and both newcallp and sq->socketp
3953 * are non-null, then this thread will process the call, and the
3954 * idle server thread will start listening on this threads socket.
3957 if (rx_enable_hot_thread && newcallp && sq->socketp) {
3960 *sq->socketp = socket;
3961 clock_GetTime(&call->startTime);
3962 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
3966 if (call->flags & RX_CALL_WAIT_PROC) {
3967 /* Conservative: I don't think this should happen */
3968 call->flags &= ~RX_CALL_WAIT_PROC;
3969 if (queue_IsOnQueue(call)) {
3971 MUTEX_ENTER(&rx_stats_mutex);
3973 MUTEX_EXIT(&rx_stats_mutex);
3976 call->state = RX_STATE_ACTIVE;
3977 call->mode = RX_MODE_RECEIVING;
3978 #ifdef RX_KERNEL_TRACE
3980 int glockOwner = ISAFS_GLOCK();
3983 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
3984 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
3990 if (call->flags & RX_CALL_CLEARED) {
3991 /* send an ack now to start the packet flow up again */
3992 call->flags &= ~RX_CALL_CLEARED;
3993 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
3995 #ifdef RX_ENABLE_LOCKS
3998 service->nRequestsRunning++;
3999 if (service->nRequestsRunning <= service->minProcs)
4005 MUTEX_EXIT(&rx_serverPool_lock);
4008 /* Delay the sending of an acknowledge event for a short while, while
4009 * a new call is being prepared (in the case of a client) or a reply
4010 * is being prepared (in the case of a server). Rather than sending
4011 * an ack packet, an ACKALL packet is sent. */
4013 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4015 #ifdef RX_ENABLE_LOCKS
4017 MUTEX_ENTER(&call->lock);
4018 call->delayedAckEvent = NULL;
4019 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4021 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4022 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4024 MUTEX_EXIT(&call->lock);
4025 #else /* RX_ENABLE_LOCKS */
4027 call->delayedAckEvent = NULL;
4028 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4029 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4030 #endif /* RX_ENABLE_LOCKS */
4034 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4037 #ifdef RX_ENABLE_LOCKS
4039 MUTEX_ENTER(&call->lock);
4040 if (event == call->delayedAckEvent)
4041 call->delayedAckEvent = NULL;
4042 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4044 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4046 MUTEX_EXIT(&call->lock);
4047 #else /* RX_ENABLE_LOCKS */
4049 call->delayedAckEvent = NULL;
4050 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4051 #endif /* RX_ENABLE_LOCKS */
4055 #ifdef RX_ENABLE_LOCKS
4056 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4057 * clearing them out.
4060 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4062 register struct rx_packet *p, *tp;
4065 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4068 p->flags |= RX_PKTFLAG_ACKED;
4072 call->flags |= RX_CALL_TQ_CLEARME;
4073 call->flags |= RX_CALL_TQ_SOME_ACKED;
4076 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4077 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4078 call->tfirst = call->tnext;
4079 call->nSoftAcked = 0;
4081 if (call->flags & RX_CALL_FAST_RECOVER) {
4082 call->flags &= ~RX_CALL_FAST_RECOVER;
4083 call->cwind = call->nextCwind;
4084 call->nextCwind = 0;
4087 CV_SIGNAL(&call->cv_twind);
4089 #endif /* RX_ENABLE_LOCKS */
4091 /* Clear out the transmit queue for the current call (all packets have
4092 * been received by peer) */
4094 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4096 register struct rx_packet *p, *tp;
4098 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4099 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4101 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4104 p->flags |= RX_PKTFLAG_ACKED;
4108 call->flags |= RX_CALL_TQ_CLEARME;
4109 call->flags |= RX_CALL_TQ_SOME_ACKED;
4112 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4113 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4119 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4120 call->flags &= ~RX_CALL_TQ_CLEARME;
4122 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4124 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4125 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4126 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4127 call->nSoftAcked = 0;
4129 if (call->flags & RX_CALL_FAST_RECOVER) {
4130 call->flags &= ~RX_CALL_FAST_RECOVER;
4131 call->cwind = call->nextCwind;
4133 #ifdef RX_ENABLE_LOCKS
4134 CV_SIGNAL(&call->cv_twind);
4136 osi_rxWakeup(&call->twind);
4141 rxi_ClearReceiveQueue(register struct rx_call *call)
4143 register struct rx_packet *p, *tp;
4144 if (queue_IsNotEmpty(&call->rq)) {
4145 for (queue_Scan(&call->rq, p, tp, rx_packet)) {
4150 rx_packetReclaims++;
4152 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4154 if (call->state == RX_STATE_PRECALL) {
4155 call->flags |= RX_CALL_CLEARED;
4159 /* Send an abort packet for the specified call */
4161 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4162 int istack, int force)
4170 /* Clients should never delay abort messages */
4171 if (rx_IsClientConn(call->conn))
4174 if (call->abortCode != call->error) {
4175 call->abortCode = call->error;
4176 call->abortCount = 0;
4179 if (force || rxi_callAbortThreshhold == 0
4180 || call->abortCount < rxi_callAbortThreshhold) {
4181 if (call->delayedAbortEvent) {
4182 rxevent_Cancel(call->delayedAbortEvent, call,
4183 RX_CALL_REFCOUNT_ABORT);
4185 error = htonl(call->error);
4188 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4189 (char *)&error, sizeof(error), istack);
4190 } else if (!call->delayedAbortEvent) {
4191 clock_GetTime(&when);
4192 clock_Addmsec(&when, rxi_callAbortDelay);
4193 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4194 call->delayedAbortEvent =
4195 rxevent_Post(&when, rxi_SendDelayedCallAbort, call, 0);
4200 /* Send an abort packet for the specified connection. Packet is an
4201 * optional pointer to a packet that can be used to send the abort.
4202 * Once the number of abort messages reaches the threshhold, an
4203 * event is scheduled to send the abort. Setting the force flag
4204 * overrides sending delayed abort messages.
4206 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4207 * to send the abort packet.
4210 rxi_SendConnectionAbort(register struct rx_connection *conn,
4211 struct rx_packet *packet, int istack, int force)
4219 /* Clients should never delay abort messages */
4220 if (rx_IsClientConn(conn))
4223 if (force || rxi_connAbortThreshhold == 0
4224 || conn->abortCount < rxi_connAbortThreshhold) {
4225 if (conn->delayedAbortEvent) {
4226 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4228 error = htonl(conn->error);
4230 MUTEX_EXIT(&conn->conn_data_lock);
4232 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4233 RX_PACKET_TYPE_ABORT, (char *)&error,
4234 sizeof(error), istack);
4235 MUTEX_ENTER(&conn->conn_data_lock);
4236 } else if (!conn->delayedAbortEvent) {
4237 clock_GetTime(&when);
4238 clock_Addmsec(&when, rxi_connAbortDelay);
4239 conn->delayedAbortEvent =
4240 rxevent_Post(&when, rxi_SendDelayedConnAbort, conn, 0);
4245 /* Associate an error all of the calls owned by a connection. Called
4246 * with error non-zero. This is only for really fatal things, like
4247 * bad authentication responses. The connection itself is set in
4248 * error at this point, so that future packets received will be
4251 rxi_ConnectionError(register struct rx_connection *conn,
4252 register afs_int32 error)
4256 MUTEX_ENTER(&conn->conn_data_lock);
4257 if (conn->challengeEvent)
4258 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4259 if (conn->checkReachEvent) {
4260 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4261 conn->checkReachEvent = 0;
4262 conn->flags &= ~RX_CONN_ATTACHWAIT;
4265 MUTEX_EXIT(&conn->conn_data_lock);
4266 for (i = 0; i < RX_MAXCALLS; i++) {
4267 struct rx_call *call = conn->call[i];
4269 MUTEX_ENTER(&call->lock);
4270 rxi_CallError(call, error);
4271 MUTEX_EXIT(&call->lock);
4274 conn->error = error;
4275 MUTEX_ENTER(&rx_stats_mutex);
4276 rx_stats.fatalErrors++;
4277 MUTEX_EXIT(&rx_stats_mutex);
4282 rxi_CallError(register struct rx_call *call, afs_int32 error)
4285 error = call->error;
4286 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4287 if (!(call->flags & RX_CALL_TQ_BUSY)) {
4288 rxi_ResetCall(call, 0);
4291 rxi_ResetCall(call, 0);
4293 call->error = error;
4294 call->mode = RX_MODE_ERROR;
4297 /* Reset various fields in a call structure, and wakeup waiting
4298 * processes. Some fields aren't changed: state & mode are not
4299 * touched (these must be set by the caller), and bufptr, nLeft, and
4300 * nFree are not reset, since these fields are manipulated by
4301 * unprotected macros, and may only be reset by non-interrupting code.
4304 /* this code requires that call->conn be set properly as a pre-condition. */
4305 #endif /* ADAPT_WINDOW */
4308 rxi_ResetCall(register struct rx_call *call, register int newcall)
4311 register struct rx_peer *peer;
4312 struct rx_packet *packet;
4314 /* Notify anyone who is waiting for asynchronous packet arrival */
4315 if (call->arrivalProc) {
4316 (*call->arrivalProc) (call, call->arrivalProcHandle,
4317 call->arrivalProcArg);
4318 call->arrivalProc = (void (*)())0;
4321 if (call->delayedAbortEvent) {
4322 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4323 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4325 rxi_SendCallAbort(call, packet, 0, 1);
4326 rxi_FreePacket(packet);
4331 * Update the peer with the congestion information in this call
4332 * so other calls on this connection can pick up where this call
4333 * left off. If the congestion sequence numbers don't match then
4334 * another call experienced a retransmission.
4336 peer = call->conn->peer;
4337 MUTEX_ENTER(&peer->peer_lock);
4339 if (call->congestSeq == peer->congestSeq) {
4340 peer->cwind = MAX(peer->cwind, call->cwind);
4341 peer->MTU = MAX(peer->MTU, call->MTU);
4342 peer->nDgramPackets =
4343 MAX(peer->nDgramPackets, call->nDgramPackets);
4346 call->abortCode = 0;
4347 call->abortCount = 0;
4349 if (peer->maxDgramPackets > 1) {
4350 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4352 call->MTU = peer->MTU;
4354 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4355 call->ssthresh = rx_maxSendWindow;
4356 call->nDgramPackets = peer->nDgramPackets;
4357 call->congestSeq = peer->congestSeq;
4358 MUTEX_EXIT(&peer->peer_lock);
4360 flags = call->flags;
4361 rxi_ClearReceiveQueue(call);
4362 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4363 if (call->flags & RX_CALL_TQ_BUSY) {
4364 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4365 call->flags |= (flags & RX_CALL_TQ_WAIT);
4367 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4369 rxi_ClearTransmitQueue(call, 0);
4370 queue_Init(&call->tq);
4373 queue_Init(&call->rq);
4375 call->rwind = rx_initReceiveWindow;
4376 call->twind = rx_initSendWindow;
4377 call->nSoftAcked = 0;
4378 call->nextCwind = 0;
4381 call->nCwindAcks = 0;
4382 call->nSoftAcks = 0;
4383 call->nHardAcks = 0;
4385 call->tfirst = call->rnext = call->tnext = 1;
4387 call->lastAcked = 0;
4388 call->localStatus = call->remoteStatus = 0;
4390 if (flags & RX_CALL_READER_WAIT) {
4391 #ifdef RX_ENABLE_LOCKS
4392 CV_BROADCAST(&call->cv_rq);
4394 osi_rxWakeup(&call->rq);
4397 if (flags & RX_CALL_WAIT_PACKETS) {
4398 MUTEX_ENTER(&rx_freePktQ_lock);
4399 rxi_PacketsUnWait(); /* XXX */
4400 MUTEX_EXIT(&rx_freePktQ_lock);
4402 #ifdef RX_ENABLE_LOCKS
4403 CV_SIGNAL(&call->cv_twind);
4405 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4406 osi_rxWakeup(&call->twind);
4409 #ifdef RX_ENABLE_LOCKS
4410 /* The following ensures that we don't mess with any queue while some
4411 * other thread might also be doing so. The call_queue_lock field is
4412 * is only modified under the call lock. If the call is in the process
4413 * of being removed from a queue, the call is not locked until the
4414 * the queue lock is dropped and only then is the call_queue_lock field
4415 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4416 * Note that any other routine which removes a call from a queue has to
4417 * obtain the queue lock before examing the queue and removing the call.
4419 if (call->call_queue_lock) {
4420 MUTEX_ENTER(call->call_queue_lock);
4421 if (queue_IsOnQueue(call)) {
4423 if (flags & RX_CALL_WAIT_PROC) {
4424 MUTEX_ENTER(&rx_stats_mutex);
4426 MUTEX_EXIT(&rx_stats_mutex);
4429 MUTEX_EXIT(call->call_queue_lock);
4430 CLEAR_CALL_QUEUE_LOCK(call);
4432 #else /* RX_ENABLE_LOCKS */
4433 if (queue_IsOnQueue(call)) {
4435 if (flags & RX_CALL_WAIT_PROC)
4438 #endif /* RX_ENABLE_LOCKS */
4440 rxi_KeepAliveOff(call);
4441 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4444 /* Send an acknowledge for the indicated packet (seq,serial) of the
4445 * indicated call, for the indicated reason (reason). This
4446 * acknowledge will specifically acknowledge receiving the packet, and
4447 * will also specify which other packets for this call have been
4448 * received. This routine returns the packet that was used to the
4449 * caller. The caller is responsible for freeing it or re-using it.
4450 * This acknowledgement also returns the highest sequence number
4451 * actually read out by the higher level to the sender; the sender
4452 * promises to keep around packets that have not been read by the
4453 * higher level yet (unless, of course, the sender decides to abort
4454 * the call altogether). Any of p, seq, serial, pflags, or reason may
4455 * be set to zero without ill effect. That is, if they are zero, they
4456 * will not convey any information.
4457 * NOW there is a trailer field, after the ack where it will safely be
4458 * ignored by mundanes, which indicates the maximum size packet this
4459 * host can swallow. */
4461 register struct rx_packet *optionalPacket; use to send ack (or null)
4462 int seq; Sequence number of the packet we are acking
4463 int serial; Serial number of the packet
4464 int pflags; Flags field from packet header
4465 int reason; Reason an acknowledge was prompted
4469 rxi_SendAck(register struct rx_call *call,
4470 register struct rx_packet *optionalPacket, int serial, int reason,
4473 struct rx_ackPacket *ap;
4474 register struct rx_packet *rqp;
4475 register struct rx_packet *nxp; /* For queue_Scan */
4476 register struct rx_packet *p;
4481 * Open the receive window once a thread starts reading packets
4483 if (call->rnext > 1) {
4484 call->rwind = rx_maxReceiveWindow;
4487 call->nHardAcks = 0;
4488 call->nSoftAcks = 0;
4489 if (call->rnext > call->lastAcked)
4490 call->lastAcked = call->rnext;
4494 rx_computelen(p, p->length); /* reset length, you never know */
4495 } /* where that's been... */
4496 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4497 /* We won't send the ack, but don't panic. */
4498 return optionalPacket;
4502 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4505 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL)) {
4506 if (!optionalPacket)
4508 return optionalPacket;
4510 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4511 if (rx_Contiguous(p) < templ) {
4512 if (!optionalPacket)
4514 return optionalPacket;
4519 /* MTUXXX failing to send an ack is very serious. We should */
4520 /* try as hard as possible to send even a partial ack; it's */
4521 /* better than nothing. */
4522 ap = (struct rx_ackPacket *)rx_DataOf(p);
4523 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4524 ap->reason = reason;
4526 /* The skew computation used to be bogus, I think it's better now. */
4527 /* We should start paying attention to skew. XXX */
4528 ap->serial = htonl(serial);
4529 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4531 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4532 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4534 /* No fear of running out of ack packet here because there can only be at most
4535 * one window full of unacknowledged packets. The window size must be constrained
4536 * to be less than the maximum ack size, of course. Also, an ack should always
4537 * fit into a single packet -- it should not ever be fragmented. */
4538 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4539 if (!rqp || !call->rq.next
4540 || (rqp->header.seq > (call->rnext + call->rwind))) {
4541 if (!optionalPacket)
4543 rxi_CallError(call, RX_CALL_DEAD);
4544 return optionalPacket;
4547 while (rqp->header.seq > call->rnext + offset)
4548 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4549 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4551 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4552 if (!optionalPacket)
4554 rxi_CallError(call, RX_CALL_DEAD);
4555 return optionalPacket;
4560 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4562 /* these are new for AFS 3.3 */
4563 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4564 templ = htonl(templ);
4565 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4566 templ = htonl(call->conn->peer->ifMTU);
4567 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4568 sizeof(afs_int32), &templ);
4570 /* new for AFS 3.4 */
4571 templ = htonl(call->rwind);
4572 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4573 sizeof(afs_int32), &templ);
4575 /* new for AFS 3.5 */
4576 templ = htonl(call->conn->peer->ifDgramPackets);
4577 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4578 sizeof(afs_int32), &templ);
4580 p->header.serviceId = call->conn->serviceId;
4581 p->header.cid = (call->conn->cid | call->channel);
4582 p->header.callNumber = *call->callNumber;
4584 p->header.securityIndex = call->conn->securityIndex;
4585 p->header.epoch = call->conn->epoch;
4586 p->header.type = RX_PACKET_TYPE_ACK;
4587 p->header.flags = RX_SLOW_START_OK;
4588 if (reason == RX_ACK_PING) {
4589 p->header.flags |= RX_REQUEST_ACK;
4591 clock_GetTime(&call->pingRequestTime);
4594 if (call->conn->type == RX_CLIENT_CONNECTION)
4595 p->header.flags |= RX_CLIENT_INITIATED;
4599 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u",
4600 ap->reason, ntohl(ap->previousPacket),
4601 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4603 for (offset = 0; offset < ap->nAcks; offset++)
4604 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4612 register int i, nbytes = p->length;
4614 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4615 if (nbytes <= p->wirevec[i].iov_len) {
4616 register int savelen, saven;
4618 savelen = p->wirevec[i].iov_len;
4620 p->wirevec[i].iov_len = nbytes;
4622 rxi_Send(call, p, istack);
4623 p->wirevec[i].iov_len = savelen;
4627 nbytes -= p->wirevec[i].iov_len;
4630 MUTEX_ENTER(&rx_stats_mutex);
4631 rx_stats.ackPacketsSent++;
4632 MUTEX_EXIT(&rx_stats_mutex);
4633 if (!optionalPacket)
4635 return optionalPacket; /* Return packet for re-use by caller */
4638 /* Send all of the packets in the list in single datagram */
4640 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4641 int istack, int moreFlag, struct clock *now,
4642 struct clock *retryTime, int resending)
4647 struct rx_connection *conn = call->conn;
4648 struct rx_peer *peer = conn->peer;
4650 MUTEX_ENTER(&peer->peer_lock);
4653 peer->reSends += len;
4654 MUTEX_ENTER(&rx_stats_mutex);
4655 rx_stats.dataPacketsSent += len;
4656 MUTEX_EXIT(&rx_stats_mutex);
4657 MUTEX_EXIT(&peer->peer_lock);
4659 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4663 /* Set the packet flags and schedule the resend events */
4664 /* Only request an ack for the last packet in the list */
4665 for (i = 0; i < len; i++) {
4666 list[i]->retryTime = *retryTime;
4667 if (list[i]->header.serial) {
4668 /* Exponentially backoff retry times */
4669 if (list[i]->backoff < MAXBACKOFF) {
4670 /* so it can't stay == 0 */
4671 list[i]->backoff = (list[i]->backoff << 1) + 1;
4674 clock_Addmsec(&(list[i]->retryTime),
4675 ((afs_uint32) list[i]->backoff) << 8);
4678 /* Wait a little extra for the ack on the last packet */
4679 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4680 clock_Addmsec(&(list[i]->retryTime), 400);
4683 /* Record the time sent */
4684 list[i]->timeSent = *now;
4686 /* Ask for an ack on retransmitted packets, on every other packet
4687 * if the peer doesn't support slow start. Ask for an ack on every
4688 * packet until the congestion window reaches the ack rate. */
4689 if (list[i]->header.serial) {
4691 MUTEX_ENTER(&rx_stats_mutex);
4692 rx_stats.dataPacketsReSent++;
4693 MUTEX_EXIT(&rx_stats_mutex);
4695 /* improved RTO calculation- not Karn */
4696 list[i]->firstSent = *now;
4697 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4698 || (!(call->flags & RX_CALL_SLOW_START_OK)
4699 && (list[i]->header.seq & 1)))) {
4704 MUTEX_ENTER(&peer->peer_lock);
4708 MUTEX_ENTER(&rx_stats_mutex);
4709 rx_stats.dataPacketsSent++;
4710 MUTEX_EXIT(&rx_stats_mutex);
4711 MUTEX_EXIT(&peer->peer_lock);
4713 /* Tag this packet as not being the last in this group,
4714 * for the receiver's benefit */
4715 if (i < len - 1 || moreFlag) {
4716 list[i]->header.flags |= RX_MORE_PACKETS;
4719 /* Install the new retransmit time for the packet, and
4720 * record the time sent */
4721 list[i]->timeSent = *now;
4725 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4728 /* Since we're about to send a data packet to the peer, it's
4729 * safe to nuke any scheduled end-of-packets ack */
4730 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4732 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4733 MUTEX_EXIT(&call->lock);
4735 rxi_SendPacketList(call, conn, list, len, istack);
4737 rxi_SendPacket(call, conn, list[0], istack);
4739 MUTEX_ENTER(&call->lock);
4740 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4742 /* Update last send time for this call (for keep-alive
4743 * processing), and for the connection (so that we can discover
4744 * idle connections) */
4745 conn->lastSendTime = call->lastSendTime = clock_Sec();
4748 /* When sending packets we need to follow these rules:
4749 * 1. Never send more than maxDgramPackets in a jumbogram.
4750 * 2. Never send a packet with more than two iovecs in a jumbogram.
4751 * 3. Never send a retransmitted packet in a jumbogram.
4752 * 4. Never send more than cwind/4 packets in a jumbogram
4753 * We always keep the last list we should have sent so we
4754 * can set the RX_MORE_PACKETS flags correctly.
4757 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4758 int istack, struct clock *now, struct clock *retryTime,
4761 int i, cnt, lastCnt = 0;
4762 struct rx_packet **listP, **lastP = 0;
4763 struct rx_peer *peer = call->conn->peer;
4764 int morePackets = 0;
4766 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4767 /* Does the current packet force us to flush the current list? */
4769 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4770 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4772 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
4774 /* If the call enters an error state stop sending, or if
4775 * we entered congestion recovery mode, stop sending */
4776 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4784 /* Add the current packet to the list if it hasn't been acked.
4785 * Otherwise adjust the list pointer to skip the current packet. */
4786 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
4788 /* Do we need to flush the list? */
4789 if (cnt >= (int)peer->maxDgramPackets
4790 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
4791 || list[i]->header.serial
4792 || list[i]->length != RX_JUMBOBUFFERSIZE) {
4794 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
4795 retryTime, resending);
4796 /* If the call enters an error state stop sending, or if
4797 * we entered congestion recovery mode, stop sending */
4799 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4804 listP = &list[i + 1];
4809 osi_Panic("rxi_SendList error");
4811 listP = &list[i + 1];
4815 /* Send the whole list when the call is in receive mode, when
4816 * the call is in eof mode, when we are in fast recovery mode,
4817 * and when we have the last packet */
4818 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
4819 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
4820 || (call->flags & RX_CALL_FAST_RECOVER)) {
4821 /* Check for the case where the current list contains
4822 * an acked packet. Since we always send retransmissions
4823 * in a separate packet, we only need to check the first
4824 * packet in the list */
4825 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
4829 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
4830 retryTime, resending);
4831 /* If the call enters an error state stop sending, or if
4832 * we entered congestion recovery mode, stop sending */
4833 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4837 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
4840 } else if (lastCnt > 0) {
4841 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
4846 #ifdef RX_ENABLE_LOCKS
4847 /* Call rxi_Start, below, but with the call lock held. */
4849 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
4850 void *arg1, int istack)
4852 MUTEX_ENTER(&call->lock);
4853 rxi_Start(event, call, arg1, istack);
4854 MUTEX_EXIT(&call->lock);
4856 #endif /* RX_ENABLE_LOCKS */
4858 /* This routine is called when new packets are readied for
4859 * transmission and when retransmission may be necessary, or when the
4860 * transmission window or burst count are favourable. This should be
4861 * better optimized for new packets, the usual case, now that we've
4862 * got rid of queues of send packets. XXXXXXXXXXX */
4864 rxi_Start(struct rxevent *event, register struct rx_call *call,
4865 void *arg1, int istack)
4867 struct rx_packet *p;
4868 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
4869 struct rx_peer *peer = call->conn->peer;
4870 struct clock now, retryTime;
4874 struct rx_packet **xmitList;
4877 /* If rxi_Start is being called as a result of a resend event,
4878 * then make sure that the event pointer is removed from the call
4879 * structure, since there is no longer a per-call retransmission
4881 if (event && event == call->resendEvent) {
4882 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
4883 call->resendEvent = NULL;
4885 if (queue_IsEmpty(&call->tq)) {
4889 /* Timeouts trigger congestion recovery */
4890 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4891 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4892 /* someone else is waiting to start recovery */
4895 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4896 while (call->flags & RX_CALL_TQ_BUSY) {
4897 call->flags |= RX_CALL_TQ_WAIT;
4898 #ifdef RX_ENABLE_LOCKS
4899 CV_WAIT(&call->cv_tq, &call->lock);
4900 #else /* RX_ENABLE_LOCKS */
4901 osi_rxSleep(&call->tq);
4902 #endif /* RX_ENABLE_LOCKS */
4904 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4905 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4906 call->flags |= RX_CALL_FAST_RECOVER;
4907 if (peer->maxDgramPackets > 1) {
4908 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4910 call->MTU = MIN(peer->natMTU, peer->maxMTU);
4912 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4913 call->nDgramPackets = 1;
4915 call->nextCwind = 1;
4918 MUTEX_ENTER(&peer->peer_lock);
4919 peer->MTU = call->MTU;
4920 peer->cwind = call->cwind;
4921 peer->nDgramPackets = 1;
4923 call->congestSeq = peer->congestSeq;
4924 MUTEX_EXIT(&peer->peer_lock);
4925 /* Clear retry times on packets. Otherwise, it's possible for
4926 * some packets in the queue to force resends at rates faster
4927 * than recovery rates.
4929 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
4930 if (!(p->flags & RX_PKTFLAG_ACKED)) {
4931 clock_Zero(&p->retryTime);
4936 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4937 MUTEX_ENTER(&rx_stats_mutex);
4938 rx_tq_debug.rxi_start_in_error++;
4939 MUTEX_EXIT(&rx_stats_mutex);
4944 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
4945 /* Get clock to compute the re-transmit time for any packets
4946 * in this burst. Note, if we back off, it's reasonable to
4947 * back off all of the packets in the same manner, even if
4948 * some of them have been retransmitted more times than more
4949 * recent additions */
4950 clock_GetTime(&now);
4951 retryTime = now; /* initialize before use */
4952 MUTEX_ENTER(&peer->peer_lock);
4953 clock_Add(&retryTime, &peer->timeout);
4954 MUTEX_EXIT(&peer->peer_lock);
4956 /* Send (or resend) any packets that need it, subject to
4957 * window restrictions and congestion burst control
4958 * restrictions. Ask for an ack on the last packet sent in
4959 * this burst. For now, we're relying upon the window being
4960 * considerably bigger than the largest number of packets that
4961 * are typically sent at once by one initial call to
4962 * rxi_Start. This is probably bogus (perhaps we should ask
4963 * for an ack when we're half way through the current
4964 * window?). Also, for non file transfer applications, this
4965 * may end up asking for an ack for every packet. Bogus. XXXX
4968 * But check whether we're here recursively, and let the other guy
4971 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4972 if (!(call->flags & RX_CALL_TQ_BUSY)) {
4973 call->flags |= RX_CALL_TQ_BUSY;
4975 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4977 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4978 call->flags &= ~RX_CALL_NEED_START;
4979 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4981 maxXmitPackets = MIN(call->twind, call->cwind);
4982 xmitList = (struct rx_packet **)
4983 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
4984 if (xmitList == NULL)
4985 osi_Panic("rxi_Start, failed to allocate xmit list");
4986 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
4987 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4988 /* We shouldn't be sending packets if a thread is waiting
4989 * to initiate congestion recovery */
4993 && (call->flags & RX_CALL_FAST_RECOVER)) {
4994 /* Only send one packet during fast recovery */
4997 if ((p->flags & RX_PKTFLAG_FREE)
4998 || (!queue_IsEnd(&call->tq, nxp)
4999 && (nxp->flags & RX_PKTFLAG_FREE))
5000 || (p == (struct rx_packet *)&rx_freePacketQueue)
5001 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5002 osi_Panic("rxi_Start: xmit queue clobbered");
5004 if (p->flags & RX_PKTFLAG_ACKED) {
5005 MUTEX_ENTER(&rx_stats_mutex);
5006 rx_stats.ignoreAckedPacket++;
5007 MUTEX_EXIT(&rx_stats_mutex);
5008 continue; /* Ignore this packet if it has been acknowledged */
5011 /* Turn off all flags except these ones, which are the same
5012 * on each transmission */
5013 p->header.flags &= RX_PRESET_FLAGS;
5015 if (p->header.seq >=
5016 call->tfirst + MIN((int)call->twind,
5017 (int)(call->nSoftAcked +
5019 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5020 /* Note: if we're waiting for more window space, we can
5021 * still send retransmits; hence we don't return here, but
5022 * break out to schedule a retransmit event */
5023 dpf(("call %d waiting for window",
5024 *(call->callNumber)));
5028 /* Transmit the packet if it needs to be sent. */
5029 if (!clock_Lt(&now, &p->retryTime)) {
5030 if (nXmitPackets == maxXmitPackets) {
5031 rxi_SendXmitList(call, xmitList, nXmitPackets,
5032 istack, &now, &retryTime,
5034 osi_Free(xmitList, maxXmitPackets *
5035 sizeof(struct rx_packet *));
5038 xmitList[nXmitPackets++] = p;
5042 /* xmitList now hold pointers to all of the packets that are
5043 * ready to send. Now we loop to send the packets */
5044 if (nXmitPackets > 0) {
5045 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5046 &now, &retryTime, resending);
5049 maxXmitPackets * sizeof(struct rx_packet *));
5051 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5053 * TQ references no longer protected by this flag; they must remain
5054 * protected by the global lock.
5056 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5057 call->flags &= ~RX_CALL_TQ_BUSY;
5058 if (call->flags & RX_CALL_TQ_WAIT) {
5059 call->flags &= ~RX_CALL_TQ_WAIT;
5060 #ifdef RX_ENABLE_LOCKS
5061 CV_BROADCAST(&call->cv_tq);
5062 #else /* RX_ENABLE_LOCKS */
5063 osi_rxWakeup(&call->tq);
5064 #endif /* RX_ENABLE_LOCKS */
5069 /* We went into the error state while sending packets. Now is
5070 * the time to reset the call. This will also inform the using
5071 * process that the call is in an error state.
5073 MUTEX_ENTER(&rx_stats_mutex);
5074 rx_tq_debug.rxi_start_aborted++;
5075 MUTEX_EXIT(&rx_stats_mutex);
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 */
5085 rxi_CallError(call, call->error);
5088 #ifdef RX_ENABLE_LOCKS
5089 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5090 register int missing;
5091 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5092 /* Some packets have received acks. If they all have, we can clear
5093 * the transmit queue.
5096 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5097 if (p->header.seq < call->tfirst
5098 && (p->flags & RX_PKTFLAG_ACKED)) {
5105 call->flags |= RX_CALL_TQ_CLEARME;
5107 #endif /* RX_ENABLE_LOCKS */
5108 /* Don't bother doing retransmits if the TQ is cleared. */
5109 if (call->flags & RX_CALL_TQ_CLEARME) {
5110 rxi_ClearTransmitQueue(call, 1);
5112 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5115 /* Always post a resend event, if there is anything in the
5116 * queue, and resend is possible. There should be at least
5117 * one unacknowledged packet in the queue ... otherwise none
5118 * of these packets should be on the queue in the first place.
5120 if (call->resendEvent) {
5121 /* Cancel the existing event and post a new one */
5122 rxevent_Cancel(call->resendEvent, call,
5123 RX_CALL_REFCOUNT_RESEND);
5126 /* The retry time is the retry time on the first unacknowledged
5127 * packet inside the current window */
5129 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5130 /* Don't set timers for packets outside the window */
5131 if (p->header.seq >= call->tfirst + call->twind) {
5135 if (!(p->flags & RX_PKTFLAG_ACKED)
5136 && !clock_IsZero(&p->retryTime)) {
5138 retryTime = p->retryTime;
5143 /* Post a new event to re-run rxi_Start when retries may be needed */
5144 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5145 #ifdef RX_ENABLE_LOCKS
5146 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5148 rxevent_Post2(&retryTime, rxi_StartUnlocked,
5149 (void *)call, 0, istack);
5150 #else /* RX_ENABLE_LOCKS */
5152 rxevent_Post2(&retryTime, rxi_Start, (void *)call,
5154 #endif /* RX_ENABLE_LOCKS */
5157 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5158 } while (call->flags & RX_CALL_NEED_START);
5160 * TQ references no longer protected by this flag; they must remain
5161 * protected by the global lock.
5163 call->flags &= ~RX_CALL_TQ_BUSY;
5164 if (call->flags & RX_CALL_TQ_WAIT) {
5165 call->flags &= ~RX_CALL_TQ_WAIT;
5166 #ifdef RX_ENABLE_LOCKS
5167 CV_BROADCAST(&call->cv_tq);
5168 #else /* RX_ENABLE_LOCKS */
5169 osi_rxWakeup(&call->tq);
5170 #endif /* RX_ENABLE_LOCKS */
5173 call->flags |= RX_CALL_NEED_START;
5175 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5177 if (call->resendEvent) {
5178 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5183 /* Also adjusts the keep alive parameters for the call, to reflect
5184 * that we have just sent a packet (so keep alives aren't sent
5187 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5190 register struct rx_connection *conn = call->conn;
5192 /* Stamp each packet with the user supplied status */
5193 p->header.userStatus = call->localStatus;
5195 /* Allow the security object controlling this call's security to
5196 * make any last-minute changes to the packet */
5197 RXS_SendPacket(conn->securityObject, call, p);
5199 /* Since we're about to send SOME sort of packet to the peer, it's
5200 * safe to nuke any scheduled end-of-packets ack */
5201 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5203 /* Actually send the packet, filling in more connection-specific fields */
5204 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5205 MUTEX_EXIT(&call->lock);
5206 rxi_SendPacket(call, conn, p, istack);
5207 MUTEX_ENTER(&call->lock);
5208 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5210 /* Update last send time for this call (for keep-alive
5211 * processing), and for the connection (so that we can discover
5212 * idle connections) */
5213 conn->lastSendTime = call->lastSendTime = clock_Sec();
5217 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5218 * that things are fine. Also called periodically to guarantee that nothing
5219 * falls through the cracks (e.g. (error + dally) connections have keepalive
5220 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5222 * haveCTLock Set if calling from rxi_ReapConnections
5224 #ifdef RX_ENABLE_LOCKS
5226 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5227 #else /* RX_ENABLE_LOCKS */
5229 rxi_CheckCall(register struct rx_call *call)
5230 #endif /* RX_ENABLE_LOCKS */
5232 register struct rx_connection *conn = call->conn;
5234 afs_uint32 deadTime;
5236 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5237 if (call->flags & RX_CALL_TQ_BUSY) {
5238 /* Call is active and will be reset by rxi_Start if it's
5239 * in an error state.
5244 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5246 (((afs_uint32) conn->secondsUntilDead << 10) +
5247 ((afs_uint32) conn->peer->rtt >> 3) +
5248 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5250 /* These are computed to the second (+- 1 second). But that's
5251 * good enough for these values, which should be a significant
5252 * number of seconds. */
5253 if (now > (call->lastReceiveTime + deadTime)) {
5254 if (call->state == RX_STATE_ACTIVE) {
5255 rxi_CallError(call, RX_CALL_DEAD);
5258 #ifdef RX_ENABLE_LOCKS
5259 /* Cancel pending events */
5260 rxevent_Cancel(call->delayedAckEvent, call,
5261 RX_CALL_REFCOUNT_DELAY);
5262 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5263 rxevent_Cancel(call->keepAliveEvent, call,
5264 RX_CALL_REFCOUNT_ALIVE);
5265 if (call->refCount == 0) {
5266 rxi_FreeCall(call, haveCTLock);
5270 #else /* RX_ENABLE_LOCKS */
5273 #endif /* RX_ENABLE_LOCKS */
5275 /* Non-active calls are destroyed if they are not responding
5276 * to pings; active calls are simply flagged in error, so the
5277 * attached process can die reasonably gracefully. */
5279 /* see if we have a non-activity timeout */
5280 if (call->startWait && conn->idleDeadTime
5281 && ((call->startWait + conn->idleDeadTime) < now)) {
5282 if (call->state == RX_STATE_ACTIVE) {
5283 rxi_CallError(call, RX_CALL_TIMEOUT);
5287 /* see if we have a hard timeout */
5288 if (conn->hardDeadTime
5289 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5290 if (call->state == RX_STATE_ACTIVE)
5291 rxi_CallError(call, RX_CALL_TIMEOUT);
5298 /* When a call is in progress, this routine is called occasionally to
5299 * make sure that some traffic has arrived (or been sent to) the peer.
5300 * If nothing has arrived in a reasonable amount of time, the call is
5301 * declared dead; if nothing has been sent for a while, we send a
5302 * keep-alive packet (if we're actually trying to keep the call alive)
5305 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5308 struct rx_connection *conn;
5311 MUTEX_ENTER(&call->lock);
5312 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5313 if (event == call->keepAliveEvent)
5314 call->keepAliveEvent = NULL;
5317 #ifdef RX_ENABLE_LOCKS
5318 if (rxi_CheckCall(call, 0)) {
5319 MUTEX_EXIT(&call->lock);
5322 #else /* RX_ENABLE_LOCKS */
5323 if (rxi_CheckCall(call))
5325 #endif /* RX_ENABLE_LOCKS */
5327 /* Don't try to keep alive dallying calls */
5328 if (call->state == RX_STATE_DALLY) {
5329 MUTEX_EXIT(&call->lock);
5334 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5335 /* Don't try to send keepalives if there is unacknowledged data */
5336 /* the rexmit code should be good enough, this little hack
5337 * doesn't quite work XXX */
5338 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5340 rxi_ScheduleKeepAliveEvent(call);
5341 MUTEX_EXIT(&call->lock);
5346 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5348 if (!call->keepAliveEvent) {
5350 clock_GetTime(&when);
5351 when.sec += call->conn->secondsUntilPing;
5352 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5353 call->keepAliveEvent =
5354 rxevent_Post(&when, rxi_KeepAliveEvent, call, 0);
5358 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5360 rxi_KeepAliveOn(register struct rx_call *call)
5362 /* Pretend last packet received was received now--i.e. if another
5363 * packet isn't received within the keep alive time, then the call
5364 * will die; Initialize last send time to the current time--even
5365 * if a packet hasn't been sent yet. This will guarantee that a
5366 * keep-alive is sent within the ping time */
5367 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5368 rxi_ScheduleKeepAliveEvent(call);
5371 /* This routine is called to send connection abort messages
5372 * that have been delayed to throttle looping clients. */
5374 rxi_SendDelayedConnAbort(struct rxevent *event,
5375 register struct rx_connection *conn, char *dummy)
5378 struct rx_packet *packet;
5380 MUTEX_ENTER(&conn->conn_data_lock);
5381 conn->delayedAbortEvent = NULL;
5382 error = htonl(conn->error);
5384 MUTEX_EXIT(&conn->conn_data_lock);
5385 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5388 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5389 RX_PACKET_TYPE_ABORT, (char *)&error,
5391 rxi_FreePacket(packet);
5395 /* This routine is called to send call abort messages
5396 * that have been delayed to throttle looping clients. */
5398 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5402 struct rx_packet *packet;
5404 MUTEX_ENTER(&call->lock);
5405 call->delayedAbortEvent = NULL;
5406 error = htonl(call->error);
5408 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5411 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5412 (char *)&error, sizeof(error), 0);
5413 rxi_FreePacket(packet);
5415 MUTEX_EXIT(&call->lock);
5418 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5419 * seconds) to ask the client to authenticate itself. The routine
5420 * issues a challenge to the client, which is obtained from the
5421 * security object associated with the connection */
5423 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5424 void *arg1, int tries)
5426 conn->challengeEvent = NULL;
5427 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5428 register struct rx_packet *packet;
5432 /* We've failed to authenticate for too long.
5433 * Reset any calls waiting for authentication;
5434 * they are all in RX_STATE_PRECALL.
5438 MUTEX_ENTER(&conn->conn_call_lock);
5439 for (i = 0; i < RX_MAXCALLS; i++) {
5440 struct rx_call *call = conn->call[i];
5442 MUTEX_ENTER(&call->lock);
5443 if (call->state == RX_STATE_PRECALL) {
5444 rxi_CallError(call, RX_CALL_DEAD);
5445 rxi_SendCallAbort(call, NULL, 0, 0);
5447 MUTEX_EXIT(&call->lock);
5450 MUTEX_EXIT(&conn->conn_call_lock);
5454 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5456 /* If there's no packet available, do this later. */
5457 RXS_GetChallenge(conn->securityObject, conn, packet);
5458 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5459 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5460 rxi_FreePacket(packet);
5462 clock_GetTime(&when);
5463 when.sec += RX_CHALLENGE_TIMEOUT;
5464 conn->challengeEvent =
5465 rxevent_Post2(&when, rxi_ChallengeEvent, conn, 0,
5470 /* Call this routine to start requesting the client to authenticate
5471 * itself. This will continue until authentication is established,
5472 * the call times out, or an invalid response is returned. The
5473 * security object associated with the connection is asked to create
5474 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5475 * defined earlier. */
5477 rxi_ChallengeOn(register struct rx_connection *conn)
5479 if (!conn->challengeEvent) {
5480 RXS_CreateChallenge(conn->securityObject, conn);
5481 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5486 /* Compute round trip time of the packet provided, in *rttp.
5489 /* rxi_ComputeRoundTripTime is called with peer locked. */
5490 /* sentp and/or peer may be null */
5492 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5493 register struct clock *sentp,
5494 register struct rx_peer *peer)
5496 struct clock thisRtt, *rttp = &thisRtt;
5498 register int rtt_timeout;
5500 clock_GetTime(rttp);
5502 if (clock_Lt(rttp, sentp)) {
5504 return; /* somebody set the clock back, don't count this time. */
5506 clock_Sub(rttp, sentp);
5507 MUTEX_ENTER(&rx_stats_mutex);
5508 if (clock_Lt(rttp, &rx_stats.minRtt))
5509 rx_stats.minRtt = *rttp;
5510 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5511 if (rttp->sec > 60) {
5512 MUTEX_EXIT(&rx_stats_mutex);
5513 return; /* somebody set the clock ahead */
5515 rx_stats.maxRtt = *rttp;
5517 clock_Add(&rx_stats.totalRtt, rttp);
5518 rx_stats.nRttSamples++;
5519 MUTEX_EXIT(&rx_stats_mutex);
5521 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5523 /* Apply VanJacobson round-trip estimations */
5528 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5529 * srtt is stored as fixed point with 3 bits after the binary
5530 * point (i.e., scaled by 8). The following magic is
5531 * equivalent to the smoothing algorithm in rfc793 with an
5532 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5533 * srtt*8 = srtt*8 + rtt - srtt
5534 * srtt = srtt + rtt/8 - srtt/8
5537 delta = MSEC(rttp) - (peer->rtt >> 3);
5541 * We accumulate a smoothed rtt variance (actually, a smoothed
5542 * mean difference), then set the retransmit timer to smoothed
5543 * rtt + 4 times the smoothed variance (was 2x in van's original
5544 * paper, but 4x works better for me, and apparently for him as
5546 * rttvar is stored as
5547 * fixed point with 2 bits after the binary point (scaled by
5548 * 4). The following is equivalent to rfc793 smoothing with
5549 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5550 * replaces rfc793's wired-in beta.
5551 * dev*4 = dev*4 + (|actual - expected| - dev)
5557 delta -= (peer->rtt_dev >> 2);
5558 peer->rtt_dev += delta;
5560 /* I don't have a stored RTT so I start with this value. Since I'm
5561 * probably just starting a call, and will be pushing more data down
5562 * this, I expect congestion to increase rapidly. So I fudge a
5563 * little, and I set deviance to half the rtt. In practice,
5564 * deviance tends to approach something a little less than
5565 * half the smoothed rtt. */
5566 peer->rtt = (MSEC(rttp) << 3) + 8;
5567 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5569 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5570 * the other of these connections is usually in a user process, and can
5571 * be switched and/or swapped out. So on fast, reliable networks, the
5572 * timeout would otherwise be too short.
5574 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5575 clock_Zero(&(peer->timeout));
5576 clock_Addmsec(&(peer->timeout), rtt_timeout);
5578 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)));
5582 /* Find all server connections that have not been active for a long time, and
5585 rxi_ReapConnections(void)
5588 clock_GetTime(&now);
5590 /* Find server connection structures that haven't been used for
5591 * greater than rx_idleConnectionTime */
5593 struct rx_connection **conn_ptr, **conn_end;
5594 int i, havecalls = 0;
5595 MUTEX_ENTER(&rx_connHashTable_lock);
5596 for (conn_ptr = &rx_connHashTable[0], conn_end =
5597 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5599 struct rx_connection *conn, *next;
5600 struct rx_call *call;
5604 for (conn = *conn_ptr; conn; conn = next) {
5605 /* XXX -- Shouldn't the connection be locked? */
5608 for (i = 0; i < RX_MAXCALLS; i++) {
5609 call = conn->call[i];
5612 MUTEX_ENTER(&call->lock);
5613 #ifdef RX_ENABLE_LOCKS
5614 result = rxi_CheckCall(call, 1);
5615 #else /* RX_ENABLE_LOCKS */
5616 result = rxi_CheckCall(call);
5617 #endif /* RX_ENABLE_LOCKS */
5618 MUTEX_EXIT(&call->lock);
5620 /* If CheckCall freed the call, it might
5621 * have destroyed the connection as well,
5622 * which screws up the linked lists.
5628 if (conn->type == RX_SERVER_CONNECTION) {
5629 /* This only actually destroys the connection if
5630 * there are no outstanding calls */
5631 MUTEX_ENTER(&conn->conn_data_lock);
5632 if (!havecalls && !conn->refCount
5633 && ((conn->lastSendTime + rx_idleConnectionTime) <
5635 conn->refCount++; /* it will be decr in rx_DestroyConn */
5636 MUTEX_EXIT(&conn->conn_data_lock);
5637 #ifdef RX_ENABLE_LOCKS
5638 rxi_DestroyConnectionNoLock(conn);
5639 #else /* RX_ENABLE_LOCKS */
5640 rxi_DestroyConnection(conn);
5641 #endif /* RX_ENABLE_LOCKS */
5643 #ifdef RX_ENABLE_LOCKS
5645 MUTEX_EXIT(&conn->conn_data_lock);
5647 #endif /* RX_ENABLE_LOCKS */
5651 #ifdef RX_ENABLE_LOCKS
5652 while (rx_connCleanup_list) {
5653 struct rx_connection *conn;
5654 conn = rx_connCleanup_list;
5655 rx_connCleanup_list = rx_connCleanup_list->next;
5656 MUTEX_EXIT(&rx_connHashTable_lock);
5657 rxi_CleanupConnection(conn);
5658 MUTEX_ENTER(&rx_connHashTable_lock);
5660 MUTEX_EXIT(&rx_connHashTable_lock);
5661 #endif /* RX_ENABLE_LOCKS */
5664 /* Find any peer structures that haven't been used (haven't had an
5665 * associated connection) for greater than rx_idlePeerTime */
5667 struct rx_peer **peer_ptr, **peer_end;
5669 MUTEX_ENTER(&rx_rpc_stats);
5670 MUTEX_ENTER(&rx_peerHashTable_lock);
5671 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5672 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5674 struct rx_peer *peer, *next, *prev;
5675 for (prev = peer = *peer_ptr; peer; peer = next) {
5677 code = MUTEX_TRYENTER(&peer->peer_lock);
5678 if ((code) && (peer->refCount == 0)
5679 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5680 rx_interface_stat_p rpc_stat, nrpc_stat;
5682 MUTEX_EXIT(&peer->peer_lock);
5683 MUTEX_DESTROY(&peer->peer_lock);
5685 (&peer->rpcStats, rpc_stat, nrpc_stat,
5686 rx_interface_stat)) {
5687 unsigned int num_funcs;
5690 queue_Remove(&rpc_stat->queue_header);
5691 queue_Remove(&rpc_stat->all_peers);
5692 num_funcs = rpc_stat->stats[0].func_total;
5694 sizeof(rx_interface_stat_t) +
5695 rpc_stat->stats[0].func_total *
5696 sizeof(rx_function_entry_v1_t);
5698 rxi_Free(rpc_stat, space);
5699 rxi_rpc_peer_stat_cnt -= num_funcs;
5702 MUTEX_ENTER(&rx_stats_mutex);
5703 rx_stats.nPeerStructs--;
5704 MUTEX_EXIT(&rx_stats_mutex);
5705 if (peer == *peer_ptr) {
5712 MUTEX_EXIT(&peer->peer_lock);
5718 MUTEX_EXIT(&rx_peerHashTable_lock);
5719 MUTEX_EXIT(&rx_rpc_stats);
5722 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5723 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5724 * GC, just below. Really, we shouldn't have to keep moving packets from
5725 * one place to another, but instead ought to always know if we can
5726 * afford to hold onto a packet in its particular use. */
5727 MUTEX_ENTER(&rx_freePktQ_lock);
5728 if (rx_waitingForPackets) {
5729 rx_waitingForPackets = 0;
5730 #ifdef RX_ENABLE_LOCKS
5731 CV_BROADCAST(&rx_waitingForPackets_cv);
5733 osi_rxWakeup(&rx_waitingForPackets);
5736 MUTEX_EXIT(&rx_freePktQ_lock);
5738 now.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5739 rxevent_Post(&now, rxi_ReapConnections, 0, 0);
5743 /* rxs_Release - This isn't strictly necessary but, since the macro name from
5744 * rx.h is sort of strange this is better. This is called with a security
5745 * object before it is discarded. Each connection using a security object has
5746 * its own refcount to the object so it won't actually be freed until the last
5747 * connection is destroyed.
5749 * This is the only rxs module call. A hold could also be written but no one
5753 rxs_Release(struct rx_securityClass *aobj)
5755 return RXS_Close(aobj);
5759 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
5760 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
5761 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
5762 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
5764 /* Adjust our estimate of the transmission rate to this peer, given
5765 * that the packet p was just acked. We can adjust peer->timeout and
5766 * call->twind. Pragmatically, this is called
5767 * only with packets of maximal length.
5768 * Called with peer and call locked.
5772 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
5773 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
5775 afs_int32 xferSize, xferMs;
5776 register afs_int32 minTime;
5779 /* Count down packets */
5780 if (peer->rateFlag > 0)
5782 /* Do nothing until we're enabled */
5783 if (peer->rateFlag != 0)
5788 /* Count only when the ack seems legitimate */
5789 switch (ackReason) {
5790 case RX_ACK_REQUESTED:
5792 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
5796 case RX_ACK_PING_RESPONSE:
5797 if (p) /* want the response to ping-request, not data send */
5799 clock_GetTime(&newTO);
5800 if (clock_Gt(&newTO, &call->pingRequestTime)) {
5801 clock_Sub(&newTO, &call->pingRequestTime);
5802 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
5806 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
5813 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));
5815 /* Track only packets that are big enough. */
5816 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
5820 /* absorb RTT data (in milliseconds) for these big packets */
5821 if (peer->smRtt == 0) {
5822 peer->smRtt = xferMs;
5824 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
5829 if (peer->countDown) {
5833 peer->countDown = 10; /* recalculate only every so often */
5835 /* In practice, we can measure only the RTT for full packets,
5836 * because of the way Rx acks the data that it receives. (If it's
5837 * smaller than a full packet, it often gets implicitly acked
5838 * either by the call response (from a server) or by the next call
5839 * (from a client), and either case confuses transmission times
5840 * with processing times.) Therefore, replace the above
5841 * more-sophisticated processing with a simpler version, where the
5842 * smoothed RTT is kept for full-size packets, and the time to
5843 * transmit a windowful of full-size packets is simply RTT *
5844 * windowSize. Again, we take two steps:
5845 - ensure the timeout is large enough for a single packet's RTT;
5846 - ensure that the window is small enough to fit in the desired timeout.*/
5848 /* First, the timeout check. */
5849 minTime = peer->smRtt;
5850 /* Get a reasonable estimate for a timeout period */
5852 newTO.sec = minTime / 1000;
5853 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
5855 /* Increase the timeout period so that we can always do at least
5856 * one packet exchange */
5857 if (clock_Gt(&newTO, &peer->timeout)) {
5859 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));
5861 peer->timeout = newTO;
5864 /* Now, get an estimate for the transmit window size. */
5865 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
5866 /* Now, convert to the number of full packets that could fit in a
5867 * reasonable fraction of that interval */
5868 minTime /= (peer->smRtt << 1);
5869 xferSize = minTime; /* (make a copy) */
5871 /* Now clamp the size to reasonable bounds. */
5874 else if (minTime > rx_Window)
5875 minTime = rx_Window;
5876 /* if (minTime != peer->maxWindow) {
5877 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
5878 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
5879 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
5881 peer->maxWindow = minTime;
5882 elide... call->twind = minTime;
5886 /* Cut back on the peer timeout if it had earlier grown unreasonably.
5887 * Discern this by calculating the timeout necessary for rx_Window
5889 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
5890 /* calculate estimate for transmission interval in milliseconds */
5891 minTime = rx_Window * peer->smRtt;
5892 if (minTime < 1000) {
5893 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
5894 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
5895 peer->timeout.usec, peer->smRtt, peer->packetSize));
5897 newTO.sec = 0; /* cut back on timeout by half a second */
5898 newTO.usec = 500000;
5899 clock_Sub(&peer->timeout, &newTO);
5904 } /* end of rxi_ComputeRate */
5905 #endif /* ADAPT_WINDOW */
5913 /* Don't call this debugging routine directly; use dpf */
5915 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
5916 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
5920 clock_GetTime(&now);
5921 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
5922 (unsigned int)now.usec / 1000);
5923 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
5931 * This function is used to process the rx_stats structure that is local
5932 * to a process as well as an rx_stats structure received from a remote
5933 * process (via rxdebug). Therefore, it needs to do minimal version
5937 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
5938 afs_int32 freePackets, char version)
5942 if (size != sizeof(struct rx_stats)) {
5944 "Unexpected size of stats structure: was %d, expected %d\n",
5945 size, sizeof(struct rx_stats));
5948 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
5951 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
5952 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
5953 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
5954 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
5955 s->specialPktAllocFailures);
5957 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
5958 s->receivePktAllocFailures, s->sendPktAllocFailures,
5959 s->specialPktAllocFailures);
5963 " greedy %d, " "bogusReads %d (last from host %x), "
5964 "noPackets %d, " "noBuffers %d, " "selects %d, "
5965 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
5966 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
5967 s->selects, s->sendSelects);
5969 fprintf(file, " packets read: ");
5970 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
5971 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
5973 fprintf(file, "\n");
5976 " other read counters: data %d, " "ack %d, " "dup %d "
5977 "spurious %d " "dally %d\n", s->dataPacketsRead,
5978 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
5979 s->ignorePacketDally);
5981 fprintf(file, " packets sent: ");
5982 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
5983 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
5985 fprintf(file, "\n");
5988 " other send counters: ack %d, " "data %d (not resends), "
5989 "resends %d, " "pushed %d, " "acked&ignored %d\n",
5990 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
5991 s->dataPacketsPushed, s->ignoreAckedPacket);
5994 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
5995 s->netSendFailures, (int)s->fatalErrors);
5997 if (s->nRttSamples) {
5998 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
5999 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6001 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6002 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6006 " %d server connections, " "%d client connections, "
6007 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6008 s->nServerConns, s->nClientConns, s->nPeerStructs,
6009 s->nCallStructs, s->nFreeCallStructs);
6011 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6012 fprintf(file, " %d clock updates\n", clock_nUpdates);
6017 /* for backward compatibility */
6019 rx_PrintStats(FILE * file)
6021 MUTEX_ENTER(&rx_stats_mutex);
6022 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6024 MUTEX_EXIT(&rx_stats_mutex);
6028 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6030 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6031 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6032 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6035 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6036 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6037 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6040 " Packet size %d, " "max in packet skew %d, "
6041 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6042 (int)peer->outPacketSkew);
6045 #ifdef AFS_PTHREAD_ENV
6047 * This mutex protects the following static variables:
6051 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6052 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6054 #define LOCK_RX_DEBUG
6055 #define UNLOCK_RX_DEBUG
6056 #endif /* AFS_PTHREAD_ENV */
6059 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6060 u_char type, void *inputData, size_t inputLength,
6061 void *outputData, size_t outputLength)
6063 static afs_int32 counter = 100;
6065 struct rx_header theader;
6067 register afs_int32 code;
6069 struct sockaddr_in taddr, faddr;
6074 endTime = time(0) + 20; /* try for 20 seconds */
6078 tp = &tbuffer[sizeof(struct rx_header)];
6079 taddr.sin_family = AF_INET;
6080 taddr.sin_port = remotePort;
6081 taddr.sin_addr.s_addr = remoteAddr;
6082 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6083 taddr.sin_len = sizeof(struct sockaddr_in);
6086 memset(&theader, 0, sizeof(theader));
6087 theader.epoch = htonl(999);
6089 theader.callNumber = htonl(counter);
6092 theader.type = type;
6093 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6094 theader.serviceId = 0;
6096 memcpy(tbuffer, &theader, sizeof(theader));
6097 memcpy(tp, inputData, inputLength);
6099 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6100 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6102 /* see if there's a packet available */
6104 FD_SET(socket, &imask);
6107 code = select(socket + 1, &imask, 0, 0, &tv);
6108 if (code == 1 && FD_ISSET(socket, &imask)) {
6109 /* now receive a packet */
6110 faddrLen = sizeof(struct sockaddr_in);
6112 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6113 (struct sockaddr *)&faddr, &faddrLen);
6116 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6117 if (counter == ntohl(theader.callNumber))
6122 /* see if we've timed out */
6123 if (endTime < time(0))
6126 code -= sizeof(struct rx_header);
6127 if (code > outputLength)
6128 code = outputLength;
6129 memcpy(outputData, tp, code);
6134 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6135 afs_uint16 remotePort, struct rx_debugStats * stat,
6136 afs_uint32 * supportedValues)
6138 struct rx_debugIn in;
6141 *supportedValues = 0;
6142 in.type = htonl(RX_DEBUGI_GETSTATS);
6145 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6146 &in, sizeof(in), stat, sizeof(*stat));
6149 * If the call was successful, fixup the version and indicate
6150 * what contents of the stat structure are valid.
6151 * Also do net to host conversion of fields here.
6155 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6156 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6158 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6159 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6161 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6162 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6164 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6165 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6167 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6168 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6170 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6171 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6173 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6174 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6176 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6177 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6180 stat->nFreePackets = ntohl(stat->nFreePackets);
6181 stat->packetReclaims = ntohl(stat->packetReclaims);
6182 stat->callsExecuted = ntohl(stat->callsExecuted);
6183 stat->nWaiting = ntohl(stat->nWaiting);
6184 stat->idleThreads = ntohl(stat->idleThreads);
6191 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6192 afs_uint16 remotePort, struct rx_stats * stat,
6193 afs_uint32 * supportedValues)
6195 struct rx_debugIn in;
6196 afs_int32 *lp = (afs_int32 *) stat;
6201 * supportedValues is currently unused, but added to allow future
6202 * versioning of this function.
6205 *supportedValues = 0;
6206 in.type = htonl(RX_DEBUGI_RXSTATS);
6208 memset(stat, 0, sizeof(*stat));
6210 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6211 &in, sizeof(in), stat, sizeof(*stat));
6216 * Do net to host conversion here
6219 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6228 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6229 afs_uint16 remotePort, size_t version_length,
6233 return MakeDebugCall(socket, remoteAddr, remotePort,
6234 RX_PACKET_TYPE_VERSION, a, 1, version,
6239 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6240 afs_uint16 remotePort, afs_int32 * nextConnection,
6241 int allConnections, afs_uint32 debugSupportedValues,
6242 struct rx_debugConn * conn,
6243 afs_uint32 * supportedValues)
6245 struct rx_debugIn in;
6250 * supportedValues is currently unused, but added to allow future
6251 * versioning of this function.
6254 *supportedValues = 0;
6255 if (allConnections) {
6256 in.type = htonl(RX_DEBUGI_GETALLCONN);
6258 in.type = htonl(RX_DEBUGI_GETCONN);
6260 in.index = htonl(*nextConnection);
6261 memset(conn, 0, sizeof(*conn));
6263 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6264 &in, sizeof(in), conn, sizeof(*conn));
6267 *nextConnection += 1;
6270 * Convert old connection format to new structure.
6273 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6274 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6275 #define MOVEvL(a) (conn->a = vL->a)
6277 /* any old or unrecognized version... */
6278 for (i = 0; i < RX_MAXCALLS; i++) {
6279 MOVEvL(callState[i]);
6280 MOVEvL(callMode[i]);
6281 MOVEvL(callFlags[i]);
6282 MOVEvL(callOther[i]);
6284 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6285 MOVEvL(secStats.type);
6286 MOVEvL(secStats.level);
6287 MOVEvL(secStats.flags);
6288 MOVEvL(secStats.expires);
6289 MOVEvL(secStats.packetsReceived);
6290 MOVEvL(secStats.packetsSent);
6291 MOVEvL(secStats.bytesReceived);
6292 MOVEvL(secStats.bytesSent);
6297 * Do net to host conversion here
6299 * I don't convert host or port since we are most likely
6300 * going to want these in NBO.
6302 conn->cid = ntohl(conn->cid);
6303 conn->serial = ntohl(conn->serial);
6304 for (i = 0; i < RX_MAXCALLS; i++) {
6305 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6307 conn->error = ntohl(conn->error);
6308 conn->secStats.flags = ntohl(conn->secStats.flags);
6309 conn->secStats.expires = ntohl(conn->secStats.expires);
6310 conn->secStats.packetsReceived =
6311 ntohl(conn->secStats.packetsReceived);
6312 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6313 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6314 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6315 conn->epoch = ntohl(conn->epoch);
6316 conn->natMTU = ntohl(conn->natMTU);
6323 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6324 afs_uint16 remotePort, afs_int32 * nextPeer,
6325 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6326 afs_uint32 * supportedValues)
6328 struct rx_debugIn in;
6332 * supportedValues is currently unused, but added to allow future
6333 * versioning of this function.
6336 *supportedValues = 0;
6337 in.type = htonl(RX_DEBUGI_GETPEER);
6338 in.index = htonl(*nextPeer);
6339 memset(peer, 0, sizeof(*peer));
6341 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6342 &in, sizeof(in), peer, sizeof(*peer));
6348 * Do net to host conversion here
6350 * I don't convert host or port since we are most likely
6351 * going to want these in NBO.
6353 peer->ifMTU = ntohs(peer->ifMTU);
6354 peer->idleWhen = ntohl(peer->idleWhen);
6355 peer->refCount = ntohs(peer->refCount);
6356 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6357 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6358 peer->rtt = ntohl(peer->rtt);
6359 peer->rtt_dev = ntohl(peer->rtt_dev);
6360 peer->timeout.sec = ntohl(peer->timeout.sec);
6361 peer->timeout.usec = ntohl(peer->timeout.usec);
6362 peer->nSent = ntohl(peer->nSent);
6363 peer->reSends = ntohl(peer->reSends);
6364 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6365 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6366 peer->rateFlag = ntohl(peer->rateFlag);
6367 peer->natMTU = ntohs(peer->natMTU);
6368 peer->maxMTU = ntohs(peer->maxMTU);
6369 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6370 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6371 peer->MTU = ntohs(peer->MTU);
6372 peer->cwind = ntohs(peer->cwind);
6373 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6374 peer->congestSeq = ntohs(peer->congestSeq);
6375 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6376 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6377 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6378 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6383 #endif /* RXDEBUG */
6388 struct rx_serverQueueEntry *np;
6391 register struct rx_call *call;
6392 register struct rx_serverQueueEntry *sq;
6396 if (rxinit_status == 1) {
6398 return; /* Already shutdown. */
6402 #ifndef AFS_PTHREAD_ENV
6403 FD_ZERO(&rx_selectMask);
6404 #endif /* AFS_PTHREAD_ENV */
6405 rxi_dataQuota = RX_MAX_QUOTA;
6406 #ifndef AFS_PTHREAD_ENV
6408 #endif /* AFS_PTHREAD_ENV */
6411 #ifndef AFS_PTHREAD_ENV
6412 #ifndef AFS_USE_GETTIMEOFDAY
6414 #endif /* AFS_USE_GETTIMEOFDAY */
6415 #endif /* AFS_PTHREAD_ENV */
6417 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6418 call = queue_First(&rx_freeCallQueue, rx_call);
6420 rxi_Free(call, sizeof(struct rx_call));
6423 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6424 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6430 struct rx_peer **peer_ptr, **peer_end;
6431 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6432 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6434 struct rx_peer *peer, *next;
6435 for (peer = *peer_ptr; peer; peer = next) {
6436 rx_interface_stat_p rpc_stat, nrpc_stat;
6439 (&peer->rpcStats, rpc_stat, nrpc_stat,
6440 rx_interface_stat)) {
6441 unsigned int num_funcs;
6444 queue_Remove(&rpc_stat->queue_header);
6445 queue_Remove(&rpc_stat->all_peers);
6446 num_funcs = rpc_stat->stats[0].func_total;
6448 sizeof(rx_interface_stat_t) +
6449 rpc_stat->stats[0].func_total *
6450 sizeof(rx_function_entry_v1_t);
6452 rxi_Free(rpc_stat, space);
6453 MUTEX_ENTER(&rx_rpc_stats);
6454 rxi_rpc_peer_stat_cnt -= num_funcs;
6455 MUTEX_EXIT(&rx_rpc_stats);
6459 MUTEX_ENTER(&rx_stats_mutex);
6460 rx_stats.nPeerStructs--;
6461 MUTEX_EXIT(&rx_stats_mutex);
6465 for (i = 0; i < RX_MAX_SERVICES; i++) {
6467 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6469 for (i = 0; i < rx_hashTableSize; i++) {
6470 register struct rx_connection *tc, *ntc;
6471 MUTEX_ENTER(&rx_connHashTable_lock);
6472 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6474 for (j = 0; j < RX_MAXCALLS; j++) {
6476 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6479 rxi_Free(tc, sizeof(*tc));
6481 MUTEX_EXIT(&rx_connHashTable_lock);
6484 MUTEX_ENTER(&freeSQEList_lock);
6486 while ((np = rx_FreeSQEList)) {
6487 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6488 MUTEX_DESTROY(&np->lock);
6489 rxi_Free(np, sizeof(*np));
6492 MUTEX_EXIT(&freeSQEList_lock);
6493 MUTEX_DESTROY(&freeSQEList_lock);
6494 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6495 MUTEX_DESTROY(&rx_connHashTable_lock);
6496 MUTEX_DESTROY(&rx_peerHashTable_lock);
6497 MUTEX_DESTROY(&rx_serverPool_lock);
6499 osi_Free(rx_connHashTable,
6500 rx_hashTableSize * sizeof(struct rx_connection *));
6501 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6503 UNPIN(rx_connHashTable,
6504 rx_hashTableSize * sizeof(struct rx_connection *));
6505 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6507 rxi_FreeAllPackets();
6509 MUTEX_ENTER(&rx_stats_mutex);
6510 rxi_dataQuota = RX_MAX_QUOTA;
6511 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6512 MUTEX_EXIT(&rx_stats_mutex);
6518 #ifdef RX_ENABLE_LOCKS
6520 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6522 if (!MUTEX_ISMINE(lockaddr))
6523 osi_Panic("Lock not held: %s", msg);
6525 #endif /* RX_ENABLE_LOCKS */
6530 * Routines to implement connection specific data.
6534 rx_KeyCreate(rx_destructor_t rtn)
6537 MUTEX_ENTER(&rxi_keyCreate_lock);
6538 key = rxi_keyCreate_counter++;
6539 rxi_keyCreate_destructor = (rx_destructor_t *)
6540 realloc((void *)rxi_keyCreate_destructor,
6541 (key + 1) * sizeof(rx_destructor_t));
6542 rxi_keyCreate_destructor[key] = rtn;
6543 MUTEX_EXIT(&rxi_keyCreate_lock);
6548 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6551 MUTEX_ENTER(&conn->conn_data_lock);
6552 if (!conn->specific) {
6553 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6554 for (i = 0; i < key; i++)
6555 conn->specific[i] = NULL;
6556 conn->nSpecific = key + 1;
6557 conn->specific[key] = ptr;
6558 } else if (key >= conn->nSpecific) {
6559 conn->specific = (void **)
6560 realloc(conn->specific, (key + 1) * sizeof(void *));
6561 for (i = conn->nSpecific; i < key; i++)
6562 conn->specific[i] = NULL;
6563 conn->nSpecific = key + 1;
6564 conn->specific[key] = ptr;
6566 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6567 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6568 conn->specific[key] = ptr;
6570 MUTEX_EXIT(&conn->conn_data_lock);
6574 rx_GetSpecific(struct rx_connection *conn, int key)
6577 MUTEX_ENTER(&conn->conn_data_lock);
6578 if (key >= conn->nSpecific)
6581 ptr = conn->specific[key];
6582 MUTEX_EXIT(&conn->conn_data_lock);
6586 #endif /* !KERNEL */
6589 * processStats is a queue used to store the statistics for the local
6590 * process. Its contents are similar to the contents of the rpcStats
6591 * queue on a rx_peer structure, but the actual data stored within
6592 * this queue contains totals across the lifetime of the process (assuming
6593 * the stats have not been reset) - unlike the per peer structures
6594 * which can come and go based upon the peer lifetime.
6597 static struct rx_queue processStats = { &processStats, &processStats };
6600 * peerStats is a queue used to store the statistics for all peer structs.
6601 * Its contents are the union of all the peer rpcStats queues.
6604 static struct rx_queue peerStats = { &peerStats, &peerStats };
6607 * rxi_monitor_processStats is used to turn process wide stat collection
6611 static int rxi_monitor_processStats = 0;
6614 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6617 static int rxi_monitor_peerStats = 0;
6620 * rxi_AddRpcStat - given all of the information for a particular rpc
6621 * call, create (if needed) and update the stat totals for the rpc.
6625 * IN stats - the queue of stats that will be updated with the new value
6627 * IN rxInterface - a unique number that identifies the rpc interface
6629 * IN currentFunc - the index of the function being invoked
6631 * IN totalFunc - the total number of functions in this interface
6633 * IN queueTime - the amount of time this function waited for a thread
6635 * IN execTime - the amount of time this function invocation took to execute
6637 * IN bytesSent - the number bytes sent by this invocation
6639 * IN bytesRcvd - the number bytes received by this invocation
6641 * IN isServer - if true, this invocation was made to a server
6643 * IN remoteHost - the ip address of the remote host
6645 * IN remotePort - the port of the remote host
6647 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6649 * INOUT counter - if a new stats structure is allocated, the counter will
6650 * be updated with the new number of allocated stat structures
6658 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6659 afs_uint32 currentFunc, afs_uint32 totalFunc,
6660 struct clock *queueTime, struct clock *execTime,
6661 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6662 afs_uint32 remoteHost, afs_uint32 remotePort,
6663 int addToPeerList, unsigned int *counter)
6666 rx_interface_stat_p rpc_stat, nrpc_stat;
6669 * See if there's already a structure for this interface
6672 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6673 if ((rpc_stat->stats[0].interfaceId == rxInterface)
6674 && (rpc_stat->stats[0].remote_is_server == isServer))
6679 * Didn't find a match so allocate a new structure and add it to the
6683 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
6684 || (rpc_stat->stats[0].interfaceId != rxInterface)
6685 || (rpc_stat->stats[0].remote_is_server != isServer)) {
6690 sizeof(rx_interface_stat_t) +
6691 totalFunc * sizeof(rx_function_entry_v1_t);
6693 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
6694 if (rpc_stat == NULL) {
6698 *counter += totalFunc;
6699 for (i = 0; i < totalFunc; i++) {
6700 rpc_stat->stats[i].remote_peer = remoteHost;
6701 rpc_stat->stats[i].remote_port = remotePort;
6702 rpc_stat->stats[i].remote_is_server = isServer;
6703 rpc_stat->stats[i].interfaceId = rxInterface;
6704 rpc_stat->stats[i].func_total = totalFunc;
6705 rpc_stat->stats[i].func_index = i;
6706 hzero(rpc_stat->stats[i].invocations);
6707 hzero(rpc_stat->stats[i].bytes_sent);
6708 hzero(rpc_stat->stats[i].bytes_rcvd);
6709 rpc_stat->stats[i].queue_time_sum.sec = 0;
6710 rpc_stat->stats[i].queue_time_sum.usec = 0;
6711 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
6712 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
6713 rpc_stat->stats[i].queue_time_min.sec = 9999999;
6714 rpc_stat->stats[i].queue_time_min.usec = 9999999;
6715 rpc_stat->stats[i].queue_time_max.sec = 0;
6716 rpc_stat->stats[i].queue_time_max.usec = 0;
6717 rpc_stat->stats[i].execution_time_sum.sec = 0;
6718 rpc_stat->stats[i].execution_time_sum.usec = 0;
6719 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
6720 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
6721 rpc_stat->stats[i].execution_time_min.sec = 9999999;
6722 rpc_stat->stats[i].execution_time_min.usec = 9999999;
6723 rpc_stat->stats[i].execution_time_max.sec = 0;
6724 rpc_stat->stats[i].execution_time_max.usec = 0;
6726 queue_Prepend(stats, rpc_stat);
6727 if (addToPeerList) {
6728 queue_Prepend(&peerStats, &rpc_stat->all_peers);
6733 * Increment the stats for this function
6736 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
6737 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
6738 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
6739 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
6740 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
6741 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
6742 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
6744 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
6745 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
6747 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
6748 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
6750 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
6751 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
6753 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
6754 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
6762 * rx_IncrementTimeAndCount - increment the times and count for a particular
6767 * IN peer - the peer who invoked the rpc
6769 * IN rxInterface - a unique number that identifies the rpc interface
6771 * IN currentFunc - the index of the function being invoked
6773 * IN totalFunc - the total number of functions in this interface
6775 * IN queueTime - the amount of time this function waited for a thread
6777 * IN execTime - the amount of time this function invocation took to execute
6779 * IN bytesSent - the number bytes sent by this invocation
6781 * IN bytesRcvd - the number bytes received by this invocation
6783 * IN isServer - if true, this invocation was made to a server
6791 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
6792 afs_uint32 currentFunc, afs_uint32 totalFunc,
6793 struct clock *queueTime, struct clock *execTime,
6794 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
6798 MUTEX_ENTER(&rx_rpc_stats);
6799 MUTEX_ENTER(&peer->peer_lock);
6801 if (rxi_monitor_peerStats) {
6802 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
6803 queueTime, execTime, bytesSent, bytesRcvd, isServer,
6804 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
6807 if (rxi_monitor_processStats) {
6808 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
6809 queueTime, execTime, bytesSent, bytesRcvd, isServer,
6810 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
6813 MUTEX_EXIT(&peer->peer_lock);
6814 MUTEX_EXIT(&rx_rpc_stats);
6819 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
6823 * IN callerVersion - the rpc stat version of the caller.
6825 * IN count - the number of entries to marshall.
6827 * IN stats - pointer to stats to be marshalled.
6829 * OUT ptr - Where to store the marshalled data.
6836 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
6837 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
6843 * We only support the first version
6845 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
6846 *(ptr++) = stats->remote_peer;
6847 *(ptr++) = stats->remote_port;
6848 *(ptr++) = stats->remote_is_server;
6849 *(ptr++) = stats->interfaceId;
6850 *(ptr++) = stats->func_total;
6851 *(ptr++) = stats->func_index;
6852 *(ptr++) = hgethi(stats->invocations);
6853 *(ptr++) = hgetlo(stats->invocations);
6854 *(ptr++) = hgethi(stats->bytes_sent);
6855 *(ptr++) = hgetlo(stats->bytes_sent);
6856 *(ptr++) = hgethi(stats->bytes_rcvd);
6857 *(ptr++) = hgetlo(stats->bytes_rcvd);
6858 *(ptr++) = stats->queue_time_sum.sec;
6859 *(ptr++) = stats->queue_time_sum.usec;
6860 *(ptr++) = stats->queue_time_sum_sqr.sec;
6861 *(ptr++) = stats->queue_time_sum_sqr.usec;
6862 *(ptr++) = stats->queue_time_min.sec;
6863 *(ptr++) = stats->queue_time_min.usec;
6864 *(ptr++) = stats->queue_time_max.sec;
6865 *(ptr++) = stats->queue_time_max.usec;
6866 *(ptr++) = stats->execution_time_sum.sec;
6867 *(ptr++) = stats->execution_time_sum.usec;
6868 *(ptr++) = stats->execution_time_sum_sqr.sec;
6869 *(ptr++) = stats->execution_time_sum_sqr.usec;
6870 *(ptr++) = stats->execution_time_min.sec;
6871 *(ptr++) = stats->execution_time_min.usec;
6872 *(ptr++) = stats->execution_time_max.sec;
6873 *(ptr++) = stats->execution_time_max.usec;
6879 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
6884 * IN callerVersion - the rpc stat version of the caller
6886 * OUT myVersion - the rpc stat version of this function
6888 * OUT clock_sec - local time seconds
6890 * OUT clock_usec - local time microseconds
6892 * OUT allocSize - the number of bytes allocated to contain stats
6894 * OUT statCount - the number stats retrieved from this process.
6896 * OUT stats - the actual stats retrieved from this process.
6900 * Returns void. If successful, stats will != NULL.
6904 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
6905 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
6906 size_t * allocSize, afs_uint32 * statCount,
6907 afs_uint32 ** stats)
6917 *myVersion = RX_STATS_RETRIEVAL_VERSION;
6920 * Check to see if stats are enabled
6923 MUTEX_ENTER(&rx_rpc_stats);
6924 if (!rxi_monitor_processStats) {
6925 MUTEX_EXIT(&rx_rpc_stats);
6929 clock_GetTime(&now);
6930 *clock_sec = now.sec;
6931 *clock_usec = now.usec;
6934 * Allocate the space based upon the caller version
6936 * If the client is at an older version than we are,
6937 * we return the statistic data in the older data format, but
6938 * we still return our version number so the client knows we
6939 * are maintaining more data than it can retrieve.
6942 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
6943 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
6944 *statCount = rxi_rpc_process_stat_cnt;
6947 * This can't happen yet, but in the future version changes
6948 * can be handled by adding additional code here
6952 if (space > (size_t) 0) {
6954 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
6957 rx_interface_stat_p rpc_stat, nrpc_stat;
6961 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6963 * Copy the data based upon the caller version
6965 rx_MarshallProcessRPCStats(callerVersion,
6966 rpc_stat->stats[0].func_total,
6967 rpc_stat->stats, &ptr);
6973 MUTEX_EXIT(&rx_rpc_stats);
6978 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
6982 * IN callerVersion - the rpc stat version of the caller
6984 * OUT myVersion - the rpc stat version of this function
6986 * OUT clock_sec - local time seconds
6988 * OUT clock_usec - local time microseconds
6990 * OUT allocSize - the number of bytes allocated to contain stats
6992 * OUT statCount - the number of stats retrieved from the individual
6995 * OUT stats - the actual stats retrieved from the individual peer structures.
6999 * Returns void. If successful, stats will != NULL.
7003 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7004 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7005 size_t * allocSize, afs_uint32 * statCount,
7006 afs_uint32 ** stats)
7016 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7019 * Check to see if stats are enabled
7022 MUTEX_ENTER(&rx_rpc_stats);
7023 if (!rxi_monitor_peerStats) {
7024 MUTEX_EXIT(&rx_rpc_stats);
7028 clock_GetTime(&now);
7029 *clock_sec = now.sec;
7030 *clock_usec = now.usec;
7033 * Allocate the space based upon the caller version
7035 * If the client is at an older version than we are,
7036 * we return the statistic data in the older data format, but
7037 * we still return our version number so the client knows we
7038 * are maintaining more data than it can retrieve.
7041 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7042 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7043 *statCount = rxi_rpc_peer_stat_cnt;
7046 * This can't happen yet, but in the future version changes
7047 * can be handled by adding additional code here
7051 if (space > (size_t) 0) {
7053 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7056 rx_interface_stat_p rpc_stat, nrpc_stat;
7060 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7062 * We have to fix the offset of rpc_stat since we are
7063 * keeping this structure on two rx_queues. The rx_queue
7064 * package assumes that the rx_queue member is the first
7065 * member of the structure. That is, rx_queue assumes that
7066 * any one item is only on one queue at a time. We are
7067 * breaking that assumption and so we have to do a little
7068 * math to fix our pointers.
7071 fix_offset = (char *)rpc_stat;
7072 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7073 rpc_stat = (rx_interface_stat_p) fix_offset;
7076 * Copy the data based upon the caller version
7078 rx_MarshallProcessRPCStats(callerVersion,
7079 rpc_stat->stats[0].func_total,
7080 rpc_stat->stats, &ptr);
7086 MUTEX_EXIT(&rx_rpc_stats);
7091 * rx_FreeRPCStats - free memory allocated by
7092 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7096 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7097 * rx_RetrievePeerRPCStats
7099 * IN allocSize - the number of bytes in stats.
7107 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7109 rxi_Free(stats, allocSize);
7113 * rx_queryProcessRPCStats - see if process rpc stat collection is
7114 * currently enabled.
7120 * Returns 0 if stats are not enabled != 0 otherwise
7124 rx_queryProcessRPCStats(void)
7127 MUTEX_ENTER(&rx_rpc_stats);
7128 rc = rxi_monitor_processStats;
7129 MUTEX_EXIT(&rx_rpc_stats);
7134 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7140 * Returns 0 if stats are not enabled != 0 otherwise
7144 rx_queryPeerRPCStats(void)
7147 MUTEX_ENTER(&rx_rpc_stats);
7148 rc = rxi_monitor_peerStats;
7149 MUTEX_EXIT(&rx_rpc_stats);
7154 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7164 rx_enableProcessRPCStats(void)
7166 MUTEX_ENTER(&rx_rpc_stats);
7167 rx_enable_stats = 1;
7168 rxi_monitor_processStats = 1;
7169 MUTEX_EXIT(&rx_rpc_stats);
7173 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7183 rx_enablePeerRPCStats(void)
7185 MUTEX_ENTER(&rx_rpc_stats);
7186 rx_enable_stats = 1;
7187 rxi_monitor_peerStats = 1;
7188 MUTEX_EXIT(&rx_rpc_stats);
7192 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7202 rx_disableProcessRPCStats(void)
7204 rx_interface_stat_p rpc_stat, nrpc_stat;
7207 MUTEX_ENTER(&rx_rpc_stats);
7210 * Turn off process statistics and if peer stats is also off, turn
7214 rxi_monitor_processStats = 0;
7215 if (rxi_monitor_peerStats == 0) {
7216 rx_enable_stats = 0;
7219 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7220 unsigned int num_funcs = 0;
7223 queue_Remove(rpc_stat);
7224 num_funcs = rpc_stat->stats[0].func_total;
7226 sizeof(rx_interface_stat_t) +
7227 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7229 rxi_Free(rpc_stat, space);
7230 rxi_rpc_process_stat_cnt -= num_funcs;
7232 MUTEX_EXIT(&rx_rpc_stats);
7236 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7246 rx_disablePeerRPCStats(void)
7248 struct rx_peer **peer_ptr, **peer_end;
7251 MUTEX_ENTER(&rx_rpc_stats);
7254 * Turn off peer statistics and if process stats is also off, turn
7258 rxi_monitor_peerStats = 0;
7259 if (rxi_monitor_processStats == 0) {
7260 rx_enable_stats = 0;
7263 MUTEX_ENTER(&rx_peerHashTable_lock);
7264 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7265 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7267 struct rx_peer *peer, *next, *prev;
7268 for (prev = peer = *peer_ptr; peer; peer = next) {
7270 code = MUTEX_TRYENTER(&peer->peer_lock);
7272 rx_interface_stat_p rpc_stat, nrpc_stat;
7275 (&peer->rpcStats, rpc_stat, nrpc_stat,
7276 rx_interface_stat)) {
7277 unsigned int num_funcs = 0;
7280 queue_Remove(&rpc_stat->queue_header);
7281 queue_Remove(&rpc_stat->all_peers);
7282 num_funcs = rpc_stat->stats[0].func_total;
7284 sizeof(rx_interface_stat_t) +
7285 rpc_stat->stats[0].func_total *
7286 sizeof(rx_function_entry_v1_t);
7288 rxi_Free(rpc_stat, space);
7289 rxi_rpc_peer_stat_cnt -= num_funcs;
7291 MUTEX_EXIT(&peer->peer_lock);
7292 if (prev == *peer_ptr) {
7302 MUTEX_EXIT(&rx_peerHashTable_lock);
7303 MUTEX_EXIT(&rx_rpc_stats);
7307 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7312 * IN clearFlag - flag indicating which stats to clear
7320 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7322 rx_interface_stat_p rpc_stat, nrpc_stat;
7324 MUTEX_ENTER(&rx_rpc_stats);
7326 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7327 unsigned int num_funcs = 0, i;
7328 num_funcs = rpc_stat->stats[0].func_total;
7329 for (i = 0; i < num_funcs; i++) {
7330 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7331 hzero(rpc_stat->stats[i].invocations);
7333 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7334 hzero(rpc_stat->stats[i].bytes_sent);
7336 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7337 hzero(rpc_stat->stats[i].bytes_rcvd);
7339 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7340 rpc_stat->stats[i].queue_time_sum.sec = 0;
7341 rpc_stat->stats[i].queue_time_sum.usec = 0;
7343 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7344 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7345 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7347 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7348 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7349 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7351 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7352 rpc_stat->stats[i].queue_time_max.sec = 0;
7353 rpc_stat->stats[i].queue_time_max.usec = 0;
7355 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7356 rpc_stat->stats[i].execution_time_sum.sec = 0;
7357 rpc_stat->stats[i].execution_time_sum.usec = 0;
7359 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7360 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7361 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7363 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7364 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7365 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7367 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7368 rpc_stat->stats[i].execution_time_max.sec = 0;
7369 rpc_stat->stats[i].execution_time_max.usec = 0;
7374 MUTEX_EXIT(&rx_rpc_stats);
7378 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7383 * IN clearFlag - flag indicating which stats to clear
7391 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7393 rx_interface_stat_p rpc_stat, nrpc_stat;
7395 MUTEX_ENTER(&rx_rpc_stats);
7397 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7398 unsigned int num_funcs = 0, i;
7401 * We have to fix the offset of rpc_stat since we are
7402 * keeping this structure on two rx_queues. The rx_queue
7403 * package assumes that the rx_queue member is the first
7404 * member of the structure. That is, rx_queue assumes that
7405 * any one item is only on one queue at a time. We are
7406 * breaking that assumption and so we have to do a little
7407 * math to fix our pointers.
7410 fix_offset = (char *)rpc_stat;
7411 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7412 rpc_stat = (rx_interface_stat_p) fix_offset;
7414 num_funcs = rpc_stat->stats[0].func_total;
7415 for (i = 0; i < num_funcs; i++) {
7416 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7417 hzero(rpc_stat->stats[i].invocations);
7419 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7420 hzero(rpc_stat->stats[i].bytes_sent);
7422 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7423 hzero(rpc_stat->stats[i].bytes_rcvd);
7425 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7426 rpc_stat->stats[i].queue_time_sum.sec = 0;
7427 rpc_stat->stats[i].queue_time_sum.usec = 0;
7429 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7430 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7431 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7433 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7434 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7435 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7437 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7438 rpc_stat->stats[i].queue_time_max.sec = 0;
7439 rpc_stat->stats[i].queue_time_max.usec = 0;
7441 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7442 rpc_stat->stats[i].execution_time_sum.sec = 0;
7443 rpc_stat->stats[i].execution_time_sum.usec = 0;
7445 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7446 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7447 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7449 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7450 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7451 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7453 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7454 rpc_stat->stats[i].execution_time_max.sec = 0;
7455 rpc_stat->stats[i].execution_time_max.usec = 0;
7460 MUTEX_EXIT(&rx_rpc_stats);
7464 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7465 * is authorized to enable/disable/clear RX statistics.
7467 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7470 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7472 rxi_rxstat_userok = proc;
7476 rx_RxStatUserOk(struct rx_call *call)
7478 if (!rxi_rxstat_userok)
7480 return rxi_rxstat_userok(call);