2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
14 #include "afs/param.h"
16 #include <afs/param.h>
23 #include "afs/sysincludes.h"
24 #include "afsincludes.h"
30 #include <net/net_globals.h>
31 #endif /* AFS_OSF_ENV */
32 #ifdef AFS_LINUX20_ENV
35 #include "netinet/in.h"
36 #include "afs/afs_args.h"
37 #include "afs/afs_osi.h"
38 #ifdef RX_KERNEL_TRACE
39 #include "rx_kcommon.h"
41 #if (defined(AFS_AUX_ENV) || defined(AFS_AIX_ENV))
45 #undef RXDEBUG /* turn off debugging */
47 #if defined(AFS_SGI_ENV)
48 #include "sys/debug.h"
57 #endif /* AFS_ALPHA_ENV */
59 #include "afs/sysincludes.h"
60 #include "afsincludes.h"
63 #include "rx_kmutex.h"
64 #include "rx_kernel.h"
68 #include "rx_globals.h"
70 #define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
71 #define AFSOP_STOP_AFS 211 /* Stop AFS process */
72 #define AFSOP_STOP_BKG 212 /* Stop BKG process */
74 extern afs_int32 afs_termState;
76 #include "sys/lockl.h"
77 #include "sys/lock_def.h"
78 #endif /* AFS_AIX41_ENV */
79 # include "rxgen_consts.h"
81 # include <sys/types.h>
86 # include <afs/afsutil.h>
88 # include <sys/socket.h>
89 # include <sys/file.h>
91 # include <sys/stat.h>
92 # include <netinet/in.h>
93 # include <sys/time.h>
103 # include "rx_user.h"
104 # include "rx_clock.h"
105 # include "rx_queue.h"
106 # include "rx_globals.h"
107 # include "rx_trace.h"
108 # include <afs/rxgen_consts.h>
111 int (*registerProgram) () = 0;
112 int (*swapNameProgram) () = 0;
114 /* Local static routines */
115 static void rxi_DestroyConnectionNoLock(register struct rx_connection *conn);
116 #ifdef RX_ENABLE_LOCKS
117 static void rxi_SetAcksInTransmitQueue(register struct rx_call *call);
120 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
122 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
123 afs_int32 rxi_start_in_error;
125 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
128 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
129 * currently allocated within rx. This number is used to allocate the
130 * memory required to return the statistics when queried.
133 static unsigned int rxi_rpc_peer_stat_cnt;
136 * rxi_rpc_process_stat_cnt counts the total number of local process stat
137 * structures currently allocated within rx. The number is used to allocate
138 * the memory required to return the statistics when queried.
141 static unsigned int rxi_rpc_process_stat_cnt;
143 #if !defined(offsetof)
144 #include <stddef.h> /* for definition of offsetof() */
147 #ifdef AFS_PTHREAD_ENV
151 * Use procedural initialization of mutexes/condition variables
155 extern pthread_mutex_t rx_stats_mutex;
156 extern pthread_mutex_t rxkad_stats_mutex;
157 extern pthread_mutex_t des_init_mutex;
158 extern pthread_mutex_t des_random_mutex;
159 extern pthread_mutex_t rx_clock_mutex;
160 extern pthread_mutex_t rxi_connCacheMutex;
161 extern pthread_mutex_t rx_event_mutex;
162 extern pthread_mutex_t osi_malloc_mutex;
163 extern pthread_mutex_t event_handler_mutex;
164 extern pthread_mutex_t listener_mutex;
165 extern pthread_mutex_t rx_if_init_mutex;
166 extern pthread_mutex_t rx_if_mutex;
167 extern pthread_mutex_t rxkad_client_uid_mutex;
168 extern pthread_mutex_t rxkad_random_mutex;
170 extern pthread_cond_t rx_event_handler_cond;
171 extern pthread_cond_t rx_listener_cond;
173 static pthread_mutex_t epoch_mutex;
174 static pthread_mutex_t rx_init_mutex;
175 static pthread_mutex_t rx_debug_mutex;
178 rxi_InitPthread(void)
180 assert(pthread_mutex_init(&rx_clock_mutex, (const pthread_mutexattr_t *)0)
182 assert(pthread_mutex_init(&rx_stats_mutex, (const pthread_mutexattr_t *)0)
184 assert(pthread_mutex_init
185 (&rxi_connCacheMutex, (const pthread_mutexattr_t *)0) == 0);
186 assert(pthread_mutex_init(&rx_init_mutex, (const pthread_mutexattr_t *)0)
188 assert(pthread_mutex_init(&epoch_mutex, (const pthread_mutexattr_t *)0) ==
190 assert(pthread_mutex_init(&rx_event_mutex, (const pthread_mutexattr_t *)0)
192 assert(pthread_mutex_init(&des_init_mutex, (const pthread_mutexattr_t *)0)
194 assert(pthread_mutex_init
195 (&des_random_mutex, (const pthread_mutexattr_t *)0) == 0);
196 assert(pthread_mutex_init
197 (&osi_malloc_mutex, (const pthread_mutexattr_t *)0) == 0);
198 assert(pthread_mutex_init
199 (&event_handler_mutex, (const pthread_mutexattr_t *)0) == 0);
200 assert(pthread_mutex_init(&listener_mutex, (const pthread_mutexattr_t *)0)
202 assert(pthread_mutex_init
203 (&rx_if_init_mutex, (const pthread_mutexattr_t *)0) == 0);
204 assert(pthread_mutex_init(&rx_if_mutex, (const pthread_mutexattr_t *)0) ==
206 assert(pthread_mutex_init
207 (&rxkad_client_uid_mutex, (const pthread_mutexattr_t *)0) == 0);
208 assert(pthread_mutex_init
209 (&rxkad_random_mutex, (const pthread_mutexattr_t *)0) == 0);
210 assert(pthread_mutex_init
211 (&rxkad_stats_mutex, (const pthread_mutexattr_t *)0) == 0);
212 assert(pthread_mutex_init(&rx_debug_mutex, (const pthread_mutexattr_t *)0)
215 assert(pthread_cond_init
216 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
217 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
219 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
222 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
223 #define INIT_PTHREAD_LOCKS \
224 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0);
226 * The rx_stats_mutex mutex protects the following global variables:
231 * rxi_lowConnRefCount
232 * rxi_lowPeerRefCount
241 #define INIT_PTHREAD_LOCKS
245 /* Variables for handling the minProcs implementation. availProcs gives the
246 * number of threads available in the pool at this moment (not counting dudes
247 * executing right now). totalMin gives the total number of procs required
248 * for handling all minProcs requests. minDeficit is a dynamic variable
249 * tracking the # of procs required to satisfy all of the remaining minProcs
251 * For fine grain locking to work, the quota check and the reservation of
252 * a server thread has to come while rxi_availProcs and rxi_minDeficit
253 * are locked. To this end, the code has been modified under #ifdef
254 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
255 * same time. A new function, ReturnToServerPool() returns the allocation.
257 * A call can be on several queue's (but only one at a time). When
258 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
259 * that no one else is touching the queue. To this end, we store the address
260 * of the queue lock in the call structure (under the call lock) when we
261 * put the call on a queue, and we clear the call_queue_lock when the
262 * call is removed from a queue (once the call lock has been obtained).
263 * This allows rxi_ResetCall to safely synchronize with others wishing
264 * to manipulate the queue.
267 #ifdef RX_ENABLE_LOCKS
268 static afs_kmutex_t rx_rpc_stats;
269 void rxi_StartUnlocked();
272 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
273 ** pretty good that the next packet coming in is from the same connection
274 ** as the last packet, since we're send multiple packets in a transmit window.
276 struct rx_connection *rxLastConn = 0;
278 #ifdef RX_ENABLE_LOCKS
279 /* The locking hierarchy for rx fine grain locking is composed of these
282 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
283 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
284 * call->lock - locks call data fields.
285 * These are independent of each other:
286 * rx_freeCallQueue_lock
291 * serverQueueEntry->lock
293 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
294 * peer->lock - locks peer data fields.
295 * conn_data_lock - that more than one thread is not updating a conn data
296 * field at the same time.
304 * Do we need a lock to protect the peer field in the conn structure?
305 * conn->peer was previously a constant for all intents and so has no
306 * lock protecting this field. The multihomed client delta introduced
307 * a RX code change : change the peer field in the connection structure
308 * to that remote inetrface from which the last packet for this
309 * connection was sent out. This may become an issue if further changes
312 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
313 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
315 /* rxdb_fileID is used to identify the lock location, along with line#. */
316 static int rxdb_fileID = RXDB_FILE_RX;
317 #endif /* RX_LOCKS_DB */
318 #else /* RX_ENABLE_LOCKS */
319 #define SET_CALL_QUEUE_LOCK(C, L)
320 #define CLEAR_CALL_QUEUE_LOCK(C)
321 #endif /* RX_ENABLE_LOCKS */
322 struct rx_serverQueueEntry *rx_waitForPacket = 0;
323 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
325 /* ------------Exported Interfaces------------- */
327 /* This function allows rxkad to set the epoch to a suitably random number
328 * which rx_NewConnection will use in the future. The principle purpose is to
329 * get rxnull connections to use the same epoch as the rxkad connections do, at
330 * least once the first rxkad connection is established. This is important now
331 * that the host/port addresses aren't used in FindConnection: the uniqueness
332 * of epoch/cid matters and the start time won't do. */
334 #ifdef AFS_PTHREAD_ENV
336 * This mutex protects the following global variables:
340 #define LOCK_EPOCH assert(pthread_mutex_lock(&epoch_mutex)==0);
341 #define UNLOCK_EPOCH assert(pthread_mutex_unlock(&epoch_mutex)==0);
345 #endif /* AFS_PTHREAD_ENV */
348 rx_SetEpoch(afs_uint32 epoch)
350 LOCK_EPOCH rx_epoch = epoch;
353 /* Initialize rx. A port number may be mentioned, in which case this
354 * becomes the default port number for any service installed later.
355 * If 0 is provided for the port number, a random port will be chosen
356 * by the kernel. Whether this will ever overlap anything in
357 * /etc/services is anybody's guess... Returns 0 on success, -1 on
359 static int rxinit_status = 1;
360 #ifdef AFS_PTHREAD_ENV
362 * This mutex protects the following global variables:
366 #define LOCK_RX_INIT assert(pthread_mutex_lock(&rx_init_mutex)==0);
367 #define UNLOCK_RX_INIT assert(pthread_mutex_unlock(&rx_init_mutex)==0);
370 #define UNLOCK_RX_INIT
374 rx_InitHost(u_int host, u_int port)
381 char *htable, *ptable;
384 #if defined(AFS_DJGPP_ENV) && !defined(DEBUG)
385 __djgpp_set_quiet_socket(1);
390 INIT_PTHREAD_LOCKS LOCK_RX_INIT if (rxinit_status == 0) {
391 tmp_status = rxinit_status;
392 UNLOCK_RX_INIT return tmp_status; /* Already started; return previous error code. */
395 if (afs_winsockInit() < 0)
401 * Initialize anything necessary to provide a non-premptive threading
404 rxi_InitializeThreadSupport();
407 /* Allocate and initialize a socket for client and perhaps server
410 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
411 if (rx_socket == OSI_NULLSOCKET) {
412 UNLOCK_RX_INIT return RX_ADDRINUSE;
414 #ifdef RX_ENABLE_LOCKS
417 #endif /* RX_LOCKS_DB */
418 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
419 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
420 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
421 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
422 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
424 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
426 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
428 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
430 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
432 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
434 #if defined(KERNEL) && defined(AFS_HPUX110_ENV)
436 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
437 #endif /* KERNEL && AFS_HPUX110_ENV */
438 #else /* RX_ENABLE_LOCKS */
439 #if defined(KERNEL) && defined(AFS_GLOBAL_SUNLOCK) && !defined(AFS_HPUX_ENV) && !defined(AFS_OBSD_ENV)
440 mutex_init(&afs_rxglobal_lock, "afs_rxglobal_lock", MUTEX_DEFAULT, NULL);
441 #endif /* AFS_GLOBAL_SUNLOCK */
442 #endif /* RX_ENABLE_LOCKS */
445 rx_connDeadTime = 12;
446 rx_tranquil = 0; /* reset flag */
447 memset((char *)&rx_stats, 0, sizeof(struct rx_stats));
449 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
450 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
451 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
452 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
453 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
454 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
456 /* Malloc up a bunch of packets & buffers */
458 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
459 queue_Init(&rx_freePacketQueue);
460 rxi_NeedMorePackets = FALSE;
461 rxi_MorePackets(rx_nPackets);
469 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
470 tv.tv_sec = clock_now.sec;
471 tv.tv_usec = clock_now.usec;
472 srand((unsigned int)tv.tv_usec);
479 #if defined(KERNEL) && !defined(UKERNEL)
480 /* Really, this should never happen in a real kernel */
483 struct sockaddr_in addr;
484 int addrlen = sizeof(addr);
485 if (getsockname((int)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
489 rx_port = addr.sin_port;
492 rx_stats.minRtt.sec = 9999999;
494 rx_SetEpoch(tv.tv_sec | 0x80000000);
496 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
497 * will provide a randomer value. */
499 MUTEX_ENTER(&rx_stats_mutex);
500 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
501 MUTEX_EXIT(&rx_stats_mutex);
502 /* *Slightly* random start time for the cid. This is just to help
503 * out with the hashing function at the peer */
504 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
505 rx_connHashTable = (struct rx_connection **)htable;
506 rx_peerHashTable = (struct rx_peer **)ptable;
508 rx_lastAckDelay.sec = 0;
509 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
510 rx_hardAckDelay.sec = 0;
511 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
512 rx_softAckDelay.sec = 0;
513 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
515 rxevent_Init(20, rxi_ReScheduleEvents);
517 /* Initialize various global queues */
518 queue_Init(&rx_idleServerQueue);
519 queue_Init(&rx_incomingCallQueue);
520 queue_Init(&rx_freeCallQueue);
522 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
523 /* Initialize our list of usable IP addresses. */
527 /* Start listener process (exact function is dependent on the
528 * implementation environment--kernel or user space) */
533 tmp_status = rxinit_status = 0;
534 UNLOCK_RX_INIT return tmp_status;
537 int rx_Init(u_int port)
539 return rx_InitHost(htonl(INADDR_ANY), port);
542 /* called with unincremented nRequestsRunning to see if it is OK to start
543 * a new thread in this service. Could be "no" for two reasons: over the
544 * max quota, or would prevent others from reaching their min quota.
546 #ifdef RX_ENABLE_LOCKS
547 /* This verion of QuotaOK reserves quota if it's ok while the
548 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
551 QuotaOK(register struct rx_service *aservice)
553 /* check if over max quota */
554 if (aservice->nRequestsRunning >= aservice->maxProcs) {
558 /* under min quota, we're OK */
559 /* otherwise, can use only if there are enough to allow everyone
560 * to go to their min quota after this guy starts.
562 MUTEX_ENTER(&rx_stats_mutex);
563 if ((aservice->nRequestsRunning < aservice->minProcs)
564 || (rxi_availProcs > rxi_minDeficit)) {
565 aservice->nRequestsRunning++;
566 /* just started call in minProcs pool, need fewer to maintain
568 if (aservice->nRequestsRunning <= aservice->minProcs)
571 MUTEX_EXIT(&rx_stats_mutex);
574 MUTEX_EXIT(&rx_stats_mutex);
580 ReturnToServerPool(register struct rx_service *aservice)
582 aservice->nRequestsRunning--;
583 MUTEX_ENTER(&rx_stats_mutex);
584 if (aservice->nRequestsRunning < aservice->minProcs)
587 MUTEX_EXIT(&rx_stats_mutex);
590 #else /* RX_ENABLE_LOCKS */
592 QuotaOK(register struct rx_service *aservice)
595 /* under min quota, we're OK */
596 if (aservice->nRequestsRunning < aservice->minProcs)
599 /* check if over max quota */
600 if (aservice->nRequestsRunning >= aservice->maxProcs)
603 /* otherwise, can use only if there are enough to allow everyone
604 * to go to their min quota after this guy starts.
606 if (rxi_availProcs > rxi_minDeficit)
610 #endif /* RX_ENABLE_LOCKS */
613 /* Called by rx_StartServer to start up lwp's to service calls.
614 NExistingProcs gives the number of procs already existing, and which
615 therefore needn't be created. */
617 rxi_StartServerProcs(int nExistingProcs)
619 register struct rx_service *service;
624 /* For each service, reserve N processes, where N is the "minimum"
625 * number of processes that MUST be able to execute a request in parallel,
626 * at any time, for that process. Also compute the maximum difference
627 * between any service's maximum number of processes that can run
628 * (i.e. the maximum number that ever will be run, and a guarantee
629 * that this number will run if other services aren't running), and its
630 * minimum number. The result is the extra number of processes that
631 * we need in order to provide the latter guarantee */
632 for (i = 0; i < RX_MAX_SERVICES; i++) {
634 service = rx_services[i];
635 if (service == (struct rx_service *)0)
637 nProcs += service->minProcs;
638 diff = service->maxProcs - service->minProcs;
642 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
643 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
644 for (i = 0; i < nProcs; i++) {
645 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
650 /* This routine must be called if any services are exported. If the
651 * donateMe flag is set, the calling process is donated to the server
654 rx_StartServer(int donateMe)
656 register struct rx_service *service;
657 register int i, nProcs = 0;
663 /* Start server processes, if necessary (exact function is dependent
664 * on the implementation environment--kernel or user space). DonateMe
665 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
666 * case, one less new proc will be created rx_StartServerProcs.
668 rxi_StartServerProcs(donateMe);
670 /* count up the # of threads in minProcs, and add set the min deficit to
671 * be that value, too.
673 for (i = 0; i < RX_MAX_SERVICES; i++) {
674 service = rx_services[i];
675 if (service == (struct rx_service *)0)
677 MUTEX_ENTER(&rx_stats_mutex);
678 rxi_totalMin += service->minProcs;
679 /* below works even if a thread is running, since minDeficit would
680 * still have been decremented and later re-incremented.
682 rxi_minDeficit += service->minProcs;
683 MUTEX_EXIT(&rx_stats_mutex);
686 /* Turn on reaping of idle server connections */
687 rxi_ReapConnections();
696 #ifdef AFS_PTHREAD_ENV
698 pid = (pid_t) pthread_self();
699 #else /* AFS_PTHREAD_ENV */
701 LWP_CurrentProcess(&pid);
702 #endif /* AFS_PTHREAD_ENV */
704 sprintf(name, "srv_%d", ++nProcs);
706 (*registerProgram) (pid, name);
708 #endif /* AFS_NT40_ENV */
709 rx_ServerProc(); /* Never returns */
714 /* Create a new client connection to the specified service, using the
715 * specified security object to implement the security model for this
717 struct rx_connection *
718 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
719 register struct rx_securityClass *securityObject,
720 int serviceSecurityIndex)
724 register struct rx_connection *conn;
729 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", shost, sport, sservice, securityObject, serviceSecurityIndex));
731 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
732 * the case of kmem_alloc? */
733 conn = rxi_AllocConnection();
734 #ifdef RX_ENABLE_LOCKS
735 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
736 MUTEX_INIT(&conn->conn_data_lock, "conn call lock", MUTEX_DEFAULT, 0);
737 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
741 MUTEX_ENTER(&rx_connHashTable_lock);
742 cid = (rx_nextCid += RX_MAXCALLS);
743 conn->type = RX_CLIENT_CONNECTION;
745 conn->epoch = rx_epoch;
746 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
747 conn->serviceId = sservice;
748 conn->securityObject = securityObject;
749 /* This doesn't work in all compilers with void (they're buggy), so fake it
751 conn->securityData = (VOID *) 0;
752 conn->securityIndex = serviceSecurityIndex;
753 rx_SetConnDeadTime(conn, rx_connDeadTime);
754 conn->ackRate = RX_FAST_ACK_RATE;
756 conn->specific = NULL;
757 conn->challengeEvent = NULL;
758 conn->delayedAbortEvent = NULL;
759 conn->abortCount = 0;
762 RXS_NewConnection(securityObject, conn);
764 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
766 conn->refCount++; /* no lock required since only this thread knows... */
767 conn->next = rx_connHashTable[hashindex];
768 rx_connHashTable[hashindex] = conn;
769 MUTEX_ENTER(&rx_stats_mutex);
770 rx_stats.nClientConns++;
771 MUTEX_EXIT(&rx_stats_mutex);
773 MUTEX_EXIT(&rx_connHashTable_lock);
780 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
782 /* The idea is to set the dead time to a value that allows several
783 * keepalives to be dropped without timing out the connection. */
784 conn->secondsUntilDead = MAX(seconds, 6);
785 conn->secondsUntilPing = conn->secondsUntilDead / 6;
788 int rxi_lowPeerRefCount = 0;
789 int rxi_lowConnRefCount = 0;
792 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
793 * NOTE: must not be called with rx_connHashTable_lock held.
796 rxi_CleanupConnection(struct rx_connection *conn)
798 /* Notify the service exporter, if requested, that this connection
799 * is being destroyed */
800 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
801 (*conn->service->destroyConnProc) (conn);
803 /* Notify the security module that this connection is being destroyed */
804 RXS_DestroyConnection(conn->securityObject, conn);
806 /* If this is the last connection using the rx_peer struct, set its
807 * idle time to now. rxi_ReapConnections will reap it if it's still
808 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
810 MUTEX_ENTER(&rx_peerHashTable_lock);
811 if (--conn->peer->refCount <= 0) {
812 conn->peer->idleWhen = clock_Sec();
813 if (conn->peer->refCount < 0) {
814 conn->peer->refCount = 0;
815 MUTEX_ENTER(&rx_stats_mutex);
816 rxi_lowPeerRefCount++;
817 MUTEX_EXIT(&rx_stats_mutex);
820 MUTEX_EXIT(&rx_peerHashTable_lock);
822 MUTEX_ENTER(&rx_stats_mutex);
823 if (conn->type == RX_SERVER_CONNECTION)
824 rx_stats.nServerConns--;
826 rx_stats.nClientConns--;
827 MUTEX_EXIT(&rx_stats_mutex);
830 if (conn->specific) {
832 for (i = 0; i < conn->nSpecific; i++) {
833 if (conn->specific[i] && rxi_keyCreate_destructor[i])
834 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
835 conn->specific[i] = NULL;
837 free(conn->specific);
839 conn->specific = NULL;
843 MUTEX_DESTROY(&conn->conn_call_lock);
844 MUTEX_DESTROY(&conn->conn_data_lock);
845 CV_DESTROY(&conn->conn_call_cv);
847 rxi_FreeConnection(conn);
850 /* Destroy the specified connection */
852 rxi_DestroyConnection(register struct rx_connection *conn)
854 MUTEX_ENTER(&rx_connHashTable_lock);
855 rxi_DestroyConnectionNoLock(conn);
856 /* conn should be at the head of the cleanup list */
857 if (conn == rx_connCleanup_list) {
858 rx_connCleanup_list = rx_connCleanup_list->next;
859 MUTEX_EXIT(&rx_connHashTable_lock);
860 rxi_CleanupConnection(conn);
862 #ifdef RX_ENABLE_LOCKS
864 MUTEX_EXIT(&rx_connHashTable_lock);
866 #endif /* RX_ENABLE_LOCKS */
870 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
872 register struct rx_connection **conn_ptr;
873 register int havecalls = 0;
874 struct rx_packet *packet;
881 MUTEX_ENTER(&conn->conn_data_lock);
882 if (conn->refCount > 0)
885 MUTEX_ENTER(&rx_stats_mutex);
886 rxi_lowConnRefCount++;
887 MUTEX_EXIT(&rx_stats_mutex);
890 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
891 /* Busy; wait till the last guy before proceeding */
892 MUTEX_EXIT(&conn->conn_data_lock);
897 /* If the client previously called rx_NewCall, but it is still
898 * waiting, treat this as a running call, and wait to destroy the
899 * connection later when the call completes. */
900 if ((conn->type == RX_CLIENT_CONNECTION)
901 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
902 conn->flags |= RX_CONN_DESTROY_ME;
903 MUTEX_EXIT(&conn->conn_data_lock);
907 MUTEX_EXIT(&conn->conn_data_lock);
909 /* Check for extant references to this connection */
910 for (i = 0; i < RX_MAXCALLS; i++) {
911 register struct rx_call *call = conn->call[i];
914 if (conn->type == RX_CLIENT_CONNECTION) {
915 MUTEX_ENTER(&call->lock);
916 if (call->delayedAckEvent) {
917 /* Push the final acknowledgment out now--there
918 * won't be a subsequent call to acknowledge the
919 * last reply packets */
920 rxevent_Cancel(call->delayedAckEvent, call,
921 RX_CALL_REFCOUNT_DELAY);
922 if (call->state == RX_STATE_PRECALL
923 || call->state == RX_STATE_ACTIVE) {
924 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
926 rxi_AckAll(NULL, call, 0);
929 MUTEX_EXIT(&call->lock);
933 #ifdef RX_ENABLE_LOCKS
935 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
936 MUTEX_EXIT(&conn->conn_data_lock);
938 /* Someone is accessing a packet right now. */
942 #endif /* RX_ENABLE_LOCKS */
945 /* Don't destroy the connection if there are any call
946 * structures still in use */
947 MUTEX_ENTER(&conn->conn_data_lock);
948 conn->flags |= RX_CONN_DESTROY_ME;
949 MUTEX_EXIT(&conn->conn_data_lock);
954 if (conn->delayedAbortEvent) {
955 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
956 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
958 MUTEX_ENTER(&conn->conn_data_lock);
959 rxi_SendConnectionAbort(conn, packet, 0, 1);
960 MUTEX_EXIT(&conn->conn_data_lock);
961 rxi_FreePacket(packet);
965 /* Remove from connection hash table before proceeding */
967 &rx_connHashTable[CONN_HASH
968 (peer->host, peer->port, conn->cid, conn->epoch,
970 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
971 if (*conn_ptr == conn) {
972 *conn_ptr = conn->next;
976 /* if the conn that we are destroying was the last connection, then we
977 * clear rxLastConn as well */
978 if (rxLastConn == conn)
981 /* Make sure the connection is completely reset before deleting it. */
982 /* get rid of pending events that could zap us later */
983 if (conn->challengeEvent)
984 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
985 if (conn->checkReachEvent)
986 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
988 /* Add the connection to the list of destroyed connections that
989 * need to be cleaned up. This is necessary to avoid deadlocks
990 * in the routines we call to inform others that this connection is
991 * being destroyed. */
992 conn->next = rx_connCleanup_list;
993 rx_connCleanup_list = conn;
996 /* Externally available version */
998 rx_DestroyConnection(register struct rx_connection *conn)
1004 rxi_DestroyConnection(conn);
1009 /* Start a new rx remote procedure call, on the specified connection.
1010 * If wait is set to 1, wait for a free call channel; otherwise return
1011 * 0. Maxtime gives the maximum number of seconds this call may take,
1012 * after rx_MakeCall returns. After this time interval, a call to any
1013 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1014 * For fine grain locking, we hold the conn_call_lock in order to
1015 * to ensure that we don't get signalle after we found a call in an active
1016 * state and before we go to sleep.
1019 rx_NewCall(register struct rx_connection *conn)
1022 register struct rx_call *call;
1023 struct clock queueTime;
1027 dpf(("rx_MakeCall(conn %x)\n", conn));
1030 clock_GetTime(&queueTime);
1032 MUTEX_ENTER(&conn->conn_call_lock);
1035 * Check if there are others waiting for a new call.
1036 * If so, let them go first to avoid starving them.
1037 * This is a fairly simple scheme, and might not be
1038 * a complete solution for large numbers of waiters.
1040 if (conn->makeCallWaiters) {
1041 #ifdef RX_ENABLE_LOCKS
1042 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1049 for (i = 0; i < RX_MAXCALLS; i++) {
1050 call = conn->call[i];
1052 MUTEX_ENTER(&call->lock);
1053 if (call->state == RX_STATE_DALLY) {
1054 rxi_ResetCall(call, 0);
1055 (*call->callNumber)++;
1058 MUTEX_EXIT(&call->lock);
1060 call = rxi_NewCall(conn, i);
1064 if (i < RX_MAXCALLS) {
1067 MUTEX_ENTER(&conn->conn_data_lock);
1068 conn->flags |= RX_CONN_MAKECALL_WAITING;
1069 MUTEX_EXIT(&conn->conn_data_lock);
1071 conn->makeCallWaiters++;
1072 #ifdef RX_ENABLE_LOCKS
1073 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1077 conn->makeCallWaiters--;
1080 * Wake up anyone else who might be giving us a chance to
1081 * run (see code above that avoids resource starvation).
1083 #ifdef RX_ENABLE_LOCKS
1084 CV_BROADCAST(&conn->conn_call_cv);
1089 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1091 /* Client is initially in send mode */
1092 call->state = RX_STATE_ACTIVE;
1093 call->mode = RX_MODE_SENDING;
1095 /* remember start time for call in case we have hard dead time limit */
1096 call->queueTime = queueTime;
1097 clock_GetTime(&call->startTime);
1098 hzero(call->bytesSent);
1099 hzero(call->bytesRcvd);
1101 /* Turn on busy protocol. */
1102 rxi_KeepAliveOn(call);
1104 MUTEX_EXIT(&call->lock);
1105 MUTEX_EXIT(&conn->conn_call_lock);
1109 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1110 /* Now, if TQ wasn't cleared earlier, do it now. */
1112 MUTEX_ENTER(&call->lock);
1113 while (call->flags & RX_CALL_TQ_BUSY) {
1114 call->flags |= RX_CALL_TQ_WAIT;
1115 #ifdef RX_ENABLE_LOCKS
1116 CV_WAIT(&call->cv_tq, &call->lock);
1117 #else /* RX_ENABLE_LOCKS */
1118 osi_rxSleep(&call->tq);
1119 #endif /* RX_ENABLE_LOCKS */
1121 if (call->flags & RX_CALL_TQ_CLEARME) {
1122 rxi_ClearTransmitQueue(call, 0);
1123 queue_Init(&call->tq);
1125 MUTEX_EXIT(&call->lock);
1127 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1133 rxi_HasActiveCalls(register struct rx_connection *aconn)
1136 register struct rx_call *tcall;
1140 for (i = 0; i < RX_MAXCALLS; i++) {
1141 if ((tcall = aconn->call[i])) {
1142 if ((tcall->state == RX_STATE_ACTIVE)
1143 || (tcall->state == RX_STATE_PRECALL)) {
1154 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1155 register afs_int32 * aint32s)
1158 register struct rx_call *tcall;
1162 for (i = 0; i < RX_MAXCALLS; i++) {
1163 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1164 aint32s[i] = aconn->callNumber[i] + 1;
1166 aint32s[i] = aconn->callNumber[i];
1173 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1174 register afs_int32 * aint32s)
1177 register struct rx_call *tcall;
1181 for (i = 0; i < RX_MAXCALLS; i++) {
1182 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1183 aconn->callNumber[i] = aint32s[i] - 1;
1185 aconn->callNumber[i] = aint32s[i];
1191 /* Advertise a new service. A service is named locally by a UDP port
1192 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1195 char *serviceName; Name for identification purposes (e.g. the
1196 service name might be used for probing for
1199 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1200 struct rx_securityClass **securityObjects, int nSecurityObjects,
1201 afs_int32(*serviceProc) (struct rx_call * acall))
1203 osi_socket socket = OSI_NULLSOCKET;
1204 register struct rx_service *tservice;
1210 if (serviceId == 0) {
1212 "rx_NewService: service id for service %s is not non-zero.\n",
1219 "rx_NewService: A non-zero port must be specified on this call if a non-zero port was not provided at Rx initialization (service %s).\n",
1227 tservice = rxi_AllocService();
1230 for (i = 0; i < RX_MAX_SERVICES; i++) {
1231 register struct rx_service *service = rx_services[i];
1233 if (port == service->servicePort) {
1234 if (service->serviceId == serviceId) {
1235 /* The identical service has already been
1236 * installed; if the caller was intending to
1237 * change the security classes used by this
1238 * service, he/she loses. */
1240 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1241 serviceName, serviceId, service->serviceName);
1244 rxi_FreeService(tservice);
1247 /* Different service, same port: re-use the socket
1248 * which is bound to the same port */
1249 socket = service->socket;
1252 if (socket == OSI_NULLSOCKET) {
1253 /* If we don't already have a socket (from another
1254 * service on same port) get a new one */
1255 socket = rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
1256 if (socket == OSI_NULLSOCKET) {
1259 rxi_FreeService(tservice);
1264 service->socket = socket;
1265 service->servicePort = port;
1266 service->serviceId = serviceId;
1267 service->serviceName = serviceName;
1268 service->nSecurityObjects = nSecurityObjects;
1269 service->securityObjects = securityObjects;
1270 service->minProcs = 0;
1271 service->maxProcs = 1;
1272 service->idleDeadTime = 60;
1273 service->connDeadTime = rx_connDeadTime;
1274 service->executeRequestProc = serviceProc;
1275 service->checkReach = 0;
1276 rx_services[i] = service; /* not visible until now */
1284 rxi_FreeService(tservice);
1285 (osi_Msg "rx_NewService: cannot support > %d services\n",
1290 /* Generic request processing loop. This routine should be called
1291 * by the implementation dependent rx_ServerProc. If socketp is
1292 * non-null, it will be set to the file descriptor that this thread
1293 * is now listening on. If socketp is null, this routine will never
1296 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1298 register struct rx_call *call;
1299 register afs_int32 code;
1300 register struct rx_service *tservice = NULL;
1307 call = rx_GetCall(threadID, tservice, socketp);
1308 if (socketp && *socketp != OSI_NULLSOCKET) {
1309 /* We are now a listener thread */
1314 /* if server is restarting( typically smooth shutdown) then do not
1315 * allow any new calls.
1318 if (rx_tranquil && (call != NULL)) {
1323 MUTEX_ENTER(&call->lock);
1325 rxi_CallError(call, RX_RESTARTING);
1326 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1328 MUTEX_EXIT(&call->lock);
1333 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1334 #ifdef RX_ENABLE_LOCKS
1336 #endif /* RX_ENABLE_LOCKS */
1337 afs_termState = AFSOP_STOP_AFS;
1338 afs_osi_Wakeup(&afs_termState);
1339 #ifdef RX_ENABLE_LOCKS
1341 #endif /* RX_ENABLE_LOCKS */
1346 tservice = call->conn->service;
1348 if (tservice->beforeProc)
1349 (*tservice->beforeProc) (call);
1351 code = call->conn->service->executeRequestProc(call);
1353 if (tservice->afterProc)
1354 (*tservice->afterProc) (call, code);
1356 rx_EndCall(call, code);
1357 MUTEX_ENTER(&rx_stats_mutex);
1359 MUTEX_EXIT(&rx_stats_mutex);
1365 rx_WakeupServerProcs(void)
1367 struct rx_serverQueueEntry *np, *tqp;
1372 MUTEX_ENTER(&rx_serverPool_lock);
1374 #ifdef RX_ENABLE_LOCKS
1375 if (rx_waitForPacket)
1376 CV_BROADCAST(&rx_waitForPacket->cv);
1377 #else /* RX_ENABLE_LOCKS */
1378 if (rx_waitForPacket)
1379 osi_rxWakeup(rx_waitForPacket);
1380 #endif /* RX_ENABLE_LOCKS */
1381 MUTEX_ENTER(&freeSQEList_lock);
1382 for (np = rx_FreeSQEList; np; np = tqp) {
1383 tqp = *(struct rx_serverQueueEntry **)np;
1384 #ifdef RX_ENABLE_LOCKS
1385 CV_BROADCAST(&np->cv);
1386 #else /* RX_ENABLE_LOCKS */
1388 #endif /* RX_ENABLE_LOCKS */
1390 MUTEX_EXIT(&freeSQEList_lock);
1391 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1392 #ifdef RX_ENABLE_LOCKS
1393 CV_BROADCAST(&np->cv);
1394 #else /* RX_ENABLE_LOCKS */
1396 #endif /* RX_ENABLE_LOCKS */
1398 MUTEX_EXIT(&rx_serverPool_lock);
1404 * One thing that seems to happen is that all the server threads get
1405 * tied up on some empty or slow call, and then a whole bunch of calls
1406 * arrive at once, using up the packet pool, so now there are more
1407 * empty calls. The most critical resources here are server threads
1408 * and the free packet pool. The "doreclaim" code seems to help in
1409 * general. I think that eventually we arrive in this state: there
1410 * are lots of pending calls which do have all their packets present,
1411 * so they won't be reclaimed, are multi-packet calls, so they won't
1412 * be scheduled until later, and thus are tying up most of the free
1413 * packet pool for a very long time.
1415 * 1. schedule multi-packet calls if all the packets are present.
1416 * Probably CPU-bound operation, useful to return packets to pool.
1417 * Do what if there is a full window, but the last packet isn't here?
1418 * 3. preserve one thread which *only* runs "best" calls, otherwise
1419 * it sleeps and waits for that type of call.
1420 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1421 * the current dataquota business is badly broken. The quota isn't adjusted
1422 * to reflect how many packets are presently queued for a running call.
1423 * So, when we schedule a queued call with a full window of packets queued
1424 * up for it, that *should* free up a window full of packets for other 2d-class
1425 * calls to be able to use from the packet pool. But it doesn't.
1427 * NB. Most of the time, this code doesn't run -- since idle server threads
1428 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1429 * as a new call arrives.
1431 /* Sleep until a call arrives. Returns a pointer to the call, ready
1432 * for an rx_Read. */
1433 #ifdef RX_ENABLE_LOCKS
1435 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1437 struct rx_serverQueueEntry *sq;
1438 register struct rx_call *call = (struct rx_call *)0;
1439 struct rx_service *service = NULL;
1442 MUTEX_ENTER(&freeSQEList_lock);
1444 if ((sq = rx_FreeSQEList)) {
1445 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1446 MUTEX_EXIT(&freeSQEList_lock);
1447 } else { /* otherwise allocate a new one and return that */
1448 MUTEX_EXIT(&freeSQEList_lock);
1449 sq = (struct rx_serverQueueEntry *)
1450 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1451 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1452 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1455 MUTEX_ENTER(&rx_serverPool_lock);
1456 if (cur_service != NULL) {
1457 ReturnToServerPool(cur_service);
1460 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1461 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1463 /* Scan for eligible incoming calls. A call is not eligible
1464 * if the maximum number of calls for its service type are
1465 * already executing */
1466 /* One thread will process calls FCFS (to prevent starvation),
1467 * while the other threads may run ahead looking for calls which
1468 * have all their input data available immediately. This helps
1469 * keep threads from blocking, waiting for data from the client. */
1470 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1471 service = tcall->conn->service;
1472 if (!QuotaOK(service)) {
1475 if (tno == rxi_fcfs_thread_num
1476 || !tcall->queue_item_header.next) {
1477 /* If we're the fcfs thread , then we'll just use
1478 * this call. If we haven't been able to find an optimal
1479 * choice, and we're at the end of the list, then use a
1480 * 2d choice if one has been identified. Otherwise... */
1481 call = (choice2 ? choice2 : tcall);
1482 service = call->conn->service;
1483 } else if (!queue_IsEmpty(&tcall->rq)) {
1484 struct rx_packet *rp;
1485 rp = queue_First(&tcall->rq, rx_packet);
1486 if (rp->header.seq == 1) {
1488 || (rp->header.flags & RX_LAST_PACKET)) {
1490 } else if (rxi_2dchoice && !choice2
1491 && !(tcall->flags & RX_CALL_CLEARED)
1492 && (tcall->rprev > rxi_HardAckRate)) {
1501 ReturnToServerPool(service);
1508 MUTEX_EXIT(&rx_serverPool_lock);
1509 MUTEX_ENTER(&call->lock);
1511 if (call->flags & RX_CALL_WAIT_PROC) {
1512 call->flags &= ~RX_CALL_WAIT_PROC;
1513 MUTEX_ENTER(&rx_stats_mutex);
1515 MUTEX_EXIT(&rx_stats_mutex);
1518 if (call->state != RX_STATE_PRECALL || call->error) {
1519 MUTEX_EXIT(&call->lock);
1520 MUTEX_ENTER(&rx_serverPool_lock);
1521 ReturnToServerPool(service);
1526 if (queue_IsEmpty(&call->rq)
1527 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1528 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1530 CLEAR_CALL_QUEUE_LOCK(call);
1533 /* If there are no eligible incoming calls, add this process
1534 * to the idle server queue, to wait for one */
1538 *socketp = OSI_NULLSOCKET;
1540 sq->socketp = socketp;
1541 queue_Append(&rx_idleServerQueue, sq);
1542 #ifndef AFS_AIX41_ENV
1543 rx_waitForPacket = sq;
1545 rx_waitingForPacket = sq;
1546 #endif /* AFS_AIX41_ENV */
1548 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1550 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1551 MUTEX_EXIT(&rx_serverPool_lock);
1552 return (struct rx_call *)0;
1555 } while (!(call = sq->newcall)
1556 && !(socketp && *socketp != OSI_NULLSOCKET));
1557 MUTEX_EXIT(&rx_serverPool_lock);
1559 MUTEX_ENTER(&call->lock);
1565 MUTEX_ENTER(&freeSQEList_lock);
1566 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1567 rx_FreeSQEList = sq;
1568 MUTEX_EXIT(&freeSQEList_lock);
1571 clock_GetTime(&call->startTime);
1572 call->state = RX_STATE_ACTIVE;
1573 call->mode = RX_MODE_RECEIVING;
1574 #ifdef RX_KERNEL_TRACE
1575 if (ICL_SETACTIVE(afs_iclSetp)) {
1576 int glockOwner = ISAFS_GLOCK();
1579 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1580 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1587 rxi_calltrace(RX_CALL_START, call);
1588 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1589 call->conn->service->servicePort, call->conn->service->serviceId,
1592 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1593 MUTEX_EXIT(&call->lock);
1595 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1600 #else /* RX_ENABLE_LOCKS */
1602 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1604 struct rx_serverQueueEntry *sq;
1605 register struct rx_call *call = (struct rx_call *)0, *choice2;
1606 struct rx_service *service = NULL;
1611 MUTEX_ENTER(&freeSQEList_lock);
1613 if ((sq = rx_FreeSQEList)) {
1614 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1615 MUTEX_EXIT(&freeSQEList_lock);
1616 } else { /* otherwise allocate a new one and return that */
1617 MUTEX_EXIT(&freeSQEList_lock);
1618 sq = (struct rx_serverQueueEntry *)
1619 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1620 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1621 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1623 MUTEX_ENTER(&sq->lock);
1625 if (cur_service != NULL) {
1626 cur_service->nRequestsRunning--;
1627 if (cur_service->nRequestsRunning < cur_service->minProcs)
1631 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1632 register struct rx_call *tcall, *ncall;
1633 /* Scan for eligible incoming calls. A call is not eligible
1634 * if the maximum number of calls for its service type are
1635 * already executing */
1636 /* One thread will process calls FCFS (to prevent starvation),
1637 * while the other threads may run ahead looking for calls which
1638 * have all their input data available immediately. This helps
1639 * keep threads from blocking, waiting for data from the client. */
1640 choice2 = (struct rx_call *)0;
1641 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1642 service = tcall->conn->service;
1643 if (QuotaOK(service)) {
1644 if (tno == rxi_fcfs_thread_num
1645 || !tcall->queue_item_header.next) {
1646 /* If we're the fcfs thread, then we'll just use
1647 * this call. If we haven't been able to find an optimal
1648 * choice, and we're at the end of the list, then use a
1649 * 2d choice if one has been identified. Otherwise... */
1650 call = (choice2 ? choice2 : tcall);
1651 service = call->conn->service;
1652 } else if (!queue_IsEmpty(&tcall->rq)) {
1653 struct rx_packet *rp;
1654 rp = queue_First(&tcall->rq, rx_packet);
1655 if (rp->header.seq == 1
1657 || (rp->header.flags & RX_LAST_PACKET))) {
1659 } else if (rxi_2dchoice && !choice2
1660 && !(tcall->flags & RX_CALL_CLEARED)
1661 && (tcall->rprev > rxi_HardAckRate)) {
1674 /* we can't schedule a call if there's no data!!! */
1675 /* send an ack if there's no data, if we're missing the
1676 * first packet, or we're missing something between first
1677 * and last -- there's a "hole" in the incoming data. */
1678 if (queue_IsEmpty(&call->rq)
1679 || queue_First(&call->rq, rx_packet)->header.seq != 1
1680 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1681 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1683 call->flags &= (~RX_CALL_WAIT_PROC);
1684 service->nRequestsRunning++;
1685 /* just started call in minProcs pool, need fewer to maintain
1687 if (service->nRequestsRunning <= service->minProcs)
1691 /* MUTEX_EXIT(&call->lock); */
1693 /* If there are no eligible incoming calls, add this process
1694 * to the idle server queue, to wait for one */
1697 *socketp = OSI_NULLSOCKET;
1699 sq->socketp = socketp;
1700 queue_Append(&rx_idleServerQueue, sq);
1704 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1707 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1708 return (struct rx_call *)0;
1711 } while (!(call = sq->newcall)
1712 && !(socketp && *socketp != OSI_NULLSOCKET));
1714 MUTEX_EXIT(&sq->lock);
1716 MUTEX_ENTER(&freeSQEList_lock);
1717 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1718 rx_FreeSQEList = sq;
1719 MUTEX_EXIT(&freeSQEList_lock);
1722 clock_GetTime(&call->startTime);
1723 call->state = RX_STATE_ACTIVE;
1724 call->mode = RX_MODE_RECEIVING;
1725 #ifdef RX_KERNEL_TRACE
1726 if (ICL_SETACTIVE(afs_iclSetp)) {
1727 int glockOwner = ISAFS_GLOCK();
1730 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1731 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1738 rxi_calltrace(RX_CALL_START, call);
1739 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1740 call->conn->service->servicePort, call->conn->service->serviceId,
1743 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1751 #endif /* RX_ENABLE_LOCKS */
1755 /* Establish a procedure to be called when a packet arrives for a
1756 * call. This routine will be called at most once after each call,
1757 * and will also be called if there is an error condition on the or
1758 * the call is complete. Used by multi rx to build a selection
1759 * function which determines which of several calls is likely to be a
1760 * good one to read from.
1761 * NOTE: the way this is currently implemented it is probably only a
1762 * good idea to (1) use it immediately after a newcall (clients only)
1763 * and (2) only use it once. Other uses currently void your warranty
1766 rx_SetArrivalProc(register struct rx_call *call,
1767 register VOID(*proc) (register struct rx_call * call,
1768 register struct multi_handle * mh,
1769 register int index),
1770 register VOID * handle, register VOID * arg)
1772 call->arrivalProc = proc;
1773 call->arrivalProcHandle = handle;
1774 call->arrivalProcArg = arg;
1777 /* Call is finished (possibly prematurely). Return rc to the peer, if
1778 * appropriate, and return the final error code from the conversation
1782 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1784 register struct rx_connection *conn = call->conn;
1785 register struct rx_service *service;
1786 register struct rx_packet *tp; /* Temporary packet pointer */
1787 register struct rx_packet *nxp; /* Next packet pointer, for queue_Scan */
1791 dpf(("rx_EndCall(call %x)\n", call));
1795 MUTEX_ENTER(&call->lock);
1797 if (rc == 0 && call->error == 0) {
1798 call->abortCode = 0;
1799 call->abortCount = 0;
1802 call->arrivalProc = (VOID(*)())0;
1803 if (rc && call->error == 0) {
1804 rxi_CallError(call, rc);
1805 /* Send an abort message to the peer if this error code has
1806 * only just been set. If it was set previously, assume the
1807 * peer has already been sent the error code or will request it
1809 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1811 if (conn->type == RX_SERVER_CONNECTION) {
1812 /* Make sure reply or at least dummy reply is sent */
1813 if (call->mode == RX_MODE_RECEIVING) {
1814 rxi_WriteProc(call, 0, 0);
1816 if (call->mode == RX_MODE_SENDING) {
1817 rxi_FlushWrite(call);
1819 service = conn->service;
1820 rxi_calltrace(RX_CALL_END, call);
1821 /* Call goes to hold state until reply packets are acknowledged */
1822 if (call->tfirst + call->nSoftAcked < call->tnext) {
1823 call->state = RX_STATE_HOLD;
1825 call->state = RX_STATE_DALLY;
1826 rxi_ClearTransmitQueue(call, 0);
1827 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1828 rxevent_Cancel(call->keepAliveEvent, call,
1829 RX_CALL_REFCOUNT_ALIVE);
1831 } else { /* Client connection */
1833 /* Make sure server receives input packets, in the case where
1834 * no reply arguments are expected */
1835 if ((call->mode == RX_MODE_SENDING)
1836 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1837 (void)rxi_ReadProc(call, &dummy, 1);
1840 /* If we had an outstanding delayed ack, be nice to the server
1841 * and force-send it now.
1843 if (call->delayedAckEvent) {
1844 rxevent_Cancel(call->delayedAckEvent, call,
1845 RX_CALL_REFCOUNT_DELAY);
1846 call->delayedAckEvent = NULL;
1847 rxi_SendDelayedAck(NULL, call, NULL);
1850 /* We need to release the call lock since it's lower than the
1851 * conn_call_lock and we don't want to hold the conn_call_lock
1852 * over the rx_ReadProc call. The conn_call_lock needs to be held
1853 * here for the case where rx_NewCall is perusing the calls on
1854 * the connection structure. We don't want to signal until
1855 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1856 * have checked this call, found it active and by the time it
1857 * goes to sleep, will have missed the signal.
1859 MUTEX_EXIT(&call->lock);
1860 MUTEX_ENTER(&conn->conn_call_lock);
1861 MUTEX_ENTER(&call->lock);
1862 MUTEX_ENTER(&conn->conn_data_lock);
1863 conn->flags |= RX_CONN_BUSY;
1864 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1865 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1866 MUTEX_EXIT(&conn->conn_data_lock);
1867 #ifdef RX_ENABLE_LOCKS
1868 CV_BROADCAST(&conn->conn_call_cv);
1873 #ifdef RX_ENABLE_LOCKS
1875 MUTEX_EXIT(&conn->conn_data_lock);
1877 #endif /* RX_ENABLE_LOCKS */
1878 call->state = RX_STATE_DALLY;
1880 error = call->error;
1882 /* currentPacket, nLeft, and NFree must be zeroed here, because
1883 * ResetCall cannot: ResetCall may be called at splnet(), in the
1884 * kernel version, and may interrupt the macros rx_Read or
1885 * rx_Write, which run at normal priority for efficiency. */
1886 if (call->currentPacket) {
1887 rxi_FreePacket(call->currentPacket);
1888 call->currentPacket = (struct rx_packet *)0;
1889 call->nLeft = call->nFree = call->curlen = 0;
1891 call->nLeft = call->nFree = call->curlen = 0;
1893 /* Free any packets from the last call to ReadvProc/WritevProc */
1894 for (queue_Scan(&call->iovq, tp, nxp, rx_packet)) {
1899 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1900 MUTEX_EXIT(&call->lock);
1901 if (conn->type == RX_CLIENT_CONNECTION) {
1902 MUTEX_EXIT(&conn->conn_call_lock);
1903 conn->flags &= ~RX_CONN_BUSY;
1908 * Map errors to the local host's errno.h format.
1910 error = ntoh_syserr_conv(error);
1914 #if !defined(KERNEL)
1916 /* Call this routine when shutting down a server or client (especially
1917 * clients). This will allow Rx to gracefully garbage collect server
1918 * connections, and reduce the number of retries that a server might
1919 * make to a dead client.
1920 * This is not quite right, since some calls may still be ongoing and
1921 * we can't lock them to destroy them. */
1925 register struct rx_connection **conn_ptr, **conn_end;
1927 INIT_PTHREAD_LOCKS LOCK_RX_INIT if (rxinit_status == 1) {
1928 UNLOCK_RX_INIT return; /* Already shutdown. */
1930 rxi_DeleteCachedConnections();
1931 if (rx_connHashTable) {
1932 MUTEX_ENTER(&rx_connHashTable_lock);
1933 for (conn_ptr = &rx_connHashTable[0], conn_end =
1934 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
1936 struct rx_connection *conn, *next;
1937 for (conn = *conn_ptr; conn; conn = next) {
1939 if (conn->type == RX_CLIENT_CONNECTION) {
1940 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
1942 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
1943 #ifdef RX_ENABLE_LOCKS
1944 rxi_DestroyConnectionNoLock(conn);
1945 #else /* RX_ENABLE_LOCKS */
1946 rxi_DestroyConnection(conn);
1947 #endif /* RX_ENABLE_LOCKS */
1951 #ifdef RX_ENABLE_LOCKS
1952 while (rx_connCleanup_list) {
1953 struct rx_connection *conn;
1954 conn = rx_connCleanup_list;
1955 rx_connCleanup_list = rx_connCleanup_list->next;
1956 MUTEX_EXIT(&rx_connHashTable_lock);
1957 rxi_CleanupConnection(conn);
1958 MUTEX_ENTER(&rx_connHashTable_lock);
1960 MUTEX_EXIT(&rx_connHashTable_lock);
1961 #endif /* RX_ENABLE_LOCKS */
1969 /* if we wakeup packet waiter too often, can get in loop with two
1970 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
1972 rxi_PacketsUnWait(void)
1974 if (!rx_waitingForPackets) {
1978 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
1979 return; /* still over quota */
1982 rx_waitingForPackets = 0;
1983 #ifdef RX_ENABLE_LOCKS
1984 CV_BROADCAST(&rx_waitingForPackets_cv);
1986 osi_rxWakeup(&rx_waitingForPackets);
1992 /* ------------------Internal interfaces------------------------- */
1994 /* Return this process's service structure for the
1995 * specified socket and service */
1997 rxi_FindService(register osi_socket socket, register u_short serviceId)
1999 register struct rx_service **sp;
2000 for (sp = &rx_services[0]; *sp; sp++) {
2001 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2007 /* Allocate a call structure, for the indicated channel of the
2008 * supplied connection. The mode and state of the call must be set by
2009 * the caller. Returns the call with mutex locked. */
2011 rxi_NewCall(register struct rx_connection *conn, register int channel)
2013 register struct rx_call *call;
2014 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2015 register struct rx_call *cp; /* Call pointer temp */
2016 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2017 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2019 /* Grab an existing call structure, or allocate a new one.
2020 * Existing call structures are assumed to have been left reset by
2022 MUTEX_ENTER(&rx_freeCallQueue_lock);
2024 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2026 * EXCEPT that the TQ might not yet be cleared out.
2027 * Skip over those with in-use TQs.
2030 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2031 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2037 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2038 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2039 call = queue_First(&rx_freeCallQueue, rx_call);
2040 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2042 MUTEX_ENTER(&rx_stats_mutex);
2043 rx_stats.nFreeCallStructs--;
2044 MUTEX_EXIT(&rx_stats_mutex);
2045 MUTEX_EXIT(&rx_freeCallQueue_lock);
2046 MUTEX_ENTER(&call->lock);
2047 CLEAR_CALL_QUEUE_LOCK(call);
2048 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2049 /* Now, if TQ wasn't cleared earlier, do it now. */
2050 if (call->flags & RX_CALL_TQ_CLEARME) {
2051 rxi_ClearTransmitQueue(call, 0);
2052 queue_Init(&call->tq);
2054 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2055 /* Bind the call to its connection structure */
2057 rxi_ResetCall(call, 1);
2059 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2061 MUTEX_EXIT(&rx_freeCallQueue_lock);
2062 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2063 MUTEX_ENTER(&call->lock);
2064 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2065 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2066 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2068 MUTEX_ENTER(&rx_stats_mutex);
2069 rx_stats.nCallStructs++;
2070 MUTEX_EXIT(&rx_stats_mutex);
2071 /* Initialize once-only items */
2072 queue_Init(&call->tq);
2073 queue_Init(&call->rq);
2074 queue_Init(&call->iovq);
2075 /* Bind the call to its connection structure (prereq for reset) */
2077 rxi_ResetCall(call, 1);
2079 call->channel = channel;
2080 call->callNumber = &conn->callNumber[channel];
2081 /* Note that the next expected call number is retained (in
2082 * conn->callNumber[i]), even if we reallocate the call structure
2084 conn->call[channel] = call;
2085 /* if the channel's never been used (== 0), we should start at 1, otherwise
2086 * the call number is valid from the last time this channel was used */
2087 if (*call->callNumber == 0)
2088 *call->callNumber = 1;
2093 /* A call has been inactive long enough that so we can throw away
2094 * state, including the call structure, which is placed on the call
2096 * Call is locked upon entry.
2097 * haveCTLock set if called from rxi_ReapConnections
2099 #ifdef RX_ENABLE_LOCKS
2101 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2102 #else /* RX_ENABLE_LOCKS */
2104 rxi_FreeCall(register struct rx_call *call)
2105 #endif /* RX_ENABLE_LOCKS */
2107 register int channel = call->channel;
2108 register struct rx_connection *conn = call->conn;
2111 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2112 (*call->callNumber)++;
2113 rxi_ResetCall(call, 0);
2114 call->conn->call[channel] = (struct rx_call *)0;
2116 MUTEX_ENTER(&rx_freeCallQueue_lock);
2117 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2118 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2119 /* A call may be free even though its transmit queue is still in use.
2120 * Since we search the call list from head to tail, put busy calls at
2121 * the head of the list, and idle calls at the tail.
2123 if (call->flags & RX_CALL_TQ_BUSY)
2124 queue_Prepend(&rx_freeCallQueue, call);
2126 queue_Append(&rx_freeCallQueue, call);
2127 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2128 queue_Append(&rx_freeCallQueue, call);
2129 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2130 MUTEX_ENTER(&rx_stats_mutex);
2131 rx_stats.nFreeCallStructs++;
2132 MUTEX_EXIT(&rx_stats_mutex);
2134 MUTEX_EXIT(&rx_freeCallQueue_lock);
2136 /* Destroy the connection if it was previously slated for
2137 * destruction, i.e. the Rx client code previously called
2138 * rx_DestroyConnection (client connections), or
2139 * rxi_ReapConnections called the same routine (server
2140 * connections). Only do this, however, if there are no
2141 * outstanding calls. Note that for fine grain locking, there appears
2142 * to be a deadlock in that rxi_FreeCall has a call locked and
2143 * DestroyConnectionNoLock locks each call in the conn. But note a
2144 * few lines up where we have removed this call from the conn.
2145 * If someone else destroys a connection, they either have no
2146 * call lock held or are going through this section of code.
2148 if (conn->flags & RX_CONN_DESTROY_ME) {
2149 MUTEX_ENTER(&conn->conn_data_lock);
2151 MUTEX_EXIT(&conn->conn_data_lock);
2152 #ifdef RX_ENABLE_LOCKS
2154 rxi_DestroyConnectionNoLock(conn);
2156 rxi_DestroyConnection(conn);
2157 #else /* RX_ENABLE_LOCKS */
2158 rxi_DestroyConnection(conn);
2159 #endif /* RX_ENABLE_LOCKS */
2163 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2165 rxi_Alloc(register size_t size)
2169 #if defined(AFS_AIX41_ENV) && defined(KERNEL)
2170 /* Grab the AFS filesystem lock. See afs/osi.h for the lock
2173 int glockOwner = ISAFS_GLOCK();
2177 MUTEX_ENTER(&rx_stats_mutex);
2179 rxi_Allocsize += size;
2180 MUTEX_EXIT(&rx_stats_mutex);
2181 #if (defined(AFS_AIX32_ENV) || defined(AFS_HPUX_ENV)) && !defined(AFS_HPUX100_ENV) && defined(KERNEL)
2182 if (size > AFS_SMALLOCSIZ) {
2183 p = (char *)osi_AllocMediumSpace(size);
2185 p = (char *)osi_AllocSmall(size, 1);
2186 #if defined(AFS_AIX41_ENV) && defined(KERNEL)
2191 p = (char *)osi_Alloc(size);
2194 osi_Panic("rxi_Alloc error");
2200 rxi_Free(void *addr, register size_t size)
2202 #if defined(AFS_AIX41_ENV) && defined(KERNEL)
2203 /* Grab the AFS filesystem lock. See afs/osi.h for the lock
2206 int glockOwner = ISAFS_GLOCK();
2210 MUTEX_ENTER(&rx_stats_mutex);
2212 rxi_Allocsize -= size;
2213 MUTEX_EXIT(&rx_stats_mutex);
2214 #if (defined(AFS_AIX32_ENV) || defined(AFS_HPUX_ENV)) && !defined(AFS_HPUX100_ENV) && defined(KERNEL)
2215 if (size > AFS_SMALLOCSIZ)
2216 osi_FreeMediumSpace(addr);
2218 osi_FreeSmall(addr);
2219 #if defined(AFS_AIX41_ENV) && defined(KERNEL)
2224 osi_Free(addr, size);
2228 /* Find the peer process represented by the supplied (host,port)
2229 * combination. If there is no appropriate active peer structure, a
2230 * new one will be allocated and initialized
2231 * The origPeer, if set, is a pointer to a peer structure on which the
2232 * refcount will be be decremented. This is used to replace the peer
2233 * structure hanging off a connection structure */
2235 rxi_FindPeer(register afs_uint32 host, register u_short port,
2236 struct rx_peer *origPeer, int create)
2238 register struct rx_peer *pp;
2240 hashIndex = PEER_HASH(host, port);
2241 MUTEX_ENTER(&rx_peerHashTable_lock);
2242 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2243 if ((pp->host == host) && (pp->port == port))
2248 pp = rxi_AllocPeer(); /* This bzero's *pp */
2249 pp->host = host; /* set here or in InitPeerParams is zero */
2251 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2252 queue_Init(&pp->congestionQueue);
2253 queue_Init(&pp->rpcStats);
2254 pp->next = rx_peerHashTable[hashIndex];
2255 rx_peerHashTable[hashIndex] = pp;
2256 rxi_InitPeerParams(pp);
2257 MUTEX_ENTER(&rx_stats_mutex);
2258 rx_stats.nPeerStructs++;
2259 MUTEX_EXIT(&rx_stats_mutex);
2266 origPeer->refCount--;
2267 MUTEX_EXIT(&rx_peerHashTable_lock);
2272 /* Find the connection at (host, port) started at epoch, and with the
2273 * given connection id. Creates the server connection if necessary.
2274 * The type specifies whether a client connection or a server
2275 * connection is desired. In both cases, (host, port) specify the
2276 * peer's (host, pair) pair. Client connections are not made
2277 * automatically by this routine. The parameter socket gives the
2278 * socket descriptor on which the packet was received. This is used,
2279 * in the case of server connections, to check that *new* connections
2280 * come via a valid (port, serviceId). Finally, the securityIndex
2281 * parameter must match the existing index for the connection. If a
2282 * server connection is created, it will be created using the supplied
2283 * index, if the index is valid for this service */
2284 struct rx_connection *
2285 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2286 register u_short port, u_short serviceId, afs_uint32 cid,
2287 afs_uint32 epoch, int type, u_int securityIndex)
2289 int hashindex, flag;
2290 register struct rx_connection *conn;
2291 hashindex = CONN_HASH(host, port, cid, epoch, type);
2292 MUTEX_ENTER(&rx_connHashTable_lock);
2293 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2294 rx_connHashTable[hashindex],
2297 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2298 && (epoch == conn->epoch)) {
2299 register struct rx_peer *pp = conn->peer;
2300 if (securityIndex != conn->securityIndex) {
2301 /* this isn't supposed to happen, but someone could forge a packet
2302 * like this, and there seems to be some CM bug that makes this
2303 * happen from time to time -- in which case, the fileserver
2305 MUTEX_EXIT(&rx_connHashTable_lock);
2306 return (struct rx_connection *)0;
2308 if (pp->host == host && pp->port == port)
2310 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2312 /* So what happens when it's a callback connection? */
2313 if (/*type == RX_CLIENT_CONNECTION &&*/ (conn->epoch & 0x80000000))
2317 /* the connection rxLastConn that was used the last time is not the
2318 ** one we are looking for now. Hence, start searching in the hash */
2320 conn = rx_connHashTable[hashindex];
2325 struct rx_service *service;
2326 if (type == RX_CLIENT_CONNECTION) {
2327 MUTEX_EXIT(&rx_connHashTable_lock);
2328 return (struct rx_connection *)0;
2330 service = rxi_FindService(socket, serviceId);
2331 if (!service || (securityIndex >= service->nSecurityObjects)
2332 || (service->securityObjects[securityIndex] == 0)) {
2333 MUTEX_EXIT(&rx_connHashTable_lock);
2334 return (struct rx_connection *)0;
2336 conn = rxi_AllocConnection(); /* This bzero's the connection */
2337 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2338 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2339 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2340 conn->next = rx_connHashTable[hashindex];
2341 rx_connHashTable[hashindex] = conn;
2342 conn->peer = rxi_FindPeer(host, port, 0, 1);
2343 conn->type = RX_SERVER_CONNECTION;
2344 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2345 conn->epoch = epoch;
2346 conn->cid = cid & RX_CIDMASK;
2347 /* conn->serial = conn->lastSerial = 0; */
2348 /* conn->timeout = 0; */
2349 conn->ackRate = RX_FAST_ACK_RATE;
2350 conn->service = service;
2351 conn->serviceId = serviceId;
2352 conn->securityIndex = securityIndex;
2353 conn->securityObject = service->securityObjects[securityIndex];
2354 conn->nSpecific = 0;
2355 conn->specific = NULL;
2356 rx_SetConnDeadTime(conn, service->connDeadTime);
2357 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2358 /* Notify security object of the new connection */
2359 RXS_NewConnection(conn->securityObject, conn);
2360 /* XXXX Connection timeout? */
2361 if (service->newConnProc)
2362 (*service->newConnProc) (conn);
2363 MUTEX_ENTER(&rx_stats_mutex);
2364 rx_stats.nServerConns++;
2365 MUTEX_EXIT(&rx_stats_mutex);
2368 MUTEX_ENTER(&conn->conn_data_lock);
2370 MUTEX_EXIT(&conn->conn_data_lock);
2372 rxLastConn = conn; /* store this connection as the last conn used */
2373 MUTEX_EXIT(&rx_connHashTable_lock);
2377 /* There are two packet tracing routines available for testing and monitoring
2378 * Rx. One is called just after every packet is received and the other is
2379 * called just before every packet is sent. Received packets, have had their
2380 * headers decoded, and packets to be sent have not yet had their headers
2381 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2382 * containing the network address. Both can be modified. The return value, if
2383 * non-zero, indicates that the packet should be dropped. */
2385 int (*rx_justReceived) () = 0;
2386 int (*rx_almostSent) () = 0;
2388 /* A packet has been received off the interface. Np is the packet, socket is
2389 * the socket number it was received from (useful in determining which service
2390 * this packet corresponds to), and (host, port) reflect the host,port of the
2391 * sender. This call returns the packet to the caller if it is finished with
2392 * it, rather than de-allocating it, just as a small performance hack */
2395 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2396 afs_uint32 host, u_short port, int *tnop,
2397 struct rx_call **newcallp)
2399 register struct rx_call *call;
2400 register struct rx_connection *conn;
2402 afs_uint32 currentCallNumber;
2408 struct rx_packet *tnp;
2411 /* We don't print out the packet until now because (1) the time may not be
2412 * accurate enough until now in the lwp implementation (rx_Listener only gets
2413 * the time after the packet is read) and (2) from a protocol point of view,
2414 * this is the first time the packet has been seen */
2415 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2416 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2417 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2418 np->header.serial, packetType, host, port, np->header.serviceId,
2419 np->header.epoch, np->header.cid, np->header.callNumber,
2420 np->header.seq, np->header.flags, np));
2423 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2424 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2427 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2428 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2431 /* If an input tracer function is defined, call it with the packet and
2432 * network address. Note this function may modify its arguments. */
2433 if (rx_justReceived) {
2434 struct sockaddr_in addr;
2436 addr.sin_family = AF_INET;
2437 addr.sin_port = port;
2438 addr.sin_addr.s_addr = host;
2439 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2440 addr.sin_len = sizeof(addr);
2441 #endif /* AFS_OSF_ENV */
2442 drop = (*rx_justReceived) (np, &addr);
2443 /* drop packet if return value is non-zero */
2446 port = addr.sin_port; /* in case fcn changed addr */
2447 host = addr.sin_addr.s_addr;
2451 /* If packet was not sent by the client, then *we* must be the client */
2452 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2453 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2455 /* Find the connection (or fabricate one, if we're the server & if
2456 * necessary) associated with this packet */
2458 rxi_FindConnection(socket, host, port, np->header.serviceId,
2459 np->header.cid, np->header.epoch, type,
2460 np->header.securityIndex);
2463 /* If no connection found or fabricated, just ignore the packet.
2464 * (An argument could be made for sending an abort packet for
2469 MUTEX_ENTER(&conn->conn_data_lock);
2470 if (conn->maxSerial < np->header.serial)
2471 conn->maxSerial = np->header.serial;
2472 MUTEX_EXIT(&conn->conn_data_lock);
2474 /* If the connection is in an error state, send an abort packet and ignore
2475 * the incoming packet */
2477 /* Don't respond to an abort packet--we don't want loops! */
2478 MUTEX_ENTER(&conn->conn_data_lock);
2479 if (np->header.type != RX_PACKET_TYPE_ABORT)
2480 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2482 MUTEX_EXIT(&conn->conn_data_lock);
2486 /* Check for connection-only requests (i.e. not call specific). */
2487 if (np->header.callNumber == 0) {
2488 switch (np->header.type) {
2489 case RX_PACKET_TYPE_ABORT:
2490 /* What if the supplied error is zero? */
2491 rxi_ConnectionError(conn, ntohl(rx_GetInt32(np, 0)));
2492 MUTEX_ENTER(&conn->conn_data_lock);
2494 MUTEX_EXIT(&conn->conn_data_lock);
2496 case RX_PACKET_TYPE_CHALLENGE:
2497 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2498 MUTEX_ENTER(&conn->conn_data_lock);
2500 MUTEX_EXIT(&conn->conn_data_lock);
2502 case RX_PACKET_TYPE_RESPONSE:
2503 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2504 MUTEX_ENTER(&conn->conn_data_lock);
2506 MUTEX_EXIT(&conn->conn_data_lock);
2508 case RX_PACKET_TYPE_PARAMS:
2509 case RX_PACKET_TYPE_PARAMS + 1:
2510 case RX_PACKET_TYPE_PARAMS + 2:
2511 /* ignore these packet types for now */
2512 MUTEX_ENTER(&conn->conn_data_lock);
2514 MUTEX_EXIT(&conn->conn_data_lock);
2519 /* Should not reach here, unless the peer is broken: send an
2521 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2522 MUTEX_ENTER(&conn->conn_data_lock);
2523 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2525 MUTEX_EXIT(&conn->conn_data_lock);
2530 channel = np->header.cid & RX_CHANNELMASK;
2531 call = conn->call[channel];
2532 #ifdef RX_ENABLE_LOCKS
2534 MUTEX_ENTER(&call->lock);
2535 /* Test to see if call struct is still attached to conn. */
2536 if (call != conn->call[channel]) {
2538 MUTEX_EXIT(&call->lock);
2539 if (type == RX_SERVER_CONNECTION) {
2540 call = conn->call[channel];
2541 /* If we started with no call attached and there is one now,
2542 * another thread is also running this routine and has gotten
2543 * the connection channel. We should drop this packet in the tests
2544 * below. If there was a call on this connection and it's now
2545 * gone, then we'll be making a new call below.
2546 * If there was previously a call and it's now different then
2547 * the old call was freed and another thread running this routine
2548 * has created a call on this channel. One of these two threads
2549 * has a packet for the old call and the code below handles those
2553 MUTEX_ENTER(&call->lock);
2555 /* This packet can't be for this call. If the new call address is
2556 * 0 then no call is running on this channel. If there is a call
2557 * then, since this is a client connection we're getting data for
2558 * it must be for the previous call.
2560 MUTEX_ENTER(&rx_stats_mutex);
2561 rx_stats.spuriousPacketsRead++;
2562 MUTEX_EXIT(&rx_stats_mutex);
2563 MUTEX_ENTER(&conn->conn_data_lock);
2565 MUTEX_EXIT(&conn->conn_data_lock);
2570 currentCallNumber = conn->callNumber[channel];
2572 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2573 if (np->header.callNumber < currentCallNumber) {
2574 MUTEX_ENTER(&rx_stats_mutex);
2575 rx_stats.spuriousPacketsRead++;
2576 MUTEX_EXIT(&rx_stats_mutex);
2577 #ifdef RX_ENABLE_LOCKS
2579 MUTEX_EXIT(&call->lock);
2581 MUTEX_ENTER(&conn->conn_data_lock);
2583 MUTEX_EXIT(&conn->conn_data_lock);
2587 MUTEX_ENTER(&conn->conn_call_lock);
2588 call = rxi_NewCall(conn, channel);
2589 MUTEX_EXIT(&conn->conn_call_lock);
2590 *call->callNumber = np->header.callNumber;
2591 call->state = RX_STATE_PRECALL;
2592 clock_GetTime(&call->queueTime);
2593 hzero(call->bytesSent);
2594 hzero(call->bytesRcvd);
2595 rxi_KeepAliveOn(call);
2596 } else if (np->header.callNumber != currentCallNumber) {
2597 /* Wait until the transmit queue is idle before deciding
2598 * whether to reset the current call. Chances are that the
2599 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2602 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2603 while ((call->state == RX_STATE_ACTIVE)
2604 && (call->flags & RX_CALL_TQ_BUSY)) {
2605 call->flags |= RX_CALL_TQ_WAIT;
2606 #ifdef RX_ENABLE_LOCKS
2607 CV_WAIT(&call->cv_tq, &call->lock);
2608 #else /* RX_ENABLE_LOCKS */
2609 osi_rxSleep(&call->tq);
2610 #endif /* RX_ENABLE_LOCKS */
2612 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2613 /* If the new call cannot be taken right now send a busy and set
2614 * the error condition in this call, so that it terminates as
2615 * quickly as possible */
2616 if (call->state == RX_STATE_ACTIVE) {
2617 struct rx_packet *tp;
2619 rxi_CallError(call, RX_CALL_DEAD);
2620 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2622 MUTEX_EXIT(&call->lock);
2623 MUTEX_ENTER(&conn->conn_data_lock);
2625 MUTEX_EXIT(&conn->conn_data_lock);
2628 rxi_ResetCall(call, 0);
2629 *call->callNumber = np->header.callNumber;
2630 call->state = RX_STATE_PRECALL;
2631 clock_GetTime(&call->queueTime);
2632 hzero(call->bytesSent);
2633 hzero(call->bytesRcvd);
2635 * If the number of queued calls exceeds the overload
2636 * threshold then abort this call.
2638 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2639 struct rx_packet *tp;
2641 rxi_CallError(call, rx_BusyError);
2642 tp = rxi_SendCallAbort(call, np, 1, 0);
2643 MUTEX_EXIT(&call->lock);
2644 MUTEX_ENTER(&conn->conn_data_lock);
2646 MUTEX_EXIT(&conn->conn_data_lock);
2649 rxi_KeepAliveOn(call);
2651 /* Continuing call; do nothing here. */
2653 } else { /* we're the client */
2654 /* Ignore all incoming acknowledgements for calls in DALLY state */
2655 if (call && (call->state == RX_STATE_DALLY)
2656 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2657 MUTEX_ENTER(&rx_stats_mutex);
2658 rx_stats.ignorePacketDally++;
2659 MUTEX_EXIT(&rx_stats_mutex);
2660 #ifdef RX_ENABLE_LOCKS
2662 MUTEX_EXIT(&call->lock);
2665 MUTEX_ENTER(&conn->conn_data_lock);
2667 MUTEX_EXIT(&conn->conn_data_lock);
2671 /* Ignore anything that's not relevant to the current call. If there
2672 * isn't a current call, then no packet is relevant. */
2673 if (!call || (np->header.callNumber != currentCallNumber)) {
2674 MUTEX_ENTER(&rx_stats_mutex);
2675 rx_stats.spuriousPacketsRead++;
2676 MUTEX_EXIT(&rx_stats_mutex);
2677 #ifdef RX_ENABLE_LOCKS
2679 MUTEX_EXIT(&call->lock);
2682 MUTEX_ENTER(&conn->conn_data_lock);
2684 MUTEX_EXIT(&conn->conn_data_lock);
2687 /* If the service security object index stamped in the packet does not
2688 * match the connection's security index, ignore the packet */
2689 if (np->header.securityIndex != conn->securityIndex) {
2690 #ifdef RX_ENABLE_LOCKS
2691 MUTEX_EXIT(&call->lock);
2693 MUTEX_ENTER(&conn->conn_data_lock);
2695 MUTEX_EXIT(&conn->conn_data_lock);
2699 /* If we're receiving the response, then all transmit packets are
2700 * implicitly acknowledged. Get rid of them. */
2701 if (np->header.type == RX_PACKET_TYPE_DATA) {
2702 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2703 /* XXX Hack. Because we must release the global rx lock when
2704 * sending packets (osi_NetSend) we drop all acks while we're
2705 * traversing the tq in rxi_Start sending packets out because
2706 * packets may move to the freePacketQueue as result of being here!
2707 * So we drop these packets until we're safely out of the
2708 * traversing. Really ugly!
2709 * For fine grain RX locking, we set the acked field in the
2710 * packets and let rxi_Start remove them from the transmit queue.
2712 if (call->flags & RX_CALL_TQ_BUSY) {
2713 #ifdef RX_ENABLE_LOCKS
2714 rxi_SetAcksInTransmitQueue(call);
2717 return np; /* xmitting; drop packet */
2720 rxi_ClearTransmitQueue(call, 0);
2722 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2723 rxi_ClearTransmitQueue(call, 0);
2724 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2726 if (np->header.type == RX_PACKET_TYPE_ACK) {
2727 /* now check to see if this is an ack packet acknowledging that the
2728 * server actually *lost* some hard-acked data. If this happens we
2729 * ignore this packet, as it may indicate that the server restarted in
2730 * the middle of a call. It is also possible that this is an old ack
2731 * packet. We don't abort the connection in this case, because this
2732 * *might* just be an old ack packet. The right way to detect a server
2733 * restart in the midst of a call is to notice that the server epoch
2735 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2736 * XXX unacknowledged. I think that this is off-by-one, but
2737 * XXX I don't dare change it just yet, since it will
2738 * XXX interact badly with the server-restart detection
2739 * XXX code in receiveackpacket. */
2740 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2741 MUTEX_ENTER(&rx_stats_mutex);
2742 rx_stats.spuriousPacketsRead++;
2743 MUTEX_EXIT(&rx_stats_mutex);
2744 MUTEX_EXIT(&call->lock);
2745 MUTEX_ENTER(&conn->conn_data_lock);
2747 MUTEX_EXIT(&conn->conn_data_lock);
2751 } /* else not a data packet */
2754 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2755 /* Set remote user defined status from packet */
2756 call->remoteStatus = np->header.userStatus;
2758 /* Note the gap between the expected next packet and the actual
2759 * packet that arrived, when the new packet has a smaller serial number
2760 * than expected. Rioses frequently reorder packets all by themselves,
2761 * so this will be quite important with very large window sizes.
2762 * Skew is checked against 0 here to avoid any dependence on the type of
2763 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2765 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2766 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2767 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2769 MUTEX_ENTER(&conn->conn_data_lock);
2770 skew = conn->lastSerial - np->header.serial;
2771 conn->lastSerial = np->header.serial;
2772 MUTEX_EXIT(&conn->conn_data_lock);
2774 register struct rx_peer *peer;
2776 if (skew > peer->inPacketSkew) {
2777 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2779 peer->inPacketSkew = skew;
2783 /* Now do packet type-specific processing */
2784 switch (np->header.type) {
2785 case RX_PACKET_TYPE_DATA:
2786 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2789 case RX_PACKET_TYPE_ACK:
2790 /* Respond immediately to ack packets requesting acknowledgement
2792 if (np->header.flags & RX_REQUEST_ACK) {
2794 (void)rxi_SendCallAbort(call, 0, 1, 0);
2796 (void)rxi_SendAck(call, 0, np->header.serial,
2797 RX_ACK_PING_RESPONSE, 1);
2799 np = rxi_ReceiveAckPacket(call, np, 1);
2801 case RX_PACKET_TYPE_ABORT:
2802 /* An abort packet: reset the connection, passing the error up to
2804 /* What if error is zero? */
2805 rxi_CallError(call, ntohl(*(afs_int32 *) rx_DataOf(np)));
2807 case RX_PACKET_TYPE_BUSY:
2810 case RX_PACKET_TYPE_ACKALL:
2811 /* All packets acknowledged, so we can drop all packets previously
2812 * readied for sending */
2813 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2814 /* XXX Hack. We because we can't release the global rx lock when
2815 * sending packets (osi_NetSend) we drop all ack pkts while we're
2816 * traversing the tq in rxi_Start sending packets out because
2817 * packets may move to the freePacketQueue as result of being
2818 * here! So we drop these packets until we're safely out of the
2819 * traversing. Really ugly!
2820 * For fine grain RX locking, we set the acked field in the packets
2821 * and let rxi_Start remove the packets from the transmit queue.
2823 if (call->flags & RX_CALL_TQ_BUSY) {
2824 #ifdef RX_ENABLE_LOCKS
2825 rxi_SetAcksInTransmitQueue(call);
2827 #else /* RX_ENABLE_LOCKS */
2829 return np; /* xmitting; drop packet */
2830 #endif /* RX_ENABLE_LOCKS */
2832 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2833 rxi_ClearTransmitQueue(call, 0);
2836 /* Should not reach here, unless the peer is broken: send an abort
2838 rxi_CallError(call, RX_PROTOCOL_ERROR);
2839 np = rxi_SendCallAbort(call, np, 1, 0);
2842 /* Note when this last legitimate packet was received, for keep-alive
2843 * processing. Note, we delay getting the time until now in the hope that
2844 * the packet will be delivered to the user before any get time is required
2845 * (if not, then the time won't actually be re-evaluated here). */
2846 call->lastReceiveTime = clock_Sec();
2847 MUTEX_EXIT(&call->lock);
2848 MUTEX_ENTER(&conn->conn_data_lock);
2850 MUTEX_EXIT(&conn->conn_data_lock);
2854 /* return true if this is an "interesting" connection from the point of view
2855 of someone trying to debug the system */
2857 rxi_IsConnInteresting(struct rx_connection *aconn)
2860 register struct rx_call *tcall;
2862 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2864 for (i = 0; i < RX_MAXCALLS; i++) {
2865 tcall = aconn->call[i];
2867 if ((tcall->state == RX_STATE_PRECALL)
2868 || (tcall->state == RX_STATE_ACTIVE))
2870 if ((tcall->mode == RX_MODE_SENDING)
2871 || (tcall->mode == RX_MODE_RECEIVING))
2879 /* if this is one of the last few packets AND it wouldn't be used by the
2880 receiving call to immediately satisfy a read request, then drop it on
2881 the floor, since accepting it might prevent a lock-holding thread from
2882 making progress in its reading. If a call has been cleared while in
2883 the precall state then ignore all subsequent packets until the call
2884 is assigned to a thread. */
2887 TooLow(struct rx_packet *ap, struct rx_call *acall)
2890 MUTEX_ENTER(&rx_stats_mutex);
2891 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
2892 && (acall->state == RX_STATE_PRECALL))
2893 || ((rx_nFreePackets < rxi_dataQuota + 2)
2894 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
2895 && (acall->flags & RX_CALL_READER_WAIT)))) {
2898 MUTEX_EXIT(&rx_stats_mutex);
2904 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
2905 struct rx_call *acall)
2907 struct rx_call *call = acall;
2911 MUTEX_ENTER(&conn->conn_data_lock);
2912 conn->checkReachEvent = NULL;
2913 waiting = conn->flags & RX_CONN_ATTACHWAIT;
2916 MUTEX_EXIT(&conn->conn_data_lock);
2920 MUTEX_ENTER(&conn->conn_call_lock);
2921 MUTEX_ENTER(&conn->conn_data_lock);
2922 for (i = 0; i < RX_MAXCALLS; i++) {
2923 struct rx_call *tc = conn->call[i];
2924 if (tc && tc->state == RX_STATE_PRECALL) {
2930 /* Indicate that rxi_CheckReachEvent is no longer running by
2931 * clearing the flag. Must be atomic under conn_data_lock to
2932 * avoid a new call slipping by: rxi_CheckConnReach holds
2933 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
2935 conn->flags &= ~RX_CONN_ATTACHWAIT;
2936 MUTEX_EXIT(&conn->conn_data_lock);
2937 MUTEX_EXIT(&conn->conn_call_lock);
2942 MUTEX_ENTER(&call->lock);
2943 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
2945 MUTEX_EXIT(&call->lock);
2947 clock_GetTime(&when);
2948 when.sec += RX_CHECKREACH_TIMEOUT;
2949 MUTEX_ENTER(&conn->conn_data_lock);
2950 if (!conn->checkReachEvent) {
2952 conn->checkReachEvent =
2953 rxevent_Post(&when, rxi_CheckReachEvent, conn, NULL);
2955 MUTEX_EXIT(&conn->conn_data_lock);
2961 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
2963 struct rx_service *service = conn->service;
2964 struct rx_peer *peer = conn->peer;
2965 afs_uint32 now, lastReach;
2967 if (service->checkReach == 0)
2971 MUTEX_ENTER(&peer->peer_lock);
2972 lastReach = peer->lastReachTime;
2973 MUTEX_EXIT(&peer->peer_lock);
2974 if (now - lastReach < RX_CHECKREACH_TTL)
2977 MUTEX_ENTER(&conn->conn_data_lock);
2978 if (conn->flags & RX_CONN_ATTACHWAIT) {
2979 MUTEX_EXIT(&conn->conn_data_lock);
2982 conn->flags |= RX_CONN_ATTACHWAIT;
2983 MUTEX_EXIT(&conn->conn_data_lock);
2984 if (!conn->checkReachEvent)
2985 rxi_CheckReachEvent(NULL, conn, call);
2990 /* try to attach call, if authentication is complete */
2992 TryAttach(register struct rx_call *acall, register osi_socket socket,
2993 register int *tnop, register struct rx_call **newcallp,
2996 struct rx_connection *conn = acall->conn;
2998 if (conn->type == RX_SERVER_CONNECTION
2999 && acall->state == RX_STATE_PRECALL) {
3000 /* Don't attach until we have any req'd. authentication. */
3001 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3002 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3003 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3004 /* Note: this does not necessarily succeed; there
3005 * may not any proc available
3008 rxi_ChallengeOn(acall->conn);
3013 /* A data packet has been received off the interface. This packet is
3014 * appropriate to the call (the call is in the right state, etc.). This
3015 * routine can return a packet to the caller, for re-use */
3018 rxi_ReceiveDataPacket(register struct rx_call *call,
3019 register struct rx_packet *np, int istack,
3020 osi_socket socket, afs_uint32 host, u_short port,
3021 int *tnop, struct rx_call **newcallp)
3023 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3027 afs_uint32 seq, serial, flags;
3029 struct rx_packet *tnp;
3031 MUTEX_ENTER(&rx_stats_mutex);
3032 rx_stats.dataPacketsRead++;
3033 MUTEX_EXIT(&rx_stats_mutex);
3036 /* If there are no packet buffers, drop this new packet, unless we can find
3037 * packet buffers from inactive calls */
3039 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3040 MUTEX_ENTER(&rx_freePktQ_lock);
3041 rxi_NeedMorePackets = TRUE;
3042 MUTEX_EXIT(&rx_freePktQ_lock);
3043 MUTEX_ENTER(&rx_stats_mutex);
3044 rx_stats.noPacketBuffersOnRead++;
3045 MUTEX_EXIT(&rx_stats_mutex);
3046 call->rprev = np->header.serial;
3047 rxi_calltrace(RX_TRACE_DROP, call);
3048 dpf(("packet %x dropped on receipt - quota problems", np));
3050 rxi_ClearReceiveQueue(call);
3051 clock_GetTime(&when);
3052 clock_Add(&when, &rx_softAckDelay);
3053 if (!call->delayedAckEvent
3054 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3055 rxevent_Cancel(call->delayedAckEvent, call,
3056 RX_CALL_REFCOUNT_DELAY);
3057 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3058 call->delayedAckEvent =
3059 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3061 /* we've damaged this call already, might as well do it in. */
3067 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3068 * packet is one of several packets transmitted as a single
3069 * datagram. Do not send any soft or hard acks until all packets
3070 * in a jumbogram have been processed. Send negative acks right away.
3072 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3073 /* tnp is non-null when there are more packets in the
3074 * current jumbo gram */
3081 seq = np->header.seq;
3082 serial = np->header.serial;
3083 flags = np->header.flags;
3085 /* If the call is in an error state, send an abort message */
3087 return rxi_SendCallAbort(call, np, istack, 0);
3089 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3090 * AFS 3.5 jumbogram. */
3091 if (flags & RX_JUMBO_PACKET) {
3092 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3097 if (np->header.spare != 0) {
3098 MUTEX_ENTER(&call->conn->conn_data_lock);
3099 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3100 MUTEX_EXIT(&call->conn->conn_data_lock);
3103 /* The usual case is that this is the expected next packet */
3104 if (seq == call->rnext) {
3106 /* Check to make sure it is not a duplicate of one already queued */
3107 if (queue_IsNotEmpty(&call->rq)
3108 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3109 MUTEX_ENTER(&rx_stats_mutex);
3110 rx_stats.dupPacketsRead++;
3111 MUTEX_EXIT(&rx_stats_mutex);
3112 dpf(("packet %x dropped on receipt - duplicate", np));
3113 rxevent_Cancel(call->delayedAckEvent, call,
3114 RX_CALL_REFCOUNT_DELAY);
3115 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3121 /* It's the next packet. Stick it on the receive queue
3122 * for this call. Set newPackets to make sure we wake
3123 * the reader once all packets have been processed */
3124 queue_Prepend(&call->rq, np);
3126 np = NULL; /* We can't use this anymore */
3129 /* If an ack is requested then set a flag to make sure we
3130 * send an acknowledgement for this packet */
3131 if (flags & RX_REQUEST_ACK) {
3132 ackNeeded = RX_ACK_REQUESTED;
3135 /* Keep track of whether we have received the last packet */
3136 if (flags & RX_LAST_PACKET) {
3137 call->flags |= RX_CALL_HAVE_LAST;
3141 /* Check whether we have all of the packets for this call */
3142 if (call->flags & RX_CALL_HAVE_LAST) {
3143 afs_uint32 tseq; /* temporary sequence number */
3144 struct rx_packet *tp; /* Temporary packet pointer */
3145 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3147 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3148 if (tseq != tp->header.seq)
3150 if (tp->header.flags & RX_LAST_PACKET) {
3151 call->flags |= RX_CALL_RECEIVE_DONE;
3158 /* Provide asynchronous notification for those who want it
3159 * (e.g. multi rx) */
3160 if (call->arrivalProc) {
3161 (*call->arrivalProc) (call, call->arrivalProcHandle,
3162 (int)call->arrivalProcArg);
3163 call->arrivalProc = (VOID(*)())0;
3166 /* Update last packet received */
3169 /* If there is no server process serving this call, grab
3170 * one, if available. We only need to do this once. If a
3171 * server thread is available, this thread becomes a server
3172 * thread and the server thread becomes a listener thread. */
3174 TryAttach(call, socket, tnop, newcallp, 0);
3177 /* This is not the expected next packet. */
3179 /* Determine whether this is a new or old packet, and if it's
3180 * a new one, whether it fits into the current receive window.
3181 * Also figure out whether the packet was delivered in sequence.
3182 * We use the prev variable to determine whether the new packet
3183 * is the successor of its immediate predecessor in the
3184 * receive queue, and the missing flag to determine whether
3185 * any of this packets predecessors are missing. */
3187 afs_uint32 prev; /* "Previous packet" sequence number */
3188 struct rx_packet *tp; /* Temporary packet pointer */
3189 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3190 int missing; /* Are any predecessors missing? */
3192 /* If the new packet's sequence number has been sent to the
3193 * application already, then this is a duplicate */
3194 if (seq < call->rnext) {
3195 MUTEX_ENTER(&rx_stats_mutex);
3196 rx_stats.dupPacketsRead++;
3197 MUTEX_EXIT(&rx_stats_mutex);
3198 rxevent_Cancel(call->delayedAckEvent, call,
3199 RX_CALL_REFCOUNT_DELAY);
3200 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3206 /* If the sequence number is greater than what can be
3207 * accomodated by the current window, then send a negative
3208 * acknowledge and drop the packet */
3209 if ((call->rnext + call->rwind) <= seq) {
3210 rxevent_Cancel(call->delayedAckEvent, call,
3211 RX_CALL_REFCOUNT_DELAY);
3212 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3219 /* Look for the packet in the queue of old received packets */
3220 for (prev = call->rnext - 1, missing =
3221 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3222 /*Check for duplicate packet */
3223 if (seq == tp->header.seq) {
3224 MUTEX_ENTER(&rx_stats_mutex);
3225 rx_stats.dupPacketsRead++;
3226 MUTEX_EXIT(&rx_stats_mutex);
3227 rxevent_Cancel(call->delayedAckEvent, call,
3228 RX_CALL_REFCOUNT_DELAY);
3229 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3235 /* If we find a higher sequence packet, break out and
3236 * insert the new packet here. */
3237 if (seq < tp->header.seq)
3239 /* Check for missing packet */
3240 if (tp->header.seq != prev + 1) {
3244 prev = tp->header.seq;
3247 /* Keep track of whether we have received the last packet. */
3248 if (flags & RX_LAST_PACKET) {
3249 call->flags |= RX_CALL_HAVE_LAST;
3252 /* It's within the window: add it to the the receive queue.
3253 * tp is left by the previous loop either pointing at the
3254 * packet before which to insert the new packet, or at the
3255 * queue head if the queue is empty or the packet should be
3257 queue_InsertBefore(tp, np);
3261 /* Check whether we have all of the packets for this call */
3262 if ((call->flags & RX_CALL_HAVE_LAST)
3263 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3264 afs_uint32 tseq; /* temporary sequence number */
3267 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3268 if (tseq != tp->header.seq)
3270 if (tp->header.flags & RX_LAST_PACKET) {
3271 call->flags |= RX_CALL_RECEIVE_DONE;
3278 /* We need to send an ack of the packet is out of sequence,
3279 * or if an ack was requested by the peer. */
3280 if (seq != prev + 1 || missing || (flags & RX_REQUEST_ACK)) {
3281 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3284 /* Acknowledge the last packet for each call */
3285 if (flags & RX_LAST_PACKET) {
3296 * If the receiver is waiting for an iovec, fill the iovec
3297 * using the data from the receive queue */
3298 if (call->flags & RX_CALL_IOVEC_WAIT) {
3299 didHardAck = rxi_FillReadVec(call, serial);
3300 /* the call may have been aborted */
3309 /* Wakeup the reader if any */
3310 if ((call->flags & RX_CALL_READER_WAIT)
3311 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3312 || (call->iovNext >= call->iovMax)
3313 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3314 call->flags &= ~RX_CALL_READER_WAIT;
3315 #ifdef RX_ENABLE_LOCKS
3316 CV_BROADCAST(&call->cv_rq);
3318 osi_rxWakeup(&call->rq);
3324 * Send an ack when requested by the peer, or once every
3325 * rxi_SoftAckRate packets until the last packet has been
3326 * received. Always send a soft ack for the last packet in
3327 * the server's reply. */
3329 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3330 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3331 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3332 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3333 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3334 } else if (call->nSoftAcks) {
3335 clock_GetTime(&when);
3336 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3337 clock_Add(&when, &rx_lastAckDelay);
3339 clock_Add(&when, &rx_softAckDelay);
3341 if (!call->delayedAckEvent
3342 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3343 rxevent_Cancel(call->delayedAckEvent, call,
3344 RX_CALL_REFCOUNT_DELAY);
3345 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3346 call->delayedAckEvent =
3347 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3349 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3350 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3357 static void rxi_ComputeRate();
3361 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3363 struct rx_peer *peer = conn->peer;
3365 MUTEX_ENTER(&peer->peer_lock);
3366 peer->lastReachTime = clock_Sec();
3367 MUTEX_EXIT(&peer->peer_lock);
3369 MUTEX_ENTER(&conn->conn_data_lock);
3370 if (conn->flags & RX_CONN_ATTACHWAIT) {
3373 conn->flags &= ~RX_CONN_ATTACHWAIT;
3374 MUTEX_EXIT(&conn->conn_data_lock);
3376 for (i = 0; i < RX_MAXCALLS; i++) {
3377 struct rx_call *call = conn->call[i];
3380 MUTEX_ENTER(&call->lock);
3381 /* tnop can be null if newcallp is null */
3382 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3384 MUTEX_EXIT(&call->lock);
3388 MUTEX_EXIT(&conn->conn_data_lock);
3391 /* rxi_ComputePeerNetStats
3393 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3394 * estimates (like RTT and throughput) based on ack packets. Caller
3395 * must ensure that the packet in question is the right one (i.e.
3396 * serial number matches).
3399 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3400 struct rx_ackPacket *ap, struct rx_packet *np)
3402 struct rx_peer *peer = call->conn->peer;
3404 /* Use RTT if not delayed by client. */
3405 if (ap->reason != RX_ACK_DELAY)
3406 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3408 rxi_ComputeRate(peer, call, p, np, ap->reason);
3412 /* The real smarts of the whole thing. */
3414 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3417 struct rx_ackPacket *ap;
3419 register struct rx_packet *tp;
3420 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3421 register struct rx_connection *conn = call->conn;
3422 struct rx_peer *peer = conn->peer;
3425 /* because there are CM's that are bogus, sending weird values for this. */
3426 afs_uint32 skew = 0;
3431 int newAckCount = 0;
3432 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3433 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3435 MUTEX_ENTER(&rx_stats_mutex);
3436 rx_stats.ackPacketsRead++;
3437 MUTEX_EXIT(&rx_stats_mutex);
3438 ap = (struct rx_ackPacket *)rx_DataOf(np);
3439 nbytes = rx_Contiguous(np) - ((ap->acks) - (u_char *) ap);
3441 return np; /* truncated ack packet */
3443 /* depends on ack packet struct */
3444 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3445 first = ntohl(ap->firstPacket);
3446 serial = ntohl(ap->serial);
3447 /* temporarily disabled -- needs to degrade over time
3448 * skew = ntohs(ap->maxSkew); */
3450 /* Ignore ack packets received out of order */
3451 if (first < call->tfirst) {
3455 if (np->header.flags & RX_SLOW_START_OK) {
3456 call->flags |= RX_CALL_SLOW_START_OK;
3459 if (ap->reason == RX_ACK_PING_RESPONSE)
3460 rxi_UpdatePeerReach(conn, call);
3465 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3466 ap->reason, ntohl(ap->previousPacket),
3467 (unsigned int)np->header.seq, (unsigned int)serial,
3468 (unsigned int)skew, ntohl(ap->firstPacket));
3471 for (offset = 0; offset < nAcks; offset++)
3472 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3479 /* Update the outgoing packet skew value to the latest value of
3480 * the peer's incoming packet skew value. The ack packet, of
3481 * course, could arrive out of order, but that won't affect things
3483 MUTEX_ENTER(&peer->peer_lock);
3484 peer->outPacketSkew = skew;
3486 /* Check for packets that no longer need to be transmitted, and
3487 * discard them. This only applies to packets positively
3488 * acknowledged as having been sent to the peer's upper level.
3489 * All other packets must be retained. So only packets with
3490 * sequence numbers < ap->firstPacket are candidates. */
3491 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3492 if (tp->header.seq >= first)
3494 call->tfirst = tp->header.seq + 1;
3496 && (tp->header.serial == serial || tp->firstSerial == serial))
3497 rxi_ComputePeerNetStats(call, tp, ap, np);
3498 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3499 /* XXX Hack. Because we have to release the global rx lock when sending
3500 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3501 * in rxi_Start sending packets out because packets may move to the
3502 * freePacketQueue as result of being here! So we drop these packets until
3503 * we're safely out of the traversing. Really ugly!
3504 * To make it even uglier, if we're using fine grain locking, we can
3505 * set the ack bits in the packets and have rxi_Start remove the packets
3506 * when it's done transmitting.
3508 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3511 if (call->flags & RX_CALL_TQ_BUSY) {
3512 #ifdef RX_ENABLE_LOCKS
3513 tp->flags |= RX_PKTFLAG_ACKED;
3514 call->flags |= RX_CALL_TQ_SOME_ACKED;
3515 #else /* RX_ENABLE_LOCKS */
3517 #endif /* RX_ENABLE_LOCKS */
3519 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3522 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3527 /* Give rate detector a chance to respond to ping requests */
3528 if (ap->reason == RX_ACK_PING_RESPONSE) {
3529 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3533 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3535 /* Now go through explicit acks/nacks and record the results in
3536 * the waiting packets. These are packets that can't be released
3537 * yet, even with a positive acknowledge. This positive
3538 * acknowledge only means the packet has been received by the
3539 * peer, not that it will be retained long enough to be sent to
3540 * the peer's upper level. In addition, reset the transmit timers
3541 * of any missing packets (those packets that must be missing
3542 * because this packet was out of sequence) */
3544 call->nSoftAcked = 0;
3545 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3546 /* Update round trip time if the ack was stimulated on receipt
3548 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3549 #ifdef RX_ENABLE_LOCKS
3550 if (tp->header.seq >= first)
3551 #endif /* RX_ENABLE_LOCKS */
3552 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3554 && (tp->header.serial == serial || tp->firstSerial == serial))
3555 rxi_ComputePeerNetStats(call, tp, ap, np);
3557 /* Set the acknowledge flag per packet based on the
3558 * information in the ack packet. An acknowlegded packet can
3559 * be downgraded when the server has discarded a packet it
3560 * soacked previously, or when an ack packet is received
3561 * out of sequence. */
3562 if (tp->header.seq < first) {
3563 /* Implicit ack information */
3564 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3567 tp->flags |= RX_PKTFLAG_ACKED;
3568 } else if (tp->header.seq < first + nAcks) {
3569 /* Explicit ack information: set it in the packet appropriately */
3570 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3571 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3573 tp->flags |= RX_PKTFLAG_ACKED;
3581 tp->flags &= ~RX_PKTFLAG_ACKED;
3585 tp->flags &= ~RX_PKTFLAG_ACKED;
3589 /* If packet isn't yet acked, and it has been transmitted at least
3590 * once, reset retransmit time using latest timeout
3591 * ie, this should readjust the retransmit timer for all outstanding
3592 * packets... So we don't just retransmit when we should know better*/
3594 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3595 tp->retryTime = tp->timeSent;
3596 clock_Add(&tp->retryTime, &peer->timeout);
3597 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3598 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3602 /* If the window has been extended by this acknowledge packet,
3603 * then wakeup a sender waiting in alloc for window space, or try
3604 * sending packets now, if he's been sitting on packets due to
3605 * lack of window space */
3606 if (call->tnext < (call->tfirst + call->twind)) {
3607 #ifdef RX_ENABLE_LOCKS
3608 CV_SIGNAL(&call->cv_twind);
3610 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3611 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3612 osi_rxWakeup(&call->twind);
3615 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3616 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3620 /* if the ack packet has a receivelen field hanging off it,
3621 * update our state */
3622 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3625 /* If the ack packet has a "recommended" size that is less than
3626 * what I am using now, reduce my size to match */
3627 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3628 sizeof(afs_int32), &tSize);
3629 tSize = (afs_uint32) ntohl(tSize);
3630 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3632 /* Get the maximum packet size to send to this peer */
3633 rx_packetread(np, rx_AckDataSize(ap->nAcks), sizeof(afs_int32),
3635 tSize = (afs_uint32) ntohl(tSize);
3636 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3637 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3639 /* sanity check - peer might have restarted with different params.
3640 * If peer says "send less", dammit, send less... Peer should never
3641 * be unable to accept packets of the size that prior AFS versions would
3642 * send without asking. */
3643 if (peer->maxMTU != tSize) {
3644 peer->maxMTU = tSize;
3645 peer->MTU = MIN(tSize, peer->MTU);
3646 call->MTU = MIN(call->MTU, tSize);
3650 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3653 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3654 sizeof(afs_int32), &tSize);
3655 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3656 if (tSize < call->twind) { /* smaller than our send */
3657 call->twind = tSize; /* window, we must send less... */
3658 call->ssthresh = MIN(call->twind, call->ssthresh);
3661 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3662 * network MTU confused with the loopback MTU. Calculate the
3663 * maximum MTU here for use in the slow start code below.
3665 maxMTU = peer->maxMTU;
3666 /* Did peer restart with older RX version? */
3667 if (peer->maxDgramPackets > 1) {
3668 peer->maxDgramPackets = 1;
3670 } else if (np->length >=
3671 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3674 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3675 sizeof(afs_int32), &tSize);
3676 tSize = (afs_uint32) ntohl(tSize);
3678 * As of AFS 3.5 we set the send window to match the receive window.
3680 if (tSize < call->twind) {
3681 call->twind = tSize;
3682 call->ssthresh = MIN(call->twind, call->ssthresh);
3683 } else if (tSize > call->twind) {
3684 call->twind = tSize;
3688 * As of AFS 3.5, a jumbogram is more than one fixed size
3689 * packet transmitted in a single UDP datagram. If the remote
3690 * MTU is smaller than our local MTU then never send a datagram
3691 * larger than the natural MTU.
3694 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3695 sizeof(afs_int32), &tSize);
3696 maxDgramPackets = (afs_uint32) ntohl(tSize);
3697 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3699 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
3700 maxDgramPackets = MIN(maxDgramPackets, tSize);
3701 if (maxDgramPackets > 1) {
3702 peer->maxDgramPackets = maxDgramPackets;
3703 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3705 peer->maxDgramPackets = 1;
3706 call->MTU = peer->natMTU;
3708 } else if (peer->maxDgramPackets > 1) {
3709 /* Restarted with lower version of RX */
3710 peer->maxDgramPackets = 1;
3712 } else if (peer->maxDgramPackets > 1
3713 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3714 /* Restarted with lower version of RX */
3715 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3716 peer->natMTU = OLD_MAX_PACKET_SIZE;
3717 peer->MTU = OLD_MAX_PACKET_SIZE;
3718 peer->maxDgramPackets = 1;
3719 peer->nDgramPackets = 1;
3721 call->MTU = OLD_MAX_PACKET_SIZE;
3726 * Calculate how many datagrams were successfully received after
3727 * the first missing packet and adjust the negative ack counter
3732 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3733 if (call->nNacks < nNacked) {
3734 call->nNacks = nNacked;
3743 if (call->flags & RX_CALL_FAST_RECOVER) {
3745 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3747 call->flags &= ~RX_CALL_FAST_RECOVER;
3748 call->cwind = call->nextCwind;
3749 call->nextCwind = 0;
3752 call->nCwindAcks = 0;
3753 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3754 /* Three negative acks in a row trigger congestion recovery */
3755 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3756 MUTEX_EXIT(&peer->peer_lock);
3757 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3758 /* someone else is waiting to start recovery */
3761 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3762 while (call->flags & RX_CALL_TQ_BUSY) {
3763 call->flags |= RX_CALL_TQ_WAIT;
3764 #ifdef RX_ENABLE_LOCKS
3765 CV_WAIT(&call->cv_tq, &call->lock);
3766 #else /* RX_ENABLE_LOCKS */
3767 osi_rxSleep(&call->tq);
3768 #endif /* RX_ENABLE_LOCKS */
3770 MUTEX_ENTER(&peer->peer_lock);
3771 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3772 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3773 call->flags |= RX_CALL_FAST_RECOVER;
3774 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3776 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3777 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3778 call->nextCwind = call->ssthresh;
3781 peer->MTU = call->MTU;
3782 peer->cwind = call->nextCwind;
3783 peer->nDgramPackets = call->nDgramPackets;
3785 call->congestSeq = peer->congestSeq;
3786 /* Reset the resend times on the packets that were nacked
3787 * so we will retransmit as soon as the window permits*/
3788 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3790 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3791 clock_Zero(&tp->retryTime);
3793 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3798 /* If cwind is smaller than ssthresh, then increase
3799 * the window one packet for each ack we receive (exponential
3801 * If cwind is greater than or equal to ssthresh then increase
3802 * the congestion window by one packet for each cwind acks we
3803 * receive (linear growth). */
3804 if (call->cwind < call->ssthresh) {
3806 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3807 call->nCwindAcks = 0;
3809 call->nCwindAcks += newAckCount;
3810 if (call->nCwindAcks >= call->cwind) {
3811 call->nCwindAcks = 0;
3812 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3816 * If we have received several acknowledgements in a row then
3817 * it is time to increase the size of our datagrams
3819 if ((int)call->nAcks > rx_nDgramThreshold) {
3820 if (peer->maxDgramPackets > 1) {
3821 if (call->nDgramPackets < peer->maxDgramPackets) {
3822 call->nDgramPackets++;
3824 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3825 } else if (call->MTU < peer->maxMTU) {
3826 call->MTU += peer->natMTU;
3827 call->MTU = MIN(call->MTU, peer->maxMTU);
3833 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3835 /* Servers need to hold the call until all response packets have
3836 * been acknowledged. Soft acks are good enough since clients
3837 * are not allowed to clear their receive queues. */
3838 if (call->state == RX_STATE_HOLD
3839 && call->tfirst + call->nSoftAcked >= call->tnext) {
3840 call->state = RX_STATE_DALLY;
3841 rxi_ClearTransmitQueue(call, 0);
3842 } else if (!queue_IsEmpty(&call->tq)) {
3843 rxi_Start(0, call, istack);
3848 /* Received a response to a challenge packet */
3850 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
3851 register struct rx_packet *np, int istack)
3855 /* Ignore the packet if we're the client */
3856 if (conn->type == RX_CLIENT_CONNECTION)
3859 /* If already authenticated, ignore the packet (it's probably a retry) */
3860 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
3863 /* Otherwise, have the security object evaluate the response packet */
3864 error = RXS_CheckResponse(conn->securityObject, conn, np);
3866 /* If the response is invalid, reset the connection, sending
3867 * an abort to the peer */
3871 rxi_ConnectionError(conn, error);
3872 MUTEX_ENTER(&conn->conn_data_lock);
3873 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3874 MUTEX_EXIT(&conn->conn_data_lock);
3877 /* If the response is valid, any calls waiting to attach
3878 * servers can now do so */
3881 for (i = 0; i < RX_MAXCALLS; i++) {
3882 struct rx_call *call = conn->call[i];
3884 MUTEX_ENTER(&call->lock);
3885 if (call->state == RX_STATE_PRECALL)
3886 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
3887 /* tnop can be null if newcallp is null */
3888 MUTEX_EXIT(&call->lock);
3892 /* Update the peer reachability information, just in case
3893 * some calls went into attach-wait while we were waiting
3894 * for authentication..
3896 rxi_UpdatePeerReach(conn, NULL);
3901 /* A client has received an authentication challenge: the security
3902 * object is asked to cough up a respectable response packet to send
3903 * back to the server. The server is responsible for retrying the
3904 * challenge if it fails to get a response. */
3907 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
3908 register struct rx_packet *np, int istack)
3912 /* Ignore the challenge if we're the server */
3913 if (conn->type == RX_SERVER_CONNECTION)
3916 /* Ignore the challenge if the connection is otherwise idle; someone's
3917 * trying to use us as an oracle. */
3918 if (!rxi_HasActiveCalls(conn))
3921 /* Send the security object the challenge packet. It is expected to fill
3922 * in the response. */
3923 error = RXS_GetResponse(conn->securityObject, conn, np);
3925 /* If the security object is unable to return a valid response, reset the
3926 * connection and send an abort to the peer. Otherwise send the response
3927 * packet to the peer connection. */
3929 rxi_ConnectionError(conn, error);
3930 MUTEX_ENTER(&conn->conn_data_lock);
3931 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3932 MUTEX_EXIT(&conn->conn_data_lock);
3934 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
3935 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
3941 /* Find an available server process to service the current request in
3942 * the given call structure. If one isn't available, queue up this
3943 * call so it eventually gets one */
3945 rxi_AttachServerProc(register struct rx_call *call,
3946 register osi_socket socket, register int *tnop,
3947 register struct rx_call **newcallp)
3949 register struct rx_serverQueueEntry *sq;
3950 register struct rx_service *service = call->conn->service;
3951 register int haveQuota = 0;
3953 /* May already be attached */
3954 if (call->state == RX_STATE_ACTIVE)
3957 MUTEX_ENTER(&rx_serverPool_lock);
3959 haveQuota = QuotaOK(service);
3960 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
3961 /* If there are no processes available to service this call,
3962 * put the call on the incoming call queue (unless it's
3963 * already on the queue).
3965 #ifdef RX_ENABLE_LOCKS
3967 ReturnToServerPool(service);
3968 #endif /* RX_ENABLE_LOCKS */
3970 if (!(call->flags & RX_CALL_WAIT_PROC)) {
3971 call->flags |= RX_CALL_WAIT_PROC;
3972 MUTEX_ENTER(&rx_stats_mutex);
3974 MUTEX_EXIT(&rx_stats_mutex);
3975 rxi_calltrace(RX_CALL_ARRIVAL, call);
3976 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
3977 queue_Append(&rx_incomingCallQueue, call);
3980 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
3982 /* If hot threads are enabled, and both newcallp and sq->socketp
3983 * are non-null, then this thread will process the call, and the
3984 * idle server thread will start listening on this threads socket.
3987 if (rx_enable_hot_thread && newcallp && sq->socketp) {
3990 *sq->socketp = socket;
3991 clock_GetTime(&call->startTime);
3992 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
3996 if (call->flags & RX_CALL_WAIT_PROC) {
3997 /* Conservative: I don't think this should happen */
3998 call->flags &= ~RX_CALL_WAIT_PROC;
3999 if (queue_IsOnQueue(call)) {
4001 MUTEX_ENTER(&rx_stats_mutex);
4003 MUTEX_EXIT(&rx_stats_mutex);
4006 call->state = RX_STATE_ACTIVE;
4007 call->mode = RX_MODE_RECEIVING;
4008 #ifdef RX_KERNEL_TRACE
4010 int glockOwner = ISAFS_GLOCK();
4013 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4014 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4020 if (call->flags & RX_CALL_CLEARED) {
4021 /* send an ack now to start the packet flow up again */
4022 call->flags &= ~RX_CALL_CLEARED;
4023 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4025 #ifdef RX_ENABLE_LOCKS
4028 service->nRequestsRunning++;
4029 if (service->nRequestsRunning <= service->minProcs)
4035 MUTEX_EXIT(&rx_serverPool_lock);
4038 /* Delay the sending of an acknowledge event for a short while, while
4039 * a new call is being prepared (in the case of a client) or a reply
4040 * is being prepared (in the case of a server). Rather than sending
4041 * an ack packet, an ACKALL packet is sent. */
4043 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4045 #ifdef RX_ENABLE_LOCKS
4047 MUTEX_ENTER(&call->lock);
4048 call->delayedAckEvent = NULL;
4049 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4051 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4052 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4054 MUTEX_EXIT(&call->lock);
4055 #else /* RX_ENABLE_LOCKS */
4057 call->delayedAckEvent = NULL;
4058 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4059 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4060 #endif /* RX_ENABLE_LOCKS */
4064 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4067 #ifdef RX_ENABLE_LOCKS
4069 MUTEX_ENTER(&call->lock);
4070 if (event == call->delayedAckEvent)
4071 call->delayedAckEvent = NULL;
4072 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4074 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4076 MUTEX_EXIT(&call->lock);
4077 #else /* RX_ENABLE_LOCKS */
4079 call->delayedAckEvent = NULL;
4080 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4081 #endif /* RX_ENABLE_LOCKS */
4085 #ifdef RX_ENABLE_LOCKS
4086 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4087 * clearing them out.
4090 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4092 register struct rx_packet *p, *tp;
4095 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4098 p->flags |= RX_PKTFLAG_ACKED;
4102 call->flags |= RX_CALL_TQ_CLEARME;
4103 call->flags |= RX_CALL_TQ_SOME_ACKED;
4106 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4107 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4108 call->tfirst = call->tnext;
4109 call->nSoftAcked = 0;
4111 if (call->flags & RX_CALL_FAST_RECOVER) {
4112 call->flags &= ~RX_CALL_FAST_RECOVER;
4113 call->cwind = call->nextCwind;
4114 call->nextCwind = 0;
4117 CV_SIGNAL(&call->cv_twind);
4119 #endif /* RX_ENABLE_LOCKS */
4121 /* Clear out the transmit queue for the current call (all packets have
4122 * been received by peer) */
4124 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4126 register struct rx_packet *p, *tp;
4128 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4129 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4131 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4134 p->flags |= RX_PKTFLAG_ACKED;
4138 call->flags |= RX_CALL_TQ_CLEARME;
4139 call->flags |= RX_CALL_TQ_SOME_ACKED;
4142 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4143 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4149 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4150 call->flags &= ~RX_CALL_TQ_CLEARME;
4152 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4154 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4155 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4156 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4157 call->nSoftAcked = 0;
4159 if (call->flags & RX_CALL_FAST_RECOVER) {
4160 call->flags &= ~RX_CALL_FAST_RECOVER;
4161 call->cwind = call->nextCwind;
4163 #ifdef RX_ENABLE_LOCKS
4164 CV_SIGNAL(&call->cv_twind);
4166 osi_rxWakeup(&call->twind);
4171 rxi_ClearReceiveQueue(register struct rx_call *call)
4173 register struct rx_packet *p, *tp;
4174 if (queue_IsNotEmpty(&call->rq)) {
4175 for (queue_Scan(&call->rq, p, tp, rx_packet)) {
4180 rx_packetReclaims++;
4182 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4184 if (call->state == RX_STATE_PRECALL) {
4185 call->flags |= RX_CALL_CLEARED;
4189 /* Send an abort packet for the specified call */
4191 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4192 int istack, int force)
4200 /* Clients should never delay abort messages */
4201 if (rx_IsClientConn(call->conn))
4204 if (call->abortCode != call->error) {
4205 call->abortCode = call->error;
4206 call->abortCount = 0;
4209 if (force || rxi_callAbortThreshhold == 0
4210 || call->abortCount < rxi_callAbortThreshhold) {
4211 if (call->delayedAbortEvent) {
4212 rxevent_Cancel(call->delayedAbortEvent, call,
4213 RX_CALL_REFCOUNT_ABORT);
4215 error = htonl(call->error);
4218 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4219 (char *)&error, sizeof(error), istack);
4220 } else if (!call->delayedAbortEvent) {
4221 clock_GetTime(&when);
4222 clock_Addmsec(&when, rxi_callAbortDelay);
4223 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4224 call->delayedAbortEvent =
4225 rxevent_Post(&when, rxi_SendDelayedCallAbort, call, 0);
4230 /* Send an abort packet for the specified connection. Packet is an
4231 * optional pointer to a packet that can be used to send the abort.
4232 * Once the number of abort messages reaches the threshhold, an
4233 * event is scheduled to send the abort. Setting the force flag
4234 * overrides sending delayed abort messages.
4236 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4237 * to send the abort packet.
4240 rxi_SendConnectionAbort(register struct rx_connection *conn,
4241 struct rx_packet *packet, int istack, int force)
4249 /* Clients should never delay abort messages */
4250 if (rx_IsClientConn(conn))
4253 if (force || rxi_connAbortThreshhold == 0
4254 || conn->abortCount < rxi_connAbortThreshhold) {
4255 if (conn->delayedAbortEvent) {
4256 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4258 error = htonl(conn->error);
4260 MUTEX_EXIT(&conn->conn_data_lock);
4262 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4263 RX_PACKET_TYPE_ABORT, (char *)&error,
4264 sizeof(error), istack);
4265 MUTEX_ENTER(&conn->conn_data_lock);
4266 } else if (!conn->delayedAbortEvent) {
4267 clock_GetTime(&when);
4268 clock_Addmsec(&when, rxi_connAbortDelay);
4269 conn->delayedAbortEvent =
4270 rxevent_Post(&when, rxi_SendDelayedConnAbort, conn, 0);
4275 /* Associate an error all of the calls owned by a connection. Called
4276 * with error non-zero. This is only for really fatal things, like
4277 * bad authentication responses. The connection itself is set in
4278 * error at this point, so that future packets received will be
4281 rxi_ConnectionError(register struct rx_connection *conn,
4282 register afs_int32 error)
4286 MUTEX_ENTER(&conn->conn_data_lock);
4287 if (conn->challengeEvent)
4288 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4289 if (conn->checkReachEvent) {
4290 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4291 conn->checkReachEvent = 0;
4292 conn->flags &= ~RX_CONN_ATTACHWAIT;
4295 MUTEX_EXIT(&conn->conn_data_lock);
4296 for (i = 0; i < RX_MAXCALLS; i++) {
4297 struct rx_call *call = conn->call[i];
4299 MUTEX_ENTER(&call->lock);
4300 rxi_CallError(call, error);
4301 MUTEX_EXIT(&call->lock);
4304 conn->error = error;
4305 MUTEX_ENTER(&rx_stats_mutex);
4306 rx_stats.fatalErrors++;
4307 MUTEX_EXIT(&rx_stats_mutex);
4312 rxi_CallError(register struct rx_call *call, afs_int32 error)
4315 error = call->error;
4316 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4317 if (!(call->flags & RX_CALL_TQ_BUSY)) {
4318 rxi_ResetCall(call, 0);
4321 rxi_ResetCall(call, 0);
4323 call->error = error;
4324 call->mode = RX_MODE_ERROR;
4327 /* Reset various fields in a call structure, and wakeup waiting
4328 * processes. Some fields aren't changed: state & mode are not
4329 * touched (these must be set by the caller), and bufptr, nLeft, and
4330 * nFree are not reset, since these fields are manipulated by
4331 * unprotected macros, and may only be reset by non-interrupting code.
4334 /* this code requires that call->conn be set properly as a pre-condition. */
4335 #endif /* ADAPT_WINDOW */
4338 rxi_ResetCall(register struct rx_call *call, register int newcall)
4341 register struct rx_peer *peer;
4342 struct rx_packet *packet;
4344 /* Notify anyone who is waiting for asynchronous packet arrival */
4345 if (call->arrivalProc) {
4346 (*call->arrivalProc) (call, call->arrivalProcHandle,
4347 (int)call->arrivalProcArg);
4348 call->arrivalProc = (VOID(*)())0;
4351 if (call->delayedAbortEvent) {
4352 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4353 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4355 rxi_SendCallAbort(call, packet, 0, 1);
4356 rxi_FreePacket(packet);
4361 * Update the peer with the congestion information in this call
4362 * so other calls on this connection can pick up where this call
4363 * left off. If the congestion sequence numbers don't match then
4364 * another call experienced a retransmission.
4366 peer = call->conn->peer;
4367 MUTEX_ENTER(&peer->peer_lock);
4369 if (call->congestSeq == peer->congestSeq) {
4370 peer->cwind = MAX(peer->cwind, call->cwind);
4371 peer->MTU = MAX(peer->MTU, call->MTU);
4372 peer->nDgramPackets =
4373 MAX(peer->nDgramPackets, call->nDgramPackets);
4376 call->abortCode = 0;
4377 call->abortCount = 0;
4379 if (peer->maxDgramPackets > 1) {
4380 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4382 call->MTU = peer->MTU;
4384 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4385 call->ssthresh = rx_maxSendWindow;
4386 call->nDgramPackets = peer->nDgramPackets;
4387 call->congestSeq = peer->congestSeq;
4388 MUTEX_EXIT(&peer->peer_lock);
4390 flags = call->flags;
4391 rxi_ClearReceiveQueue(call);
4392 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4393 if (call->flags & RX_CALL_TQ_BUSY) {
4394 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4395 call->flags |= (flags & RX_CALL_TQ_WAIT);
4397 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4399 rxi_ClearTransmitQueue(call, 0);
4400 queue_Init(&call->tq);
4403 queue_Init(&call->rq);
4405 call->rwind = rx_initReceiveWindow;
4406 call->twind = rx_initSendWindow;
4407 call->nSoftAcked = 0;
4408 call->nextCwind = 0;
4411 call->nCwindAcks = 0;
4412 call->nSoftAcks = 0;
4413 call->nHardAcks = 0;
4415 call->tfirst = call->rnext = call->tnext = 1;
4417 call->lastAcked = 0;
4418 call->localStatus = call->remoteStatus = 0;
4420 if (flags & RX_CALL_READER_WAIT) {
4421 #ifdef RX_ENABLE_LOCKS
4422 CV_BROADCAST(&call->cv_rq);
4424 osi_rxWakeup(&call->rq);
4427 if (flags & RX_CALL_WAIT_PACKETS) {
4428 MUTEX_ENTER(&rx_freePktQ_lock);
4429 rxi_PacketsUnWait(); /* XXX */
4430 MUTEX_EXIT(&rx_freePktQ_lock);
4432 #ifdef RX_ENABLE_LOCKS
4433 CV_SIGNAL(&call->cv_twind);
4435 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4436 osi_rxWakeup(&call->twind);
4439 #ifdef RX_ENABLE_LOCKS
4440 /* The following ensures that we don't mess with any queue while some
4441 * other thread might also be doing so. The call_queue_lock field is
4442 * is only modified under the call lock. If the call is in the process
4443 * of being removed from a queue, the call is not locked until the
4444 * the queue lock is dropped and only then is the call_queue_lock field
4445 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4446 * Note that any other routine which removes a call from a queue has to
4447 * obtain the queue lock before examing the queue and removing the call.
4449 if (call->call_queue_lock) {
4450 MUTEX_ENTER(call->call_queue_lock);
4451 if (queue_IsOnQueue(call)) {
4453 if (flags & RX_CALL_WAIT_PROC) {
4454 MUTEX_ENTER(&rx_stats_mutex);
4456 MUTEX_EXIT(&rx_stats_mutex);
4459 MUTEX_EXIT(call->call_queue_lock);
4460 CLEAR_CALL_QUEUE_LOCK(call);
4462 #else /* RX_ENABLE_LOCKS */
4463 if (queue_IsOnQueue(call)) {
4465 if (flags & RX_CALL_WAIT_PROC)
4468 #endif /* RX_ENABLE_LOCKS */
4470 rxi_KeepAliveOff(call);
4471 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4474 /* Send an acknowledge for the indicated packet (seq,serial) of the
4475 * indicated call, for the indicated reason (reason). This
4476 * acknowledge will specifically acknowledge receiving the packet, and
4477 * will also specify which other packets for this call have been
4478 * received. This routine returns the packet that was used to the
4479 * caller. The caller is responsible for freeing it or re-using it.
4480 * This acknowledgement also returns the highest sequence number
4481 * actually read out by the higher level to the sender; the sender
4482 * promises to keep around packets that have not been read by the
4483 * higher level yet (unless, of course, the sender decides to abort
4484 * the call altogether). Any of p, seq, serial, pflags, or reason may
4485 * be set to zero without ill effect. That is, if they are zero, they
4486 * will not convey any information.
4487 * NOW there is a trailer field, after the ack where it will safely be
4488 * ignored by mundanes, which indicates the maximum size packet this
4489 * host can swallow. */
4491 register struct rx_packet *optionalPacket; use to send ack (or null)
4492 int seq; Sequence number of the packet we are acking
4493 int serial; Serial number of the packet
4494 int pflags; Flags field from packet header
4495 int reason; Reason an acknowledge was prompted
4499 rxi_SendAck(register struct rx_call *call,
4500 register struct rx_packet *optionalPacket, int serial, int reason,
4503 struct rx_ackPacket *ap;
4504 register struct rx_packet *rqp;
4505 register struct rx_packet *nxp; /* For queue_Scan */
4506 register struct rx_packet *p;
4511 * Open the receive window once a thread starts reading packets
4513 if (call->rnext > 1) {
4514 call->rwind = rx_maxReceiveWindow;
4517 call->nHardAcks = 0;
4518 call->nSoftAcks = 0;
4519 if (call->rnext > call->lastAcked)
4520 call->lastAcked = call->rnext;
4524 rx_computelen(p, p->length); /* reset length, you never know */
4525 } /* where that's been... */
4526 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4527 /* We won't send the ack, but don't panic. */
4528 return optionalPacket;
4532 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4535 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL)) {
4536 if (!optionalPacket)
4538 return optionalPacket;
4540 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4541 if (rx_Contiguous(p) < templ) {
4542 if (!optionalPacket)
4544 return optionalPacket;
4549 /* MTUXXX failing to send an ack is very serious. We should */
4550 /* try as hard as possible to send even a partial ack; it's */
4551 /* better than nothing. */
4552 ap = (struct rx_ackPacket *)rx_DataOf(p);
4553 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4554 ap->reason = reason;
4556 /* The skew computation used to be bogus, I think it's better now. */
4557 /* We should start paying attention to skew. XXX */
4558 ap->serial = htonl(serial);
4559 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4561 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4562 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4564 /* No fear of running out of ack packet here because there can only be at most
4565 * one window full of unacknowledged packets. The window size must be constrained
4566 * to be less than the maximum ack size, of course. Also, an ack should always
4567 * fit into a single packet -- it should not ever be fragmented. */
4568 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4569 if (!rqp || !call->rq.next
4570 || (rqp->header.seq > (call->rnext + call->rwind))) {
4571 if (!optionalPacket)
4573 rxi_CallError(call, RX_CALL_DEAD);
4574 return optionalPacket;
4577 while (rqp->header.seq > call->rnext + offset)
4578 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4579 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4581 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4582 if (!optionalPacket)
4584 rxi_CallError(call, RX_CALL_DEAD);
4585 return optionalPacket;
4590 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4592 /* these are new for AFS 3.3 */
4593 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4594 templ = htonl(templ);
4595 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4596 templ = htonl(call->conn->peer->ifMTU);
4597 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4598 sizeof(afs_int32), &templ);
4600 /* new for AFS 3.4 */
4601 templ = htonl(call->rwind);
4602 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4603 sizeof(afs_int32), &templ);
4605 /* new for AFS 3.5 */
4606 templ = htonl(call->conn->peer->ifDgramPackets);
4607 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4608 sizeof(afs_int32), &templ);
4610 p->header.serviceId = call->conn->serviceId;
4611 p->header.cid = (call->conn->cid | call->channel);
4612 p->header.callNumber = *call->callNumber;
4614 p->header.securityIndex = call->conn->securityIndex;
4615 p->header.epoch = call->conn->epoch;
4616 p->header.type = RX_PACKET_TYPE_ACK;
4617 p->header.flags = RX_SLOW_START_OK;
4618 if (reason == RX_ACK_PING) {
4619 p->header.flags |= RX_REQUEST_ACK;
4621 clock_GetTime(&call->pingRequestTime);
4624 if (call->conn->type == RX_CLIENT_CONNECTION)
4625 p->header.flags |= RX_CLIENT_INITIATED;
4629 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u",
4630 ap->reason, ntohl(ap->previousPacket),
4631 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4633 for (offset = 0; offset < ap->nAcks; offset++)
4634 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4642 register int i, nbytes = p->length;
4644 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4645 if (nbytes <= p->wirevec[i].iov_len) {
4646 register int savelen, saven;
4648 savelen = p->wirevec[i].iov_len;
4650 p->wirevec[i].iov_len = nbytes;
4652 rxi_Send(call, p, istack);
4653 p->wirevec[i].iov_len = savelen;
4657 nbytes -= p->wirevec[i].iov_len;
4660 MUTEX_ENTER(&rx_stats_mutex);
4661 rx_stats.ackPacketsSent++;
4662 MUTEX_EXIT(&rx_stats_mutex);
4663 if (!optionalPacket)
4665 return optionalPacket; /* Return packet for re-use by caller */
4668 /* Send all of the packets in the list in single datagram */
4670 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4671 int istack, int moreFlag, struct clock *now,
4672 struct clock *retryTime, int resending)
4677 struct rx_connection *conn = call->conn;
4678 struct rx_peer *peer = conn->peer;
4680 MUTEX_ENTER(&peer->peer_lock);
4683 peer->reSends += len;
4684 MUTEX_ENTER(&rx_stats_mutex);
4685 rx_stats.dataPacketsSent += len;
4686 MUTEX_EXIT(&rx_stats_mutex);
4687 MUTEX_EXIT(&peer->peer_lock);
4689 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4693 /* Set the packet flags and schedule the resend events */
4694 /* Only request an ack for the last packet in the list */
4695 for (i = 0; i < len; i++) {
4696 list[i]->retryTime = *retryTime;
4697 if (list[i]->header.serial) {
4698 /* Exponentially backoff retry times */
4699 if (list[i]->backoff < MAXBACKOFF) {
4700 /* so it can't stay == 0 */
4701 list[i]->backoff = (list[i]->backoff << 1) + 1;
4704 clock_Addmsec(&(list[i]->retryTime),
4705 ((afs_uint32) list[i]->backoff) << 8);
4708 /* Wait a little extra for the ack on the last packet */
4709 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4710 clock_Addmsec(&(list[i]->retryTime), 400);
4713 /* Record the time sent */
4714 list[i]->timeSent = *now;
4716 /* Ask for an ack on retransmitted packets, on every other packet
4717 * if the peer doesn't support slow start. Ask for an ack on every
4718 * packet until the congestion window reaches the ack rate. */
4719 if (list[i]->header.serial) {
4721 MUTEX_ENTER(&rx_stats_mutex);
4722 rx_stats.dataPacketsReSent++;
4723 MUTEX_EXIT(&rx_stats_mutex);
4725 /* improved RTO calculation- not Karn */
4726 list[i]->firstSent = *now;
4727 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4728 || (!(call->flags & RX_CALL_SLOW_START_OK)
4729 && (list[i]->header.seq & 1)))) {
4734 MUTEX_ENTER(&peer->peer_lock);
4738 MUTEX_ENTER(&rx_stats_mutex);
4739 rx_stats.dataPacketsSent++;
4740 MUTEX_EXIT(&rx_stats_mutex);
4741 MUTEX_EXIT(&peer->peer_lock);
4743 /* Tag this packet as not being the last in this group,
4744 * for the receiver's benefit */
4745 if (i < len - 1 || moreFlag) {
4746 list[i]->header.flags |= RX_MORE_PACKETS;
4749 /* Install the new retransmit time for the packet, and
4750 * record the time sent */
4751 list[i]->timeSent = *now;
4755 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4758 /* Since we're about to send a data packet to the peer, it's
4759 * safe to nuke any scheduled end-of-packets ack */
4760 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4762 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4763 MUTEX_EXIT(&call->lock);
4765 rxi_SendPacketList(call, conn, list, len, istack);
4767 rxi_SendPacket(call, conn, list[0], istack);
4769 MUTEX_ENTER(&call->lock);
4770 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4772 /* Update last send time for this call (for keep-alive
4773 * processing), and for the connection (so that we can discover
4774 * idle connections) */
4775 conn->lastSendTime = call->lastSendTime = clock_Sec();
4778 /* When sending packets we need to follow these rules:
4779 * 1. Never send more than maxDgramPackets in a jumbogram.
4780 * 2. Never send a packet with more than two iovecs in a jumbogram.
4781 * 3. Never send a retransmitted packet in a jumbogram.
4782 * 4. Never send more than cwind/4 packets in a jumbogram
4783 * We always keep the last list we should have sent so we
4784 * can set the RX_MORE_PACKETS flags correctly.
4787 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4788 int istack, struct clock *now, struct clock *retryTime,
4791 int i, cnt, lastCnt = 0;
4792 struct rx_packet **listP, **lastP = 0;
4793 struct rx_peer *peer = call->conn->peer;
4794 int morePackets = 0;
4796 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4797 /* Does the current packet force us to flush the current list? */
4799 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4800 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4802 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
4804 /* If the call enters an error state stop sending, or if
4805 * we entered congestion recovery mode, stop sending */
4806 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4814 /* Add the current packet to the list if it hasn't been acked.
4815 * Otherwise adjust the list pointer to skip the current packet. */
4816 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
4818 /* Do we need to flush the list? */
4819 if (cnt >= (int)peer->maxDgramPackets
4820 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
4821 || list[i]->header.serial
4822 || list[i]->length != RX_JUMBOBUFFERSIZE) {
4824 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
4825 retryTime, resending);
4826 /* If the call enters an error state stop sending, or if
4827 * we entered congestion recovery mode, stop sending */
4829 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4834 listP = &list[i + 1];
4839 osi_Panic("rxi_SendList error");
4841 listP = &list[i + 1];
4845 /* Send the whole list when the call is in receive mode, when
4846 * the call is in eof mode, when we are in fast recovery mode,
4847 * and when we have the last packet */
4848 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
4849 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
4850 || (call->flags & RX_CALL_FAST_RECOVER)) {
4851 /* Check for the case where the current list contains
4852 * an acked packet. Since we always send retransmissions
4853 * in a separate packet, we only need to check the first
4854 * packet in the list */
4855 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
4859 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
4860 retryTime, resending);
4861 /* If the call enters an error state stop sending, or if
4862 * we entered congestion recovery mode, stop sending */
4863 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4867 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
4870 } else if (lastCnt > 0) {
4871 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
4876 #ifdef RX_ENABLE_LOCKS
4877 /* Call rxi_Start, below, but with the call lock held. */
4879 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
4882 MUTEX_ENTER(&call->lock);
4883 rxi_Start(event, call, istack);
4884 MUTEX_EXIT(&call->lock);
4886 #endif /* RX_ENABLE_LOCKS */
4888 /* This routine is called when new packets are readied for
4889 * transmission and when retransmission may be necessary, or when the
4890 * transmission window or burst count are favourable. This should be
4891 * better optimized for new packets, the usual case, now that we've
4892 * got rid of queues of send packets. XXXXXXXXXXX */
4894 rxi_Start(struct rxevent *event, register struct rx_call *call, int istack)
4896 struct rx_packet *p;
4897 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
4898 struct rx_peer *peer = call->conn->peer;
4899 struct clock now, retryTime;
4903 struct rx_packet **xmitList;
4906 /* If rxi_Start is being called as a result of a resend event,
4907 * then make sure that the event pointer is removed from the call
4908 * structure, since there is no longer a per-call retransmission
4910 if (event && event == call->resendEvent) {
4911 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
4912 call->resendEvent = NULL;
4914 if (queue_IsEmpty(&call->tq)) {
4918 /* Timeouts trigger congestion recovery */
4919 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4920 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4921 /* someone else is waiting to start recovery */
4924 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4925 while (call->flags & RX_CALL_TQ_BUSY) {
4926 call->flags |= RX_CALL_TQ_WAIT;
4927 #ifdef RX_ENABLE_LOCKS
4928 CV_WAIT(&call->cv_tq, &call->lock);
4929 #else /* RX_ENABLE_LOCKS */
4930 osi_rxSleep(&call->tq);
4931 #endif /* RX_ENABLE_LOCKS */
4933 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4934 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4935 call->flags |= RX_CALL_FAST_RECOVER;
4936 if (peer->maxDgramPackets > 1) {
4937 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4939 call->MTU = MIN(peer->natMTU, peer->maxMTU);
4941 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4942 call->nDgramPackets = 1;
4944 call->nextCwind = 1;
4947 MUTEX_ENTER(&peer->peer_lock);
4948 peer->MTU = call->MTU;
4949 peer->cwind = call->cwind;
4950 peer->nDgramPackets = 1;
4952 call->congestSeq = peer->congestSeq;
4953 MUTEX_EXIT(&peer->peer_lock);
4954 /* Clear retry times on packets. Otherwise, it's possible for
4955 * some packets in the queue to force resends at rates faster
4956 * than recovery rates.
4958 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
4959 if (!(p->flags & RX_PKTFLAG_ACKED)) {
4960 clock_Zero(&p->retryTime);
4965 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4966 MUTEX_ENTER(&rx_stats_mutex);
4967 rx_tq_debug.rxi_start_in_error++;
4968 MUTEX_EXIT(&rx_stats_mutex);
4973 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
4974 /* Get clock to compute the re-transmit time for any packets
4975 * in this burst. Note, if we back off, it's reasonable to
4976 * back off all of the packets in the same manner, even if
4977 * some of them have been retransmitted more times than more
4978 * recent additions */
4979 clock_GetTime(&now);
4980 retryTime = now; /* initialize before use */
4981 MUTEX_ENTER(&peer->peer_lock);
4982 clock_Add(&retryTime, &peer->timeout);
4983 MUTEX_EXIT(&peer->peer_lock);
4985 /* Send (or resend) any packets that need it, subject to
4986 * window restrictions and congestion burst control
4987 * restrictions. Ask for an ack on the last packet sent in
4988 * this burst. For now, we're relying upon the window being
4989 * considerably bigger than the largest number of packets that
4990 * are typically sent at once by one initial call to
4991 * rxi_Start. This is probably bogus (perhaps we should ask
4992 * for an ack when we're half way through the current
4993 * window?). Also, for non file transfer applications, this
4994 * may end up asking for an ack for every packet. Bogus. XXXX
4997 * But check whether we're here recursively, and let the other guy
5000 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5001 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5002 call->flags |= RX_CALL_TQ_BUSY;
5004 call->flags &= ~RX_CALL_NEED_START;
5005 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5007 maxXmitPackets = MIN(call->twind, call->cwind);
5008 xmitList = (struct rx_packet **)
5009 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5010 if (xmitList == NULL)
5011 osi_Panic("rxi_Start, failed to allocate xmit list");
5012 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5013 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5014 /* We shouldn't be sending packets if a thread is waiting
5015 * to initiate congestion recovery */
5019 && (call->flags & RX_CALL_FAST_RECOVER)) {
5020 /* Only send one packet during fast recovery */
5023 if ((p->flags & RX_PKTFLAG_FREE)
5024 || (!queue_IsEnd(&call->tq, nxp)
5025 && (nxp->flags & RX_PKTFLAG_FREE))
5026 || (p == (struct rx_packet *)&rx_freePacketQueue)
5027 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5028 osi_Panic("rxi_Start: xmit queue clobbered");
5030 if (p->flags & RX_PKTFLAG_ACKED) {
5031 MUTEX_ENTER(&rx_stats_mutex);
5032 rx_stats.ignoreAckedPacket++;
5033 MUTEX_EXIT(&rx_stats_mutex);
5034 continue; /* Ignore this packet if it has been acknowledged */
5037 /* Turn off all flags except these ones, which are the same
5038 * on each transmission */
5039 p->header.flags &= RX_PRESET_FLAGS;
5041 if (p->header.seq >=
5042 call->tfirst + MIN((int)call->twind,
5043 (int)(call->nSoftAcked +
5045 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5046 /* Note: if we're waiting for more window space, we can
5047 * still send retransmits; hence we don't return here, but
5048 * break out to schedule a retransmit event */
5049 dpf(("call %d waiting for window",
5050 *(call->callNumber)));
5054 /* Transmit the packet if it needs to be sent. */
5055 if (!clock_Lt(&now, &p->retryTime)) {
5056 if (nXmitPackets == maxXmitPackets) {
5057 osi_Panic("rxi_Start: xmit list overflowed");
5059 xmitList[nXmitPackets++] = p;
5063 /* xmitList now hold pointers to all of the packets that are
5064 * ready to send. Now we loop to send the packets */
5065 if (nXmitPackets > 0) {
5066 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5067 &now, &retryTime, resending);
5070 maxXmitPackets * sizeof(struct rx_packet *));
5072 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5074 * TQ references no longer protected by this flag; they must remain
5075 * protected by the global lock.
5077 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5078 call->flags &= ~RX_CALL_TQ_BUSY;
5079 if (call->flags & RX_CALL_TQ_WAIT) {
5080 call->flags &= ~RX_CALL_TQ_WAIT;
5081 #ifdef RX_ENABLE_LOCKS
5082 CV_BROADCAST(&call->cv_tq);
5083 #else /* RX_ENABLE_LOCKS */
5084 osi_rxWakeup(&call->tq);
5085 #endif /* RX_ENABLE_LOCKS */
5090 /* We went into the error state while sending packets. Now is
5091 * the time to reset the call. This will also inform the using
5092 * process that the call is in an error state.
5094 MUTEX_ENTER(&rx_stats_mutex);
5095 rx_tq_debug.rxi_start_aborted++;
5096 MUTEX_EXIT(&rx_stats_mutex);
5097 call->flags &= ~RX_CALL_TQ_BUSY;
5098 if (call->flags & RX_CALL_TQ_WAIT) {
5099 call->flags &= ~RX_CALL_TQ_WAIT;
5100 #ifdef RX_ENABLE_LOCKS
5101 CV_BROADCAST(&call->cv_tq);
5102 #else /* RX_ENABLE_LOCKS */
5103 osi_rxWakeup(&call->tq);
5104 #endif /* RX_ENABLE_LOCKS */
5106 rxi_CallError(call, call->error);
5109 #ifdef RX_ENABLE_LOCKS
5110 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5111 register int missing;
5112 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5113 /* Some packets have received acks. If they all have, we can clear
5114 * the transmit queue.
5117 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5118 if (p->header.seq < call->tfirst
5119 && (p->flags & RX_PKTFLAG_ACKED)) {
5126 call->flags |= RX_CALL_TQ_CLEARME;
5128 #endif /* RX_ENABLE_LOCKS */
5129 /* Don't bother doing retransmits if the TQ is cleared. */
5130 if (call->flags & RX_CALL_TQ_CLEARME) {
5131 rxi_ClearTransmitQueue(call, 1);
5133 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5136 /* Always post a resend event, if there is anything in the
5137 * queue, and resend is possible. There should be at least
5138 * one unacknowledged packet in the queue ... otherwise none
5139 * of these packets should be on the queue in the first place.
5141 if (call->resendEvent) {
5142 /* Cancel the existing event and post a new one */
5143 rxevent_Cancel(call->resendEvent, call,
5144 RX_CALL_REFCOUNT_RESEND);
5147 /* The retry time is the retry time on the first unacknowledged
5148 * packet inside the current window */
5150 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5151 /* Don't set timers for packets outside the window */
5152 if (p->header.seq >= call->tfirst + call->twind) {
5156 if (!(p->flags & RX_PKTFLAG_ACKED)
5157 && !clock_IsZero(&p->retryTime)) {
5159 retryTime = p->retryTime;
5164 /* Post a new event to re-run rxi_Start when retries may be needed */
5165 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5166 #ifdef RX_ENABLE_LOCKS
5167 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5169 rxevent_Post(&retryTime, rxi_StartUnlocked,
5170 (void *)call, (void *)istack);
5171 #else /* RX_ENABLE_LOCKS */
5173 rxevent_Post(&retryTime, rxi_Start, (void *)call,
5175 #endif /* RX_ENABLE_LOCKS */
5178 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5179 } while (call->flags & RX_CALL_NEED_START);
5181 * TQ references no longer protected by this flag; they must remain
5182 * protected by the global lock.
5184 call->flags &= ~RX_CALL_TQ_BUSY;
5185 if (call->flags & RX_CALL_TQ_WAIT) {
5186 call->flags &= ~RX_CALL_TQ_WAIT;
5187 #ifdef RX_ENABLE_LOCKS
5188 CV_BROADCAST(&call->cv_tq);
5189 #else /* RX_ENABLE_LOCKS */
5190 osi_rxWakeup(&call->tq);
5191 #endif /* RX_ENABLE_LOCKS */
5194 call->flags |= RX_CALL_NEED_START;
5196 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5198 if (call->resendEvent) {
5199 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5204 /* Also adjusts the keep alive parameters for the call, to reflect
5205 * that we have just sent a packet (so keep alives aren't sent
5208 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5211 register struct rx_connection *conn = call->conn;
5213 /* Stamp each packet with the user supplied status */
5214 p->header.userStatus = call->localStatus;
5216 /* Allow the security object controlling this call's security to
5217 * make any last-minute changes to the packet */
5218 RXS_SendPacket(conn->securityObject, call, p);
5220 /* Since we're about to send SOME sort of packet to the peer, it's
5221 * safe to nuke any scheduled end-of-packets ack */
5222 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5224 /* Actually send the packet, filling in more connection-specific fields */
5225 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5226 MUTEX_EXIT(&call->lock);
5227 rxi_SendPacket(call, conn, p, istack);
5228 MUTEX_ENTER(&call->lock);
5229 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5231 /* Update last send time for this call (for keep-alive
5232 * processing), and for the connection (so that we can discover
5233 * idle connections) */
5234 conn->lastSendTime = call->lastSendTime = clock_Sec();
5238 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5239 * that things are fine. Also called periodically to guarantee that nothing
5240 * falls through the cracks (e.g. (error + dally) connections have keepalive
5241 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5243 * haveCTLock Set if calling from rxi_ReapConnections
5245 #ifdef RX_ENABLE_LOCKS
5247 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5248 #else /* RX_ENABLE_LOCKS */
5250 rxi_CheckCall(register struct rx_call *call)
5251 #endif /* RX_ENABLE_LOCKS */
5253 register struct rx_connection *conn = call->conn;
5255 afs_uint32 deadTime;
5257 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5258 if (call->flags & RX_CALL_TQ_BUSY) {
5259 /* Call is active and will be reset by rxi_Start if it's
5260 * in an error state.
5265 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5267 (((afs_uint32) conn->secondsUntilDead << 10) +
5268 ((afs_uint32) conn->peer->rtt >> 3) +
5269 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5271 /* These are computed to the second (+- 1 second). But that's
5272 * good enough for these values, which should be a significant
5273 * number of seconds. */
5274 if (now > (call->lastReceiveTime + deadTime)) {
5275 if (call->state == RX_STATE_ACTIVE) {
5276 rxi_CallError(call, RX_CALL_DEAD);
5279 #ifdef RX_ENABLE_LOCKS
5280 /* Cancel pending events */
5281 rxevent_Cancel(call->delayedAckEvent, call,
5282 RX_CALL_REFCOUNT_DELAY);
5283 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5284 rxevent_Cancel(call->keepAliveEvent, call,
5285 RX_CALL_REFCOUNT_ALIVE);
5286 if (call->refCount == 0) {
5287 rxi_FreeCall(call, haveCTLock);
5291 #else /* RX_ENABLE_LOCKS */
5294 #endif /* RX_ENABLE_LOCKS */
5296 /* Non-active calls are destroyed if they are not responding
5297 * to pings; active calls are simply flagged in error, so the
5298 * attached process can die reasonably gracefully. */
5300 /* see if we have a non-activity timeout */
5301 if (call->startWait && conn->idleDeadTime
5302 && ((call->startWait + conn->idleDeadTime) < now)) {
5303 if (call->state == RX_STATE_ACTIVE) {
5304 rxi_CallError(call, RX_CALL_TIMEOUT);
5308 /* see if we have a hard timeout */
5309 if (conn->hardDeadTime
5310 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5311 if (call->state == RX_STATE_ACTIVE)
5312 rxi_CallError(call, RX_CALL_TIMEOUT);
5319 /* When a call is in progress, this routine is called occasionally to
5320 * make sure that some traffic has arrived (or been sent to) the peer.
5321 * If nothing has arrived in a reasonable amount of time, the call is
5322 * declared dead; if nothing has been sent for a while, we send a
5323 * keep-alive packet (if we're actually trying to keep the call alive)
5326 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5329 struct rx_connection *conn;
5332 MUTEX_ENTER(&call->lock);
5333 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5334 if (event == call->keepAliveEvent)
5335 call->keepAliveEvent = NULL;
5338 #ifdef RX_ENABLE_LOCKS
5339 if (rxi_CheckCall(call, 0)) {
5340 MUTEX_EXIT(&call->lock);
5343 #else /* RX_ENABLE_LOCKS */
5344 if (rxi_CheckCall(call))
5346 #endif /* RX_ENABLE_LOCKS */
5348 /* Don't try to keep alive dallying calls */
5349 if (call->state == RX_STATE_DALLY) {
5350 MUTEX_EXIT(&call->lock);
5355 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5356 /* Don't try to send keepalives if there is unacknowledged data */
5357 /* the rexmit code should be good enough, this little hack
5358 * doesn't quite work XXX */
5359 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5361 rxi_ScheduleKeepAliveEvent(call);
5362 MUTEX_EXIT(&call->lock);
5367 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5369 if (!call->keepAliveEvent) {
5371 clock_GetTime(&when);
5372 when.sec += call->conn->secondsUntilPing;
5373 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5374 call->keepAliveEvent =
5375 rxevent_Post(&when, rxi_KeepAliveEvent, call, 0);
5379 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5381 rxi_KeepAliveOn(register struct rx_call *call)
5383 /* Pretend last packet received was received now--i.e. if another
5384 * packet isn't received within the keep alive time, then the call
5385 * will die; Initialize last send time to the current time--even
5386 * if a packet hasn't been sent yet. This will guarantee that a
5387 * keep-alive is sent within the ping time */
5388 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5389 rxi_ScheduleKeepAliveEvent(call);
5392 /* This routine is called to send connection abort messages
5393 * that have been delayed to throttle looping clients. */
5395 rxi_SendDelayedConnAbort(struct rxevent *event,
5396 register struct rx_connection *conn, char *dummy)
5399 struct rx_packet *packet;
5401 MUTEX_ENTER(&conn->conn_data_lock);
5402 conn->delayedAbortEvent = NULL;
5403 error = htonl(conn->error);
5405 MUTEX_EXIT(&conn->conn_data_lock);
5406 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5409 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5410 RX_PACKET_TYPE_ABORT, (char *)&error,
5412 rxi_FreePacket(packet);
5416 /* This routine is called to send call abort messages
5417 * that have been delayed to throttle looping clients. */
5419 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5423 struct rx_packet *packet;
5425 MUTEX_ENTER(&call->lock);
5426 call->delayedAbortEvent = NULL;
5427 error = htonl(call->error);
5429 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5432 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5433 (char *)&error, sizeof(error), 0);
5434 rxi_FreePacket(packet);
5436 MUTEX_EXIT(&call->lock);
5439 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5440 * seconds) to ask the client to authenticate itself. The routine
5441 * issues a challenge to the client, which is obtained from the
5442 * security object associated with the connection */
5444 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5447 int tries = (int)atries;
5448 conn->challengeEvent = NULL;
5449 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5450 register struct rx_packet *packet;
5454 /* We've failed to authenticate for too long.
5455 * Reset any calls waiting for authentication;
5456 * they are all in RX_STATE_PRECALL.
5460 MUTEX_ENTER(&conn->conn_call_lock);
5461 for (i = 0; i < RX_MAXCALLS; i++) {
5462 struct rx_call *call = conn->call[i];
5464 MUTEX_ENTER(&call->lock);
5465 if (call->state == RX_STATE_PRECALL) {
5466 rxi_CallError(call, RX_CALL_DEAD);
5467 rxi_SendCallAbort(call, NULL, 0, 0);
5469 MUTEX_EXIT(&call->lock);
5472 MUTEX_EXIT(&conn->conn_call_lock);
5476 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5478 /* If there's no packet available, do this later. */
5479 RXS_GetChallenge(conn->securityObject, conn, packet);
5480 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5481 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5482 rxi_FreePacket(packet);
5484 clock_GetTime(&when);
5485 when.sec += RX_CHALLENGE_TIMEOUT;
5486 conn->challengeEvent =
5487 rxevent_Post(&when, rxi_ChallengeEvent, conn,
5488 (void *)(tries - 1));
5492 /* Call this routine to start requesting the client to authenticate
5493 * itself. This will continue until authentication is established,
5494 * the call times out, or an invalid response is returned. The
5495 * security object associated with the connection is asked to create
5496 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5497 * defined earlier. */
5499 rxi_ChallengeOn(register struct rx_connection *conn)
5501 if (!conn->challengeEvent) {
5502 RXS_CreateChallenge(conn->securityObject, conn);
5503 rxi_ChallengeEvent(NULL, conn, (void *)RX_CHALLENGE_MAXTRIES);
5508 /* Compute round trip time of the packet provided, in *rttp.
5511 /* rxi_ComputeRoundTripTime is called with peer locked. */
5512 /* sentp and/or peer may be null */
5514 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5515 register struct clock *sentp,
5516 register struct rx_peer *peer)
5518 struct clock thisRtt, *rttp = &thisRtt;
5520 #if defined(AFS_ALPHA_LINUX22_ENV) && defined(AFS_PTHREAD_ENV) && !defined(KERNEL)
5521 /* making year 2038 bugs to get this running now - stroucki */
5522 struct timeval temptime;
5524 register int rtt_timeout;
5526 #if defined(AFS_ALPHA_LINUX20_ENV) && defined(AFS_PTHREAD_ENV) && !defined(KERNEL)
5527 /* yet again. This was the worst Heisenbug of the port - stroucki */
5528 clock_GetTime(&temptime);
5529 rttp->sec = (afs_int32) temptime.tv_sec;
5530 rttp->usec = (afs_int32) temptime.tv_usec;
5532 clock_GetTime(rttp);
5534 if (clock_Lt(rttp, sentp)) {
5536 return; /* somebody set the clock back, don't count this time. */
5538 clock_Sub(rttp, sentp);
5539 MUTEX_ENTER(&rx_stats_mutex);
5540 if (clock_Lt(rttp, &rx_stats.minRtt))
5541 rx_stats.minRtt = *rttp;
5542 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5543 if (rttp->sec > 60) {
5544 MUTEX_EXIT(&rx_stats_mutex);
5545 return; /* somebody set the clock ahead */
5547 rx_stats.maxRtt = *rttp;
5549 clock_Add(&rx_stats.totalRtt, rttp);
5550 rx_stats.nRttSamples++;
5551 MUTEX_EXIT(&rx_stats_mutex);
5553 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5555 /* Apply VanJacobson round-trip estimations */
5560 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5561 * srtt is stored as fixed point with 3 bits after the binary
5562 * point (i.e., scaled by 8). The following magic is
5563 * equivalent to the smoothing algorithm in rfc793 with an
5564 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5565 * srtt*8 = srtt*8 + rtt - srtt
5566 * srtt = srtt + rtt/8 - srtt/8
5569 delta = MSEC(rttp) - (peer->rtt >> 3);
5573 * We accumulate a smoothed rtt variance (actually, a smoothed
5574 * mean difference), then set the retransmit timer to smoothed
5575 * rtt + 4 times the smoothed variance (was 2x in van's original
5576 * paper, but 4x works better for me, and apparently for him as
5578 * rttvar is stored as
5579 * fixed point with 2 bits after the binary point (scaled by
5580 * 4). The following is equivalent to rfc793 smoothing with
5581 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5582 * replaces rfc793's wired-in beta.
5583 * dev*4 = dev*4 + (|actual - expected| - dev)
5589 delta -= (peer->rtt_dev >> 2);
5590 peer->rtt_dev += delta;
5592 /* I don't have a stored RTT so I start with this value. Since I'm
5593 * probably just starting a call, and will be pushing more data down
5594 * this, I expect congestion to increase rapidly. So I fudge a
5595 * little, and I set deviance to half the rtt. In practice,
5596 * deviance tends to approach something a little less than
5597 * half the smoothed rtt. */
5598 peer->rtt = (MSEC(rttp) << 3) + 8;
5599 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5601 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5602 * the other of these connections is usually in a user process, and can
5603 * be switched and/or swapped out. So on fast, reliable networks, the
5604 * timeout would otherwise be too short.
5606 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5607 clock_Zero(&(peer->timeout));
5608 clock_Addmsec(&(peer->timeout), rtt_timeout);
5610 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)));
5614 /* Find all server connections that have not been active for a long time, and
5617 rxi_ReapConnections(void)
5620 clock_GetTime(&now);
5622 /* Find server connection structures that haven't been used for
5623 * greater than rx_idleConnectionTime */
5625 struct rx_connection **conn_ptr, **conn_end;
5626 int i, havecalls = 0;
5627 MUTEX_ENTER(&rx_connHashTable_lock);
5628 for (conn_ptr = &rx_connHashTable[0], conn_end =
5629 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5631 struct rx_connection *conn, *next;
5632 struct rx_call *call;
5636 for (conn = *conn_ptr; conn; conn = next) {
5637 /* XXX -- Shouldn't the connection be locked? */
5640 for (i = 0; i < RX_MAXCALLS; i++) {
5641 call = conn->call[i];
5644 MUTEX_ENTER(&call->lock);
5645 #ifdef RX_ENABLE_LOCKS
5646 result = rxi_CheckCall(call, 1);
5647 #else /* RX_ENABLE_LOCKS */
5648 result = rxi_CheckCall(call);
5649 #endif /* RX_ENABLE_LOCKS */
5650 MUTEX_EXIT(&call->lock);
5652 /* If CheckCall freed the call, it might
5653 * have destroyed the connection as well,
5654 * which screws up the linked lists.
5660 if (conn->type == RX_SERVER_CONNECTION) {
5661 /* This only actually destroys the connection if
5662 * there are no outstanding calls */
5663 MUTEX_ENTER(&conn->conn_data_lock);
5664 if (!havecalls && !conn->refCount
5665 && ((conn->lastSendTime + rx_idleConnectionTime) <
5667 conn->refCount++; /* it will be decr in rx_DestroyConn */
5668 MUTEX_EXIT(&conn->conn_data_lock);
5669 #ifdef RX_ENABLE_LOCKS
5670 rxi_DestroyConnectionNoLock(conn);
5671 #else /* RX_ENABLE_LOCKS */
5672 rxi_DestroyConnection(conn);
5673 #endif /* RX_ENABLE_LOCKS */
5675 #ifdef RX_ENABLE_LOCKS
5677 MUTEX_EXIT(&conn->conn_data_lock);
5679 #endif /* RX_ENABLE_LOCKS */
5683 #ifdef RX_ENABLE_LOCKS
5684 while (rx_connCleanup_list) {
5685 struct rx_connection *conn;
5686 conn = rx_connCleanup_list;
5687 rx_connCleanup_list = rx_connCleanup_list->next;
5688 MUTEX_EXIT(&rx_connHashTable_lock);
5689 rxi_CleanupConnection(conn);
5690 MUTEX_ENTER(&rx_connHashTable_lock);
5692 MUTEX_EXIT(&rx_connHashTable_lock);
5693 #endif /* RX_ENABLE_LOCKS */
5696 /* Find any peer structures that haven't been used (haven't had an
5697 * associated connection) for greater than rx_idlePeerTime */
5699 struct rx_peer **peer_ptr, **peer_end;
5701 MUTEX_ENTER(&rx_rpc_stats);
5702 MUTEX_ENTER(&rx_peerHashTable_lock);
5703 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5704 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5706 struct rx_peer *peer, *next, *prev;
5707 for (prev = peer = *peer_ptr; peer; peer = next) {
5709 code = MUTEX_TRYENTER(&peer->peer_lock);
5710 if ((code) && (peer->refCount == 0)
5711 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5712 rx_interface_stat_p rpc_stat, nrpc_stat;
5714 MUTEX_EXIT(&peer->peer_lock);
5715 MUTEX_DESTROY(&peer->peer_lock);
5717 (&peer->rpcStats, rpc_stat, nrpc_stat,
5718 rx_interface_stat)) {
5719 unsigned int num_funcs;
5722 queue_Remove(&rpc_stat->queue_header);
5723 queue_Remove(&rpc_stat->all_peers);
5724 num_funcs = rpc_stat->stats[0].func_total;
5726 sizeof(rx_interface_stat_t) +
5727 rpc_stat->stats[0].func_total *
5728 sizeof(rx_function_entry_v1_t);
5730 rxi_Free(rpc_stat, space);
5731 rxi_rpc_peer_stat_cnt -= num_funcs;
5734 MUTEX_ENTER(&rx_stats_mutex);
5735 rx_stats.nPeerStructs--;
5736 MUTEX_EXIT(&rx_stats_mutex);
5737 if (prev == *peer_ptr) {
5744 MUTEX_EXIT(&peer->peer_lock);
5750 MUTEX_EXIT(&rx_peerHashTable_lock);
5751 MUTEX_EXIT(&rx_rpc_stats);
5754 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5755 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5756 * GC, just below. Really, we shouldn't have to keep moving packets from
5757 * one place to another, but instead ought to always know if we can
5758 * afford to hold onto a packet in its particular use. */
5759 MUTEX_ENTER(&rx_freePktQ_lock);
5760 if (rx_waitingForPackets) {
5761 rx_waitingForPackets = 0;
5762 #ifdef RX_ENABLE_LOCKS
5763 CV_BROADCAST(&rx_waitingForPackets_cv);
5765 osi_rxWakeup(&rx_waitingForPackets);
5768 MUTEX_EXIT(&rx_freePktQ_lock);
5770 now.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5771 rxevent_Post(&now, rxi_ReapConnections, 0, 0);
5775 /* rxs_Release - This isn't strictly necessary but, since the macro name from
5776 * rx.h is sort of strange this is better. This is called with a security
5777 * object before it is discarded. Each connection using a security object has
5778 * its own refcount to the object so it won't actually be freed until the last
5779 * connection is destroyed.
5781 * This is the only rxs module call. A hold could also be written but no one
5785 rxs_Release(struct rx_securityClass *aobj)
5787 return RXS_Close(aobj);
5791 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
5792 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
5793 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
5794 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
5796 /* Adjust our estimate of the transmission rate to this peer, given
5797 * that the packet p was just acked. We can adjust peer->timeout and
5798 * call->twind. Pragmatically, this is called
5799 * only with packets of maximal length.
5800 * Called with peer and call locked.
5804 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
5805 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
5807 afs_int32 xferSize, xferMs;
5808 register afs_int32 minTime;
5811 /* Count down packets */
5812 if (peer->rateFlag > 0)
5814 /* Do nothing until we're enabled */
5815 if (peer->rateFlag != 0)
5820 /* Count only when the ack seems legitimate */
5821 switch (ackReason) {
5822 case RX_ACK_REQUESTED:
5824 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
5828 case RX_ACK_PING_RESPONSE:
5829 if (p) /* want the response to ping-request, not data send */
5831 clock_GetTime(&newTO);
5832 if (clock_Gt(&newTO, &call->pingRequestTime)) {
5833 clock_Sub(&newTO, &call->pingRequestTime);
5834 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
5838 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
5845 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));
5847 /* Track only packets that are big enough. */
5848 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
5852 /* absorb RTT data (in milliseconds) for these big packets */
5853 if (peer->smRtt == 0) {
5854 peer->smRtt = xferMs;
5856 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
5861 if (peer->countDown) {
5865 peer->countDown = 10; /* recalculate only every so often */
5867 /* In practice, we can measure only the RTT for full packets,
5868 * because of the way Rx acks the data that it receives. (If it's
5869 * smaller than a full packet, it often gets implicitly acked
5870 * either by the call response (from a server) or by the next call
5871 * (from a client), and either case confuses transmission times
5872 * with processing times.) Therefore, replace the above
5873 * more-sophisticated processing with a simpler version, where the
5874 * smoothed RTT is kept for full-size packets, and the time to
5875 * transmit a windowful of full-size packets is simply RTT *
5876 * windowSize. Again, we take two steps:
5877 - ensure the timeout is large enough for a single packet's RTT;
5878 - ensure that the window is small enough to fit in the desired timeout.*/
5880 /* First, the timeout check. */
5881 minTime = peer->smRtt;
5882 /* Get a reasonable estimate for a timeout period */
5884 newTO.sec = minTime / 1000;
5885 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
5887 /* Increase the timeout period so that we can always do at least
5888 * one packet exchange */
5889 if (clock_Gt(&newTO, &peer->timeout)) {
5891 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));
5893 peer->timeout = newTO;
5896 /* Now, get an estimate for the transmit window size. */
5897 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
5898 /* Now, convert to the number of full packets that could fit in a
5899 * reasonable fraction of that interval */
5900 minTime /= (peer->smRtt << 1);
5901 xferSize = minTime; /* (make a copy) */
5903 /* Now clamp the size to reasonable bounds. */
5906 else if (minTime > rx_Window)
5907 minTime = rx_Window;
5908 /* if (minTime != peer->maxWindow) {
5909 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
5910 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
5911 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
5913 peer->maxWindow = minTime;
5914 elide... call->twind = minTime;
5918 /* Cut back on the peer timeout if it had earlier grown unreasonably.
5919 * Discern this by calculating the timeout necessary for rx_Window
5921 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
5922 /* calculate estimate for transmission interval in milliseconds */
5923 minTime = rx_Window * peer->smRtt;
5924 if (minTime < 1000) {
5925 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
5926 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
5927 peer->timeout.usec, peer->smRtt, peer->packetSize));
5929 newTO.sec = 0; /* cut back on timeout by half a second */
5930 newTO.usec = 500000;
5931 clock_Sub(&peer->timeout, &newTO);
5936 } /* end of rxi_ComputeRate */
5937 #endif /* ADAPT_WINDOW */
5945 /* Don't call this debugging routine directly; use dpf */
5947 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
5948 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
5952 clock_GetTime(&now);
5953 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
5954 (unsigned int)now.usec / 1000);
5955 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
5963 * This function is used to process the rx_stats structure that is local
5964 * to a process as well as an rx_stats structure received from a remote
5965 * process (via rxdebug). Therefore, it needs to do minimal version
5969 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
5970 afs_int32 freePackets, char version)
5974 if (size != sizeof(struct rx_stats)) {
5976 "Unexpected size of stats structure: was %d, expected %d\n",
5977 size, sizeof(struct rx_stats));
5980 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
5983 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
5984 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
5985 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
5986 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
5987 s->specialPktAllocFailures);
5989 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
5990 s->receivePktAllocFailures, s->sendPktAllocFailures,
5991 s->specialPktAllocFailures);
5995 " greedy %d, " "bogusReads %d (last from host %x), "
5996 "noPackets %d, " "noBuffers %d, " "selects %d, "
5997 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
5998 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
5999 s->selects, s->sendSelects);
6001 fprintf(file, " packets read: ");
6002 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6003 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6005 fprintf(file, "\n");
6008 " other read counters: data %d, " "ack %d, " "dup %d "
6009 "spurious %d " "dally %d\n", s->dataPacketsRead,
6010 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6011 s->ignorePacketDally);
6013 fprintf(file, " packets sent: ");
6014 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6015 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6017 fprintf(file, "\n");
6020 " other send counters: ack %d, " "data %d (not resends), "
6021 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6022 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6023 s->dataPacketsPushed, s->ignoreAckedPacket);
6026 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6027 s->netSendFailures, (int)s->fatalErrors);
6029 if (s->nRttSamples) {
6030 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6031 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6033 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6034 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6038 " %d server connections, " "%d client connections, "
6039 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6040 s->nServerConns, s->nClientConns, s->nPeerStructs,
6041 s->nCallStructs, s->nFreeCallStructs);
6043 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6044 fprintf(file, " %d clock updates\n", clock_nUpdates);
6049 /* for backward compatibility */
6051 rx_PrintStats(FILE * file)
6053 MUTEX_ENTER(&rx_stats_mutex);
6054 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6056 MUTEX_EXIT(&rx_stats_mutex);
6060 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6062 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6063 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6064 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6067 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6068 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6069 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6072 " Packet size %d, " "max in packet skew %d, "
6073 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6074 (int)peer->outPacketSkew);
6077 #ifdef AFS_PTHREAD_ENV
6079 * This mutex protects the following static variables:
6083 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0);
6084 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0);
6086 #define LOCK_RX_DEBUG
6087 #define UNLOCK_RX_DEBUG
6088 #endif /* AFS_PTHREAD_ENV */
6091 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6092 u_char type, void *inputData, size_t inputLength,
6093 void *outputData, size_t outputLength)
6095 static afs_int32 counter = 100;
6097 struct rx_header theader;
6099 register afs_int32 code;
6101 struct sockaddr_in taddr, faddr;
6106 endTime = time(0) + 20; /* try for 20 seconds */
6107 LOCK_RX_DEBUG counter++;
6108 UNLOCK_RX_DEBUG tp = &tbuffer[sizeof(struct rx_header)];
6109 taddr.sin_family = AF_INET;
6110 taddr.sin_port = remotePort;
6111 taddr.sin_addr.s_addr = remoteAddr;
6112 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6113 taddr.sin_len = sizeof(struct sockaddr_in);
6116 memset(&theader, 0, sizeof(theader));
6117 theader.epoch = htonl(999);
6119 theader.callNumber = htonl(counter);
6122 theader.type = type;
6123 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6124 theader.serviceId = 0;
6126 memcpy(tbuffer, &theader, sizeof(theader));
6127 memcpy(tp, inputData, inputLength);
6129 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6130 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6132 /* see if there's a packet available */
6134 FD_SET(socket, &imask);
6137 code = select(socket + 1, &imask, 0, 0, &tv);
6138 if (code == 1 && FD_ISSET(socket,&imask)) {
6139 /* now receive a packet */
6140 faddrLen = sizeof(struct sockaddr_in);
6142 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6143 (struct sockaddr *)&faddr, &faddrLen);
6146 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6147 if (counter == ntohl(theader.callNumber))
6152 /* see if we've timed out */
6153 if (endTime < time(0))
6156 code -= sizeof(struct rx_header);
6157 if (code > outputLength)
6158 code = outputLength;
6159 memcpy(outputData, tp, code);
6164 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6165 afs_uint16 remotePort, struct rx_debugStats * stat,
6166 afs_uint32 * supportedValues)
6168 struct rx_debugIn in;
6171 *supportedValues = 0;
6172 in.type = htonl(RX_DEBUGI_GETSTATS);
6175 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6176 &in, sizeof(in), stat, sizeof(*stat));
6179 * If the call was successful, fixup the version and indicate
6180 * what contents of the stat structure are valid.
6181 * Also do net to host conversion of fields here.
6185 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6186 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6188 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6189 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6191 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6192 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6194 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6195 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6197 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6198 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6200 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6201 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6203 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6204 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6207 stat->nFreePackets = ntohl(stat->nFreePackets);
6208 stat->packetReclaims = ntohl(stat->packetReclaims);
6209 stat->callsExecuted = ntohl(stat->callsExecuted);
6210 stat->nWaiting = ntohl(stat->nWaiting);
6211 stat->idleThreads = ntohl(stat->idleThreads);
6218 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6219 afs_uint16 remotePort, struct rx_stats * stat,
6220 afs_uint32 * supportedValues)
6222 struct rx_debugIn in;
6223 afs_int32 *lp = (afs_int32 *) stat;
6228 * supportedValues is currently unused, but added to allow future
6229 * versioning of this function.
6232 *supportedValues = 0;
6233 in.type = htonl(RX_DEBUGI_RXSTATS);
6235 memset(stat, 0, sizeof(*stat));
6237 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6238 &in, sizeof(in), stat, sizeof(*stat));
6243 * Do net to host conversion here
6246 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6255 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6256 afs_uint16 remotePort, size_t version_length,
6260 return MakeDebugCall(socket, remoteAddr, remotePort,
6261 RX_PACKET_TYPE_VERSION, a, 1, version,
6266 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6267 afs_uint16 remotePort, afs_int32 * nextConnection,
6268 int allConnections, afs_uint32 debugSupportedValues,
6269 struct rx_debugConn * conn,
6270 afs_uint32 * supportedValues)
6272 struct rx_debugIn in;
6277 * supportedValues is currently unused, but added to allow future
6278 * versioning of this function.
6281 *supportedValues = 0;
6282 if (allConnections) {
6283 in.type = htonl(RX_DEBUGI_GETALLCONN);
6285 in.type = htonl(RX_DEBUGI_GETCONN);
6287 in.index = htonl(*nextConnection);
6288 memset(conn, 0, sizeof(*conn));
6290 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6291 &in, sizeof(in), conn, sizeof(*conn));
6294 *nextConnection += 1;
6297 * Convert old connection format to new structure.
6300 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6301 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6302 #define MOVEvL(a) (conn->a = vL->a)
6304 /* any old or unrecognized version... */
6305 for (i = 0; i < RX_MAXCALLS; i++) {
6306 MOVEvL(callState[i]);
6307 MOVEvL(callMode[i]);
6308 MOVEvL(callFlags[i]);
6309 MOVEvL(callOther[i]);
6311 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6312 MOVEvL(secStats.type);
6313 MOVEvL(secStats.level);
6314 MOVEvL(secStats.flags);
6315 MOVEvL(secStats.expires);
6316 MOVEvL(secStats.packetsReceived);
6317 MOVEvL(secStats.packetsSent);
6318 MOVEvL(secStats.bytesReceived);
6319 MOVEvL(secStats.bytesSent);
6324 * Do net to host conversion here
6326 * I don't convert host or port since we are most likely
6327 * going to want these in NBO.
6329 conn->cid = ntohl(conn->cid);
6330 conn->serial = ntohl(conn->serial);
6331 for (i = 0; i < RX_MAXCALLS; i++) {
6332 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6334 conn->error = ntohl(conn->error);
6335 conn->secStats.flags = ntohl(conn->secStats.flags);
6336 conn->secStats.expires = ntohl(conn->secStats.expires);
6337 conn->secStats.packetsReceived =
6338 ntohl(conn->secStats.packetsReceived);
6339 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6340 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6341 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6342 conn->epoch = ntohl(conn->epoch);
6343 conn->natMTU = ntohl(conn->natMTU);
6350 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6351 afs_uint16 remotePort, afs_int32 * nextPeer,
6352 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6353 afs_uint32 * supportedValues)
6355 struct rx_debugIn in;
6359 * supportedValues is currently unused, but added to allow future
6360 * versioning of this function.
6363 *supportedValues = 0;
6364 in.type = htonl(RX_DEBUGI_GETPEER);
6365 in.index = htonl(*nextPeer);
6366 memset(peer, 0, sizeof(*peer));
6368 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6369 &in, sizeof(in), peer, sizeof(*peer));
6375 * Do net to host conversion here
6377 * I don't convert host or port since we are most likely
6378 * going to want these in NBO.
6380 peer->ifMTU = ntohs(peer->ifMTU);
6381 peer->idleWhen = ntohl(peer->idleWhen);
6382 peer->refCount = ntohs(peer->refCount);
6383 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6384 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6385 peer->rtt = ntohl(peer->rtt);
6386 peer->rtt_dev = ntohl(peer->rtt_dev);
6387 peer->timeout.sec = ntohl(peer->timeout.sec);
6388 peer->timeout.usec = ntohl(peer->timeout.usec);
6389 peer->nSent = ntohl(peer->nSent);
6390 peer->reSends = ntohl(peer->reSends);
6391 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6392 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6393 peer->rateFlag = ntohl(peer->rateFlag);
6394 peer->natMTU = ntohs(peer->natMTU);
6395 peer->maxMTU = ntohs(peer->maxMTU);
6396 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6397 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6398 peer->MTU = ntohs(peer->MTU);
6399 peer->cwind = ntohs(peer->cwind);
6400 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6401 peer->congestSeq = ntohs(peer->congestSeq);
6402 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6403 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6404 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6405 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6410 #endif /* RXDEBUG */
6415 struct rx_serverQueueEntry *np;
6418 register struct rx_call *call;
6419 register struct rx_serverQueueEntry *sq;
6422 LOCK_RX_INIT if (rxinit_status == 1) {
6423 UNLOCK_RX_INIT return; /* Already shutdown. */
6427 #ifndef AFS_PTHREAD_ENV
6428 FD_ZERO(&rx_selectMask);
6429 #endif /* AFS_PTHREAD_ENV */
6430 rxi_dataQuota = RX_MAX_QUOTA;
6431 #ifndef AFS_PTHREAD_ENV
6433 #endif /* AFS_PTHREAD_ENV */
6436 #ifndef AFS_PTHREAD_ENV
6437 #ifndef AFS_USE_GETTIMEOFDAY
6439 #endif /* AFS_USE_GETTIMEOFDAY */
6440 #endif /* AFS_PTHREAD_ENV */
6442 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6443 call = queue_First(&rx_freeCallQueue, rx_call);
6445 rxi_Free(call, sizeof(struct rx_call));
6448 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6449 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6455 struct rx_peer **peer_ptr, **peer_end;
6456 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6457 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6459 struct rx_peer *peer, *next;
6460 for (peer = *peer_ptr; peer; peer = next) {
6461 rx_interface_stat_p rpc_stat, nrpc_stat;
6464 (&peer->rpcStats, rpc_stat, nrpc_stat,
6465 rx_interface_stat)) {
6466 unsigned int num_funcs;
6469 queue_Remove(&rpc_stat->queue_header);
6470 queue_Remove(&rpc_stat->all_peers);
6471 num_funcs = rpc_stat->stats[0].func_total;
6473 sizeof(rx_interface_stat_t) +
6474 rpc_stat->stats[0].func_total *
6475 sizeof(rx_function_entry_v1_t);
6477 rxi_Free(rpc_stat, space);
6478 MUTEX_ENTER(&rx_rpc_stats);
6479 rxi_rpc_peer_stat_cnt -= num_funcs;
6480 MUTEX_EXIT(&rx_rpc_stats);
6484 MUTEX_ENTER(&rx_stats_mutex);
6485 rx_stats.nPeerStructs--;
6486 MUTEX_EXIT(&rx_stats_mutex);
6490 for (i = 0; i < RX_MAX_SERVICES; i++) {
6492 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6494 for (i = 0; i < rx_hashTableSize; i++) {
6495 register struct rx_connection *tc, *ntc;
6496 MUTEX_ENTER(&rx_connHashTable_lock);
6497 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6499 for (j = 0; j < RX_MAXCALLS; j++) {
6501 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6504 rxi_Free(tc, sizeof(*tc));
6506 MUTEX_EXIT(&rx_connHashTable_lock);
6509 MUTEX_ENTER(&freeSQEList_lock);
6511 while ((np = rx_FreeSQEList)) {
6512 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6513 MUTEX_DESTROY(&np->lock);
6514 rxi_Free(np, sizeof(*np));
6517 MUTEX_EXIT(&freeSQEList_lock);
6518 MUTEX_DESTROY(&freeSQEList_lock);
6519 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6520 MUTEX_DESTROY(&rx_connHashTable_lock);
6521 MUTEX_DESTROY(&rx_peerHashTable_lock);
6522 MUTEX_DESTROY(&rx_serverPool_lock);
6524 osi_Free(rx_connHashTable,
6525 rx_hashTableSize * sizeof(struct rx_connection *));
6526 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6528 UNPIN(rx_connHashTable,
6529 rx_hashTableSize * sizeof(struct rx_connection *));
6530 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6532 rxi_FreeAllPackets();
6534 MUTEX_ENTER(&rx_stats_mutex);
6535 rxi_dataQuota = RX_MAX_QUOTA;
6536 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6537 MUTEX_EXIT(&rx_stats_mutex);
6542 #ifdef RX_ENABLE_LOCKS
6544 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6546 if (!MUTEX_ISMINE(lockaddr))
6547 osi_Panic("Lock not held: %s", msg);
6549 #endif /* RX_ENABLE_LOCKS */
6554 * Routines to implement connection specific data.
6558 rx_KeyCreate(rx_destructor_t rtn)
6561 MUTEX_ENTER(&rxi_keyCreate_lock);
6562 key = rxi_keyCreate_counter++;
6563 rxi_keyCreate_destructor = (rx_destructor_t *)
6564 realloc((void *)rxi_keyCreate_destructor,
6565 (key + 1) * sizeof(rx_destructor_t));
6566 rxi_keyCreate_destructor[key] = rtn;
6567 MUTEX_EXIT(&rxi_keyCreate_lock);
6572 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6575 MUTEX_ENTER(&conn->conn_data_lock);
6576 if (!conn->specific) {
6577 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6578 for (i = 0; i < key; i++)
6579 conn->specific[i] = NULL;
6580 conn->nSpecific = key + 1;
6581 conn->specific[key] = ptr;
6582 } else if (key >= conn->nSpecific) {
6583 conn->specific = (void **)
6584 realloc(conn->specific, (key + 1) * sizeof(void *));
6585 for (i = conn->nSpecific; i < key; i++)
6586 conn->specific[i] = NULL;
6587 conn->nSpecific = key + 1;
6588 conn->specific[key] = ptr;
6590 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6591 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6592 conn->specific[key] = ptr;
6594 MUTEX_EXIT(&conn->conn_data_lock);
6598 rx_GetSpecific(struct rx_connection *conn, int key)
6601 MUTEX_ENTER(&conn->conn_data_lock);
6602 if (key >= conn->nSpecific)
6605 ptr = conn->specific[key];
6606 MUTEX_EXIT(&conn->conn_data_lock);
6610 #endif /* !KERNEL */
6613 * processStats is a queue used to store the statistics for the local
6614 * process. Its contents are similar to the contents of the rpcStats
6615 * queue on a rx_peer structure, but the actual data stored within
6616 * this queue contains totals across the lifetime of the process (assuming
6617 * the stats have not been reset) - unlike the per peer structures
6618 * which can come and go based upon the peer lifetime.
6621 static struct rx_queue processStats = { &processStats, &processStats };
6624 * peerStats is a queue used to store the statistics for all peer structs.
6625 * Its contents are the union of all the peer rpcStats queues.
6628 static struct rx_queue peerStats = { &peerStats, &peerStats };
6631 * rxi_monitor_processStats is used to turn process wide stat collection
6635 static int rxi_monitor_processStats = 0;
6638 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6641 static int rxi_monitor_peerStats = 0;
6644 * rxi_AddRpcStat - given all of the information for a particular rpc
6645 * call, create (if needed) and update the stat totals for the rpc.
6649 * IN stats - the queue of stats that will be updated with the new value
6651 * IN rxInterface - a unique number that identifies the rpc interface
6653 * IN currentFunc - the index of the function being invoked
6655 * IN totalFunc - the total number of functions in this interface
6657 * IN queueTime - the amount of time this function waited for a thread
6659 * IN execTime - the amount of time this function invocation took to execute
6661 * IN bytesSent - the number bytes sent by this invocation
6663 * IN bytesRcvd - the number bytes received by this invocation
6665 * IN isServer - if true, this invocation was made to a server
6667 * IN remoteHost - the ip address of the remote host
6669 * IN remotePort - the port of the remote host
6671 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6673 * INOUT counter - if a new stats structure is allocated, the counter will
6674 * be updated with the new number of allocated stat structures
6682 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6683 afs_uint32 currentFunc, afs_uint32 totalFunc,
6684 struct clock *queueTime, struct clock *execTime,
6685 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6686 afs_uint32 remoteHost, afs_uint32 remotePort,
6687 int addToPeerList, unsigned int *counter)
6690 rx_interface_stat_p rpc_stat, nrpc_stat;
6693 * See if there's already a structure for this interface
6696 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6697 if ((rpc_stat->stats[0].interfaceId == rxInterface)
6698 && (rpc_stat->stats[0].remote_is_server == isServer))
6703 * Didn't find a match so allocate a new structure and add it to the
6707 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
6708 || (rpc_stat->stats[0].interfaceId != rxInterface)
6709 || (rpc_stat->stats[0].remote_is_server != isServer)) {
6714 sizeof(rx_interface_stat_t) +
6715 totalFunc * sizeof(rx_function_entry_v1_t);
6717 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
6718 if (rpc_stat == NULL) {
6722 *counter += totalFunc;
6723 for (i = 0; i < totalFunc; i++) {
6724 rpc_stat->stats[i].remote_peer = remoteHost;
6725 rpc_stat->stats[i].remote_port = remotePort;
6726 rpc_stat->stats[i].remote_is_server = isServer;
6727 rpc_stat->stats[i].interfaceId = rxInterface;
6728 rpc_stat->stats[i].func_total = totalFunc;
6729 rpc_stat->stats[i].func_index = i;
6730 hzero(rpc_stat->stats[i].invocations);
6731 hzero(rpc_stat->stats[i].bytes_sent);
6732 hzero(rpc_stat->stats[i].bytes_rcvd);
6733 rpc_stat->stats[i].queue_time_sum.sec = 0;
6734 rpc_stat->stats[i].queue_time_sum.usec = 0;
6735 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
6736 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
6737 rpc_stat->stats[i].queue_time_min.sec = 9999999;
6738 rpc_stat->stats[i].queue_time_min.usec = 9999999;
6739 rpc_stat->stats[i].queue_time_max.sec = 0;
6740 rpc_stat->stats[i].queue_time_max.usec = 0;
6741 rpc_stat->stats[i].execution_time_sum.sec = 0;
6742 rpc_stat->stats[i].execution_time_sum.usec = 0;
6743 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
6744 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
6745 rpc_stat->stats[i].execution_time_min.sec = 9999999;
6746 rpc_stat->stats[i].execution_time_min.usec = 9999999;
6747 rpc_stat->stats[i].execution_time_max.sec = 0;
6748 rpc_stat->stats[i].execution_time_max.usec = 0;
6750 queue_Prepend(stats, rpc_stat);
6751 if (addToPeerList) {
6752 queue_Prepend(&peerStats, &rpc_stat->all_peers);
6757 * Increment the stats for this function
6760 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
6761 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
6762 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
6763 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
6764 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
6765 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
6766 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
6768 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
6769 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
6771 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
6772 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
6774 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
6775 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
6777 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
6778 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
6786 * rx_IncrementTimeAndCount - increment the times and count for a particular
6791 * IN peer - the peer who invoked the rpc
6793 * IN rxInterface - a unique number that identifies the rpc interface
6795 * IN currentFunc - the index of the function being invoked
6797 * IN totalFunc - the total number of functions in this interface
6799 * IN queueTime - the amount of time this function waited for a thread
6801 * IN execTime - the amount of time this function invocation took to execute
6803 * IN bytesSent - the number bytes sent by this invocation
6805 * IN bytesRcvd - the number bytes received by this invocation
6807 * IN isServer - if true, this invocation was made to a server
6815 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
6816 afs_uint32 currentFunc, afs_uint32 totalFunc,
6817 struct clock *queueTime, struct clock *execTime,
6818 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
6822 MUTEX_ENTER(&rx_rpc_stats);
6823 MUTEX_ENTER(&peer->peer_lock);
6825 if (rxi_monitor_peerStats) {
6826 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
6827 queueTime, execTime, bytesSent, bytesRcvd, isServer,
6828 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
6831 if (rxi_monitor_processStats) {
6832 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
6833 queueTime, execTime, bytesSent, bytesRcvd, isServer,
6834 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
6837 MUTEX_EXIT(&peer->peer_lock);
6838 MUTEX_EXIT(&rx_rpc_stats);
6843 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
6847 * IN callerVersion - the rpc stat version of the caller.
6849 * IN count - the number of entries to marshall.
6851 * IN stats - pointer to stats to be marshalled.
6853 * OUT ptr - Where to store the marshalled data.
6860 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
6861 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
6867 * We only support the first version
6869 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
6870 *(ptr++) = stats->remote_peer;
6871 *(ptr++) = stats->remote_port;
6872 *(ptr++) = stats->remote_is_server;
6873 *(ptr++) = stats->interfaceId;
6874 *(ptr++) = stats->func_total;
6875 *(ptr++) = stats->func_index;
6876 *(ptr++) = hgethi(stats->invocations);
6877 *(ptr++) = hgetlo(stats->invocations);
6878 *(ptr++) = hgethi(stats->bytes_sent);
6879 *(ptr++) = hgetlo(stats->bytes_sent);
6880 *(ptr++) = hgethi(stats->bytes_rcvd);
6881 *(ptr++) = hgetlo(stats->bytes_rcvd);
6882 *(ptr++) = stats->queue_time_sum.sec;
6883 *(ptr++) = stats->queue_time_sum.usec;
6884 *(ptr++) = stats->queue_time_sum_sqr.sec;
6885 *(ptr++) = stats->queue_time_sum_sqr.usec;
6886 *(ptr++) = stats->queue_time_min.sec;
6887 *(ptr++) = stats->queue_time_min.usec;
6888 *(ptr++) = stats->queue_time_max.sec;
6889 *(ptr++) = stats->queue_time_max.usec;
6890 *(ptr++) = stats->execution_time_sum.sec;
6891 *(ptr++) = stats->execution_time_sum.usec;
6892 *(ptr++) = stats->execution_time_sum_sqr.sec;
6893 *(ptr++) = stats->execution_time_sum_sqr.usec;
6894 *(ptr++) = stats->execution_time_min.sec;
6895 *(ptr++) = stats->execution_time_min.usec;
6896 *(ptr++) = stats->execution_time_max.sec;
6897 *(ptr++) = stats->execution_time_max.usec;
6903 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
6908 * IN callerVersion - the rpc stat version of the caller
6910 * OUT myVersion - the rpc stat version of this function
6912 * OUT clock_sec - local time seconds
6914 * OUT clock_usec - local time microseconds
6916 * OUT allocSize - the number of bytes allocated to contain stats
6918 * OUT statCount - the number stats retrieved from this process.
6920 * OUT stats - the actual stats retrieved from this process.
6924 * Returns void. If successful, stats will != NULL.
6928 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
6929 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
6930 size_t * allocSize, afs_uint32 * statCount,
6931 afs_uint32 ** stats)
6941 *myVersion = RX_STATS_RETRIEVAL_VERSION;
6944 * Check to see if stats are enabled
6947 MUTEX_ENTER(&rx_rpc_stats);
6948 if (!rxi_monitor_processStats) {
6949 MUTEX_EXIT(&rx_rpc_stats);
6953 clock_GetTime(&now);
6954 *clock_sec = now.sec;
6955 *clock_usec = now.usec;
6958 * Allocate the space based upon the caller version
6960 * If the client is at an older version than we are,
6961 * we return the statistic data in the older data format, but
6962 * we still return our version number so the client knows we
6963 * are maintaining more data than it can retrieve.
6966 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
6967 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
6968 *statCount = rxi_rpc_process_stat_cnt;
6971 * This can't happen yet, but in the future version changes
6972 * can be handled by adding additional code here
6976 if (space > (size_t) 0) {
6978 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
6981 rx_interface_stat_p rpc_stat, nrpc_stat;
6985 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6987 * Copy the data based upon the caller version
6989 rx_MarshallProcessRPCStats(callerVersion,
6990 rpc_stat->stats[0].func_total,
6991 rpc_stat->stats, &ptr);
6997 MUTEX_EXIT(&rx_rpc_stats);
7002 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7006 * IN callerVersion - the rpc stat version of the caller
7008 * OUT myVersion - the rpc stat version of this function
7010 * OUT clock_sec - local time seconds
7012 * OUT clock_usec - local time microseconds
7014 * OUT allocSize - the number of bytes allocated to contain stats
7016 * OUT statCount - the number of stats retrieved from the individual
7019 * OUT stats - the actual stats retrieved from the individual peer structures.
7023 * Returns void. If successful, stats will != NULL.
7027 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7028 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7029 size_t * allocSize, afs_uint32 * statCount,
7030 afs_uint32 ** stats)
7040 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7043 * Check to see if stats are enabled
7046 MUTEX_ENTER(&rx_rpc_stats);
7047 if (!rxi_monitor_peerStats) {
7048 MUTEX_EXIT(&rx_rpc_stats);
7052 clock_GetTime(&now);
7053 *clock_sec = now.sec;
7054 *clock_usec = now.usec;
7057 * Allocate the space based upon the caller version
7059 * If the client is at an older version than we are,
7060 * we return the statistic data in the older data format, but
7061 * we still return our version number so the client knows we
7062 * are maintaining more data than it can retrieve.
7065 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7066 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7067 *statCount = rxi_rpc_peer_stat_cnt;
7070 * This can't happen yet, but in the future version changes
7071 * can be handled by adding additional code here
7075 if (space > (size_t) 0) {
7077 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7080 rx_interface_stat_p rpc_stat, nrpc_stat;
7084 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7086 * We have to fix the offset of rpc_stat since we are
7087 * keeping this structure on two rx_queues. The rx_queue
7088 * package assumes that the rx_queue member is the first
7089 * member of the structure. That is, rx_queue assumes that
7090 * any one item is only on one queue at a time. We are
7091 * breaking that assumption and so we have to do a little
7092 * math to fix our pointers.
7095 fix_offset = (char *)rpc_stat;
7096 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7097 rpc_stat = (rx_interface_stat_p) fix_offset;
7100 * Copy the data based upon the caller version
7102 rx_MarshallProcessRPCStats(callerVersion,
7103 rpc_stat->stats[0].func_total,
7104 rpc_stat->stats, &ptr);
7110 MUTEX_EXIT(&rx_rpc_stats);
7115 * rx_FreeRPCStats - free memory allocated by
7116 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7120 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7121 * rx_RetrievePeerRPCStats
7123 * IN allocSize - the number of bytes in stats.
7131 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7133 rxi_Free(stats, allocSize);
7137 * rx_queryProcessRPCStats - see if process rpc stat collection is
7138 * currently enabled.
7144 * Returns 0 if stats are not enabled != 0 otherwise
7148 rx_queryProcessRPCStats(void)
7151 MUTEX_ENTER(&rx_rpc_stats);
7152 rc = rxi_monitor_processStats;
7153 MUTEX_EXIT(&rx_rpc_stats);
7158 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7164 * Returns 0 if stats are not enabled != 0 otherwise
7168 rx_queryPeerRPCStats(void)
7171 MUTEX_ENTER(&rx_rpc_stats);
7172 rc = rxi_monitor_peerStats;
7173 MUTEX_EXIT(&rx_rpc_stats);
7178 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7188 rx_enableProcessRPCStats(void)
7190 MUTEX_ENTER(&rx_rpc_stats);
7191 rx_enable_stats = 1;
7192 rxi_monitor_processStats = 1;
7193 MUTEX_EXIT(&rx_rpc_stats);
7197 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7207 rx_enablePeerRPCStats(void)
7209 MUTEX_ENTER(&rx_rpc_stats);
7210 rx_enable_stats = 1;
7211 rxi_monitor_peerStats = 1;
7212 MUTEX_EXIT(&rx_rpc_stats);
7216 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7226 rx_disableProcessRPCStats(void)
7228 rx_interface_stat_p rpc_stat, nrpc_stat;
7231 MUTEX_ENTER(&rx_rpc_stats);
7234 * Turn off process statistics and if peer stats is also off, turn
7238 rxi_monitor_processStats = 0;
7239 if (rxi_monitor_peerStats == 0) {
7240 rx_enable_stats = 0;
7243 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7244 unsigned int num_funcs = 0;
7247 queue_Remove(rpc_stat);
7248 num_funcs = rpc_stat->stats[0].func_total;
7250 sizeof(rx_interface_stat_t) +
7251 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7253 rxi_Free(rpc_stat, space);
7254 rxi_rpc_process_stat_cnt -= num_funcs;
7256 MUTEX_EXIT(&rx_rpc_stats);
7260 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7270 rx_disablePeerRPCStats(void)
7272 struct rx_peer **peer_ptr, **peer_end;
7275 MUTEX_ENTER(&rx_rpc_stats);
7278 * Turn off peer statistics and if process stats is also off, turn
7282 rxi_monitor_peerStats = 0;
7283 if (rxi_monitor_processStats == 0) {
7284 rx_enable_stats = 0;
7287 MUTEX_ENTER(&rx_peerHashTable_lock);
7288 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7289 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7291 struct rx_peer *peer, *next, *prev;
7292 for (prev = peer = *peer_ptr; peer; peer = next) {
7294 code = MUTEX_TRYENTER(&peer->peer_lock);
7296 rx_interface_stat_p rpc_stat, nrpc_stat;
7299 (&peer->rpcStats, rpc_stat, nrpc_stat,
7300 rx_interface_stat)) {
7301 unsigned int num_funcs = 0;
7304 queue_Remove(&rpc_stat->queue_header);
7305 queue_Remove(&rpc_stat->all_peers);
7306 num_funcs = rpc_stat->stats[0].func_total;
7308 sizeof(rx_interface_stat_t) +
7309 rpc_stat->stats[0].func_total *
7310 sizeof(rx_function_entry_v1_t);
7312 rxi_Free(rpc_stat, space);
7313 rxi_rpc_peer_stat_cnt -= num_funcs;
7315 MUTEX_EXIT(&peer->peer_lock);
7316 if (prev == *peer_ptr) {
7326 MUTEX_EXIT(&rx_peerHashTable_lock);
7327 MUTEX_EXIT(&rx_rpc_stats);
7331 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7336 * IN clearFlag - flag indicating which stats to clear
7344 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7346 rx_interface_stat_p rpc_stat, nrpc_stat;
7348 MUTEX_ENTER(&rx_rpc_stats);
7350 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7351 unsigned int num_funcs = 0, i;
7352 num_funcs = rpc_stat->stats[0].func_total;
7353 for (i = 0; i < num_funcs; i++) {
7354 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7355 hzero(rpc_stat->stats[i].invocations);
7357 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7358 hzero(rpc_stat->stats[i].bytes_sent);
7360 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7361 hzero(rpc_stat->stats[i].bytes_rcvd);
7363 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7364 rpc_stat->stats[i].queue_time_sum.sec = 0;
7365 rpc_stat->stats[i].queue_time_sum.usec = 0;
7367 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7368 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7369 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7371 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7372 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7373 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7375 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7376 rpc_stat->stats[i].queue_time_max.sec = 0;
7377 rpc_stat->stats[i].queue_time_max.usec = 0;
7379 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7380 rpc_stat->stats[i].execution_time_sum.sec = 0;
7381 rpc_stat->stats[i].execution_time_sum.usec = 0;
7383 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7384 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7385 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7387 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7388 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7389 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7391 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7392 rpc_stat->stats[i].execution_time_max.sec = 0;
7393 rpc_stat->stats[i].execution_time_max.usec = 0;
7398 MUTEX_EXIT(&rx_rpc_stats);
7402 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7407 * IN clearFlag - flag indicating which stats to clear
7415 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7417 rx_interface_stat_p rpc_stat, nrpc_stat;
7419 MUTEX_ENTER(&rx_rpc_stats);
7421 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7422 unsigned int num_funcs = 0, i;
7425 * We have to fix the offset of rpc_stat since we are
7426 * keeping this structure on two rx_queues. The rx_queue
7427 * package assumes that the rx_queue member is the first
7428 * member of the structure. That is, rx_queue assumes that
7429 * any one item is only on one queue at a time. We are
7430 * breaking that assumption and so we have to do a little
7431 * math to fix our pointers.
7434 fix_offset = (char *)rpc_stat;
7435 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7436 rpc_stat = (rx_interface_stat_p) fix_offset;
7438 num_funcs = rpc_stat->stats[0].func_total;
7439 for (i = 0; i < num_funcs; i++) {
7440 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7441 hzero(rpc_stat->stats[i].invocations);
7443 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7444 hzero(rpc_stat->stats[i].bytes_sent);
7446 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7447 hzero(rpc_stat->stats[i].bytes_rcvd);
7449 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7450 rpc_stat->stats[i].queue_time_sum.sec = 0;
7451 rpc_stat->stats[i].queue_time_sum.usec = 0;
7453 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7454 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7455 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7457 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7458 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7459 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7461 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7462 rpc_stat->stats[i].queue_time_max.sec = 0;
7463 rpc_stat->stats[i].queue_time_max.usec = 0;
7465 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7466 rpc_stat->stats[i].execution_time_sum.sec = 0;
7467 rpc_stat->stats[i].execution_time_sum.usec = 0;
7469 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7470 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7471 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7473 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7474 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7475 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7477 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7478 rpc_stat->stats[i].execution_time_max.sec = 0;
7479 rpc_stat->stats[i].execution_time_max.usec = 0;
7484 MUTEX_EXIT(&rx_rpc_stats);
7488 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7489 * is authorized to enable/disable/clear RX statistics.
7491 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7494 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7496 rxi_rxstat_userok = proc;
7500 rx_RxStatUserOk(struct rx_call *call)
7502 if (!rxi_rxstat_userok)
7504 return rxi_rxstat_userok(call);