2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* RX: Extended Remote Procedure Call */
12 #include <afsconfig.h>
14 #include "afs/param.h"
16 #include <afs/param.h>
23 #include "afs/sysincludes.h"
24 #include "afsincludes.h"
30 #include <net/net_globals.h>
31 #endif /* AFS_OSF_ENV */
32 #ifdef AFS_LINUX20_ENV
35 #include "netinet/in.h"
36 #include "afs/afs_args.h"
37 #include "afs/afs_osi.h"
38 #ifdef RX_KERNEL_TRACE
39 #include "rx_kcommon.h"
41 #if (defined(AFS_AUX_ENV) || defined(AFS_AIX_ENV))
45 #undef RXDEBUG /* turn off debugging */
47 #if defined(AFS_SGI_ENV)
48 #include "sys/debug.h"
57 #endif /* AFS_OSF_ENV */
59 #include "afs/sysincludes.h"
60 #include "afsincludes.h"
63 #include "rx_kmutex.h"
64 #include "rx_kernel.h"
68 #include "rx_globals.h"
70 #define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
71 #define AFSOP_STOP_AFS 211 /* Stop AFS process */
72 #define AFSOP_STOP_BKG 212 /* Stop BKG process */
74 extern afs_int32 afs_termState;
76 #include "sys/lockl.h"
77 #include "sys/lock_def.h"
78 #endif /* AFS_AIX41_ENV */
79 # include "rxgen_consts.h"
81 # include <sys/types.h>
86 # include <afs/afsutil.h>
88 # include <sys/socket.h>
89 # include <sys/file.h>
91 # include <sys/stat.h>
92 # include <netinet/in.h>
93 # include <sys/time.h>
103 # include "rx_user.h"
104 # include "rx_clock.h"
105 # include "rx_queue.h"
106 # include "rx_globals.h"
107 # include "rx_trace.h"
108 # include <afs/rxgen_consts.h>
111 int (*registerProgram) () = 0;
112 int (*swapNameProgram) () = 0;
114 /* Local static routines */
115 static void rxi_DestroyConnectionNoLock(register struct rx_connection *conn);
116 #ifdef RX_ENABLE_LOCKS
117 static void rxi_SetAcksInTransmitQueue(register struct rx_call *call);
120 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
122 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
123 afs_int32 rxi_start_in_error;
125 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
128 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
129 * currently allocated within rx. This number is used to allocate the
130 * memory required to return the statistics when queried.
133 static unsigned int rxi_rpc_peer_stat_cnt;
136 * rxi_rpc_process_stat_cnt counts the total number of local process stat
137 * structures currently allocated within rx. The number is used to allocate
138 * the memory required to return the statistics when queried.
141 static unsigned int rxi_rpc_process_stat_cnt;
143 #if !defined(offsetof)
144 #include <stddef.h> /* for definition of offsetof() */
147 #ifdef AFS_PTHREAD_ENV
151 * Use procedural initialization of mutexes/condition variables
155 extern pthread_mutex_t rx_stats_mutex;
156 extern pthread_mutex_t rxkad_stats_mutex;
157 extern pthread_mutex_t des_init_mutex;
158 extern pthread_mutex_t des_random_mutex;
159 extern pthread_mutex_t rx_clock_mutex;
160 extern pthread_mutex_t rxi_connCacheMutex;
161 extern pthread_mutex_t rx_event_mutex;
162 extern pthread_mutex_t osi_malloc_mutex;
163 extern pthread_mutex_t event_handler_mutex;
164 extern pthread_mutex_t listener_mutex;
165 extern pthread_mutex_t rx_if_init_mutex;
166 extern pthread_mutex_t rx_if_mutex;
167 extern pthread_mutex_t rxkad_client_uid_mutex;
168 extern pthread_mutex_t rxkad_random_mutex;
170 extern pthread_cond_t rx_event_handler_cond;
171 extern pthread_cond_t rx_listener_cond;
173 static pthread_mutex_t epoch_mutex;
174 static pthread_mutex_t rx_init_mutex;
175 static pthread_mutex_t rx_debug_mutex;
178 rxi_InitPthread(void)
180 assert(pthread_mutex_init(&rx_clock_mutex, (const pthread_mutexattr_t *)0)
182 assert(pthread_mutex_init(&rx_stats_mutex, (const pthread_mutexattr_t *)0)
184 assert(pthread_mutex_init
185 (&rxi_connCacheMutex, (const pthread_mutexattr_t *)0) == 0);
186 assert(pthread_mutex_init(&rx_init_mutex, (const pthread_mutexattr_t *)0)
188 assert(pthread_mutex_init(&epoch_mutex, (const pthread_mutexattr_t *)0) ==
190 assert(pthread_mutex_init(&rx_event_mutex, (const pthread_mutexattr_t *)0)
192 assert(pthread_mutex_init(&des_init_mutex, (const pthread_mutexattr_t *)0)
194 assert(pthread_mutex_init
195 (&des_random_mutex, (const pthread_mutexattr_t *)0) == 0);
196 assert(pthread_mutex_init
197 (&osi_malloc_mutex, (const pthread_mutexattr_t *)0) == 0);
198 assert(pthread_mutex_init
199 (&event_handler_mutex, (const pthread_mutexattr_t *)0) == 0);
200 assert(pthread_mutex_init(&listener_mutex, (const pthread_mutexattr_t *)0)
202 assert(pthread_mutex_init
203 (&rx_if_init_mutex, (const pthread_mutexattr_t *)0) == 0);
204 assert(pthread_mutex_init(&rx_if_mutex, (const pthread_mutexattr_t *)0) ==
206 assert(pthread_mutex_init
207 (&rxkad_client_uid_mutex, (const pthread_mutexattr_t *)0) == 0);
208 assert(pthread_mutex_init
209 (&rxkad_random_mutex, (const pthread_mutexattr_t *)0) == 0);
210 assert(pthread_mutex_init
211 (&rxkad_stats_mutex, (const pthread_mutexattr_t *)0) == 0);
212 assert(pthread_mutex_init(&rx_debug_mutex, (const pthread_mutexattr_t *)0)
215 assert(pthread_cond_init
216 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
217 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
219 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
220 assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
223 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
224 #define INIT_PTHREAD_LOCKS \
225 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
227 * The rx_stats_mutex mutex protects the following global variables:
232 * rxi_lowConnRefCount
233 * rxi_lowPeerRefCount
242 #define INIT_PTHREAD_LOCKS
246 /* Variables for handling the minProcs implementation. availProcs gives the
247 * number of threads available in the pool at this moment (not counting dudes
248 * executing right now). totalMin gives the total number of procs required
249 * for handling all minProcs requests. minDeficit is a dynamic variable
250 * tracking the # of procs required to satisfy all of the remaining minProcs
252 * For fine grain locking to work, the quota check and the reservation of
253 * a server thread has to come while rxi_availProcs and rxi_minDeficit
254 * are locked. To this end, the code has been modified under #ifdef
255 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
256 * same time. A new function, ReturnToServerPool() returns the allocation.
258 * A call can be on several queue's (but only one at a time). When
259 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
260 * that no one else is touching the queue. To this end, we store the address
261 * of the queue lock in the call structure (under the call lock) when we
262 * put the call on a queue, and we clear the call_queue_lock when the
263 * call is removed from a queue (once the call lock has been obtained).
264 * This allows rxi_ResetCall to safely synchronize with others wishing
265 * to manipulate the queue.
268 #ifdef RX_ENABLE_LOCKS
269 static afs_kmutex_t rx_rpc_stats;
270 void rxi_StartUnlocked();
273 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
274 ** pretty good that the next packet coming in is from the same connection
275 ** as the last packet, since we're send multiple packets in a transmit window.
277 struct rx_connection *rxLastConn = 0;
279 #ifdef RX_ENABLE_LOCKS
280 /* The locking hierarchy for rx fine grain locking is composed of these
283 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
284 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
285 * call->lock - locks call data fields.
286 * These are independent of each other:
287 * rx_freeCallQueue_lock
292 * serverQueueEntry->lock
294 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
295 * peer->lock - locks peer data fields.
296 * conn_data_lock - that more than one thread is not updating a conn data
297 * field at the same time.
305 * Do we need a lock to protect the peer field in the conn structure?
306 * conn->peer was previously a constant for all intents and so has no
307 * lock protecting this field. The multihomed client delta introduced
308 * a RX code change : change the peer field in the connection structure
309 * to that remote inetrface from which the last packet for this
310 * connection was sent out. This may become an issue if further changes
313 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
314 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
316 /* rxdb_fileID is used to identify the lock location, along with line#. */
317 static int rxdb_fileID = RXDB_FILE_RX;
318 #endif /* RX_LOCKS_DB */
319 #else /* RX_ENABLE_LOCKS */
320 #define SET_CALL_QUEUE_LOCK(C, L)
321 #define CLEAR_CALL_QUEUE_LOCK(C)
322 #endif /* RX_ENABLE_LOCKS */
323 struct rx_serverQueueEntry *rx_waitForPacket = 0;
324 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
326 /* ------------Exported Interfaces------------- */
328 /* This function allows rxkad to set the epoch to a suitably random number
329 * which rx_NewConnection will use in the future. The principle purpose is to
330 * get rxnull connections to use the same epoch as the rxkad connections do, at
331 * least once the first rxkad connection is established. This is important now
332 * that the host/port addresses aren't used in FindConnection: the uniqueness
333 * of epoch/cid matters and the start time won't do. */
335 #ifdef AFS_PTHREAD_ENV
337 * This mutex protects the following global variables:
341 #define LOCK_EPOCH assert(pthread_mutex_lock(&epoch_mutex)==0)
342 #define UNLOCK_EPOCH assert(pthread_mutex_unlock(&epoch_mutex)==0)
346 #endif /* AFS_PTHREAD_ENV */
349 rx_SetEpoch(afs_uint32 epoch)
356 /* Initialize rx. A port number may be mentioned, in which case this
357 * becomes the default port number for any service installed later.
358 * If 0 is provided for the port number, a random port will be chosen
359 * by the kernel. Whether this will ever overlap anything in
360 * /etc/services is anybody's guess... Returns 0 on success, -1 on
362 static int rxinit_status = 1;
363 #ifdef AFS_PTHREAD_ENV
365 * This mutex protects the following global variables:
369 #define LOCK_RX_INIT assert(pthread_mutex_lock(&rx_init_mutex)==0)
370 #define UNLOCK_RX_INIT assert(pthread_mutex_unlock(&rx_init_mutex)==0)
373 #define UNLOCK_RX_INIT
377 rx_InitHost(u_int host, u_int port)
384 char *htable, *ptable;
387 #if defined(AFS_DJGPP_ENV) && !defined(DEBUG)
388 __djgpp_set_quiet_socket(1);
395 if (rxinit_status == 0) {
396 tmp_status = rxinit_status;
398 return tmp_status; /* Already started; return previous error code. */
401 if (afs_winsockInit() < 0)
407 * Initialize anything necessary to provide a non-premptive threading
410 rxi_InitializeThreadSupport();
413 /* Allocate and initialize a socket for client and perhaps server
416 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
417 if (rx_socket == OSI_NULLSOCKET) {
421 #ifdef RX_ENABLE_LOCKS
424 #endif /* RX_LOCKS_DB */
425 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
426 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
427 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
428 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
429 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
431 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
433 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
435 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
437 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
439 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
441 #if defined(KERNEL) && defined(AFS_HPUX110_ENV)
443 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
444 #endif /* KERNEL && AFS_HPUX110_ENV */
445 #else /* RX_ENABLE_LOCKS */
446 #if defined(KERNEL) && defined(AFS_GLOBAL_SUNLOCK) && !defined(AFS_HPUX_ENV) && !defined(AFS_OBSD_ENV)
447 mutex_init(&afs_rxglobal_lock, "afs_rxglobal_lock", MUTEX_DEFAULT, NULL);
448 #endif /* AFS_GLOBAL_SUNLOCK */
449 #endif /* RX_ENABLE_LOCKS */
452 rx_connDeadTime = 12;
453 rx_tranquil = 0; /* reset flag */
454 memset((char *)&rx_stats, 0, sizeof(struct rx_stats));
456 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
457 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
458 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
459 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
460 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
461 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
463 /* Malloc up a bunch of packets & buffers */
465 queue_Init(&rx_freePacketQueue);
466 rxi_NeedMorePackets = FALSE;
467 #ifdef RX_ENABLE_TSFPQ
468 rx_nPackets = 0; /* in TSFPQ version, rx_nPackets is managed by rxi_MorePackets* */
469 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
470 #else /* RX_ENABLE_TSFPQ */
471 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
472 rxi_MorePackets(rx_nPackets);
473 #endif /* RX_ENABLE_TSFPQ */
480 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
481 tv.tv_sec = clock_now.sec;
482 tv.tv_usec = clock_now.usec;
483 srand((unsigned int)tv.tv_usec);
490 #if defined(KERNEL) && !defined(UKERNEL)
491 /* Really, this should never happen in a real kernel */
494 struct sockaddr_in addr;
495 int addrlen = sizeof(addr);
496 if (getsockname((int)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
500 rx_port = addr.sin_port;
503 rx_stats.minRtt.sec = 9999999;
505 rx_SetEpoch(tv.tv_sec | 0x80000000);
507 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
508 * will provide a randomer value. */
510 MUTEX_ENTER(&rx_stats_mutex);
511 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
512 MUTEX_EXIT(&rx_stats_mutex);
513 /* *Slightly* random start time for the cid. This is just to help
514 * out with the hashing function at the peer */
515 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
516 rx_connHashTable = (struct rx_connection **)htable;
517 rx_peerHashTable = (struct rx_peer **)ptable;
519 rx_lastAckDelay.sec = 0;
520 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
521 rx_hardAckDelay.sec = 0;
522 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
523 rx_softAckDelay.sec = 0;
524 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
526 rxevent_Init(20, rxi_ReScheduleEvents);
528 /* Initialize various global queues */
529 queue_Init(&rx_idleServerQueue);
530 queue_Init(&rx_incomingCallQueue);
531 queue_Init(&rx_freeCallQueue);
533 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
534 /* Initialize our list of usable IP addresses. */
538 /* Start listener process (exact function is dependent on the
539 * implementation environment--kernel or user space) */
543 tmp_status = rxinit_status = 0;
551 return rx_InitHost(htonl(INADDR_ANY), port);
554 /* called with unincremented nRequestsRunning to see if it is OK to start
555 * a new thread in this service. Could be "no" for two reasons: over the
556 * max quota, or would prevent others from reaching their min quota.
558 #ifdef RX_ENABLE_LOCKS
559 /* This verion of QuotaOK reserves quota if it's ok while the
560 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
563 QuotaOK(register struct rx_service *aservice)
565 /* check if over max quota */
566 if (aservice->nRequestsRunning >= aservice->maxProcs) {
570 /* under min quota, we're OK */
571 /* otherwise, can use only if there are enough to allow everyone
572 * to go to their min quota after this guy starts.
574 MUTEX_ENTER(&rx_stats_mutex);
575 if ((aservice->nRequestsRunning < aservice->minProcs)
576 || (rxi_availProcs > rxi_minDeficit)) {
577 aservice->nRequestsRunning++;
578 /* just started call in minProcs pool, need fewer to maintain
580 if (aservice->nRequestsRunning <= aservice->minProcs)
583 MUTEX_EXIT(&rx_stats_mutex);
586 MUTEX_EXIT(&rx_stats_mutex);
592 ReturnToServerPool(register struct rx_service *aservice)
594 aservice->nRequestsRunning--;
595 MUTEX_ENTER(&rx_stats_mutex);
596 if (aservice->nRequestsRunning < aservice->minProcs)
599 MUTEX_EXIT(&rx_stats_mutex);
602 #else /* RX_ENABLE_LOCKS */
604 QuotaOK(register struct rx_service *aservice)
607 /* under min quota, we're OK */
608 if (aservice->nRequestsRunning < aservice->minProcs)
611 /* check if over max quota */
612 if (aservice->nRequestsRunning >= aservice->maxProcs)
615 /* otherwise, can use only if there are enough to allow everyone
616 * to go to their min quota after this guy starts.
618 if (rxi_availProcs > rxi_minDeficit)
622 #endif /* RX_ENABLE_LOCKS */
625 /* Called by rx_StartServer to start up lwp's to service calls.
626 NExistingProcs gives the number of procs already existing, and which
627 therefore needn't be created. */
629 rxi_StartServerProcs(int nExistingProcs)
631 register struct rx_service *service;
636 /* For each service, reserve N processes, where N is the "minimum"
637 * number of processes that MUST be able to execute a request in parallel,
638 * at any time, for that process. Also compute the maximum difference
639 * between any service's maximum number of processes that can run
640 * (i.e. the maximum number that ever will be run, and a guarantee
641 * that this number will run if other services aren't running), and its
642 * minimum number. The result is the extra number of processes that
643 * we need in order to provide the latter guarantee */
644 for (i = 0; i < RX_MAX_SERVICES; i++) {
646 service = rx_services[i];
647 if (service == (struct rx_service *)0)
649 nProcs += service->minProcs;
650 diff = service->maxProcs - service->minProcs;
654 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
655 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
656 for (i = 0; i < nProcs; i++) {
657 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
663 /* This routine is only required on Windows */
665 rx_StartClientThread(void)
667 #ifdef AFS_PTHREAD_ENV
669 pid = (int) pthread_self();
670 #endif /* AFS_PTHREAD_ENV */
672 #endif /* AFS_NT40_ENV */
674 /* This routine must be called if any services are exported. If the
675 * donateMe flag is set, the calling process is donated to the server
678 rx_StartServer(int donateMe)
680 register struct rx_service *service;
686 /* Start server processes, if necessary (exact function is dependent
687 * on the implementation environment--kernel or user space). DonateMe
688 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
689 * case, one less new proc will be created rx_StartServerProcs.
691 rxi_StartServerProcs(donateMe);
693 /* count up the # of threads in minProcs, and add set the min deficit to
694 * be that value, too.
696 for (i = 0; i < RX_MAX_SERVICES; i++) {
697 service = rx_services[i];
698 if (service == (struct rx_service *)0)
700 MUTEX_ENTER(&rx_stats_mutex);
701 rxi_totalMin += service->minProcs;
702 /* below works even if a thread is running, since minDeficit would
703 * still have been decremented and later re-incremented.
705 rxi_minDeficit += service->minProcs;
706 MUTEX_EXIT(&rx_stats_mutex);
709 /* Turn on reaping of idle server connections */
710 rxi_ReapConnections();
719 #ifdef AFS_PTHREAD_ENV
721 pid = (pid_t) pthread_self();
722 #else /* AFS_PTHREAD_ENV */
724 LWP_CurrentProcess(&pid);
725 #endif /* AFS_PTHREAD_ENV */
727 sprintf(name, "srv_%d", ++nProcs);
729 (*registerProgram) (pid, name);
731 #endif /* AFS_NT40_ENV */
732 rx_ServerProc(); /* Never returns */
734 #ifdef RX_ENABLE_TSFPQ
735 /* no use leaving packets around in this thread's local queue if
736 * it isn't getting donated to the server thread pool.
738 rxi_FlushLocalPacketsTSFPQ();
739 #endif /* RX_ENABLE_TSFPQ */
743 /* Create a new client connection to the specified service, using the
744 * specified security object to implement the security model for this
746 struct rx_connection *
747 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
748 register struct rx_securityClass *securityObject,
749 int serviceSecurityIndex)
753 register struct rx_connection *conn;
758 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", shost, sport, sservice, securityObject, serviceSecurityIndex));
760 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
761 * the case of kmem_alloc? */
762 conn = rxi_AllocConnection();
763 #ifdef RX_ENABLE_LOCKS
764 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
765 MUTEX_INIT(&conn->conn_data_lock, "conn call lock", MUTEX_DEFAULT, 0);
766 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
769 MUTEX_ENTER(&rx_connHashTable_lock);
770 cid = (rx_nextCid += RX_MAXCALLS);
771 conn->type = RX_CLIENT_CONNECTION;
773 conn->epoch = rx_epoch;
774 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
775 conn->serviceId = sservice;
776 conn->securityObject = securityObject;
777 /* This doesn't work in all compilers with void (they're buggy), so fake it
779 conn->securityData = (VOID *) 0;
780 conn->securityIndex = serviceSecurityIndex;
781 rx_SetConnDeadTime(conn, rx_connDeadTime);
782 conn->ackRate = RX_FAST_ACK_RATE;
784 conn->specific = NULL;
785 conn->challengeEvent = NULL;
786 conn->delayedAbortEvent = NULL;
787 conn->abortCount = 0;
790 RXS_NewConnection(securityObject, conn);
792 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
794 conn->refCount++; /* no lock required since only this thread knows... */
795 conn->next = rx_connHashTable[hashindex];
796 rx_connHashTable[hashindex] = conn;
797 MUTEX_ENTER(&rx_stats_mutex);
798 rx_stats.nClientConns++;
799 MUTEX_EXIT(&rx_stats_mutex);
801 MUTEX_EXIT(&rx_connHashTable_lock);
807 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
809 /* The idea is to set the dead time to a value that allows several
810 * keepalives to be dropped without timing out the connection. */
811 conn->secondsUntilDead = MAX(seconds, 6);
812 conn->secondsUntilPing = conn->secondsUntilDead / 6;
815 int rxi_lowPeerRefCount = 0;
816 int rxi_lowConnRefCount = 0;
819 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
820 * NOTE: must not be called with rx_connHashTable_lock held.
823 rxi_CleanupConnection(struct rx_connection *conn)
825 /* Notify the service exporter, if requested, that this connection
826 * is being destroyed */
827 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
828 (*conn->service->destroyConnProc) (conn);
830 /* Notify the security module that this connection is being destroyed */
831 RXS_DestroyConnection(conn->securityObject, conn);
833 /* If this is the last connection using the rx_peer struct, set its
834 * idle time to now. rxi_ReapConnections will reap it if it's still
835 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
837 MUTEX_ENTER(&rx_peerHashTable_lock);
838 if (conn->peer->refCount < 2) {
839 conn->peer->idleWhen = clock_Sec();
840 if (conn->peer->refCount < 1) {
841 conn->peer->refCount = 1;
842 MUTEX_ENTER(&rx_stats_mutex);
843 rxi_lowPeerRefCount++;
844 MUTEX_EXIT(&rx_stats_mutex);
847 conn->peer->refCount--;
848 MUTEX_EXIT(&rx_peerHashTable_lock);
850 MUTEX_ENTER(&rx_stats_mutex);
851 if (conn->type == RX_SERVER_CONNECTION)
852 rx_stats.nServerConns--;
854 rx_stats.nClientConns--;
855 MUTEX_EXIT(&rx_stats_mutex);
858 if (conn->specific) {
860 for (i = 0; i < conn->nSpecific; i++) {
861 if (conn->specific[i] && rxi_keyCreate_destructor[i])
862 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
863 conn->specific[i] = NULL;
865 free(conn->specific);
867 conn->specific = NULL;
871 MUTEX_DESTROY(&conn->conn_call_lock);
872 MUTEX_DESTROY(&conn->conn_data_lock);
873 CV_DESTROY(&conn->conn_call_cv);
875 rxi_FreeConnection(conn);
878 /* Destroy the specified connection */
880 rxi_DestroyConnection(register struct rx_connection *conn)
882 MUTEX_ENTER(&rx_connHashTable_lock);
883 rxi_DestroyConnectionNoLock(conn);
884 /* conn should be at the head of the cleanup list */
885 if (conn == rx_connCleanup_list) {
886 rx_connCleanup_list = rx_connCleanup_list->next;
887 MUTEX_EXIT(&rx_connHashTable_lock);
888 rxi_CleanupConnection(conn);
890 #ifdef RX_ENABLE_LOCKS
892 MUTEX_EXIT(&rx_connHashTable_lock);
894 #endif /* RX_ENABLE_LOCKS */
898 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
900 register struct rx_connection **conn_ptr;
901 register int havecalls = 0;
902 struct rx_packet *packet;
909 MUTEX_ENTER(&conn->conn_data_lock);
910 if (conn->refCount > 0)
913 MUTEX_ENTER(&rx_stats_mutex);
914 rxi_lowConnRefCount++;
915 MUTEX_EXIT(&rx_stats_mutex);
918 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
919 /* Busy; wait till the last guy before proceeding */
920 MUTEX_EXIT(&conn->conn_data_lock);
925 /* If the client previously called rx_NewCall, but it is still
926 * waiting, treat this as a running call, and wait to destroy the
927 * connection later when the call completes. */
928 if ((conn->type == RX_CLIENT_CONNECTION)
929 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
930 conn->flags |= RX_CONN_DESTROY_ME;
931 MUTEX_EXIT(&conn->conn_data_lock);
935 MUTEX_EXIT(&conn->conn_data_lock);
937 /* Check for extant references to this connection */
938 for (i = 0; i < RX_MAXCALLS; i++) {
939 register struct rx_call *call = conn->call[i];
942 if (conn->type == RX_CLIENT_CONNECTION) {
943 MUTEX_ENTER(&call->lock);
944 if (call->delayedAckEvent) {
945 /* Push the final acknowledgment out now--there
946 * won't be a subsequent call to acknowledge the
947 * last reply packets */
948 rxevent_Cancel(call->delayedAckEvent, call,
949 RX_CALL_REFCOUNT_DELAY);
950 if (call->state == RX_STATE_PRECALL
951 || call->state == RX_STATE_ACTIVE) {
952 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
954 rxi_AckAll(NULL, call, 0);
957 MUTEX_EXIT(&call->lock);
961 #ifdef RX_ENABLE_LOCKS
963 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
964 MUTEX_EXIT(&conn->conn_data_lock);
966 /* Someone is accessing a packet right now. */
970 #endif /* RX_ENABLE_LOCKS */
973 /* Don't destroy the connection if there are any call
974 * structures still in use */
975 MUTEX_ENTER(&conn->conn_data_lock);
976 conn->flags |= RX_CONN_DESTROY_ME;
977 MUTEX_EXIT(&conn->conn_data_lock);
982 if (conn->delayedAbortEvent) {
983 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
984 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
986 MUTEX_ENTER(&conn->conn_data_lock);
987 rxi_SendConnectionAbort(conn, packet, 0, 1);
988 MUTEX_EXIT(&conn->conn_data_lock);
989 rxi_FreePacket(packet);
993 /* Remove from connection hash table before proceeding */
995 &rx_connHashTable[CONN_HASH
996 (peer->host, peer->port, conn->cid, conn->epoch,
998 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
999 if (*conn_ptr == conn) {
1000 *conn_ptr = conn->next;
1004 /* if the conn that we are destroying was the last connection, then we
1005 * clear rxLastConn as well */
1006 if (rxLastConn == conn)
1009 /* Make sure the connection is completely reset before deleting it. */
1010 /* get rid of pending events that could zap us later */
1011 if (conn->challengeEvent)
1012 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1013 if (conn->checkReachEvent)
1014 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1016 /* Add the connection to the list of destroyed connections that
1017 * need to be cleaned up. This is necessary to avoid deadlocks
1018 * in the routines we call to inform others that this connection is
1019 * being destroyed. */
1020 conn->next = rx_connCleanup_list;
1021 rx_connCleanup_list = conn;
1024 /* Externally available version */
1026 rx_DestroyConnection(register struct rx_connection *conn)
1031 rxi_DestroyConnection(conn);
1036 rx_GetConnection(register struct rx_connection *conn)
1041 MUTEX_ENTER(&conn->conn_data_lock);
1043 MUTEX_EXIT(&conn->conn_data_lock);
1047 /* Start a new rx remote procedure call, on the specified connection.
1048 * If wait is set to 1, wait for a free call channel; otherwise return
1049 * 0. Maxtime gives the maximum number of seconds this call may take,
1050 * after rx_MakeCall returns. After this time interval, a call to any
1051 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1052 * For fine grain locking, we hold the conn_call_lock in order to
1053 * to ensure that we don't get signalle after we found a call in an active
1054 * state and before we go to sleep.
1057 rx_NewCall(register struct rx_connection *conn)
1060 register struct rx_call *call;
1061 struct clock queueTime;
1065 dpf(("rx_MakeCall(conn %x)\n", conn));
1068 clock_GetTime(&queueTime);
1069 MUTEX_ENTER(&conn->conn_call_lock);
1072 * Check if there are others waiting for a new call.
1073 * If so, let them go first to avoid starving them.
1074 * This is a fairly simple scheme, and might not be
1075 * a complete solution for large numbers of waiters.
1077 * makeCallWaiters keeps track of the number of
1078 * threads waiting to make calls and the
1079 * RX_CONN_MAKECALL_WAITING flag bit is used to
1080 * indicate that there are indeed calls waiting.
1081 * The flag is set when the waiter is incremented.
1082 * It is only cleared in rx_EndCall when
1083 * makeCallWaiters is 0. This prevents us from
1084 * accidently destroying the connection while it
1085 * is potentially about to be used.
1087 MUTEX_ENTER(&conn->conn_data_lock);
1088 if (conn->makeCallWaiters) {
1089 conn->flags |= RX_CONN_MAKECALL_WAITING;
1090 conn->makeCallWaiters++;
1091 MUTEX_EXIT(&conn->conn_data_lock);
1093 #ifdef RX_ENABLE_LOCKS
1094 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1098 MUTEX_ENTER(&conn->conn_data_lock);
1099 conn->makeCallWaiters--;
1101 MUTEX_EXIT(&conn->conn_data_lock);
1104 for (i = 0; i < RX_MAXCALLS; i++) {
1105 call = conn->call[i];
1107 MUTEX_ENTER(&call->lock);
1108 if (call->state == RX_STATE_DALLY) {
1109 rxi_ResetCall(call, 0);
1110 (*call->callNumber)++;
1113 MUTEX_EXIT(&call->lock);
1115 call = rxi_NewCall(conn, i);
1119 if (i < RX_MAXCALLS) {
1122 MUTEX_ENTER(&conn->conn_data_lock);
1123 conn->flags |= RX_CONN_MAKECALL_WAITING;
1124 conn->makeCallWaiters++;
1125 MUTEX_EXIT(&conn->conn_data_lock);
1127 #ifdef RX_ENABLE_LOCKS
1128 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1132 MUTEX_ENTER(&conn->conn_data_lock);
1133 conn->makeCallWaiters--;
1134 MUTEX_EXIT(&conn->conn_data_lock);
1137 * Wake up anyone else who might be giving us a chance to
1138 * run (see code above that avoids resource starvation).
1140 #ifdef RX_ENABLE_LOCKS
1141 CV_BROADCAST(&conn->conn_call_cv);
1146 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1148 /* Client is initially in send mode */
1149 call->state = RX_STATE_ACTIVE;
1150 call->mode = RX_MODE_SENDING;
1152 /* remember start time for call in case we have hard dead time limit */
1153 call->queueTime = queueTime;
1154 clock_GetTime(&call->startTime);
1155 hzero(call->bytesSent);
1156 hzero(call->bytesRcvd);
1158 /* Turn on busy protocol. */
1159 rxi_KeepAliveOn(call);
1161 MUTEX_EXIT(&call->lock);
1162 MUTEX_EXIT(&conn->conn_call_lock);
1165 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1166 /* Now, if TQ wasn't cleared earlier, do it now. */
1167 MUTEX_ENTER(&call->lock);
1168 while (call->flags & RX_CALL_TQ_BUSY) {
1169 call->flags |= RX_CALL_TQ_WAIT;
1170 #ifdef RX_ENABLE_LOCKS
1171 CV_WAIT(&call->cv_tq, &call->lock);
1172 #else /* RX_ENABLE_LOCKS */
1173 osi_rxSleep(&call->tq);
1174 #endif /* RX_ENABLE_LOCKS */
1176 if (call->flags & RX_CALL_TQ_CLEARME) {
1177 rxi_ClearTransmitQueue(call, 0);
1178 queue_Init(&call->tq);
1180 MUTEX_EXIT(&call->lock);
1181 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1187 rxi_HasActiveCalls(register struct rx_connection *aconn)
1190 register struct rx_call *tcall;
1194 for (i = 0; i < RX_MAXCALLS; i++) {
1195 if ((tcall = aconn->call[i])) {
1196 if ((tcall->state == RX_STATE_ACTIVE)
1197 || (tcall->state == RX_STATE_PRECALL)) {
1208 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1209 register afs_int32 * aint32s)
1212 register struct rx_call *tcall;
1216 for (i = 0; i < RX_MAXCALLS; i++) {
1217 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1218 aint32s[i] = aconn->callNumber[i] + 1;
1220 aint32s[i] = aconn->callNumber[i];
1227 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1228 register afs_int32 * aint32s)
1231 register struct rx_call *tcall;
1235 for (i = 0; i < RX_MAXCALLS; i++) {
1236 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1237 aconn->callNumber[i] = aint32s[i] - 1;
1239 aconn->callNumber[i] = aint32s[i];
1245 /* Advertise a new service. A service is named locally by a UDP port
1246 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1249 char *serviceName; Name for identification purposes (e.g. the
1250 service name might be used for probing for
1253 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1254 struct rx_securityClass **securityObjects, int nSecurityObjects,
1255 afs_int32(*serviceProc) (struct rx_call * acall))
1257 osi_socket socket = OSI_NULLSOCKET;
1258 register struct rx_service *tservice;
1264 if (serviceId == 0) {
1266 "rx_NewService: service id for service %s is not non-zero.\n",
1273 "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",
1281 tservice = rxi_AllocService();
1283 for (i = 0; i < RX_MAX_SERVICES; i++) {
1284 register struct rx_service *service = rx_services[i];
1286 if (port == service->servicePort) {
1287 if (service->serviceId == serviceId) {
1288 /* The identical service has already been
1289 * installed; if the caller was intending to
1290 * change the security classes used by this
1291 * service, he/she loses. */
1293 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1294 serviceName, serviceId, service->serviceName);
1296 rxi_FreeService(tservice);
1299 /* Different service, same port: re-use the socket
1300 * which is bound to the same port */
1301 socket = service->socket;
1304 if (socket == OSI_NULLSOCKET) {
1305 /* If we don't already have a socket (from another
1306 * service on same port) get a new one */
1307 socket = rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
1308 if (socket == OSI_NULLSOCKET) {
1310 rxi_FreeService(tservice);
1315 service->socket = socket;
1316 service->servicePort = port;
1317 service->serviceId = serviceId;
1318 service->serviceName = serviceName;
1319 service->nSecurityObjects = nSecurityObjects;
1320 service->securityObjects = securityObjects;
1321 service->minProcs = 0;
1322 service->maxProcs = 1;
1323 service->idleDeadTime = 60;
1324 service->connDeadTime = rx_connDeadTime;
1325 service->executeRequestProc = serviceProc;
1326 service->checkReach = 0;
1327 rx_services[i] = service; /* not visible until now */
1333 rxi_FreeService(tservice);
1334 (osi_Msg "rx_NewService: cannot support > %d services\n",
1339 /* Generic request processing loop. This routine should be called
1340 * by the implementation dependent rx_ServerProc. If socketp is
1341 * non-null, it will be set to the file descriptor that this thread
1342 * is now listening on. If socketp is null, this routine will never
1345 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1347 register struct rx_call *call;
1348 register afs_int32 code;
1349 register struct rx_service *tservice = NULL;
1356 call = rx_GetCall(threadID, tservice, socketp);
1357 if (socketp && *socketp != OSI_NULLSOCKET) {
1358 /* We are now a listener thread */
1363 /* if server is restarting( typically smooth shutdown) then do not
1364 * allow any new calls.
1367 if (rx_tranquil && (call != NULL)) {
1371 MUTEX_ENTER(&call->lock);
1373 rxi_CallError(call, RX_RESTARTING);
1374 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1376 MUTEX_EXIT(&call->lock);
1380 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1381 #ifdef RX_ENABLE_LOCKS
1383 #endif /* RX_ENABLE_LOCKS */
1384 afs_termState = AFSOP_STOP_AFS;
1385 afs_osi_Wakeup(&afs_termState);
1386 #ifdef RX_ENABLE_LOCKS
1388 #endif /* RX_ENABLE_LOCKS */
1393 tservice = call->conn->service;
1395 if (tservice->beforeProc)
1396 (*tservice->beforeProc) (call);
1398 code = call->conn->service->executeRequestProc(call);
1400 if (tservice->afterProc)
1401 (*tservice->afterProc) (call, code);
1403 rx_EndCall(call, code);
1404 MUTEX_ENTER(&rx_stats_mutex);
1406 MUTEX_EXIT(&rx_stats_mutex);
1412 rx_WakeupServerProcs(void)
1414 struct rx_serverQueueEntry *np, *tqp;
1418 MUTEX_ENTER(&rx_serverPool_lock);
1420 #ifdef RX_ENABLE_LOCKS
1421 if (rx_waitForPacket)
1422 CV_BROADCAST(&rx_waitForPacket->cv);
1423 #else /* RX_ENABLE_LOCKS */
1424 if (rx_waitForPacket)
1425 osi_rxWakeup(rx_waitForPacket);
1426 #endif /* RX_ENABLE_LOCKS */
1427 MUTEX_ENTER(&freeSQEList_lock);
1428 for (np = rx_FreeSQEList; np; np = tqp) {
1429 tqp = *(struct rx_serverQueueEntry **)np;
1430 #ifdef RX_ENABLE_LOCKS
1431 CV_BROADCAST(&np->cv);
1432 #else /* RX_ENABLE_LOCKS */
1434 #endif /* RX_ENABLE_LOCKS */
1436 MUTEX_EXIT(&freeSQEList_lock);
1437 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1438 #ifdef RX_ENABLE_LOCKS
1439 CV_BROADCAST(&np->cv);
1440 #else /* RX_ENABLE_LOCKS */
1442 #endif /* RX_ENABLE_LOCKS */
1444 MUTEX_EXIT(&rx_serverPool_lock);
1449 * One thing that seems to happen is that all the server threads get
1450 * tied up on some empty or slow call, and then a whole bunch of calls
1451 * arrive at once, using up the packet pool, so now there are more
1452 * empty calls. The most critical resources here are server threads
1453 * and the free packet pool. The "doreclaim" code seems to help in
1454 * general. I think that eventually we arrive in this state: there
1455 * are lots of pending calls which do have all their packets present,
1456 * so they won't be reclaimed, are multi-packet calls, so they won't
1457 * be scheduled until later, and thus are tying up most of the free
1458 * packet pool for a very long time.
1460 * 1. schedule multi-packet calls if all the packets are present.
1461 * Probably CPU-bound operation, useful to return packets to pool.
1462 * Do what if there is a full window, but the last packet isn't here?
1463 * 3. preserve one thread which *only* runs "best" calls, otherwise
1464 * it sleeps and waits for that type of call.
1465 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1466 * the current dataquota business is badly broken. The quota isn't adjusted
1467 * to reflect how many packets are presently queued for a running call.
1468 * So, when we schedule a queued call with a full window of packets queued
1469 * up for it, that *should* free up a window full of packets for other 2d-class
1470 * calls to be able to use from the packet pool. But it doesn't.
1472 * NB. Most of the time, this code doesn't run -- since idle server threads
1473 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1474 * as a new call arrives.
1476 /* Sleep until a call arrives. Returns a pointer to the call, ready
1477 * for an rx_Read. */
1478 #ifdef RX_ENABLE_LOCKS
1480 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1482 struct rx_serverQueueEntry *sq;
1483 register struct rx_call *call = (struct rx_call *)0;
1484 struct rx_service *service = NULL;
1487 MUTEX_ENTER(&freeSQEList_lock);
1489 if ((sq = rx_FreeSQEList)) {
1490 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1491 MUTEX_EXIT(&freeSQEList_lock);
1492 } else { /* otherwise allocate a new one and return that */
1493 MUTEX_EXIT(&freeSQEList_lock);
1494 sq = (struct rx_serverQueueEntry *)
1495 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1496 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1497 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1500 MUTEX_ENTER(&rx_serverPool_lock);
1501 if (cur_service != NULL) {
1502 ReturnToServerPool(cur_service);
1505 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1506 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1508 /* Scan for eligible incoming calls. A call is not eligible
1509 * if the maximum number of calls for its service type are
1510 * already executing */
1511 /* One thread will process calls FCFS (to prevent starvation),
1512 * while the other threads may run ahead looking for calls which
1513 * have all their input data available immediately. This helps
1514 * keep threads from blocking, waiting for data from the client. */
1515 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1516 service = tcall->conn->service;
1517 if (!QuotaOK(service)) {
1520 if (tno == rxi_fcfs_thread_num
1521 || !tcall->queue_item_header.next) {
1522 /* If we're the fcfs thread , then we'll just use
1523 * this call. If we haven't been able to find an optimal
1524 * choice, and we're at the end of the list, then use a
1525 * 2d choice if one has been identified. Otherwise... */
1526 call = (choice2 ? choice2 : tcall);
1527 service = call->conn->service;
1528 } else if (!queue_IsEmpty(&tcall->rq)) {
1529 struct rx_packet *rp;
1530 rp = queue_First(&tcall->rq, rx_packet);
1531 if (rp->header.seq == 1) {
1533 || (rp->header.flags & RX_LAST_PACKET)) {
1535 } else if (rxi_2dchoice && !choice2
1536 && !(tcall->flags & RX_CALL_CLEARED)
1537 && (tcall->rprev > rxi_HardAckRate)) {
1546 ReturnToServerPool(service);
1553 MUTEX_EXIT(&rx_serverPool_lock);
1554 MUTEX_ENTER(&call->lock);
1556 if (call->flags & RX_CALL_WAIT_PROC) {
1557 call->flags &= ~RX_CALL_WAIT_PROC;
1558 MUTEX_ENTER(&rx_stats_mutex);
1560 MUTEX_EXIT(&rx_stats_mutex);
1563 if (call->state != RX_STATE_PRECALL || call->error) {
1564 MUTEX_EXIT(&call->lock);
1565 MUTEX_ENTER(&rx_serverPool_lock);
1566 ReturnToServerPool(service);
1571 if (queue_IsEmpty(&call->rq)
1572 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1573 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1575 CLEAR_CALL_QUEUE_LOCK(call);
1578 /* If there are no eligible incoming calls, add this process
1579 * to the idle server queue, to wait for one */
1583 *socketp = OSI_NULLSOCKET;
1585 sq->socketp = socketp;
1586 queue_Append(&rx_idleServerQueue, sq);
1587 #ifndef AFS_AIX41_ENV
1588 rx_waitForPacket = sq;
1590 rx_waitingForPacket = sq;
1591 #endif /* AFS_AIX41_ENV */
1593 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1595 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1596 MUTEX_EXIT(&rx_serverPool_lock);
1597 return (struct rx_call *)0;
1600 } while (!(call = sq->newcall)
1601 && !(socketp && *socketp != OSI_NULLSOCKET));
1602 MUTEX_EXIT(&rx_serverPool_lock);
1604 MUTEX_ENTER(&call->lock);
1610 MUTEX_ENTER(&freeSQEList_lock);
1611 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1612 rx_FreeSQEList = sq;
1613 MUTEX_EXIT(&freeSQEList_lock);
1616 clock_GetTime(&call->startTime);
1617 call->state = RX_STATE_ACTIVE;
1618 call->mode = RX_MODE_RECEIVING;
1619 #ifdef RX_KERNEL_TRACE
1620 if (ICL_SETACTIVE(afs_iclSetp)) {
1621 int glockOwner = ISAFS_GLOCK();
1624 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1625 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1632 rxi_calltrace(RX_CALL_START, call);
1633 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1634 call->conn->service->servicePort, call->conn->service->serviceId,
1637 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1638 MUTEX_EXIT(&call->lock);
1640 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1645 #else /* RX_ENABLE_LOCKS */
1647 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1649 struct rx_serverQueueEntry *sq;
1650 register struct rx_call *call = (struct rx_call *)0, *choice2;
1651 struct rx_service *service = NULL;
1655 MUTEX_ENTER(&freeSQEList_lock);
1657 if ((sq = rx_FreeSQEList)) {
1658 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1659 MUTEX_EXIT(&freeSQEList_lock);
1660 } else { /* otherwise allocate a new one and return that */
1661 MUTEX_EXIT(&freeSQEList_lock);
1662 sq = (struct rx_serverQueueEntry *)
1663 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1664 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1665 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1667 MUTEX_ENTER(&sq->lock);
1669 if (cur_service != NULL) {
1670 cur_service->nRequestsRunning--;
1671 if (cur_service->nRequestsRunning < cur_service->minProcs)
1675 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1676 register struct rx_call *tcall, *ncall;
1677 /* Scan for eligible incoming calls. A call is not eligible
1678 * if the maximum number of calls for its service type are
1679 * already executing */
1680 /* One thread will process calls FCFS (to prevent starvation),
1681 * while the other threads may run ahead looking for calls which
1682 * have all their input data available immediately. This helps
1683 * keep threads from blocking, waiting for data from the client. */
1684 choice2 = (struct rx_call *)0;
1685 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1686 service = tcall->conn->service;
1687 if (QuotaOK(service)) {
1688 if (tno == rxi_fcfs_thread_num
1689 || !tcall->queue_item_header.next) {
1690 /* If we're the fcfs thread, then we'll just use
1691 * this call. If we haven't been able to find an optimal
1692 * choice, and we're at the end of the list, then use a
1693 * 2d choice if one has been identified. Otherwise... */
1694 call = (choice2 ? choice2 : tcall);
1695 service = call->conn->service;
1696 } else if (!queue_IsEmpty(&tcall->rq)) {
1697 struct rx_packet *rp;
1698 rp = queue_First(&tcall->rq, rx_packet);
1699 if (rp->header.seq == 1
1701 || (rp->header.flags & RX_LAST_PACKET))) {
1703 } else if (rxi_2dchoice && !choice2
1704 && !(tcall->flags & RX_CALL_CLEARED)
1705 && (tcall->rprev > rxi_HardAckRate)) {
1718 /* we can't schedule a call if there's no data!!! */
1719 /* send an ack if there's no data, if we're missing the
1720 * first packet, or we're missing something between first
1721 * and last -- there's a "hole" in the incoming data. */
1722 if (queue_IsEmpty(&call->rq)
1723 || queue_First(&call->rq, rx_packet)->header.seq != 1
1724 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1725 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1727 call->flags &= (~RX_CALL_WAIT_PROC);
1728 service->nRequestsRunning++;
1729 /* just started call in minProcs pool, need fewer to maintain
1731 if (service->nRequestsRunning <= service->minProcs)
1735 /* MUTEX_EXIT(&call->lock); */
1737 /* If there are no eligible incoming calls, add this process
1738 * to the idle server queue, to wait for one */
1741 *socketp = OSI_NULLSOCKET;
1743 sq->socketp = socketp;
1744 queue_Append(&rx_idleServerQueue, sq);
1748 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1750 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1751 return (struct rx_call *)0;
1754 } while (!(call = sq->newcall)
1755 && !(socketp && *socketp != OSI_NULLSOCKET));
1757 MUTEX_EXIT(&sq->lock);
1759 MUTEX_ENTER(&freeSQEList_lock);
1760 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1761 rx_FreeSQEList = sq;
1762 MUTEX_EXIT(&freeSQEList_lock);
1765 clock_GetTime(&call->startTime);
1766 call->state = RX_STATE_ACTIVE;
1767 call->mode = RX_MODE_RECEIVING;
1768 #ifdef RX_KERNEL_TRACE
1769 if (ICL_SETACTIVE(afs_iclSetp)) {
1770 int glockOwner = ISAFS_GLOCK();
1773 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1774 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1781 rxi_calltrace(RX_CALL_START, call);
1782 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1783 call->conn->service->servicePort, call->conn->service->serviceId,
1786 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1793 #endif /* RX_ENABLE_LOCKS */
1797 /* Establish a procedure to be called when a packet arrives for a
1798 * call. This routine will be called at most once after each call,
1799 * and will also be called if there is an error condition on the or
1800 * the call is complete. Used by multi rx to build a selection
1801 * function which determines which of several calls is likely to be a
1802 * good one to read from.
1803 * NOTE: the way this is currently implemented it is probably only a
1804 * good idea to (1) use it immediately after a newcall (clients only)
1805 * and (2) only use it once. Other uses currently void your warranty
1808 rx_SetArrivalProc(register struct rx_call *call,
1809 register void (*proc) (register struct rx_call * call,
1811 register int index),
1812 register VOID * handle, register int arg)
1814 call->arrivalProc = proc;
1815 call->arrivalProcHandle = handle;
1816 call->arrivalProcArg = arg;
1819 /* Call is finished (possibly prematurely). Return rc to the peer, if
1820 * appropriate, and return the final error code from the conversation
1824 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1826 register struct rx_connection *conn = call->conn;
1827 register struct rx_service *service;
1828 register struct rx_packet *tp; /* Temporary packet pointer */
1829 register struct rx_packet *nxp; /* Next packet pointer, for queue_Scan */
1833 dpf(("rx_EndCall(call %x)\n", call));
1836 MUTEX_ENTER(&call->lock);
1838 if (rc == 0 && call->error == 0) {
1839 call->abortCode = 0;
1840 call->abortCount = 0;
1843 call->arrivalProc = (void (*)())0;
1844 if (rc && call->error == 0) {
1845 rxi_CallError(call, rc);
1846 /* Send an abort message to the peer if this error code has
1847 * only just been set. If it was set previously, assume the
1848 * peer has already been sent the error code or will request it
1850 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1852 if (conn->type == RX_SERVER_CONNECTION) {
1853 /* Make sure reply or at least dummy reply is sent */
1854 if (call->mode == RX_MODE_RECEIVING) {
1855 rxi_WriteProc(call, 0, 0);
1857 if (call->mode == RX_MODE_SENDING) {
1858 rxi_FlushWrite(call);
1860 service = conn->service;
1861 rxi_calltrace(RX_CALL_END, call);
1862 /* Call goes to hold state until reply packets are acknowledged */
1863 if (call->tfirst + call->nSoftAcked < call->tnext) {
1864 call->state = RX_STATE_HOLD;
1866 call->state = RX_STATE_DALLY;
1867 rxi_ClearTransmitQueue(call, 0);
1868 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1869 rxevent_Cancel(call->keepAliveEvent, call,
1870 RX_CALL_REFCOUNT_ALIVE);
1872 } else { /* Client connection */
1874 /* Make sure server receives input packets, in the case where
1875 * no reply arguments are expected */
1876 if ((call->mode == RX_MODE_SENDING)
1877 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1878 (void)rxi_ReadProc(call, &dummy, 1);
1881 /* If we had an outstanding delayed ack, be nice to the server
1882 * and force-send it now.
1884 if (call->delayedAckEvent) {
1885 rxevent_Cancel(call->delayedAckEvent, call,
1886 RX_CALL_REFCOUNT_DELAY);
1887 call->delayedAckEvent = NULL;
1888 rxi_SendDelayedAck(NULL, call, NULL);
1891 /* We need to release the call lock since it's lower than the
1892 * conn_call_lock and we don't want to hold the conn_call_lock
1893 * over the rx_ReadProc call. The conn_call_lock needs to be held
1894 * here for the case where rx_NewCall is perusing the calls on
1895 * the connection structure. We don't want to signal until
1896 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1897 * have checked this call, found it active and by the time it
1898 * goes to sleep, will have missed the signal.
1900 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
1901 * there are threads waiting to use the conn object.
1903 MUTEX_EXIT(&call->lock);
1904 MUTEX_ENTER(&conn->conn_call_lock);
1905 MUTEX_ENTER(&call->lock);
1906 MUTEX_ENTER(&conn->conn_data_lock);
1907 conn->flags |= RX_CONN_BUSY;
1908 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1909 if (conn->makeCallWaiters == 0)
1910 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1911 MUTEX_EXIT(&conn->conn_data_lock);
1912 #ifdef RX_ENABLE_LOCKS
1913 CV_BROADCAST(&conn->conn_call_cv);
1918 #ifdef RX_ENABLE_LOCKS
1920 MUTEX_EXIT(&conn->conn_data_lock);
1922 #endif /* RX_ENABLE_LOCKS */
1923 call->state = RX_STATE_DALLY;
1925 error = call->error;
1927 /* currentPacket, nLeft, and NFree must be zeroed here, because
1928 * ResetCall cannot: ResetCall may be called at splnet(), in the
1929 * kernel version, and may interrupt the macros rx_Read or
1930 * rx_Write, which run at normal priority for efficiency. */
1931 if (call->currentPacket) {
1932 queue_Prepend(&call->iovq, call->currentPacket);
1933 call->currentPacket = (struct rx_packet *)0;
1936 call->nLeft = call->nFree = call->curlen = 0;
1938 /* Free any packets from the last call to ReadvProc/WritevProc */
1939 rxi_FreePackets(0, &call->iovq);
1941 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1942 MUTEX_EXIT(&call->lock);
1943 if (conn->type == RX_CLIENT_CONNECTION) {
1944 MUTEX_EXIT(&conn->conn_call_lock);
1945 conn->flags &= ~RX_CONN_BUSY;
1949 * Map errors to the local host's errno.h format.
1951 error = ntoh_syserr_conv(error);
1955 #if !defined(KERNEL)
1957 /* Call this routine when shutting down a server or client (especially
1958 * clients). This will allow Rx to gracefully garbage collect server
1959 * connections, and reduce the number of retries that a server might
1960 * make to a dead client.
1961 * This is not quite right, since some calls may still be ongoing and
1962 * we can't lock them to destroy them. */
1966 register struct rx_connection **conn_ptr, **conn_end;
1970 if (rxinit_status == 1) {
1972 return; /* Already shutdown. */
1974 rxi_DeleteCachedConnections();
1975 if (rx_connHashTable) {
1976 MUTEX_ENTER(&rx_connHashTable_lock);
1977 for (conn_ptr = &rx_connHashTable[0], conn_end =
1978 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
1980 struct rx_connection *conn, *next;
1981 for (conn = *conn_ptr; conn; conn = next) {
1983 if (conn->type == RX_CLIENT_CONNECTION) {
1984 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
1986 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
1987 #ifdef RX_ENABLE_LOCKS
1988 rxi_DestroyConnectionNoLock(conn);
1989 #else /* RX_ENABLE_LOCKS */
1990 rxi_DestroyConnection(conn);
1991 #endif /* RX_ENABLE_LOCKS */
1995 #ifdef RX_ENABLE_LOCKS
1996 while (rx_connCleanup_list) {
1997 struct rx_connection *conn;
1998 conn = rx_connCleanup_list;
1999 rx_connCleanup_list = rx_connCleanup_list->next;
2000 MUTEX_EXIT(&rx_connHashTable_lock);
2001 rxi_CleanupConnection(conn);
2002 MUTEX_ENTER(&rx_connHashTable_lock);
2004 MUTEX_EXIT(&rx_connHashTable_lock);
2005 #endif /* RX_ENABLE_LOCKS */
2014 /* if we wakeup packet waiter too often, can get in loop with two
2015 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2017 rxi_PacketsUnWait(void)
2019 if (!rx_waitingForPackets) {
2023 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2024 return; /* still over quota */
2027 rx_waitingForPackets = 0;
2028 #ifdef RX_ENABLE_LOCKS
2029 CV_BROADCAST(&rx_waitingForPackets_cv);
2031 osi_rxWakeup(&rx_waitingForPackets);
2037 /* ------------------Internal interfaces------------------------- */
2039 /* Return this process's service structure for the
2040 * specified socket and service */
2042 rxi_FindService(register osi_socket socket, register u_short serviceId)
2044 register struct rx_service **sp;
2045 for (sp = &rx_services[0]; *sp; sp++) {
2046 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2052 /* Allocate a call structure, for the indicated channel of the
2053 * supplied connection. The mode and state of the call must be set by
2054 * the caller. Returns the call with mutex locked. */
2056 rxi_NewCall(register struct rx_connection *conn, register int channel)
2058 register struct rx_call *call;
2059 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2060 register struct rx_call *cp; /* Call pointer temp */
2061 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2062 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2064 /* Grab an existing call structure, or allocate a new one.
2065 * Existing call structures are assumed to have been left reset by
2067 MUTEX_ENTER(&rx_freeCallQueue_lock);
2069 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2071 * EXCEPT that the TQ might not yet be cleared out.
2072 * Skip over those with in-use TQs.
2075 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2076 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2082 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2083 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2084 call = queue_First(&rx_freeCallQueue, rx_call);
2085 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2087 MUTEX_ENTER(&rx_stats_mutex);
2088 rx_stats.nFreeCallStructs--;
2089 MUTEX_EXIT(&rx_stats_mutex);
2090 MUTEX_EXIT(&rx_freeCallQueue_lock);
2091 MUTEX_ENTER(&call->lock);
2092 CLEAR_CALL_QUEUE_LOCK(call);
2093 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2094 /* Now, if TQ wasn't cleared earlier, do it now. */
2095 if (call->flags & RX_CALL_TQ_CLEARME) {
2096 rxi_ClearTransmitQueue(call, 0);
2097 queue_Init(&call->tq);
2099 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2100 /* Bind the call to its connection structure */
2102 rxi_ResetCall(call, 1);
2104 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2106 MUTEX_EXIT(&rx_freeCallQueue_lock);
2107 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2108 MUTEX_ENTER(&call->lock);
2109 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2110 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2111 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2113 MUTEX_ENTER(&rx_stats_mutex);
2114 rx_stats.nCallStructs++;
2115 MUTEX_EXIT(&rx_stats_mutex);
2116 /* Initialize once-only items */
2117 queue_Init(&call->tq);
2118 queue_Init(&call->rq);
2119 queue_Init(&call->iovq);
2120 /* Bind the call to its connection structure (prereq for reset) */
2122 rxi_ResetCall(call, 1);
2124 call->channel = channel;
2125 call->callNumber = &conn->callNumber[channel];
2126 /* Note that the next expected call number is retained (in
2127 * conn->callNumber[i]), even if we reallocate the call structure
2129 conn->call[channel] = call;
2130 /* if the channel's never been used (== 0), we should start at 1, otherwise
2131 * the call number is valid from the last time this channel was used */
2132 if (*call->callNumber == 0)
2133 *call->callNumber = 1;
2138 /* A call has been inactive long enough that so we can throw away
2139 * state, including the call structure, which is placed on the call
2141 * Call is locked upon entry.
2142 * haveCTLock set if called from rxi_ReapConnections
2144 #ifdef RX_ENABLE_LOCKS
2146 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2147 #else /* RX_ENABLE_LOCKS */
2149 rxi_FreeCall(register struct rx_call *call)
2150 #endif /* RX_ENABLE_LOCKS */
2152 register int channel = call->channel;
2153 register struct rx_connection *conn = call->conn;
2156 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2157 (*call->callNumber)++;
2158 rxi_ResetCall(call, 0);
2159 call->conn->call[channel] = (struct rx_call *)0;
2161 MUTEX_ENTER(&rx_freeCallQueue_lock);
2162 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2163 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2164 /* A call may be free even though its transmit queue is still in use.
2165 * Since we search the call list from head to tail, put busy calls at
2166 * the head of the list, and idle calls at the tail.
2168 if (call->flags & RX_CALL_TQ_BUSY)
2169 queue_Prepend(&rx_freeCallQueue, call);
2171 queue_Append(&rx_freeCallQueue, call);
2172 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2173 queue_Append(&rx_freeCallQueue, call);
2174 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2175 MUTEX_ENTER(&rx_stats_mutex);
2176 rx_stats.nFreeCallStructs++;
2177 MUTEX_EXIT(&rx_stats_mutex);
2179 MUTEX_EXIT(&rx_freeCallQueue_lock);
2181 /* Destroy the connection if it was previously slated for
2182 * destruction, i.e. the Rx client code previously called
2183 * rx_DestroyConnection (client connections), or
2184 * rxi_ReapConnections called the same routine (server
2185 * connections). Only do this, however, if there are no
2186 * outstanding calls. Note that for fine grain locking, there appears
2187 * to be a deadlock in that rxi_FreeCall has a call locked and
2188 * DestroyConnectionNoLock locks each call in the conn. But note a
2189 * few lines up where we have removed this call from the conn.
2190 * If someone else destroys a connection, they either have no
2191 * call lock held or are going through this section of code.
2193 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2194 MUTEX_ENTER(&conn->conn_data_lock);
2196 MUTEX_EXIT(&conn->conn_data_lock);
2197 #ifdef RX_ENABLE_LOCKS
2199 rxi_DestroyConnectionNoLock(conn);
2201 rxi_DestroyConnection(conn);
2202 #else /* RX_ENABLE_LOCKS */
2203 rxi_DestroyConnection(conn);
2204 #endif /* RX_ENABLE_LOCKS */
2208 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2210 rxi_Alloc(register size_t size)
2214 MUTEX_ENTER(&rx_stats_mutex);
2216 rxi_Allocsize += size;
2217 MUTEX_EXIT(&rx_stats_mutex);
2219 p = (char *)osi_Alloc(size);
2222 osi_Panic("rxi_Alloc error");
2228 rxi_Free(void *addr, register size_t size)
2230 MUTEX_ENTER(&rx_stats_mutex);
2232 rxi_Allocsize -= size;
2233 MUTEX_EXIT(&rx_stats_mutex);
2235 osi_Free(addr, size);
2238 /* Find the peer process represented by the supplied (host,port)
2239 * combination. If there is no appropriate active peer structure, a
2240 * new one will be allocated and initialized
2241 * The origPeer, if set, is a pointer to a peer structure on which the
2242 * refcount will be be decremented. This is used to replace the peer
2243 * structure hanging off a connection structure */
2245 rxi_FindPeer(register afs_uint32 host, register u_short port,
2246 struct rx_peer *origPeer, int create)
2248 register struct rx_peer *pp;
2250 hashIndex = PEER_HASH(host, port);
2251 MUTEX_ENTER(&rx_peerHashTable_lock);
2252 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2253 if ((pp->host == host) && (pp->port == port))
2258 pp = rxi_AllocPeer(); /* This bzero's *pp */
2259 pp->host = host; /* set here or in InitPeerParams is zero */
2261 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2262 queue_Init(&pp->congestionQueue);
2263 queue_Init(&pp->rpcStats);
2264 pp->next = rx_peerHashTable[hashIndex];
2265 rx_peerHashTable[hashIndex] = pp;
2266 rxi_InitPeerParams(pp);
2267 MUTEX_ENTER(&rx_stats_mutex);
2268 rx_stats.nPeerStructs++;
2269 MUTEX_EXIT(&rx_stats_mutex);
2276 origPeer->refCount--;
2277 MUTEX_EXIT(&rx_peerHashTable_lock);
2282 /* Find the connection at (host, port) started at epoch, and with the
2283 * given connection id. Creates the server connection if necessary.
2284 * The type specifies whether a client connection or a server
2285 * connection is desired. In both cases, (host, port) specify the
2286 * peer's (host, pair) pair. Client connections are not made
2287 * automatically by this routine. The parameter socket gives the
2288 * socket descriptor on which the packet was received. This is used,
2289 * in the case of server connections, to check that *new* connections
2290 * come via a valid (port, serviceId). Finally, the securityIndex
2291 * parameter must match the existing index for the connection. If a
2292 * server connection is created, it will be created using the supplied
2293 * index, if the index is valid for this service */
2294 struct rx_connection *
2295 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2296 register u_short port, u_short serviceId, afs_uint32 cid,
2297 afs_uint32 epoch, int type, u_int securityIndex)
2299 int hashindex, flag;
2300 register struct rx_connection *conn;
2301 hashindex = CONN_HASH(host, port, cid, epoch, type);
2302 MUTEX_ENTER(&rx_connHashTable_lock);
2303 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2304 rx_connHashTable[hashindex],
2307 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2308 && (epoch == conn->epoch)) {
2309 register struct rx_peer *pp = conn->peer;
2310 if (securityIndex != conn->securityIndex) {
2311 /* this isn't supposed to happen, but someone could forge a packet
2312 * like this, and there seems to be some CM bug that makes this
2313 * happen from time to time -- in which case, the fileserver
2315 MUTEX_EXIT(&rx_connHashTable_lock);
2316 return (struct rx_connection *)0;
2318 if (pp->host == host && pp->port == port)
2320 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2322 /* So what happens when it's a callback connection? */
2323 if ( /*type == RX_CLIENT_CONNECTION && */
2324 (conn->epoch & 0x80000000))
2328 /* the connection rxLastConn that was used the last time is not the
2329 ** one we are looking for now. Hence, start searching in the hash */
2331 conn = rx_connHashTable[hashindex];
2336 struct rx_service *service;
2337 if (type == RX_CLIENT_CONNECTION) {
2338 MUTEX_EXIT(&rx_connHashTable_lock);
2339 return (struct rx_connection *)0;
2341 service = rxi_FindService(socket, serviceId);
2342 if (!service || (securityIndex >= service->nSecurityObjects)
2343 || (service->securityObjects[securityIndex] == 0)) {
2344 MUTEX_EXIT(&rx_connHashTable_lock);
2345 return (struct rx_connection *)0;
2347 conn = rxi_AllocConnection(); /* This bzero's the connection */
2348 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2349 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2350 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2351 conn->next = rx_connHashTable[hashindex];
2352 rx_connHashTable[hashindex] = conn;
2353 conn->peer = rxi_FindPeer(host, port, 0, 1);
2354 conn->type = RX_SERVER_CONNECTION;
2355 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2356 conn->epoch = epoch;
2357 conn->cid = cid & RX_CIDMASK;
2358 /* conn->serial = conn->lastSerial = 0; */
2359 /* conn->timeout = 0; */
2360 conn->ackRate = RX_FAST_ACK_RATE;
2361 conn->service = service;
2362 conn->serviceId = serviceId;
2363 conn->securityIndex = securityIndex;
2364 conn->securityObject = service->securityObjects[securityIndex];
2365 conn->nSpecific = 0;
2366 conn->specific = NULL;
2367 rx_SetConnDeadTime(conn, service->connDeadTime);
2368 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2369 /* Notify security object of the new connection */
2370 RXS_NewConnection(conn->securityObject, conn);
2371 /* XXXX Connection timeout? */
2372 if (service->newConnProc)
2373 (*service->newConnProc) (conn);
2374 MUTEX_ENTER(&rx_stats_mutex);
2375 rx_stats.nServerConns++;
2376 MUTEX_EXIT(&rx_stats_mutex);
2379 MUTEX_ENTER(&conn->conn_data_lock);
2381 MUTEX_EXIT(&conn->conn_data_lock);
2383 rxLastConn = conn; /* store this connection as the last conn used */
2384 MUTEX_EXIT(&rx_connHashTable_lock);
2388 /* There are two packet tracing routines available for testing and monitoring
2389 * Rx. One is called just after every packet is received and the other is
2390 * called just before every packet is sent. Received packets, have had their
2391 * headers decoded, and packets to be sent have not yet had their headers
2392 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2393 * containing the network address. Both can be modified. The return value, if
2394 * non-zero, indicates that the packet should be dropped. */
2396 int (*rx_justReceived) () = 0;
2397 int (*rx_almostSent) () = 0;
2399 /* A packet has been received off the interface. Np is the packet, socket is
2400 * the socket number it was received from (useful in determining which service
2401 * this packet corresponds to), and (host, port) reflect the host,port of the
2402 * sender. This call returns the packet to the caller if it is finished with
2403 * it, rather than de-allocating it, just as a small performance hack */
2406 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2407 afs_uint32 host, u_short port, int *tnop,
2408 struct rx_call **newcallp)
2410 register struct rx_call *call;
2411 register struct rx_connection *conn;
2413 afs_uint32 currentCallNumber;
2419 struct rx_packet *tnp;
2422 /* We don't print out the packet until now because (1) the time may not be
2423 * accurate enough until now in the lwp implementation (rx_Listener only gets
2424 * the time after the packet is read) and (2) from a protocol point of view,
2425 * this is the first time the packet has been seen */
2426 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2427 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2428 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2429 np->header.serial, packetType, host, port, np->header.serviceId,
2430 np->header.epoch, np->header.cid, np->header.callNumber,
2431 np->header.seq, np->header.flags, np));
2434 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2435 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2438 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2439 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2442 /* If an input tracer function is defined, call it with the packet and
2443 * network address. Note this function may modify its arguments. */
2444 if (rx_justReceived) {
2445 struct sockaddr_in addr;
2447 addr.sin_family = AF_INET;
2448 addr.sin_port = port;
2449 addr.sin_addr.s_addr = host;
2450 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2451 addr.sin_len = sizeof(addr);
2452 #endif /* AFS_OSF_ENV */
2453 drop = (*rx_justReceived) (np, &addr);
2454 /* drop packet if return value is non-zero */
2457 port = addr.sin_port; /* in case fcn changed addr */
2458 host = addr.sin_addr.s_addr;
2462 /* If packet was not sent by the client, then *we* must be the client */
2463 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2464 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2466 /* Find the connection (or fabricate one, if we're the server & if
2467 * necessary) associated with this packet */
2469 rxi_FindConnection(socket, host, port, np->header.serviceId,
2470 np->header.cid, np->header.epoch, type,
2471 np->header.securityIndex);
2474 /* If no connection found or fabricated, just ignore the packet.
2475 * (An argument could be made for sending an abort packet for
2480 MUTEX_ENTER(&conn->conn_data_lock);
2481 if (conn->maxSerial < np->header.serial)
2482 conn->maxSerial = np->header.serial;
2483 MUTEX_EXIT(&conn->conn_data_lock);
2485 /* If the connection is in an error state, send an abort packet and ignore
2486 * the incoming packet */
2488 /* Don't respond to an abort packet--we don't want loops! */
2489 MUTEX_ENTER(&conn->conn_data_lock);
2490 if (np->header.type != RX_PACKET_TYPE_ABORT)
2491 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2493 MUTEX_EXIT(&conn->conn_data_lock);
2497 /* Check for connection-only requests (i.e. not call specific). */
2498 if (np->header.callNumber == 0) {
2499 switch (np->header.type) {
2500 case RX_PACKET_TYPE_ABORT:
2501 /* What if the supplied error is zero? */
2502 rxi_ConnectionError(conn, ntohl(rx_GetInt32(np, 0)));
2503 MUTEX_ENTER(&conn->conn_data_lock);
2505 MUTEX_EXIT(&conn->conn_data_lock);
2507 case RX_PACKET_TYPE_CHALLENGE:
2508 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2509 MUTEX_ENTER(&conn->conn_data_lock);
2511 MUTEX_EXIT(&conn->conn_data_lock);
2513 case RX_PACKET_TYPE_RESPONSE:
2514 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2515 MUTEX_ENTER(&conn->conn_data_lock);
2517 MUTEX_EXIT(&conn->conn_data_lock);
2519 case RX_PACKET_TYPE_PARAMS:
2520 case RX_PACKET_TYPE_PARAMS + 1:
2521 case RX_PACKET_TYPE_PARAMS + 2:
2522 /* ignore these packet types for now */
2523 MUTEX_ENTER(&conn->conn_data_lock);
2525 MUTEX_EXIT(&conn->conn_data_lock);
2530 /* Should not reach here, unless the peer is broken: send an
2532 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2533 MUTEX_ENTER(&conn->conn_data_lock);
2534 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2536 MUTEX_EXIT(&conn->conn_data_lock);
2541 channel = np->header.cid & RX_CHANNELMASK;
2542 call = conn->call[channel];
2543 #ifdef RX_ENABLE_LOCKS
2545 MUTEX_ENTER(&call->lock);
2546 /* Test to see if call struct is still attached to conn. */
2547 if (call != conn->call[channel]) {
2549 MUTEX_EXIT(&call->lock);
2550 if (type == RX_SERVER_CONNECTION) {
2551 call = conn->call[channel];
2552 /* If we started with no call attached and there is one now,
2553 * another thread is also running this routine and has gotten
2554 * the connection channel. We should drop this packet in the tests
2555 * below. If there was a call on this connection and it's now
2556 * gone, then we'll be making a new call below.
2557 * If there was previously a call and it's now different then
2558 * the old call was freed and another thread running this routine
2559 * has created a call on this channel. One of these two threads
2560 * has a packet for the old call and the code below handles those
2564 MUTEX_ENTER(&call->lock);
2566 /* This packet can't be for this call. If the new call address is
2567 * 0 then no call is running on this channel. If there is a call
2568 * then, since this is a client connection we're getting data for
2569 * it must be for the previous call.
2571 MUTEX_ENTER(&rx_stats_mutex);
2572 rx_stats.spuriousPacketsRead++;
2573 MUTEX_EXIT(&rx_stats_mutex);
2574 MUTEX_ENTER(&conn->conn_data_lock);
2576 MUTEX_EXIT(&conn->conn_data_lock);
2581 currentCallNumber = conn->callNumber[channel];
2583 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2584 if (np->header.callNumber < currentCallNumber) {
2585 MUTEX_ENTER(&rx_stats_mutex);
2586 rx_stats.spuriousPacketsRead++;
2587 MUTEX_EXIT(&rx_stats_mutex);
2588 #ifdef RX_ENABLE_LOCKS
2590 MUTEX_EXIT(&call->lock);
2592 MUTEX_ENTER(&conn->conn_data_lock);
2594 MUTEX_EXIT(&conn->conn_data_lock);
2598 MUTEX_ENTER(&conn->conn_call_lock);
2599 call = rxi_NewCall(conn, channel);
2600 MUTEX_EXIT(&conn->conn_call_lock);
2601 *call->callNumber = np->header.callNumber;
2602 call->state = RX_STATE_PRECALL;
2603 clock_GetTime(&call->queueTime);
2604 hzero(call->bytesSent);
2605 hzero(call->bytesRcvd);
2606 rxi_KeepAliveOn(call);
2607 } else if (np->header.callNumber != currentCallNumber) {
2608 /* Wait until the transmit queue is idle before deciding
2609 * whether to reset the current call. Chances are that the
2610 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2613 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2614 while ((call->state == RX_STATE_ACTIVE)
2615 && (call->flags & RX_CALL_TQ_BUSY)) {
2616 call->flags |= RX_CALL_TQ_WAIT;
2617 #ifdef RX_ENABLE_LOCKS
2618 CV_WAIT(&call->cv_tq, &call->lock);
2619 #else /* RX_ENABLE_LOCKS */
2620 osi_rxSleep(&call->tq);
2621 #endif /* RX_ENABLE_LOCKS */
2623 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2624 /* If the new call cannot be taken right now send a busy and set
2625 * the error condition in this call, so that it terminates as
2626 * quickly as possible */
2627 if (call->state == RX_STATE_ACTIVE) {
2628 struct rx_packet *tp;
2630 rxi_CallError(call, RX_CALL_DEAD);
2631 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2633 MUTEX_EXIT(&call->lock);
2634 MUTEX_ENTER(&conn->conn_data_lock);
2636 MUTEX_EXIT(&conn->conn_data_lock);
2639 rxi_ResetCall(call, 0);
2640 *call->callNumber = np->header.callNumber;
2641 call->state = RX_STATE_PRECALL;
2642 clock_GetTime(&call->queueTime);
2643 hzero(call->bytesSent);
2644 hzero(call->bytesRcvd);
2646 * If the number of queued calls exceeds the overload
2647 * threshold then abort this call.
2649 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2650 struct rx_packet *tp;
2652 rxi_CallError(call, rx_BusyError);
2653 tp = rxi_SendCallAbort(call, np, 1, 0);
2654 MUTEX_EXIT(&call->lock);
2655 MUTEX_ENTER(&conn->conn_data_lock);
2657 MUTEX_EXIT(&conn->conn_data_lock);
2658 MUTEX_ENTER(&rx_stats_mutex);
2660 MUTEX_EXIT(&rx_stats_mutex);
2663 rxi_KeepAliveOn(call);
2665 /* Continuing call; do nothing here. */
2667 } else { /* we're the client */
2668 /* Ignore all incoming acknowledgements for calls in DALLY state */
2669 if (call && (call->state == RX_STATE_DALLY)
2670 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2671 MUTEX_ENTER(&rx_stats_mutex);
2672 rx_stats.ignorePacketDally++;
2673 MUTEX_EXIT(&rx_stats_mutex);
2674 #ifdef RX_ENABLE_LOCKS
2676 MUTEX_EXIT(&call->lock);
2679 MUTEX_ENTER(&conn->conn_data_lock);
2681 MUTEX_EXIT(&conn->conn_data_lock);
2685 /* Ignore anything that's not relevant to the current call. If there
2686 * isn't a current call, then no packet is relevant. */
2687 if (!call || (np->header.callNumber != currentCallNumber)) {
2688 MUTEX_ENTER(&rx_stats_mutex);
2689 rx_stats.spuriousPacketsRead++;
2690 MUTEX_EXIT(&rx_stats_mutex);
2691 #ifdef RX_ENABLE_LOCKS
2693 MUTEX_EXIT(&call->lock);
2696 MUTEX_ENTER(&conn->conn_data_lock);
2698 MUTEX_EXIT(&conn->conn_data_lock);
2701 /* If the service security object index stamped in the packet does not
2702 * match the connection's security index, ignore the packet */
2703 if (np->header.securityIndex != conn->securityIndex) {
2704 #ifdef RX_ENABLE_LOCKS
2705 MUTEX_EXIT(&call->lock);
2707 MUTEX_ENTER(&conn->conn_data_lock);
2709 MUTEX_EXIT(&conn->conn_data_lock);
2713 /* If we're receiving the response, then all transmit packets are
2714 * implicitly acknowledged. Get rid of them. */
2715 if (np->header.type == RX_PACKET_TYPE_DATA) {
2716 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2717 /* XXX Hack. Because we must release the global rx lock when
2718 * sending packets (osi_NetSend) we drop all acks while we're
2719 * traversing the tq in rxi_Start sending packets out because
2720 * packets may move to the freePacketQueue as result of being here!
2721 * So we drop these packets until we're safely out of the
2722 * traversing. Really ugly!
2723 * For fine grain RX locking, we set the acked field in the
2724 * packets and let rxi_Start remove them from the transmit queue.
2726 if (call->flags & RX_CALL_TQ_BUSY) {
2727 #ifdef RX_ENABLE_LOCKS
2728 rxi_SetAcksInTransmitQueue(call);
2731 return np; /* xmitting; drop packet */
2734 rxi_ClearTransmitQueue(call, 0);
2736 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2737 rxi_ClearTransmitQueue(call, 0);
2738 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2740 if (np->header.type == RX_PACKET_TYPE_ACK) {
2741 /* now check to see if this is an ack packet acknowledging that the
2742 * server actually *lost* some hard-acked data. If this happens we
2743 * ignore this packet, as it may indicate that the server restarted in
2744 * the middle of a call. It is also possible that this is an old ack
2745 * packet. We don't abort the connection in this case, because this
2746 * *might* just be an old ack packet. The right way to detect a server
2747 * restart in the midst of a call is to notice that the server epoch
2749 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2750 * XXX unacknowledged. I think that this is off-by-one, but
2751 * XXX I don't dare change it just yet, since it will
2752 * XXX interact badly with the server-restart detection
2753 * XXX code in receiveackpacket. */
2754 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2755 MUTEX_ENTER(&rx_stats_mutex);
2756 rx_stats.spuriousPacketsRead++;
2757 MUTEX_EXIT(&rx_stats_mutex);
2758 MUTEX_EXIT(&call->lock);
2759 MUTEX_ENTER(&conn->conn_data_lock);
2761 MUTEX_EXIT(&conn->conn_data_lock);
2765 } /* else not a data packet */
2768 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2769 /* Set remote user defined status from packet */
2770 call->remoteStatus = np->header.userStatus;
2772 /* Note the gap between the expected next packet and the actual
2773 * packet that arrived, when the new packet has a smaller serial number
2774 * than expected. Rioses frequently reorder packets all by themselves,
2775 * so this will be quite important with very large window sizes.
2776 * Skew is checked against 0 here to avoid any dependence on the type of
2777 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2779 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2780 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2781 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2783 MUTEX_ENTER(&conn->conn_data_lock);
2784 skew = conn->lastSerial - np->header.serial;
2785 conn->lastSerial = np->header.serial;
2786 MUTEX_EXIT(&conn->conn_data_lock);
2788 register struct rx_peer *peer;
2790 if (skew > peer->inPacketSkew) {
2791 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2793 peer->inPacketSkew = skew;
2797 /* Now do packet type-specific processing */
2798 switch (np->header.type) {
2799 case RX_PACKET_TYPE_DATA:
2800 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2803 case RX_PACKET_TYPE_ACK:
2804 /* Respond immediately to ack packets requesting acknowledgement
2806 if (np->header.flags & RX_REQUEST_ACK) {
2808 (void)rxi_SendCallAbort(call, 0, 1, 0);
2810 (void)rxi_SendAck(call, 0, np->header.serial,
2811 RX_ACK_PING_RESPONSE, 1);
2813 np = rxi_ReceiveAckPacket(call, np, 1);
2815 case RX_PACKET_TYPE_ABORT:
2816 /* An abort packet: reset the connection, passing the error up to
2818 /* What if error is zero? */
2819 rxi_CallError(call, ntohl(*(afs_int32 *) rx_DataOf(np)));
2821 case RX_PACKET_TYPE_BUSY:
2824 case RX_PACKET_TYPE_ACKALL:
2825 /* All packets acknowledged, so we can drop all packets previously
2826 * readied for sending */
2827 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2828 /* XXX Hack. We because we can't release the global rx lock when
2829 * sending packets (osi_NetSend) we drop all ack pkts while we're
2830 * traversing the tq in rxi_Start sending packets out because
2831 * packets may move to the freePacketQueue as result of being
2832 * here! So we drop these packets until we're safely out of the
2833 * traversing. Really ugly!
2834 * For fine grain RX locking, we set the acked field in the packets
2835 * and let rxi_Start remove the packets from the transmit queue.
2837 if (call->flags & RX_CALL_TQ_BUSY) {
2838 #ifdef RX_ENABLE_LOCKS
2839 rxi_SetAcksInTransmitQueue(call);
2841 #else /* RX_ENABLE_LOCKS */
2843 return np; /* xmitting; drop packet */
2844 #endif /* RX_ENABLE_LOCKS */
2846 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2847 rxi_ClearTransmitQueue(call, 0);
2850 /* Should not reach here, unless the peer is broken: send an abort
2852 rxi_CallError(call, RX_PROTOCOL_ERROR);
2853 np = rxi_SendCallAbort(call, np, 1, 0);
2856 /* Note when this last legitimate packet was received, for keep-alive
2857 * processing. Note, we delay getting the time until now in the hope that
2858 * the packet will be delivered to the user before any get time is required
2859 * (if not, then the time won't actually be re-evaluated here). */
2860 call->lastReceiveTime = clock_Sec();
2861 MUTEX_EXIT(&call->lock);
2862 MUTEX_ENTER(&conn->conn_data_lock);
2864 MUTEX_EXIT(&conn->conn_data_lock);
2868 /* return true if this is an "interesting" connection from the point of view
2869 of someone trying to debug the system */
2871 rxi_IsConnInteresting(struct rx_connection *aconn)
2874 register struct rx_call *tcall;
2876 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2878 for (i = 0; i < RX_MAXCALLS; i++) {
2879 tcall = aconn->call[i];
2881 if ((tcall->state == RX_STATE_PRECALL)
2882 || (tcall->state == RX_STATE_ACTIVE))
2884 if ((tcall->mode == RX_MODE_SENDING)
2885 || (tcall->mode == RX_MODE_RECEIVING))
2893 /* if this is one of the last few packets AND it wouldn't be used by the
2894 receiving call to immediately satisfy a read request, then drop it on
2895 the floor, since accepting it might prevent a lock-holding thread from
2896 making progress in its reading. If a call has been cleared while in
2897 the precall state then ignore all subsequent packets until the call
2898 is assigned to a thread. */
2901 TooLow(struct rx_packet *ap, struct rx_call *acall)
2904 MUTEX_ENTER(&rx_stats_mutex);
2905 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
2906 && (acall->state == RX_STATE_PRECALL))
2907 || ((rx_nFreePackets < rxi_dataQuota + 2)
2908 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
2909 && (acall->flags & RX_CALL_READER_WAIT)))) {
2912 MUTEX_EXIT(&rx_stats_mutex);
2918 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
2919 struct rx_call *acall)
2921 struct rx_call *call = acall;
2925 MUTEX_ENTER(&conn->conn_data_lock);
2926 conn->checkReachEvent = NULL;
2927 waiting = conn->flags & RX_CONN_ATTACHWAIT;
2930 MUTEX_EXIT(&conn->conn_data_lock);
2934 MUTEX_ENTER(&conn->conn_call_lock);
2935 MUTEX_ENTER(&conn->conn_data_lock);
2936 for (i = 0; i < RX_MAXCALLS; i++) {
2937 struct rx_call *tc = conn->call[i];
2938 if (tc && tc->state == RX_STATE_PRECALL) {
2944 /* Indicate that rxi_CheckReachEvent is no longer running by
2945 * clearing the flag. Must be atomic under conn_data_lock to
2946 * avoid a new call slipping by: rxi_CheckConnReach holds
2947 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
2949 conn->flags &= ~RX_CONN_ATTACHWAIT;
2950 MUTEX_EXIT(&conn->conn_data_lock);
2951 MUTEX_EXIT(&conn->conn_call_lock);
2956 MUTEX_ENTER(&call->lock);
2957 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
2959 MUTEX_EXIT(&call->lock);
2961 clock_GetTime(&when);
2962 when.sec += RX_CHECKREACH_TIMEOUT;
2963 MUTEX_ENTER(&conn->conn_data_lock);
2964 if (!conn->checkReachEvent) {
2966 conn->checkReachEvent =
2967 rxevent_Post(&when, rxi_CheckReachEvent, conn, NULL);
2969 MUTEX_EXIT(&conn->conn_data_lock);
2975 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
2977 struct rx_service *service = conn->service;
2978 struct rx_peer *peer = conn->peer;
2979 afs_uint32 now, lastReach;
2981 if (service->checkReach == 0)
2985 MUTEX_ENTER(&peer->peer_lock);
2986 lastReach = peer->lastReachTime;
2987 MUTEX_EXIT(&peer->peer_lock);
2988 if (now - lastReach < RX_CHECKREACH_TTL)
2991 MUTEX_ENTER(&conn->conn_data_lock);
2992 if (conn->flags & RX_CONN_ATTACHWAIT) {
2993 MUTEX_EXIT(&conn->conn_data_lock);
2996 conn->flags |= RX_CONN_ATTACHWAIT;
2997 MUTEX_EXIT(&conn->conn_data_lock);
2998 if (!conn->checkReachEvent)
2999 rxi_CheckReachEvent(NULL, conn, call);
3004 /* try to attach call, if authentication is complete */
3006 TryAttach(register struct rx_call *acall, register osi_socket socket,
3007 register int *tnop, register struct rx_call **newcallp,
3010 struct rx_connection *conn = acall->conn;
3012 if (conn->type == RX_SERVER_CONNECTION
3013 && acall->state == RX_STATE_PRECALL) {
3014 /* Don't attach until we have any req'd. authentication. */
3015 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3016 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3017 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3018 /* Note: this does not necessarily succeed; there
3019 * may not any proc available
3022 rxi_ChallengeOn(acall->conn);
3027 /* A data packet has been received off the interface. This packet is
3028 * appropriate to the call (the call is in the right state, etc.). This
3029 * routine can return a packet to the caller, for re-use */
3032 rxi_ReceiveDataPacket(register struct rx_call *call,
3033 register struct rx_packet *np, int istack,
3034 osi_socket socket, afs_uint32 host, u_short port,
3035 int *tnop, struct rx_call **newcallp)
3037 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3041 afs_uint32 seq, serial, flags;
3043 struct rx_packet *tnp;
3045 MUTEX_ENTER(&rx_stats_mutex);
3046 rx_stats.dataPacketsRead++;
3047 MUTEX_EXIT(&rx_stats_mutex);
3050 /* If there are no packet buffers, drop this new packet, unless we can find
3051 * packet buffers from inactive calls */
3053 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3054 MUTEX_ENTER(&rx_freePktQ_lock);
3055 rxi_NeedMorePackets = TRUE;
3056 MUTEX_EXIT(&rx_freePktQ_lock);
3057 MUTEX_ENTER(&rx_stats_mutex);
3058 rx_stats.noPacketBuffersOnRead++;
3059 MUTEX_EXIT(&rx_stats_mutex);
3060 call->rprev = np->header.serial;
3061 rxi_calltrace(RX_TRACE_DROP, call);
3062 dpf(("packet %x dropped on receipt - quota problems", np));
3064 rxi_ClearReceiveQueue(call);
3065 clock_GetTime(&when);
3066 clock_Add(&when, &rx_softAckDelay);
3067 if (!call->delayedAckEvent
3068 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3069 rxevent_Cancel(call->delayedAckEvent, call,
3070 RX_CALL_REFCOUNT_DELAY);
3071 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3072 call->delayedAckEvent =
3073 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3075 /* we've damaged this call already, might as well do it in. */
3081 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3082 * packet is one of several packets transmitted as a single
3083 * datagram. Do not send any soft or hard acks until all packets
3084 * in a jumbogram have been processed. Send negative acks right away.
3086 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3087 /* tnp is non-null when there are more packets in the
3088 * current jumbo gram */
3095 seq = np->header.seq;
3096 serial = np->header.serial;
3097 flags = np->header.flags;
3099 /* If the call is in an error state, send an abort message */
3101 return rxi_SendCallAbort(call, np, istack, 0);
3103 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3104 * AFS 3.5 jumbogram. */
3105 if (flags & RX_JUMBO_PACKET) {
3106 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3111 if (np->header.spare != 0) {
3112 MUTEX_ENTER(&call->conn->conn_data_lock);
3113 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3114 MUTEX_EXIT(&call->conn->conn_data_lock);
3117 /* The usual case is that this is the expected next packet */
3118 if (seq == call->rnext) {
3120 /* Check to make sure it is not a duplicate of one already queued */
3121 if (queue_IsNotEmpty(&call->rq)
3122 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3123 MUTEX_ENTER(&rx_stats_mutex);
3124 rx_stats.dupPacketsRead++;
3125 MUTEX_EXIT(&rx_stats_mutex);
3126 dpf(("packet %x dropped on receipt - duplicate", np));
3127 rxevent_Cancel(call->delayedAckEvent, call,
3128 RX_CALL_REFCOUNT_DELAY);
3129 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3135 /* It's the next packet. Stick it on the receive queue
3136 * for this call. Set newPackets to make sure we wake
3137 * the reader once all packets have been processed */
3138 queue_Prepend(&call->rq, np);
3140 np = NULL; /* We can't use this anymore */
3143 /* If an ack is requested then set a flag to make sure we
3144 * send an acknowledgement for this packet */
3145 if (flags & RX_REQUEST_ACK) {
3146 ackNeeded = RX_ACK_REQUESTED;
3149 /* Keep track of whether we have received the last packet */
3150 if (flags & RX_LAST_PACKET) {
3151 call->flags |= RX_CALL_HAVE_LAST;
3155 /* Check whether we have all of the packets for this call */
3156 if (call->flags & RX_CALL_HAVE_LAST) {
3157 afs_uint32 tseq; /* temporary sequence number */
3158 struct rx_packet *tp; /* Temporary packet pointer */
3159 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3161 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3162 if (tseq != tp->header.seq)
3164 if (tp->header.flags & RX_LAST_PACKET) {
3165 call->flags |= RX_CALL_RECEIVE_DONE;
3172 /* Provide asynchronous notification for those who want it
3173 * (e.g. multi rx) */
3174 if (call->arrivalProc) {
3175 (*call->arrivalProc) (call, call->arrivalProcHandle,
3176 call->arrivalProcArg);
3177 call->arrivalProc = (void (*)())0;
3180 /* Update last packet received */
3183 /* If there is no server process serving this call, grab
3184 * one, if available. We only need to do this once. If a
3185 * server thread is available, this thread becomes a server
3186 * thread and the server thread becomes a listener thread. */
3188 TryAttach(call, socket, tnop, newcallp, 0);
3191 /* This is not the expected next packet. */
3193 /* Determine whether this is a new or old packet, and if it's
3194 * a new one, whether it fits into the current receive window.
3195 * Also figure out whether the packet was delivered in sequence.
3196 * We use the prev variable to determine whether the new packet
3197 * is the successor of its immediate predecessor in the
3198 * receive queue, and the missing flag to determine whether
3199 * any of this packets predecessors are missing. */
3201 afs_uint32 prev; /* "Previous packet" sequence number */
3202 struct rx_packet *tp; /* Temporary packet pointer */
3203 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3204 int missing; /* Are any predecessors missing? */
3206 /* If the new packet's sequence number has been sent to the
3207 * application already, then this is a duplicate */
3208 if (seq < call->rnext) {
3209 MUTEX_ENTER(&rx_stats_mutex);
3210 rx_stats.dupPacketsRead++;
3211 MUTEX_EXIT(&rx_stats_mutex);
3212 rxevent_Cancel(call->delayedAckEvent, call,
3213 RX_CALL_REFCOUNT_DELAY);
3214 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3220 /* If the sequence number is greater than what can be
3221 * accomodated by the current window, then send a negative
3222 * acknowledge and drop the packet */
3223 if ((call->rnext + call->rwind) <= seq) {
3224 rxevent_Cancel(call->delayedAckEvent, call,
3225 RX_CALL_REFCOUNT_DELAY);
3226 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3233 /* Look for the packet in the queue of old received packets */
3234 for (prev = call->rnext - 1, missing =
3235 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3236 /*Check for duplicate packet */
3237 if (seq == tp->header.seq) {
3238 MUTEX_ENTER(&rx_stats_mutex);
3239 rx_stats.dupPacketsRead++;
3240 MUTEX_EXIT(&rx_stats_mutex);
3241 rxevent_Cancel(call->delayedAckEvent, call,
3242 RX_CALL_REFCOUNT_DELAY);
3243 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3249 /* If we find a higher sequence packet, break out and
3250 * insert the new packet here. */
3251 if (seq < tp->header.seq)
3253 /* Check for missing packet */
3254 if (tp->header.seq != prev + 1) {
3258 prev = tp->header.seq;
3261 /* Keep track of whether we have received the last packet. */
3262 if (flags & RX_LAST_PACKET) {
3263 call->flags |= RX_CALL_HAVE_LAST;
3266 /* It's within the window: add it to the the receive queue.
3267 * tp is left by the previous loop either pointing at the
3268 * packet before which to insert the new packet, or at the
3269 * queue head if the queue is empty or the packet should be
3271 queue_InsertBefore(tp, np);
3275 /* Check whether we have all of the packets for this call */
3276 if ((call->flags & RX_CALL_HAVE_LAST)
3277 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3278 afs_uint32 tseq; /* temporary sequence number */
3281 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3282 if (tseq != tp->header.seq)
3284 if (tp->header.flags & RX_LAST_PACKET) {
3285 call->flags |= RX_CALL_RECEIVE_DONE;
3292 /* We need to send an ack of the packet is out of sequence,
3293 * or if an ack was requested by the peer. */
3294 if (seq != prev + 1 || missing || (flags & RX_REQUEST_ACK)) {
3295 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3298 /* Acknowledge the last packet for each call */
3299 if (flags & RX_LAST_PACKET) {
3310 * If the receiver is waiting for an iovec, fill the iovec
3311 * using the data from the receive queue */
3312 if (call->flags & RX_CALL_IOVEC_WAIT) {
3313 didHardAck = rxi_FillReadVec(call, serial);
3314 /* the call may have been aborted */
3323 /* Wakeup the reader if any */
3324 if ((call->flags & RX_CALL_READER_WAIT)
3325 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3326 || (call->iovNext >= call->iovMax)
3327 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3328 call->flags &= ~RX_CALL_READER_WAIT;
3329 #ifdef RX_ENABLE_LOCKS
3330 CV_BROADCAST(&call->cv_rq);
3332 osi_rxWakeup(&call->rq);
3338 * Send an ack when requested by the peer, or once every
3339 * rxi_SoftAckRate packets until the last packet has been
3340 * received. Always send a soft ack for the last packet in
3341 * the server's reply. */
3343 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3344 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3345 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3346 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3347 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3348 } else if (call->nSoftAcks) {
3349 clock_GetTime(&when);
3350 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3351 clock_Add(&when, &rx_lastAckDelay);
3353 clock_Add(&when, &rx_softAckDelay);
3355 if (!call->delayedAckEvent
3356 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3357 rxevent_Cancel(call->delayedAckEvent, call,
3358 RX_CALL_REFCOUNT_DELAY);
3359 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3360 call->delayedAckEvent =
3361 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3363 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3364 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3371 static void rxi_ComputeRate();
3375 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3377 struct rx_peer *peer = conn->peer;
3379 MUTEX_ENTER(&peer->peer_lock);
3380 peer->lastReachTime = clock_Sec();
3381 MUTEX_EXIT(&peer->peer_lock);
3383 MUTEX_ENTER(&conn->conn_data_lock);
3384 if (conn->flags & RX_CONN_ATTACHWAIT) {
3387 conn->flags &= ~RX_CONN_ATTACHWAIT;
3388 MUTEX_EXIT(&conn->conn_data_lock);
3390 for (i = 0; i < RX_MAXCALLS; i++) {
3391 struct rx_call *call = conn->call[i];
3394 MUTEX_ENTER(&call->lock);
3395 /* tnop can be null if newcallp is null */
3396 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3398 MUTEX_EXIT(&call->lock);
3402 MUTEX_EXIT(&conn->conn_data_lock);
3405 /* rxi_ComputePeerNetStats
3407 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3408 * estimates (like RTT and throughput) based on ack packets. Caller
3409 * must ensure that the packet in question is the right one (i.e.
3410 * serial number matches).
3413 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3414 struct rx_ackPacket *ap, struct rx_packet *np)
3416 struct rx_peer *peer = call->conn->peer;
3418 /* Use RTT if not delayed by client. */
3419 if (ap->reason != RX_ACK_DELAY)
3420 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3422 rxi_ComputeRate(peer, call, p, np, ap->reason);
3426 /* The real smarts of the whole thing. */
3428 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3431 struct rx_ackPacket *ap;
3433 register struct rx_packet *tp;
3434 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3435 register struct rx_connection *conn = call->conn;
3436 struct rx_peer *peer = conn->peer;
3439 /* because there are CM's that are bogus, sending weird values for this. */
3440 afs_uint32 skew = 0;
3445 int newAckCount = 0;
3446 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3447 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3449 MUTEX_ENTER(&rx_stats_mutex);
3450 rx_stats.ackPacketsRead++;
3451 MUTEX_EXIT(&rx_stats_mutex);
3452 ap = (struct rx_ackPacket *)rx_DataOf(np);
3453 nbytes = rx_Contiguous(np) - ((ap->acks) - (u_char *) ap);
3455 return np; /* truncated ack packet */
3457 /* depends on ack packet struct */
3458 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3459 first = ntohl(ap->firstPacket);
3460 serial = ntohl(ap->serial);
3461 /* temporarily disabled -- needs to degrade over time
3462 * skew = ntohs(ap->maxSkew); */
3464 /* Ignore ack packets received out of order */
3465 if (first < call->tfirst) {
3469 if (np->header.flags & RX_SLOW_START_OK) {
3470 call->flags |= RX_CALL_SLOW_START_OK;
3473 if (ap->reason == RX_ACK_PING_RESPONSE)
3474 rxi_UpdatePeerReach(conn, call);
3479 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3480 ap->reason, ntohl(ap->previousPacket),
3481 (unsigned int)np->header.seq, (unsigned int)serial,
3482 (unsigned int)skew, ntohl(ap->firstPacket));
3485 for (offset = 0; offset < nAcks; offset++)
3486 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3493 /* Update the outgoing packet skew value to the latest value of
3494 * the peer's incoming packet skew value. The ack packet, of
3495 * course, could arrive out of order, but that won't affect things
3497 MUTEX_ENTER(&peer->peer_lock);
3498 peer->outPacketSkew = skew;
3500 /* Check for packets that no longer need to be transmitted, and
3501 * discard them. This only applies to packets positively
3502 * acknowledged as having been sent to the peer's upper level.
3503 * All other packets must be retained. So only packets with
3504 * sequence numbers < ap->firstPacket are candidates. */
3505 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3506 if (tp->header.seq >= first)
3508 call->tfirst = tp->header.seq + 1;
3510 && (tp->header.serial == serial || tp->firstSerial == serial))
3511 rxi_ComputePeerNetStats(call, tp, ap, np);
3512 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3513 /* XXX Hack. Because we have to release the global rx lock when sending
3514 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3515 * in rxi_Start sending packets out because packets may move to the
3516 * freePacketQueue as result of being here! So we drop these packets until
3517 * we're safely out of the traversing. Really ugly!
3518 * To make it even uglier, if we're using fine grain locking, we can
3519 * set the ack bits in the packets and have rxi_Start remove the packets
3520 * when it's done transmitting.
3522 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3525 if (call->flags & RX_CALL_TQ_BUSY) {
3526 #ifdef RX_ENABLE_LOCKS
3527 tp->flags |= RX_PKTFLAG_ACKED;
3528 call->flags |= RX_CALL_TQ_SOME_ACKED;
3529 #else /* RX_ENABLE_LOCKS */
3531 #endif /* RX_ENABLE_LOCKS */
3533 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3536 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3541 /* Give rate detector a chance to respond to ping requests */
3542 if (ap->reason == RX_ACK_PING_RESPONSE) {
3543 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3547 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3549 /* Now go through explicit acks/nacks and record the results in
3550 * the waiting packets. These are packets that can't be released
3551 * yet, even with a positive acknowledge. This positive
3552 * acknowledge only means the packet has been received by the
3553 * peer, not that it will be retained long enough to be sent to
3554 * the peer's upper level. In addition, reset the transmit timers
3555 * of any missing packets (those packets that must be missing
3556 * because this packet was out of sequence) */
3558 call->nSoftAcked = 0;
3559 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3560 /* Update round trip time if the ack was stimulated on receipt
3562 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3563 #ifdef RX_ENABLE_LOCKS
3564 if (tp->header.seq >= first)
3565 #endif /* RX_ENABLE_LOCKS */
3566 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3568 && (tp->header.serial == serial || tp->firstSerial == serial))
3569 rxi_ComputePeerNetStats(call, tp, ap, np);
3571 /* Set the acknowledge flag per packet based on the
3572 * information in the ack packet. An acknowlegded packet can
3573 * be downgraded when the server has discarded a packet it
3574 * soacked previously, or when an ack packet is received
3575 * out of sequence. */
3576 if (tp->header.seq < first) {
3577 /* Implicit ack information */
3578 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3581 tp->flags |= RX_PKTFLAG_ACKED;
3582 } else if (tp->header.seq < first + nAcks) {
3583 /* Explicit ack information: set it in the packet appropriately */
3584 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3585 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3587 tp->flags |= RX_PKTFLAG_ACKED;
3595 tp->flags &= ~RX_PKTFLAG_ACKED;
3599 tp->flags &= ~RX_PKTFLAG_ACKED;
3603 /* If packet isn't yet acked, and it has been transmitted at least
3604 * once, reset retransmit time using latest timeout
3605 * ie, this should readjust the retransmit timer for all outstanding
3606 * packets... So we don't just retransmit when we should know better*/
3608 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3609 tp->retryTime = tp->timeSent;
3610 clock_Add(&tp->retryTime, &peer->timeout);
3611 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3612 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3616 /* If the window has been extended by this acknowledge packet,
3617 * then wakeup a sender waiting in alloc for window space, or try
3618 * sending packets now, if he's been sitting on packets due to
3619 * lack of window space */
3620 if (call->tnext < (call->tfirst + call->twind)) {
3621 #ifdef RX_ENABLE_LOCKS
3622 CV_SIGNAL(&call->cv_twind);
3624 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3625 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3626 osi_rxWakeup(&call->twind);
3629 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3630 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3634 /* if the ack packet has a receivelen field hanging off it,
3635 * update our state */
3636 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3639 /* If the ack packet has a "recommended" size that is less than
3640 * what I am using now, reduce my size to match */
3641 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3642 sizeof(afs_int32), &tSize);
3643 tSize = (afs_uint32) ntohl(tSize);
3644 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3646 /* Get the maximum packet size to send to this peer */
3647 rx_packetread(np, rx_AckDataSize(ap->nAcks), sizeof(afs_int32),
3649 tSize = (afs_uint32) ntohl(tSize);
3650 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3651 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3653 /* sanity check - peer might have restarted with different params.
3654 * If peer says "send less", dammit, send less... Peer should never
3655 * be unable to accept packets of the size that prior AFS versions would
3656 * send without asking. */
3657 if (peer->maxMTU != tSize) {
3658 peer->maxMTU = tSize;
3659 peer->MTU = MIN(tSize, peer->MTU);
3660 call->MTU = MIN(call->MTU, tSize);
3664 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3667 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3668 sizeof(afs_int32), &tSize);
3669 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3670 if (tSize < call->twind) { /* smaller than our send */
3671 call->twind = tSize; /* window, we must send less... */
3672 call->ssthresh = MIN(call->twind, call->ssthresh);
3675 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3676 * network MTU confused with the loopback MTU. Calculate the
3677 * maximum MTU here for use in the slow start code below.
3679 maxMTU = peer->maxMTU;
3680 /* Did peer restart with older RX version? */
3681 if (peer->maxDgramPackets > 1) {
3682 peer->maxDgramPackets = 1;
3684 } else if (np->length >=
3685 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3688 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3689 sizeof(afs_int32), &tSize);
3690 tSize = (afs_uint32) ntohl(tSize);
3692 * As of AFS 3.5 we set the send window to match the receive window.
3694 if (tSize < call->twind) {
3695 call->twind = tSize;
3696 call->ssthresh = MIN(call->twind, call->ssthresh);
3697 } else if (tSize > call->twind) {
3698 call->twind = tSize;
3702 * As of AFS 3.5, a jumbogram is more than one fixed size
3703 * packet transmitted in a single UDP datagram. If the remote
3704 * MTU is smaller than our local MTU then never send a datagram
3705 * larger than the natural MTU.
3708 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3709 sizeof(afs_int32), &tSize);
3710 maxDgramPackets = (afs_uint32) ntohl(tSize);
3711 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3713 MIN(maxDgramPackets, (int)(peer->ifDgramPackets));
3714 maxDgramPackets = MIN(maxDgramPackets, tSize);
3715 if (maxDgramPackets > 1) {
3716 peer->maxDgramPackets = maxDgramPackets;
3717 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3719 peer->maxDgramPackets = 1;
3720 call->MTU = peer->natMTU;
3722 } else if (peer->maxDgramPackets > 1) {
3723 /* Restarted with lower version of RX */
3724 peer->maxDgramPackets = 1;
3726 } else if (peer->maxDgramPackets > 1
3727 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3728 /* Restarted with lower version of RX */
3729 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3730 peer->natMTU = OLD_MAX_PACKET_SIZE;
3731 peer->MTU = OLD_MAX_PACKET_SIZE;
3732 peer->maxDgramPackets = 1;
3733 peer->nDgramPackets = 1;
3735 call->MTU = OLD_MAX_PACKET_SIZE;
3740 * Calculate how many datagrams were successfully received after
3741 * the first missing packet and adjust the negative ack counter
3746 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3747 if (call->nNacks < nNacked) {
3748 call->nNacks = nNacked;
3757 if (call->flags & RX_CALL_FAST_RECOVER) {
3759 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3761 call->flags &= ~RX_CALL_FAST_RECOVER;
3762 call->cwind = call->nextCwind;
3763 call->nextCwind = 0;
3766 call->nCwindAcks = 0;
3767 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3768 /* Three negative acks in a row trigger congestion recovery */
3769 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3770 MUTEX_EXIT(&peer->peer_lock);
3771 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3772 /* someone else is waiting to start recovery */
3775 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3776 while (call->flags & RX_CALL_TQ_BUSY) {
3777 call->flags |= RX_CALL_TQ_WAIT;
3778 #ifdef RX_ENABLE_LOCKS
3779 CV_WAIT(&call->cv_tq, &call->lock);
3780 #else /* RX_ENABLE_LOCKS */
3781 osi_rxSleep(&call->tq);
3782 #endif /* RX_ENABLE_LOCKS */
3784 MUTEX_ENTER(&peer->peer_lock);
3785 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3786 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3787 call->flags |= RX_CALL_FAST_RECOVER;
3788 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3790 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3791 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3792 call->nextCwind = call->ssthresh;
3795 peer->MTU = call->MTU;
3796 peer->cwind = call->nextCwind;
3797 peer->nDgramPackets = call->nDgramPackets;
3799 call->congestSeq = peer->congestSeq;
3800 /* Reset the resend times on the packets that were nacked
3801 * so we will retransmit as soon as the window permits*/
3802 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3804 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3805 clock_Zero(&tp->retryTime);
3807 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3812 /* If cwind is smaller than ssthresh, then increase
3813 * the window one packet for each ack we receive (exponential
3815 * If cwind is greater than or equal to ssthresh then increase
3816 * the congestion window by one packet for each cwind acks we
3817 * receive (linear growth). */
3818 if (call->cwind < call->ssthresh) {
3820 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3821 call->nCwindAcks = 0;
3823 call->nCwindAcks += newAckCount;
3824 if (call->nCwindAcks >= call->cwind) {
3825 call->nCwindAcks = 0;
3826 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3830 * If we have received several acknowledgements in a row then
3831 * it is time to increase the size of our datagrams
3833 if ((int)call->nAcks > rx_nDgramThreshold) {
3834 if (peer->maxDgramPackets > 1) {
3835 if (call->nDgramPackets < peer->maxDgramPackets) {
3836 call->nDgramPackets++;
3838 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3839 } else if (call->MTU < peer->maxMTU) {
3840 call->MTU += peer->natMTU;
3841 call->MTU = MIN(call->MTU, peer->maxMTU);
3847 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3849 /* Servers need to hold the call until all response packets have
3850 * been acknowledged. Soft acks are good enough since clients
3851 * are not allowed to clear their receive queues. */
3852 if (call->state == RX_STATE_HOLD
3853 && call->tfirst + call->nSoftAcked >= call->tnext) {
3854 call->state = RX_STATE_DALLY;
3855 rxi_ClearTransmitQueue(call, 0);
3856 } else if (!queue_IsEmpty(&call->tq)) {
3857 rxi_Start(0, call, 0, istack);
3862 /* Received a response to a challenge packet */
3864 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
3865 register struct rx_packet *np, int istack)
3869 /* Ignore the packet if we're the client */
3870 if (conn->type == RX_CLIENT_CONNECTION)
3873 /* If already authenticated, ignore the packet (it's probably a retry) */
3874 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
3877 /* Otherwise, have the security object evaluate the response packet */
3878 error = RXS_CheckResponse(conn->securityObject, conn, np);
3880 /* If the response is invalid, reset the connection, sending
3881 * an abort to the peer */
3885 rxi_ConnectionError(conn, error);
3886 MUTEX_ENTER(&conn->conn_data_lock);
3887 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3888 MUTEX_EXIT(&conn->conn_data_lock);
3891 /* If the response is valid, any calls waiting to attach
3892 * servers can now do so */
3895 for (i = 0; i < RX_MAXCALLS; i++) {
3896 struct rx_call *call = conn->call[i];
3898 MUTEX_ENTER(&call->lock);
3899 if (call->state == RX_STATE_PRECALL)
3900 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
3901 /* tnop can be null if newcallp is null */
3902 MUTEX_EXIT(&call->lock);
3906 /* Update the peer reachability information, just in case
3907 * some calls went into attach-wait while we were waiting
3908 * for authentication..
3910 rxi_UpdatePeerReach(conn, NULL);
3915 /* A client has received an authentication challenge: the security
3916 * object is asked to cough up a respectable response packet to send
3917 * back to the server. The server is responsible for retrying the
3918 * challenge if it fails to get a response. */
3921 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
3922 register struct rx_packet *np, int istack)
3926 /* Ignore the challenge if we're the server */
3927 if (conn->type == RX_SERVER_CONNECTION)
3930 /* Ignore the challenge if the connection is otherwise idle; someone's
3931 * trying to use us as an oracle. */
3932 if (!rxi_HasActiveCalls(conn))
3935 /* Send the security object the challenge packet. It is expected to fill
3936 * in the response. */
3937 error = RXS_GetResponse(conn->securityObject, conn, np);
3939 /* If the security object is unable to return a valid response, reset the
3940 * connection and send an abort to the peer. Otherwise send the response
3941 * packet to the peer connection. */
3943 rxi_ConnectionError(conn, error);
3944 MUTEX_ENTER(&conn->conn_data_lock);
3945 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3946 MUTEX_EXIT(&conn->conn_data_lock);
3948 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
3949 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
3955 /* Find an available server process to service the current request in
3956 * the given call structure. If one isn't available, queue up this
3957 * call so it eventually gets one */
3959 rxi_AttachServerProc(register struct rx_call *call,
3960 register osi_socket socket, register int *tnop,
3961 register struct rx_call **newcallp)
3963 register struct rx_serverQueueEntry *sq;
3964 register struct rx_service *service = call->conn->service;
3965 register int haveQuota = 0;
3967 /* May already be attached */
3968 if (call->state == RX_STATE_ACTIVE)
3971 MUTEX_ENTER(&rx_serverPool_lock);
3973 haveQuota = QuotaOK(service);
3974 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
3975 /* If there are no processes available to service this call,
3976 * put the call on the incoming call queue (unless it's
3977 * already on the queue).
3979 #ifdef RX_ENABLE_LOCKS
3981 ReturnToServerPool(service);
3982 #endif /* RX_ENABLE_LOCKS */
3984 if (!(call->flags & RX_CALL_WAIT_PROC)) {
3985 call->flags |= RX_CALL_WAIT_PROC;
3986 MUTEX_ENTER(&rx_stats_mutex);
3989 MUTEX_EXIT(&rx_stats_mutex);
3990 rxi_calltrace(RX_CALL_ARRIVAL, call);
3991 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
3992 queue_Append(&rx_incomingCallQueue, call);
3995 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
3997 /* If hot threads are enabled, and both newcallp and sq->socketp
3998 * are non-null, then this thread will process the call, and the
3999 * idle server thread will start listening on this threads socket.
4002 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4005 *sq->socketp = socket;
4006 clock_GetTime(&call->startTime);
4007 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4011 if (call->flags & RX_CALL_WAIT_PROC) {
4012 /* Conservative: I don't think this should happen */
4013 call->flags &= ~RX_CALL_WAIT_PROC;
4014 if (queue_IsOnQueue(call)) {
4016 MUTEX_ENTER(&rx_stats_mutex);
4018 MUTEX_EXIT(&rx_stats_mutex);
4021 call->state = RX_STATE_ACTIVE;
4022 call->mode = RX_MODE_RECEIVING;
4023 #ifdef RX_KERNEL_TRACE
4025 int glockOwner = ISAFS_GLOCK();
4028 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4029 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4035 if (call->flags & RX_CALL_CLEARED) {
4036 /* send an ack now to start the packet flow up again */
4037 call->flags &= ~RX_CALL_CLEARED;
4038 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4040 #ifdef RX_ENABLE_LOCKS
4043 service->nRequestsRunning++;
4044 if (service->nRequestsRunning <= service->minProcs)
4050 MUTEX_EXIT(&rx_serverPool_lock);
4053 /* Delay the sending of an acknowledge event for a short while, while
4054 * a new call is being prepared (in the case of a client) or a reply
4055 * is being prepared (in the case of a server). Rather than sending
4056 * an ack packet, an ACKALL packet is sent. */
4058 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4060 #ifdef RX_ENABLE_LOCKS
4062 MUTEX_ENTER(&call->lock);
4063 call->delayedAckEvent = NULL;
4064 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4066 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4067 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4069 MUTEX_EXIT(&call->lock);
4070 #else /* RX_ENABLE_LOCKS */
4072 call->delayedAckEvent = NULL;
4073 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4074 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4075 #endif /* RX_ENABLE_LOCKS */
4079 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4082 #ifdef RX_ENABLE_LOCKS
4084 MUTEX_ENTER(&call->lock);
4085 if (event == call->delayedAckEvent)
4086 call->delayedAckEvent = NULL;
4087 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4089 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4091 MUTEX_EXIT(&call->lock);
4092 #else /* RX_ENABLE_LOCKS */
4094 call->delayedAckEvent = NULL;
4095 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4096 #endif /* RX_ENABLE_LOCKS */
4100 #ifdef RX_ENABLE_LOCKS
4101 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4102 * clearing them out.
4105 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4107 register struct rx_packet *p, *tp;
4110 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4111 p->flags |= RX_PKTFLAG_ACKED;
4115 call->flags |= RX_CALL_TQ_CLEARME;
4116 call->flags |= RX_CALL_TQ_SOME_ACKED;
4119 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4120 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4121 call->tfirst = call->tnext;
4122 call->nSoftAcked = 0;
4124 if (call->flags & RX_CALL_FAST_RECOVER) {
4125 call->flags &= ~RX_CALL_FAST_RECOVER;
4126 call->cwind = call->nextCwind;
4127 call->nextCwind = 0;
4130 CV_SIGNAL(&call->cv_twind);
4132 #endif /* RX_ENABLE_LOCKS */
4134 /* Clear out the transmit queue for the current call (all packets have
4135 * been received by peer) */
4137 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4139 register struct rx_packet *p, *tp;
4141 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4142 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4144 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4145 p->flags |= RX_PKTFLAG_ACKED;
4149 call->flags |= RX_CALL_TQ_CLEARME;
4150 call->flags |= RX_CALL_TQ_SOME_ACKED;
4153 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4154 rxi_FreePackets(0, &call->tq);
4155 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4156 call->flags &= ~RX_CALL_TQ_CLEARME;
4158 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4160 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4161 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4162 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4163 call->nSoftAcked = 0;
4165 if (call->flags & RX_CALL_FAST_RECOVER) {
4166 call->flags &= ~RX_CALL_FAST_RECOVER;
4167 call->cwind = call->nextCwind;
4169 #ifdef RX_ENABLE_LOCKS
4170 CV_SIGNAL(&call->cv_twind);
4172 osi_rxWakeup(&call->twind);
4177 rxi_ClearReceiveQueue(register struct rx_call *call)
4179 if (queue_IsNotEmpty(&call->rq)) {
4180 rx_packetReclaims += rxi_FreePackets(0, &call->rq);
4181 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4183 if (call->state == RX_STATE_PRECALL) {
4184 call->flags |= RX_CALL_CLEARED;
4188 /* Send an abort packet for the specified call */
4190 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4191 int istack, int force)
4199 /* Clients should never delay abort messages */
4200 if (rx_IsClientConn(call->conn))
4203 if (call->abortCode != call->error) {
4204 call->abortCode = call->error;
4205 call->abortCount = 0;
4208 if (force || rxi_callAbortThreshhold == 0
4209 || call->abortCount < rxi_callAbortThreshhold) {
4210 if (call->delayedAbortEvent) {
4211 rxevent_Cancel(call->delayedAbortEvent, call,
4212 RX_CALL_REFCOUNT_ABORT);
4214 error = htonl(call->error);
4217 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4218 (char *)&error, sizeof(error), istack);
4219 } else if (!call->delayedAbortEvent) {
4220 clock_GetTime(&when);
4221 clock_Addmsec(&when, rxi_callAbortDelay);
4222 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4223 call->delayedAbortEvent =
4224 rxevent_Post(&when, rxi_SendDelayedCallAbort, call, 0);
4229 /* Send an abort packet for the specified connection. Packet is an
4230 * optional pointer to a packet that can be used to send the abort.
4231 * Once the number of abort messages reaches the threshhold, an
4232 * event is scheduled to send the abort. Setting the force flag
4233 * overrides sending delayed abort messages.
4235 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4236 * to send the abort packet.
4239 rxi_SendConnectionAbort(register struct rx_connection *conn,
4240 struct rx_packet *packet, int istack, int force)
4248 /* Clients should never delay abort messages */
4249 if (rx_IsClientConn(conn))
4252 if (force || rxi_connAbortThreshhold == 0
4253 || conn->abortCount < rxi_connAbortThreshhold) {
4254 if (conn->delayedAbortEvent) {
4255 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4257 error = htonl(conn->error);
4259 MUTEX_EXIT(&conn->conn_data_lock);
4261 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4262 RX_PACKET_TYPE_ABORT, (char *)&error,
4263 sizeof(error), istack);
4264 MUTEX_ENTER(&conn->conn_data_lock);
4265 } else if (!conn->delayedAbortEvent) {
4266 clock_GetTime(&when);
4267 clock_Addmsec(&when, rxi_connAbortDelay);
4268 conn->delayedAbortEvent =
4269 rxevent_Post(&when, rxi_SendDelayedConnAbort, conn, 0);
4274 /* Associate an error all of the calls owned by a connection. Called
4275 * with error non-zero. This is only for really fatal things, like
4276 * bad authentication responses. The connection itself is set in
4277 * error at this point, so that future packets received will be
4280 rxi_ConnectionError(register struct rx_connection *conn,
4281 register afs_int32 error)
4285 MUTEX_ENTER(&conn->conn_data_lock);
4286 if (conn->challengeEvent)
4287 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4288 if (conn->checkReachEvent) {
4289 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4290 conn->checkReachEvent = 0;
4291 conn->flags &= ~RX_CONN_ATTACHWAIT;
4294 MUTEX_EXIT(&conn->conn_data_lock);
4295 for (i = 0; i < RX_MAXCALLS; i++) {
4296 struct rx_call *call = conn->call[i];
4298 MUTEX_ENTER(&call->lock);
4299 rxi_CallError(call, error);
4300 MUTEX_EXIT(&call->lock);
4303 conn->error = error;
4304 MUTEX_ENTER(&rx_stats_mutex);
4305 rx_stats.fatalErrors++;
4306 MUTEX_EXIT(&rx_stats_mutex);
4311 rxi_CallError(register struct rx_call *call, afs_int32 error)
4314 error = call->error;
4315 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4316 if (!(call->flags & RX_CALL_TQ_BUSY)) {
4317 rxi_ResetCall(call, 0);
4320 rxi_ResetCall(call, 0);
4322 call->error = error;
4323 call->mode = RX_MODE_ERROR;
4326 /* Reset various fields in a call structure, and wakeup waiting
4327 * processes. Some fields aren't changed: state & mode are not
4328 * touched (these must be set by the caller), and bufptr, nLeft, and
4329 * nFree are not reset, since these fields are manipulated by
4330 * unprotected macros, and may only be reset by non-interrupting code.
4333 /* this code requires that call->conn be set properly as a pre-condition. */
4334 #endif /* ADAPT_WINDOW */
4337 rxi_ResetCall(register struct rx_call *call, register int newcall)
4340 register struct rx_peer *peer;
4341 struct rx_packet *packet;
4343 /* Notify anyone who is waiting for asynchronous packet arrival */
4344 if (call->arrivalProc) {
4345 (*call->arrivalProc) (call, call->arrivalProcHandle,
4346 call->arrivalProcArg);
4347 call->arrivalProc = (void (*)())0;
4350 if (call->delayedAbortEvent) {
4351 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4352 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4354 rxi_SendCallAbort(call, packet, 0, 1);
4355 rxi_FreePacket(packet);
4360 * Update the peer with the congestion information in this call
4361 * so other calls on this connection can pick up where this call
4362 * left off. If the congestion sequence numbers don't match then
4363 * another call experienced a retransmission.
4365 peer = call->conn->peer;
4366 MUTEX_ENTER(&peer->peer_lock);
4368 if (call->congestSeq == peer->congestSeq) {
4369 peer->cwind = MAX(peer->cwind, call->cwind);
4370 peer->MTU = MAX(peer->MTU, call->MTU);
4371 peer->nDgramPackets =
4372 MAX(peer->nDgramPackets, call->nDgramPackets);
4375 call->abortCode = 0;
4376 call->abortCount = 0;
4378 if (peer->maxDgramPackets > 1) {
4379 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4381 call->MTU = peer->MTU;
4383 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4384 call->ssthresh = rx_maxSendWindow;
4385 call->nDgramPackets = peer->nDgramPackets;
4386 call->congestSeq = peer->congestSeq;
4387 MUTEX_EXIT(&peer->peer_lock);
4389 flags = call->flags;
4390 rxi_ClearReceiveQueue(call);
4391 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4392 if (call->flags & RX_CALL_TQ_BUSY) {
4393 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4394 call->flags |= (flags & RX_CALL_TQ_WAIT);
4396 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4398 rxi_ClearTransmitQueue(call, 0);
4399 queue_Init(&call->tq);
4402 queue_Init(&call->rq);
4404 call->rwind = rx_initReceiveWindow;
4405 call->twind = rx_initSendWindow;
4406 call->nSoftAcked = 0;
4407 call->nextCwind = 0;
4410 call->nCwindAcks = 0;
4411 call->nSoftAcks = 0;
4412 call->nHardAcks = 0;
4414 call->tfirst = call->rnext = call->tnext = 1;
4416 call->lastAcked = 0;
4417 call->localStatus = call->remoteStatus = 0;
4419 if (flags & RX_CALL_READER_WAIT) {
4420 #ifdef RX_ENABLE_LOCKS
4421 CV_BROADCAST(&call->cv_rq);
4423 osi_rxWakeup(&call->rq);
4426 if (flags & RX_CALL_WAIT_PACKETS) {
4427 MUTEX_ENTER(&rx_freePktQ_lock);
4428 rxi_PacketsUnWait(); /* XXX */
4429 MUTEX_EXIT(&rx_freePktQ_lock);
4431 #ifdef RX_ENABLE_LOCKS
4432 CV_SIGNAL(&call->cv_twind);
4434 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4435 osi_rxWakeup(&call->twind);
4438 #ifdef RX_ENABLE_LOCKS
4439 /* The following ensures that we don't mess with any queue while some
4440 * other thread might also be doing so. The call_queue_lock field is
4441 * is only modified under the call lock. If the call is in the process
4442 * of being removed from a queue, the call is not locked until the
4443 * the queue lock is dropped and only then is the call_queue_lock field
4444 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4445 * Note that any other routine which removes a call from a queue has to
4446 * obtain the queue lock before examing the queue and removing the call.
4448 if (call->call_queue_lock) {
4449 MUTEX_ENTER(call->call_queue_lock);
4450 if (queue_IsOnQueue(call)) {
4452 if (flags & RX_CALL_WAIT_PROC) {
4453 MUTEX_ENTER(&rx_stats_mutex);
4455 MUTEX_EXIT(&rx_stats_mutex);
4458 MUTEX_EXIT(call->call_queue_lock);
4459 CLEAR_CALL_QUEUE_LOCK(call);
4461 #else /* RX_ENABLE_LOCKS */
4462 if (queue_IsOnQueue(call)) {
4464 if (flags & RX_CALL_WAIT_PROC)
4467 #endif /* RX_ENABLE_LOCKS */
4469 rxi_KeepAliveOff(call);
4470 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4473 /* Send an acknowledge for the indicated packet (seq,serial) of the
4474 * indicated call, for the indicated reason (reason). This
4475 * acknowledge will specifically acknowledge receiving the packet, and
4476 * will also specify which other packets for this call have been
4477 * received. This routine returns the packet that was used to the
4478 * caller. The caller is responsible for freeing it or re-using it.
4479 * This acknowledgement also returns the highest sequence number
4480 * actually read out by the higher level to the sender; the sender
4481 * promises to keep around packets that have not been read by the
4482 * higher level yet (unless, of course, the sender decides to abort
4483 * the call altogether). Any of p, seq, serial, pflags, or reason may
4484 * be set to zero without ill effect. That is, if they are zero, they
4485 * will not convey any information.
4486 * NOW there is a trailer field, after the ack where it will safely be
4487 * ignored by mundanes, which indicates the maximum size packet this
4488 * host can swallow. */
4490 register struct rx_packet *optionalPacket; use to send ack (or null)
4491 int seq; Sequence number of the packet we are acking
4492 int serial; Serial number of the packet
4493 int pflags; Flags field from packet header
4494 int reason; Reason an acknowledge was prompted
4498 rxi_SendAck(register struct rx_call *call,
4499 register struct rx_packet *optionalPacket, int serial, int reason,
4502 struct rx_ackPacket *ap;
4503 register struct rx_packet *rqp;
4504 register struct rx_packet *nxp; /* For queue_Scan */
4505 register struct rx_packet *p;
4508 #ifdef RX_ENABLE_TSFPQ
4509 struct rx_ts_info_t * rx_ts_info;
4513 * Open the receive window once a thread starts reading packets
4515 if (call->rnext > 1) {
4516 call->rwind = rx_maxReceiveWindow;
4519 call->nHardAcks = 0;
4520 call->nSoftAcks = 0;
4521 if (call->rnext > call->lastAcked)
4522 call->lastAcked = call->rnext;
4526 rx_computelen(p, p->length); /* reset length, you never know */
4527 } /* where that's been... */
4528 #ifdef RX_ENABLE_TSFPQ
4530 RX_TS_INFO_GET(rx_ts_info);
4531 if ((p = rx_ts_info->local_special_packet)) {
4532 rx_computelen(p, p->length);
4533 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4534 rx_ts_info->local_special_packet = p;
4535 } else { /* We won't send the ack, but don't panic. */
4536 return optionalPacket;
4540 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4541 /* We won't send the ack, but don't panic. */
4542 return optionalPacket;
4547 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4550 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4551 #ifndef RX_ENABLE_TSFPQ
4552 if (!optionalPacket)
4555 return optionalPacket;
4557 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4558 if (rx_Contiguous(p) < templ) {
4559 #ifndef RX_ENABLE_TSFPQ
4560 if (!optionalPacket)
4563 return optionalPacket;
4568 /* MTUXXX failing to send an ack is very serious. We should */
4569 /* try as hard as possible to send even a partial ack; it's */
4570 /* better than nothing. */
4571 ap = (struct rx_ackPacket *)rx_DataOf(p);
4572 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4573 ap->reason = reason;
4575 /* The skew computation used to be bogus, I think it's better now. */
4576 /* We should start paying attention to skew. XXX */
4577 ap->serial = htonl(serial);
4578 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4580 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4581 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4583 /* No fear of running out of ack packet here because there can only be at most
4584 * one window full of unacknowledged packets. The window size must be constrained
4585 * to be less than the maximum ack size, of course. Also, an ack should always
4586 * fit into a single packet -- it should not ever be fragmented. */
4587 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4588 if (!rqp || !call->rq.next
4589 || (rqp->header.seq > (call->rnext + call->rwind))) {
4590 #ifndef RX_ENABLE_TSFPQ
4591 if (!optionalPacket)
4594 rxi_CallError(call, RX_CALL_DEAD);
4595 return optionalPacket;
4598 while (rqp->header.seq > call->rnext + offset)
4599 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4600 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4602 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4603 #ifndef RX_ENABLE_TSFPQ
4604 if (!optionalPacket)
4607 rxi_CallError(call, RX_CALL_DEAD);
4608 return optionalPacket;
4613 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4615 /* these are new for AFS 3.3 */
4616 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4617 templ = htonl(templ);
4618 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4619 templ = htonl(call->conn->peer->ifMTU);
4620 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4621 sizeof(afs_int32), &templ);
4623 /* new for AFS 3.4 */
4624 templ = htonl(call->rwind);
4625 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4626 sizeof(afs_int32), &templ);
4628 /* new for AFS 3.5 */
4629 templ = htonl(call->conn->peer->ifDgramPackets);
4630 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4631 sizeof(afs_int32), &templ);
4633 p->header.serviceId = call->conn->serviceId;
4634 p->header.cid = (call->conn->cid | call->channel);
4635 p->header.callNumber = *call->callNumber;
4637 p->header.securityIndex = call->conn->securityIndex;
4638 p->header.epoch = call->conn->epoch;
4639 p->header.type = RX_PACKET_TYPE_ACK;
4640 p->header.flags = RX_SLOW_START_OK;
4641 if (reason == RX_ACK_PING) {
4642 p->header.flags |= RX_REQUEST_ACK;
4644 clock_GetTime(&call->pingRequestTime);
4647 if (call->conn->type == RX_CLIENT_CONNECTION)
4648 p->header.flags |= RX_CLIENT_INITIATED;
4652 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u",
4653 ap->reason, ntohl(ap->previousPacket),
4654 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4656 for (offset = 0; offset < ap->nAcks; offset++)
4657 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4665 register int i, nbytes = p->length;
4667 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4668 if (nbytes <= p->wirevec[i].iov_len) {
4669 register int savelen, saven;
4671 savelen = p->wirevec[i].iov_len;
4673 p->wirevec[i].iov_len = nbytes;
4675 rxi_Send(call, p, istack);
4676 p->wirevec[i].iov_len = savelen;
4680 nbytes -= p->wirevec[i].iov_len;
4683 MUTEX_ENTER(&rx_stats_mutex);
4684 rx_stats.ackPacketsSent++;
4685 MUTEX_EXIT(&rx_stats_mutex);
4686 #ifndef RX_ENABLE_TSFPQ
4687 if (!optionalPacket)
4690 return optionalPacket; /* Return packet for re-use by caller */
4693 /* Send all of the packets in the list in single datagram */
4695 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4696 int istack, int moreFlag, struct clock *now,
4697 struct clock *retryTime, int resending)
4702 struct rx_connection *conn = call->conn;
4703 struct rx_peer *peer = conn->peer;
4705 MUTEX_ENTER(&peer->peer_lock);
4708 peer->reSends += len;
4709 MUTEX_ENTER(&rx_stats_mutex);
4710 rx_stats.dataPacketsSent += len;
4711 MUTEX_EXIT(&rx_stats_mutex);
4712 MUTEX_EXIT(&peer->peer_lock);
4714 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4718 /* Set the packet flags and schedule the resend events */
4719 /* Only request an ack for the last packet in the list */
4720 for (i = 0; i < len; i++) {
4721 list[i]->retryTime = *retryTime;
4722 if (list[i]->header.serial) {
4723 /* Exponentially backoff retry times */
4724 if (list[i]->backoff < MAXBACKOFF) {
4725 /* so it can't stay == 0 */
4726 list[i]->backoff = (list[i]->backoff << 1) + 1;
4729 clock_Addmsec(&(list[i]->retryTime),
4730 ((afs_uint32) list[i]->backoff) << 8);
4733 /* Wait a little extra for the ack on the last packet */
4734 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4735 clock_Addmsec(&(list[i]->retryTime), 400);
4738 /* Record the time sent */
4739 list[i]->timeSent = *now;
4741 /* Ask for an ack on retransmitted packets, on every other packet
4742 * if the peer doesn't support slow start. Ask for an ack on every
4743 * packet until the congestion window reaches the ack rate. */
4744 if (list[i]->header.serial) {
4746 MUTEX_ENTER(&rx_stats_mutex);
4747 rx_stats.dataPacketsReSent++;
4748 MUTEX_EXIT(&rx_stats_mutex);
4750 /* improved RTO calculation- not Karn */
4751 list[i]->firstSent = *now;
4752 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4753 || (!(call->flags & RX_CALL_SLOW_START_OK)
4754 && (list[i]->header.seq & 1)))) {
4759 MUTEX_ENTER(&peer->peer_lock);
4763 MUTEX_ENTER(&rx_stats_mutex);
4764 rx_stats.dataPacketsSent++;
4765 MUTEX_EXIT(&rx_stats_mutex);
4766 MUTEX_EXIT(&peer->peer_lock);
4768 /* Tag this packet as not being the last in this group,
4769 * for the receiver's benefit */
4770 if (i < len - 1 || moreFlag) {
4771 list[i]->header.flags |= RX_MORE_PACKETS;
4774 /* Install the new retransmit time for the packet, and
4775 * record the time sent */
4776 list[i]->timeSent = *now;
4780 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4783 /* Since we're about to send a data packet to the peer, it's
4784 * safe to nuke any scheduled end-of-packets ack */
4785 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4787 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4788 MUTEX_EXIT(&call->lock);
4790 rxi_SendPacketList(call, conn, list, len, istack);
4792 rxi_SendPacket(call, conn, list[0], istack);
4794 MUTEX_ENTER(&call->lock);
4795 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4797 /* Update last send time for this call (for keep-alive
4798 * processing), and for the connection (so that we can discover
4799 * idle connections) */
4800 conn->lastSendTime = call->lastSendTime = clock_Sec();
4803 /* When sending packets we need to follow these rules:
4804 * 1. Never send more than maxDgramPackets in a jumbogram.
4805 * 2. Never send a packet with more than two iovecs in a jumbogram.
4806 * 3. Never send a retransmitted packet in a jumbogram.
4807 * 4. Never send more than cwind/4 packets in a jumbogram
4808 * We always keep the last list we should have sent so we
4809 * can set the RX_MORE_PACKETS flags correctly.
4812 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4813 int istack, struct clock *now, struct clock *retryTime,
4816 int i, cnt, lastCnt = 0;
4817 struct rx_packet **listP, **lastP = 0;
4818 struct rx_peer *peer = call->conn->peer;
4819 int morePackets = 0;
4821 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4822 /* Does the current packet force us to flush the current list? */
4824 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4825 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4827 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
4829 /* If the call enters an error state stop sending, or if
4830 * we entered congestion recovery mode, stop sending */
4831 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4839 /* Add the current packet to the list if it hasn't been acked.
4840 * Otherwise adjust the list pointer to skip the current packet. */
4841 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
4843 /* Do we need to flush the list? */
4844 if (cnt >= (int)peer->maxDgramPackets
4845 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
4846 || list[i]->header.serial
4847 || list[i]->length != RX_JUMBOBUFFERSIZE) {
4849 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
4850 retryTime, resending);
4851 /* If the call enters an error state stop sending, or if
4852 * we entered congestion recovery mode, stop sending */
4854 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4859 listP = &list[i + 1];
4864 osi_Panic("rxi_SendList error");
4866 listP = &list[i + 1];
4870 /* Send the whole list when the call is in receive mode, when
4871 * the call is in eof mode, when we are in fast recovery mode,
4872 * and when we have the last packet */
4873 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
4874 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
4875 || (call->flags & RX_CALL_FAST_RECOVER)) {
4876 /* Check for the case where the current list contains
4877 * an acked packet. Since we always send retransmissions
4878 * in a separate packet, we only need to check the first
4879 * packet in the list */
4880 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
4884 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
4885 retryTime, resending);
4886 /* If the call enters an error state stop sending, or if
4887 * we entered congestion recovery mode, stop sending */
4888 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4892 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
4895 } else if (lastCnt > 0) {
4896 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
4901 #ifdef RX_ENABLE_LOCKS
4902 /* Call rxi_Start, below, but with the call lock held. */
4904 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
4905 void *arg1, int istack)
4907 MUTEX_ENTER(&call->lock);
4908 rxi_Start(event, call, arg1, istack);
4909 MUTEX_EXIT(&call->lock);
4911 #endif /* RX_ENABLE_LOCKS */
4913 /* This routine is called when new packets are readied for
4914 * transmission and when retransmission may be necessary, or when the
4915 * transmission window or burst count are favourable. This should be
4916 * better optimized for new packets, the usual case, now that we've
4917 * got rid of queues of send packets. XXXXXXXXXXX */
4919 rxi_Start(struct rxevent *event, register struct rx_call *call,
4920 void *arg1, int istack)
4922 struct rx_packet *p;
4923 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
4924 struct rx_peer *peer = call->conn->peer;
4925 struct clock now, retryTime;
4929 struct rx_packet **xmitList;
4932 /* If rxi_Start is being called as a result of a resend event,
4933 * then make sure that the event pointer is removed from the call
4934 * structure, since there is no longer a per-call retransmission
4936 if (event && event == call->resendEvent) {
4937 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
4938 call->resendEvent = NULL;
4940 if (queue_IsEmpty(&call->tq)) {
4944 /* Timeouts trigger congestion recovery */
4945 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4946 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4947 /* someone else is waiting to start recovery */
4950 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4951 while (call->flags & RX_CALL_TQ_BUSY) {
4952 call->flags |= RX_CALL_TQ_WAIT;
4953 #ifdef RX_ENABLE_LOCKS
4954 CV_WAIT(&call->cv_tq, &call->lock);
4955 #else /* RX_ENABLE_LOCKS */
4956 osi_rxSleep(&call->tq);
4957 #endif /* RX_ENABLE_LOCKS */
4959 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4960 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4961 call->flags |= RX_CALL_FAST_RECOVER;
4962 if (peer->maxDgramPackets > 1) {
4963 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
4965 call->MTU = MIN(peer->natMTU, peer->maxMTU);
4967 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4968 call->nDgramPackets = 1;
4970 call->nextCwind = 1;
4973 MUTEX_ENTER(&peer->peer_lock);
4974 peer->MTU = call->MTU;
4975 peer->cwind = call->cwind;
4976 peer->nDgramPackets = 1;
4978 call->congestSeq = peer->congestSeq;
4979 MUTEX_EXIT(&peer->peer_lock);
4980 /* Clear retry times on packets. Otherwise, it's possible for
4981 * some packets in the queue to force resends at rates faster
4982 * than recovery rates.
4984 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
4985 if (!(p->flags & RX_PKTFLAG_ACKED)) {
4986 clock_Zero(&p->retryTime);
4991 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4992 MUTEX_ENTER(&rx_stats_mutex);
4993 rx_tq_debug.rxi_start_in_error++;
4994 MUTEX_EXIT(&rx_stats_mutex);
4999 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5000 /* Get clock to compute the re-transmit time for any packets
5001 * in this burst. Note, if we back off, it's reasonable to
5002 * back off all of the packets in the same manner, even if
5003 * some of them have been retransmitted more times than more
5004 * recent additions */
5005 clock_GetTime(&now);
5006 retryTime = now; /* initialize before use */
5007 MUTEX_ENTER(&peer->peer_lock);
5008 clock_Add(&retryTime, &peer->timeout);
5009 MUTEX_EXIT(&peer->peer_lock);
5011 /* Send (or resend) any packets that need it, subject to
5012 * window restrictions and congestion burst control
5013 * restrictions. Ask for an ack on the last packet sent in
5014 * this burst. For now, we're relying upon the window being
5015 * considerably bigger than the largest number of packets that
5016 * are typically sent at once by one initial call to
5017 * rxi_Start. This is probably bogus (perhaps we should ask
5018 * for an ack when we're half way through the current
5019 * window?). Also, for non file transfer applications, this
5020 * may end up asking for an ack for every packet. Bogus. XXXX
5023 * But check whether we're here recursively, and let the other guy
5026 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5027 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5028 call->flags |= RX_CALL_TQ_BUSY;
5030 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5032 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5033 call->flags &= ~RX_CALL_NEED_START;
5034 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5036 maxXmitPackets = MIN(call->twind, call->cwind);
5037 xmitList = (struct rx_packet **)
5038 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5039 if (xmitList == NULL)
5040 osi_Panic("rxi_Start, failed to allocate xmit list");
5041 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5042 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5043 /* We shouldn't be sending packets if a thread is waiting
5044 * to initiate congestion recovery */
5048 && (call->flags & RX_CALL_FAST_RECOVER)) {
5049 /* Only send one packet during fast recovery */
5052 if ((p->flags & RX_PKTFLAG_FREE)
5053 || (!queue_IsEnd(&call->tq, nxp)
5054 && (nxp->flags & RX_PKTFLAG_FREE))
5055 || (p == (struct rx_packet *)&rx_freePacketQueue)
5056 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5057 osi_Panic("rxi_Start: xmit queue clobbered");
5059 if (p->flags & RX_PKTFLAG_ACKED) {
5060 MUTEX_ENTER(&rx_stats_mutex);
5061 rx_stats.ignoreAckedPacket++;
5062 MUTEX_EXIT(&rx_stats_mutex);
5063 continue; /* Ignore this packet if it has been acknowledged */
5066 /* Turn off all flags except these ones, which are the same
5067 * on each transmission */
5068 p->header.flags &= RX_PRESET_FLAGS;
5070 if (p->header.seq >=
5071 call->tfirst + MIN((int)call->twind,
5072 (int)(call->nSoftAcked +
5074 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5075 /* Note: if we're waiting for more window space, we can
5076 * still send retransmits; hence we don't return here, but
5077 * break out to schedule a retransmit event */
5078 dpf(("call %d waiting for window",
5079 *(call->callNumber)));
5083 /* Transmit the packet if it needs to be sent. */
5084 if (!clock_Lt(&now, &p->retryTime)) {
5085 if (nXmitPackets == maxXmitPackets) {
5086 rxi_SendXmitList(call, xmitList, nXmitPackets,
5087 istack, &now, &retryTime,
5089 osi_Free(xmitList, maxXmitPackets *
5090 sizeof(struct rx_packet *));
5093 xmitList[nXmitPackets++] = p;
5097 /* xmitList now hold pointers to all of the packets that are
5098 * ready to send. Now we loop to send the packets */
5099 if (nXmitPackets > 0) {
5100 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5101 &now, &retryTime, resending);
5104 maxXmitPackets * sizeof(struct rx_packet *));
5106 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5108 * TQ references no longer protected by this flag; they must remain
5109 * protected by the global lock.
5111 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5112 call->flags &= ~RX_CALL_TQ_BUSY;
5113 if (call->flags & RX_CALL_TQ_WAIT) {
5114 call->flags &= ~RX_CALL_TQ_WAIT;
5115 #ifdef RX_ENABLE_LOCKS
5116 CV_BROADCAST(&call->cv_tq);
5117 #else /* RX_ENABLE_LOCKS */
5118 osi_rxWakeup(&call->tq);
5119 #endif /* RX_ENABLE_LOCKS */
5124 /* We went into the error state while sending packets. Now is
5125 * the time to reset the call. This will also inform the using
5126 * process that the call is in an error state.
5128 MUTEX_ENTER(&rx_stats_mutex);
5129 rx_tq_debug.rxi_start_aborted++;
5130 MUTEX_EXIT(&rx_stats_mutex);
5131 call->flags &= ~RX_CALL_TQ_BUSY;
5132 if (call->flags & RX_CALL_TQ_WAIT) {
5133 call->flags &= ~RX_CALL_TQ_WAIT;
5134 #ifdef RX_ENABLE_LOCKS
5135 CV_BROADCAST(&call->cv_tq);
5136 #else /* RX_ENABLE_LOCKS */
5137 osi_rxWakeup(&call->tq);
5138 #endif /* RX_ENABLE_LOCKS */
5140 rxi_CallError(call, call->error);
5143 #ifdef RX_ENABLE_LOCKS
5144 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5145 register int missing;
5146 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5147 /* Some packets have received acks. If they all have, we can clear
5148 * the transmit queue.
5151 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5152 if (p->header.seq < call->tfirst
5153 && (p->flags & RX_PKTFLAG_ACKED)) {
5160 call->flags |= RX_CALL_TQ_CLEARME;
5162 #endif /* RX_ENABLE_LOCKS */
5163 /* Don't bother doing retransmits if the TQ is cleared. */
5164 if (call->flags & RX_CALL_TQ_CLEARME) {
5165 rxi_ClearTransmitQueue(call, 1);
5167 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5170 /* Always post a resend event, if there is anything in the
5171 * queue, and resend is possible. There should be at least
5172 * one unacknowledged packet in the queue ... otherwise none
5173 * of these packets should be on the queue in the first place.
5175 if (call->resendEvent) {
5176 /* Cancel the existing event and post a new one */
5177 rxevent_Cancel(call->resendEvent, call,
5178 RX_CALL_REFCOUNT_RESEND);
5181 /* The retry time is the retry time on the first unacknowledged
5182 * packet inside the current window */
5184 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5185 /* Don't set timers for packets outside the window */
5186 if (p->header.seq >= call->tfirst + call->twind) {
5190 if (!(p->flags & RX_PKTFLAG_ACKED)
5191 && !clock_IsZero(&p->retryTime)) {
5193 retryTime = p->retryTime;
5198 /* Post a new event to re-run rxi_Start when retries may be needed */
5199 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5200 #ifdef RX_ENABLE_LOCKS
5201 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5203 rxevent_Post2(&retryTime, rxi_StartUnlocked,
5204 (void *)call, 0, istack);
5205 #else /* RX_ENABLE_LOCKS */
5207 rxevent_Post2(&retryTime, rxi_Start, (void *)call,
5209 #endif /* RX_ENABLE_LOCKS */
5212 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5213 } while (call->flags & RX_CALL_NEED_START);
5215 * TQ references no longer protected by this flag; they must remain
5216 * protected by the global lock.
5218 call->flags &= ~RX_CALL_TQ_BUSY;
5219 if (call->flags & RX_CALL_TQ_WAIT) {
5220 call->flags &= ~RX_CALL_TQ_WAIT;
5221 #ifdef RX_ENABLE_LOCKS
5222 CV_BROADCAST(&call->cv_tq);
5223 #else /* RX_ENABLE_LOCKS */
5224 osi_rxWakeup(&call->tq);
5225 #endif /* RX_ENABLE_LOCKS */
5228 call->flags |= RX_CALL_NEED_START;
5230 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5232 if (call->resendEvent) {
5233 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5238 /* Also adjusts the keep alive parameters for the call, to reflect
5239 * that we have just sent a packet (so keep alives aren't sent
5242 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5245 register struct rx_connection *conn = call->conn;
5247 /* Stamp each packet with the user supplied status */
5248 p->header.userStatus = call->localStatus;
5250 /* Allow the security object controlling this call's security to
5251 * make any last-minute changes to the packet */
5252 RXS_SendPacket(conn->securityObject, call, p);
5254 /* Since we're about to send SOME sort of packet to the peer, it's
5255 * safe to nuke any scheduled end-of-packets ack */
5256 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5258 /* Actually send the packet, filling in more connection-specific fields */
5259 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5260 MUTEX_EXIT(&call->lock);
5261 rxi_SendPacket(call, conn, p, istack);
5262 MUTEX_ENTER(&call->lock);
5263 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5265 /* Update last send time for this call (for keep-alive
5266 * processing), and for the connection (so that we can discover
5267 * idle connections) */
5268 conn->lastSendTime = call->lastSendTime = clock_Sec();
5272 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5273 * that things are fine. Also called periodically to guarantee that nothing
5274 * falls through the cracks (e.g. (error + dally) connections have keepalive
5275 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5277 * haveCTLock Set if calling from rxi_ReapConnections
5279 #ifdef RX_ENABLE_LOCKS
5281 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5282 #else /* RX_ENABLE_LOCKS */
5284 rxi_CheckCall(register struct rx_call *call)
5285 #endif /* RX_ENABLE_LOCKS */
5287 register struct rx_connection *conn = call->conn;
5289 afs_uint32 deadTime;
5291 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5292 if (call->flags & RX_CALL_TQ_BUSY) {
5293 /* Call is active and will be reset by rxi_Start if it's
5294 * in an error state.
5299 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5301 (((afs_uint32) conn->secondsUntilDead << 10) +
5302 ((afs_uint32) conn->peer->rtt >> 3) +
5303 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5305 /* These are computed to the second (+- 1 second). But that's
5306 * good enough for these values, which should be a significant
5307 * number of seconds. */
5308 if (now > (call->lastReceiveTime + deadTime)) {
5309 if (call->state == RX_STATE_ACTIVE) {
5310 rxi_CallError(call, RX_CALL_DEAD);
5313 #ifdef RX_ENABLE_LOCKS
5314 /* Cancel pending events */
5315 rxevent_Cancel(call->delayedAckEvent, call,
5316 RX_CALL_REFCOUNT_DELAY);
5317 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5318 rxevent_Cancel(call->keepAliveEvent, call,
5319 RX_CALL_REFCOUNT_ALIVE);
5320 if (call->refCount == 0) {
5321 rxi_FreeCall(call, haveCTLock);
5325 #else /* RX_ENABLE_LOCKS */
5328 #endif /* RX_ENABLE_LOCKS */
5330 /* Non-active calls are destroyed if they are not responding
5331 * to pings; active calls are simply flagged in error, so the
5332 * attached process can die reasonably gracefully. */
5334 /* see if we have a non-activity timeout */
5335 if (call->startWait && conn->idleDeadTime
5336 && ((call->startWait + conn->idleDeadTime) < now)) {
5337 if (call->state == RX_STATE_ACTIVE) {
5338 rxi_CallError(call, RX_CALL_TIMEOUT);
5342 /* see if we have a hard timeout */
5343 if (conn->hardDeadTime
5344 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5345 if (call->state == RX_STATE_ACTIVE)
5346 rxi_CallError(call, RX_CALL_TIMEOUT);
5353 /* When a call is in progress, this routine is called occasionally to
5354 * make sure that some traffic has arrived (or been sent to) the peer.
5355 * If nothing has arrived in a reasonable amount of time, the call is
5356 * declared dead; if nothing has been sent for a while, we send a
5357 * keep-alive packet (if we're actually trying to keep the call alive)
5360 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5363 struct rx_connection *conn;
5366 MUTEX_ENTER(&call->lock);
5367 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5368 if (event == call->keepAliveEvent)
5369 call->keepAliveEvent = NULL;
5372 #ifdef RX_ENABLE_LOCKS
5373 if (rxi_CheckCall(call, 0)) {
5374 MUTEX_EXIT(&call->lock);
5377 #else /* RX_ENABLE_LOCKS */
5378 if (rxi_CheckCall(call))
5380 #endif /* RX_ENABLE_LOCKS */
5382 /* Don't try to keep alive dallying calls */
5383 if (call->state == RX_STATE_DALLY) {
5384 MUTEX_EXIT(&call->lock);
5389 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5390 /* Don't try to send keepalives if there is unacknowledged data */
5391 /* the rexmit code should be good enough, this little hack
5392 * doesn't quite work XXX */
5393 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5395 rxi_ScheduleKeepAliveEvent(call);
5396 MUTEX_EXIT(&call->lock);
5401 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5403 if (!call->keepAliveEvent) {
5405 clock_GetTime(&when);
5406 when.sec += call->conn->secondsUntilPing;
5407 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5408 call->keepAliveEvent =
5409 rxevent_Post(&when, rxi_KeepAliveEvent, call, 0);
5413 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5415 rxi_KeepAliveOn(register struct rx_call *call)
5417 /* Pretend last packet received was received now--i.e. if another
5418 * packet isn't received within the keep alive time, then the call
5419 * will die; Initialize last send time to the current time--even
5420 * if a packet hasn't been sent yet. This will guarantee that a
5421 * keep-alive is sent within the ping time */
5422 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5423 rxi_ScheduleKeepAliveEvent(call);
5426 /* This routine is called to send connection abort messages
5427 * that have been delayed to throttle looping clients. */
5429 rxi_SendDelayedConnAbort(struct rxevent *event,
5430 register struct rx_connection *conn, char *dummy)
5433 struct rx_packet *packet;
5435 MUTEX_ENTER(&conn->conn_data_lock);
5436 conn->delayedAbortEvent = NULL;
5437 error = htonl(conn->error);
5439 MUTEX_EXIT(&conn->conn_data_lock);
5440 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5443 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5444 RX_PACKET_TYPE_ABORT, (char *)&error,
5446 rxi_FreePacket(packet);
5450 /* This routine is called to send call abort messages
5451 * that have been delayed to throttle looping clients. */
5453 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5457 struct rx_packet *packet;
5459 MUTEX_ENTER(&call->lock);
5460 call->delayedAbortEvent = NULL;
5461 error = htonl(call->error);
5463 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5466 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5467 (char *)&error, sizeof(error), 0);
5468 rxi_FreePacket(packet);
5470 MUTEX_EXIT(&call->lock);
5473 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5474 * seconds) to ask the client to authenticate itself. The routine
5475 * issues a challenge to the client, which is obtained from the
5476 * security object associated with the connection */
5478 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5479 void *arg1, int tries)
5481 conn->challengeEvent = NULL;
5482 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5483 register struct rx_packet *packet;
5487 /* We've failed to authenticate for too long.
5488 * Reset any calls waiting for authentication;
5489 * they are all in RX_STATE_PRECALL.
5493 MUTEX_ENTER(&conn->conn_call_lock);
5494 for (i = 0; i < RX_MAXCALLS; i++) {
5495 struct rx_call *call = conn->call[i];
5497 MUTEX_ENTER(&call->lock);
5498 if (call->state == RX_STATE_PRECALL) {
5499 rxi_CallError(call, RX_CALL_DEAD);
5500 rxi_SendCallAbort(call, NULL, 0, 0);
5502 MUTEX_EXIT(&call->lock);
5505 MUTEX_EXIT(&conn->conn_call_lock);
5509 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5511 /* If there's no packet available, do this later. */
5512 RXS_GetChallenge(conn->securityObject, conn, packet);
5513 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5514 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5515 rxi_FreePacket(packet);
5517 clock_GetTime(&when);
5518 when.sec += RX_CHALLENGE_TIMEOUT;
5519 conn->challengeEvent =
5520 rxevent_Post2(&when, rxi_ChallengeEvent, conn, 0,
5525 /* Call this routine to start requesting the client to authenticate
5526 * itself. This will continue until authentication is established,
5527 * the call times out, or an invalid response is returned. The
5528 * security object associated with the connection is asked to create
5529 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5530 * defined earlier. */
5532 rxi_ChallengeOn(register struct rx_connection *conn)
5534 if (!conn->challengeEvent) {
5535 RXS_CreateChallenge(conn->securityObject, conn);
5536 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5541 /* Compute round trip time of the packet provided, in *rttp.
5544 /* rxi_ComputeRoundTripTime is called with peer locked. */
5545 /* sentp and/or peer may be null */
5547 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5548 register struct clock *sentp,
5549 register struct rx_peer *peer)
5551 struct clock thisRtt, *rttp = &thisRtt;
5553 register int rtt_timeout;
5555 clock_GetTime(rttp);
5557 if (clock_Lt(rttp, sentp)) {
5559 return; /* somebody set the clock back, don't count this time. */
5561 clock_Sub(rttp, sentp);
5562 MUTEX_ENTER(&rx_stats_mutex);
5563 if (clock_Lt(rttp, &rx_stats.minRtt))
5564 rx_stats.minRtt = *rttp;
5565 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5566 if (rttp->sec > 60) {
5567 MUTEX_EXIT(&rx_stats_mutex);
5568 return; /* somebody set the clock ahead */
5570 rx_stats.maxRtt = *rttp;
5572 clock_Add(&rx_stats.totalRtt, rttp);
5573 rx_stats.nRttSamples++;
5574 MUTEX_EXIT(&rx_stats_mutex);
5576 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5578 /* Apply VanJacobson round-trip estimations */
5583 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5584 * srtt is stored as fixed point with 3 bits after the binary
5585 * point (i.e., scaled by 8). The following magic is
5586 * equivalent to the smoothing algorithm in rfc793 with an
5587 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5588 * srtt*8 = srtt*8 + rtt - srtt
5589 * srtt = srtt + rtt/8 - srtt/8
5592 delta = MSEC(rttp) - (peer->rtt >> 3);
5596 * We accumulate a smoothed rtt variance (actually, a smoothed
5597 * mean difference), then set the retransmit timer to smoothed
5598 * rtt + 4 times the smoothed variance (was 2x in van's original
5599 * paper, but 4x works better for me, and apparently for him as
5601 * rttvar is stored as
5602 * fixed point with 2 bits after the binary point (scaled by
5603 * 4). The following is equivalent to rfc793 smoothing with
5604 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5605 * replaces rfc793's wired-in beta.
5606 * dev*4 = dev*4 + (|actual - expected| - dev)
5612 delta -= (peer->rtt_dev >> 2);
5613 peer->rtt_dev += delta;
5615 /* I don't have a stored RTT so I start with this value. Since I'm
5616 * probably just starting a call, and will be pushing more data down
5617 * this, I expect congestion to increase rapidly. So I fudge a
5618 * little, and I set deviance to half the rtt. In practice,
5619 * deviance tends to approach something a little less than
5620 * half the smoothed rtt. */
5621 peer->rtt = (MSEC(rttp) << 3) + 8;
5622 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5624 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5625 * the other of these connections is usually in a user process, and can
5626 * be switched and/or swapped out. So on fast, reliable networks, the
5627 * timeout would otherwise be too short.
5629 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5630 clock_Zero(&(peer->timeout));
5631 clock_Addmsec(&(peer->timeout), rtt_timeout);
5633 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)));
5637 /* Find all server connections that have not been active for a long time, and
5640 rxi_ReapConnections(void)
5643 clock_GetTime(&now);
5645 /* Find server connection structures that haven't been used for
5646 * greater than rx_idleConnectionTime */
5648 struct rx_connection **conn_ptr, **conn_end;
5649 int i, havecalls = 0;
5650 MUTEX_ENTER(&rx_connHashTable_lock);
5651 for (conn_ptr = &rx_connHashTable[0], conn_end =
5652 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5654 struct rx_connection *conn, *next;
5655 struct rx_call *call;
5659 for (conn = *conn_ptr; conn; conn = next) {
5660 /* XXX -- Shouldn't the connection be locked? */
5663 for (i = 0; i < RX_MAXCALLS; i++) {
5664 call = conn->call[i];
5667 MUTEX_ENTER(&call->lock);
5668 #ifdef RX_ENABLE_LOCKS
5669 result = rxi_CheckCall(call, 1);
5670 #else /* RX_ENABLE_LOCKS */
5671 result = rxi_CheckCall(call);
5672 #endif /* RX_ENABLE_LOCKS */
5673 MUTEX_EXIT(&call->lock);
5675 /* If CheckCall freed the call, it might
5676 * have destroyed the connection as well,
5677 * which screws up the linked lists.
5683 if (conn->type == RX_SERVER_CONNECTION) {
5684 /* This only actually destroys the connection if
5685 * there are no outstanding calls */
5686 MUTEX_ENTER(&conn->conn_data_lock);
5687 if (!havecalls && !conn->refCount
5688 && ((conn->lastSendTime + rx_idleConnectionTime) <
5690 conn->refCount++; /* it will be decr in rx_DestroyConn */
5691 MUTEX_EXIT(&conn->conn_data_lock);
5692 #ifdef RX_ENABLE_LOCKS
5693 rxi_DestroyConnectionNoLock(conn);
5694 #else /* RX_ENABLE_LOCKS */
5695 rxi_DestroyConnection(conn);
5696 #endif /* RX_ENABLE_LOCKS */
5698 #ifdef RX_ENABLE_LOCKS
5700 MUTEX_EXIT(&conn->conn_data_lock);
5702 #endif /* RX_ENABLE_LOCKS */
5706 #ifdef RX_ENABLE_LOCKS
5707 while (rx_connCleanup_list) {
5708 struct rx_connection *conn;
5709 conn = rx_connCleanup_list;
5710 rx_connCleanup_list = rx_connCleanup_list->next;
5711 MUTEX_EXIT(&rx_connHashTable_lock);
5712 rxi_CleanupConnection(conn);
5713 MUTEX_ENTER(&rx_connHashTable_lock);
5715 MUTEX_EXIT(&rx_connHashTable_lock);
5716 #endif /* RX_ENABLE_LOCKS */
5719 /* Find any peer structures that haven't been used (haven't had an
5720 * associated connection) for greater than rx_idlePeerTime */
5722 struct rx_peer **peer_ptr, **peer_end;
5724 MUTEX_ENTER(&rx_rpc_stats);
5725 MUTEX_ENTER(&rx_peerHashTable_lock);
5726 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5727 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5729 struct rx_peer *peer, *next, *prev;
5730 for (prev = peer = *peer_ptr; peer; peer = next) {
5732 code = MUTEX_TRYENTER(&peer->peer_lock);
5733 if ((code) && (peer->refCount == 0)
5734 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5735 rx_interface_stat_p rpc_stat, nrpc_stat;
5737 MUTEX_EXIT(&peer->peer_lock);
5738 MUTEX_DESTROY(&peer->peer_lock);
5740 (&peer->rpcStats, rpc_stat, nrpc_stat,
5741 rx_interface_stat)) {
5742 unsigned int num_funcs;
5745 queue_Remove(&rpc_stat->queue_header);
5746 queue_Remove(&rpc_stat->all_peers);
5747 num_funcs = rpc_stat->stats[0].func_total;
5749 sizeof(rx_interface_stat_t) +
5750 rpc_stat->stats[0].func_total *
5751 sizeof(rx_function_entry_v1_t);
5753 rxi_Free(rpc_stat, space);
5754 rxi_rpc_peer_stat_cnt -= num_funcs;
5757 MUTEX_ENTER(&rx_stats_mutex);
5758 rx_stats.nPeerStructs--;
5759 MUTEX_EXIT(&rx_stats_mutex);
5760 if (peer == *peer_ptr) {
5767 MUTEX_EXIT(&peer->peer_lock);
5773 MUTEX_EXIT(&rx_peerHashTable_lock);
5774 MUTEX_EXIT(&rx_rpc_stats);
5777 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5778 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5779 * GC, just below. Really, we shouldn't have to keep moving packets from
5780 * one place to another, but instead ought to always know if we can
5781 * afford to hold onto a packet in its particular use. */
5782 MUTEX_ENTER(&rx_freePktQ_lock);
5783 if (rx_waitingForPackets) {
5784 rx_waitingForPackets = 0;
5785 #ifdef RX_ENABLE_LOCKS
5786 CV_BROADCAST(&rx_waitingForPackets_cv);
5788 osi_rxWakeup(&rx_waitingForPackets);
5791 MUTEX_EXIT(&rx_freePktQ_lock);
5793 now.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5794 rxevent_Post(&now, rxi_ReapConnections, 0, 0);
5798 /* rxs_Release - This isn't strictly necessary but, since the macro name from
5799 * rx.h is sort of strange this is better. This is called with a security
5800 * object before it is discarded. Each connection using a security object has
5801 * its own refcount to the object so it won't actually be freed until the last
5802 * connection is destroyed.
5804 * This is the only rxs module call. A hold could also be written but no one
5808 rxs_Release(struct rx_securityClass *aobj)
5810 return RXS_Close(aobj);
5814 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
5815 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
5816 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
5817 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
5819 /* Adjust our estimate of the transmission rate to this peer, given
5820 * that the packet p was just acked. We can adjust peer->timeout and
5821 * call->twind. Pragmatically, this is called
5822 * only with packets of maximal length.
5823 * Called with peer and call locked.
5827 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
5828 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
5830 afs_int32 xferSize, xferMs;
5831 register afs_int32 minTime;
5834 /* Count down packets */
5835 if (peer->rateFlag > 0)
5837 /* Do nothing until we're enabled */
5838 if (peer->rateFlag != 0)
5843 /* Count only when the ack seems legitimate */
5844 switch (ackReason) {
5845 case RX_ACK_REQUESTED:
5847 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
5851 case RX_ACK_PING_RESPONSE:
5852 if (p) /* want the response to ping-request, not data send */
5854 clock_GetTime(&newTO);
5855 if (clock_Gt(&newTO, &call->pingRequestTime)) {
5856 clock_Sub(&newTO, &call->pingRequestTime);
5857 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
5861 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
5868 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));
5870 /* Track only packets that are big enough. */
5871 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
5875 /* absorb RTT data (in milliseconds) for these big packets */
5876 if (peer->smRtt == 0) {
5877 peer->smRtt = xferMs;
5879 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
5884 if (peer->countDown) {
5888 peer->countDown = 10; /* recalculate only every so often */
5890 /* In practice, we can measure only the RTT for full packets,
5891 * because of the way Rx acks the data that it receives. (If it's
5892 * smaller than a full packet, it often gets implicitly acked
5893 * either by the call response (from a server) or by the next call
5894 * (from a client), and either case confuses transmission times
5895 * with processing times.) Therefore, replace the above
5896 * more-sophisticated processing with a simpler version, where the
5897 * smoothed RTT is kept for full-size packets, and the time to
5898 * transmit a windowful of full-size packets is simply RTT *
5899 * windowSize. Again, we take two steps:
5900 - ensure the timeout is large enough for a single packet's RTT;
5901 - ensure that the window is small enough to fit in the desired timeout.*/
5903 /* First, the timeout check. */
5904 minTime = peer->smRtt;
5905 /* Get a reasonable estimate for a timeout period */
5907 newTO.sec = minTime / 1000;
5908 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
5910 /* Increase the timeout period so that we can always do at least
5911 * one packet exchange */
5912 if (clock_Gt(&newTO, &peer->timeout)) {
5914 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));
5916 peer->timeout = newTO;
5919 /* Now, get an estimate for the transmit window size. */
5920 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
5921 /* Now, convert to the number of full packets that could fit in a
5922 * reasonable fraction of that interval */
5923 minTime /= (peer->smRtt << 1);
5924 xferSize = minTime; /* (make a copy) */
5926 /* Now clamp the size to reasonable bounds. */
5929 else if (minTime > rx_Window)
5930 minTime = rx_Window;
5931 /* if (minTime != peer->maxWindow) {
5932 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
5933 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
5934 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
5936 peer->maxWindow = minTime;
5937 elide... call->twind = minTime;
5941 /* Cut back on the peer timeout if it had earlier grown unreasonably.
5942 * Discern this by calculating the timeout necessary for rx_Window
5944 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
5945 /* calculate estimate for transmission interval in milliseconds */
5946 minTime = rx_Window * peer->smRtt;
5947 if (minTime < 1000) {
5948 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
5949 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
5950 peer->timeout.usec, peer->smRtt, peer->packetSize));
5952 newTO.sec = 0; /* cut back on timeout by half a second */
5953 newTO.usec = 500000;
5954 clock_Sub(&peer->timeout, &newTO);
5959 } /* end of rxi_ComputeRate */
5960 #endif /* ADAPT_WINDOW */
5968 /* Don't call this debugging routine directly; use dpf */
5970 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
5971 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
5975 clock_GetTime(&now);
5976 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
5977 (unsigned int)now.usec / 1000);
5978 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
5986 * This function is used to process the rx_stats structure that is local
5987 * to a process as well as an rx_stats structure received from a remote
5988 * process (via rxdebug). Therefore, it needs to do minimal version
5992 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
5993 afs_int32 freePackets, char version)
5997 if (size != sizeof(struct rx_stats)) {
5999 "Unexpected size of stats structure: was %d, expected %d\n",
6000 size, sizeof(struct rx_stats));
6003 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6006 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6007 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6008 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6009 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6010 s->specialPktAllocFailures);
6012 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6013 s->receivePktAllocFailures, s->sendPktAllocFailures,
6014 s->specialPktAllocFailures);
6018 " greedy %d, " "bogusReads %d (last from host %x), "
6019 "noPackets %d, " "noBuffers %d, " "selects %d, "
6020 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6021 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6022 s->selects, s->sendSelects);
6024 fprintf(file, " packets read: ");
6025 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6026 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6028 fprintf(file, "\n");
6031 " other read counters: data %d, " "ack %d, " "dup %d "
6032 "spurious %d " "dally %d\n", s->dataPacketsRead,
6033 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6034 s->ignorePacketDally);
6036 fprintf(file, " packets sent: ");
6037 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6038 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6040 fprintf(file, "\n");
6043 " other send counters: ack %d, " "data %d (not resends), "
6044 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6045 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6046 s->dataPacketsPushed, s->ignoreAckedPacket);
6049 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6050 s->netSendFailures, (int)s->fatalErrors);
6052 if (s->nRttSamples) {
6053 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6054 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6056 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6057 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6061 " %d server connections, " "%d client connections, "
6062 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6063 s->nServerConns, s->nClientConns, s->nPeerStructs,
6064 s->nCallStructs, s->nFreeCallStructs);
6066 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6067 fprintf(file, " %d clock updates\n", clock_nUpdates);
6072 /* for backward compatibility */
6074 rx_PrintStats(FILE * file)
6076 MUTEX_ENTER(&rx_stats_mutex);
6077 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6079 MUTEX_EXIT(&rx_stats_mutex);
6083 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6085 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6086 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6087 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6090 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6091 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6092 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6095 " Packet size %d, " "max in packet skew %d, "
6096 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6097 (int)peer->outPacketSkew);
6100 #ifdef AFS_PTHREAD_ENV
6102 * This mutex protects the following static variables:
6106 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6107 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6109 #define LOCK_RX_DEBUG
6110 #define UNLOCK_RX_DEBUG
6111 #endif /* AFS_PTHREAD_ENV */
6114 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6115 u_char type, void *inputData, size_t inputLength,
6116 void *outputData, size_t outputLength)
6118 static afs_int32 counter = 100;
6120 struct rx_header theader;
6122 register afs_int32 code;
6124 struct sockaddr_in taddr, faddr;
6129 endTime = time(0) + 20; /* try for 20 seconds */
6133 tp = &tbuffer[sizeof(struct rx_header)];
6134 taddr.sin_family = AF_INET;
6135 taddr.sin_port = remotePort;
6136 taddr.sin_addr.s_addr = remoteAddr;
6137 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6138 taddr.sin_len = sizeof(struct sockaddr_in);
6141 memset(&theader, 0, sizeof(theader));
6142 theader.epoch = htonl(999);
6144 theader.callNumber = htonl(counter);
6147 theader.type = type;
6148 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6149 theader.serviceId = 0;
6151 memcpy(tbuffer, &theader, sizeof(theader));
6152 memcpy(tp, inputData, inputLength);
6154 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6155 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6157 /* see if there's a packet available */
6159 FD_SET(socket, &imask);
6162 code = select(socket + 1, &imask, 0, 0, &tv);
6163 if (code == 1 && FD_ISSET(socket, &imask)) {
6164 /* now receive a packet */
6165 faddrLen = sizeof(struct sockaddr_in);
6167 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6168 (struct sockaddr *)&faddr, &faddrLen);
6171 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6172 if (counter == ntohl(theader.callNumber))
6177 /* see if we've timed out */
6178 if (endTime < time(0))
6181 code -= sizeof(struct rx_header);
6182 if (code > outputLength)
6183 code = outputLength;
6184 memcpy(outputData, tp, code);
6189 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6190 afs_uint16 remotePort, struct rx_debugStats * stat,
6191 afs_uint32 * supportedValues)
6193 struct rx_debugIn in;
6196 *supportedValues = 0;
6197 in.type = htonl(RX_DEBUGI_GETSTATS);
6200 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6201 &in, sizeof(in), stat, sizeof(*stat));
6204 * If the call was successful, fixup the version and indicate
6205 * what contents of the stat structure are valid.
6206 * Also do net to host conversion of fields here.
6210 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6211 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6213 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6214 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6216 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6217 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6219 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6220 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6222 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6223 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6225 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6226 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6228 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6229 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6231 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6232 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6235 stat->nFreePackets = ntohl(stat->nFreePackets);
6236 stat->packetReclaims = ntohl(stat->packetReclaims);
6237 stat->callsExecuted = ntohl(stat->callsExecuted);
6238 stat->nWaiting = ntohl(stat->nWaiting);
6239 stat->idleThreads = ntohl(stat->idleThreads);
6246 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6247 afs_uint16 remotePort, struct rx_stats * stat,
6248 afs_uint32 * supportedValues)
6250 struct rx_debugIn in;
6251 afs_int32 *lp = (afs_int32 *) stat;
6256 * supportedValues is currently unused, but added to allow future
6257 * versioning of this function.
6260 *supportedValues = 0;
6261 in.type = htonl(RX_DEBUGI_RXSTATS);
6263 memset(stat, 0, sizeof(*stat));
6265 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6266 &in, sizeof(in), stat, sizeof(*stat));
6271 * Do net to host conversion here
6274 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6283 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6284 afs_uint16 remotePort, size_t version_length,
6288 return MakeDebugCall(socket, remoteAddr, remotePort,
6289 RX_PACKET_TYPE_VERSION, a, 1, version,
6294 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6295 afs_uint16 remotePort, afs_int32 * nextConnection,
6296 int allConnections, afs_uint32 debugSupportedValues,
6297 struct rx_debugConn * conn,
6298 afs_uint32 * supportedValues)
6300 struct rx_debugIn in;
6305 * supportedValues is currently unused, but added to allow future
6306 * versioning of this function.
6309 *supportedValues = 0;
6310 if (allConnections) {
6311 in.type = htonl(RX_DEBUGI_GETALLCONN);
6313 in.type = htonl(RX_DEBUGI_GETCONN);
6315 in.index = htonl(*nextConnection);
6316 memset(conn, 0, sizeof(*conn));
6318 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6319 &in, sizeof(in), conn, sizeof(*conn));
6322 *nextConnection += 1;
6325 * Convert old connection format to new structure.
6328 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6329 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6330 #define MOVEvL(a) (conn->a = vL->a)
6332 /* any old or unrecognized version... */
6333 for (i = 0; i < RX_MAXCALLS; i++) {
6334 MOVEvL(callState[i]);
6335 MOVEvL(callMode[i]);
6336 MOVEvL(callFlags[i]);
6337 MOVEvL(callOther[i]);
6339 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6340 MOVEvL(secStats.type);
6341 MOVEvL(secStats.level);
6342 MOVEvL(secStats.flags);
6343 MOVEvL(secStats.expires);
6344 MOVEvL(secStats.packetsReceived);
6345 MOVEvL(secStats.packetsSent);
6346 MOVEvL(secStats.bytesReceived);
6347 MOVEvL(secStats.bytesSent);
6352 * Do net to host conversion here
6354 * I don't convert host or port since we are most likely
6355 * going to want these in NBO.
6357 conn->cid = ntohl(conn->cid);
6358 conn->serial = ntohl(conn->serial);
6359 for (i = 0; i < RX_MAXCALLS; i++) {
6360 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6362 conn->error = ntohl(conn->error);
6363 conn->secStats.flags = ntohl(conn->secStats.flags);
6364 conn->secStats.expires = ntohl(conn->secStats.expires);
6365 conn->secStats.packetsReceived =
6366 ntohl(conn->secStats.packetsReceived);
6367 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6368 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6369 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6370 conn->epoch = ntohl(conn->epoch);
6371 conn->natMTU = ntohl(conn->natMTU);
6378 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6379 afs_uint16 remotePort, afs_int32 * nextPeer,
6380 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6381 afs_uint32 * supportedValues)
6383 struct rx_debugIn in;
6387 * supportedValues is currently unused, but added to allow future
6388 * versioning of this function.
6391 *supportedValues = 0;
6392 in.type = htonl(RX_DEBUGI_GETPEER);
6393 in.index = htonl(*nextPeer);
6394 memset(peer, 0, sizeof(*peer));
6396 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6397 &in, sizeof(in), peer, sizeof(*peer));
6403 * Do net to host conversion here
6405 * I don't convert host or port since we are most likely
6406 * going to want these in NBO.
6408 peer->ifMTU = ntohs(peer->ifMTU);
6409 peer->idleWhen = ntohl(peer->idleWhen);
6410 peer->refCount = ntohs(peer->refCount);
6411 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6412 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6413 peer->rtt = ntohl(peer->rtt);
6414 peer->rtt_dev = ntohl(peer->rtt_dev);
6415 peer->timeout.sec = ntohl(peer->timeout.sec);
6416 peer->timeout.usec = ntohl(peer->timeout.usec);
6417 peer->nSent = ntohl(peer->nSent);
6418 peer->reSends = ntohl(peer->reSends);
6419 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6420 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6421 peer->rateFlag = ntohl(peer->rateFlag);
6422 peer->natMTU = ntohs(peer->natMTU);
6423 peer->maxMTU = ntohs(peer->maxMTU);
6424 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6425 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6426 peer->MTU = ntohs(peer->MTU);
6427 peer->cwind = ntohs(peer->cwind);
6428 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6429 peer->congestSeq = ntohs(peer->congestSeq);
6430 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6431 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6432 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6433 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6438 #endif /* RXDEBUG */
6443 struct rx_serverQueueEntry *np;
6446 register struct rx_call *call;
6447 register struct rx_serverQueueEntry *sq;
6451 if (rxinit_status == 1) {
6453 return; /* Already shutdown. */
6457 #ifndef AFS_PTHREAD_ENV
6458 FD_ZERO(&rx_selectMask);
6459 #endif /* AFS_PTHREAD_ENV */
6460 rxi_dataQuota = RX_MAX_QUOTA;
6461 #ifndef AFS_PTHREAD_ENV
6463 #endif /* AFS_PTHREAD_ENV */
6466 #ifndef AFS_PTHREAD_ENV
6467 #ifndef AFS_USE_GETTIMEOFDAY
6469 #endif /* AFS_USE_GETTIMEOFDAY */
6470 #endif /* AFS_PTHREAD_ENV */
6472 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6473 call = queue_First(&rx_freeCallQueue, rx_call);
6475 rxi_Free(call, sizeof(struct rx_call));
6478 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6479 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6485 struct rx_peer **peer_ptr, **peer_end;
6486 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6487 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6489 struct rx_peer *peer, *next;
6490 for (peer = *peer_ptr; peer; peer = next) {
6491 rx_interface_stat_p rpc_stat, nrpc_stat;
6494 (&peer->rpcStats, rpc_stat, nrpc_stat,
6495 rx_interface_stat)) {
6496 unsigned int num_funcs;
6499 queue_Remove(&rpc_stat->queue_header);
6500 queue_Remove(&rpc_stat->all_peers);
6501 num_funcs = rpc_stat->stats[0].func_total;
6503 sizeof(rx_interface_stat_t) +
6504 rpc_stat->stats[0].func_total *
6505 sizeof(rx_function_entry_v1_t);
6507 rxi_Free(rpc_stat, space);
6508 MUTEX_ENTER(&rx_rpc_stats);
6509 rxi_rpc_peer_stat_cnt -= num_funcs;
6510 MUTEX_EXIT(&rx_rpc_stats);
6514 MUTEX_ENTER(&rx_stats_mutex);
6515 rx_stats.nPeerStructs--;
6516 MUTEX_EXIT(&rx_stats_mutex);
6520 for (i = 0; i < RX_MAX_SERVICES; i++) {
6522 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6524 for (i = 0; i < rx_hashTableSize; i++) {
6525 register struct rx_connection *tc, *ntc;
6526 MUTEX_ENTER(&rx_connHashTable_lock);
6527 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6529 for (j = 0; j < RX_MAXCALLS; j++) {
6531 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6534 rxi_Free(tc, sizeof(*tc));
6536 MUTEX_EXIT(&rx_connHashTable_lock);
6539 MUTEX_ENTER(&freeSQEList_lock);
6541 while ((np = rx_FreeSQEList)) {
6542 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6543 MUTEX_DESTROY(&np->lock);
6544 rxi_Free(np, sizeof(*np));
6547 MUTEX_EXIT(&freeSQEList_lock);
6548 MUTEX_DESTROY(&freeSQEList_lock);
6549 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6550 MUTEX_DESTROY(&rx_connHashTable_lock);
6551 MUTEX_DESTROY(&rx_peerHashTable_lock);
6552 MUTEX_DESTROY(&rx_serverPool_lock);
6554 osi_Free(rx_connHashTable,
6555 rx_hashTableSize * sizeof(struct rx_connection *));
6556 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6558 UNPIN(rx_connHashTable,
6559 rx_hashTableSize * sizeof(struct rx_connection *));
6560 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6562 rxi_FreeAllPackets();
6564 MUTEX_ENTER(&rx_stats_mutex);
6565 rxi_dataQuota = RX_MAX_QUOTA;
6566 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6567 MUTEX_EXIT(&rx_stats_mutex);
6573 #ifdef RX_ENABLE_LOCKS
6575 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6577 if (!MUTEX_ISMINE(lockaddr))
6578 osi_Panic("Lock not held: %s", msg);
6580 #endif /* RX_ENABLE_LOCKS */
6585 * Routines to implement connection specific data.
6589 rx_KeyCreate(rx_destructor_t rtn)
6592 MUTEX_ENTER(&rxi_keyCreate_lock);
6593 key = rxi_keyCreate_counter++;
6594 rxi_keyCreate_destructor = (rx_destructor_t *)
6595 realloc((void *)rxi_keyCreate_destructor,
6596 (key + 1) * sizeof(rx_destructor_t));
6597 rxi_keyCreate_destructor[key] = rtn;
6598 MUTEX_EXIT(&rxi_keyCreate_lock);
6603 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6606 MUTEX_ENTER(&conn->conn_data_lock);
6607 if (!conn->specific) {
6608 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6609 for (i = 0; i < key; i++)
6610 conn->specific[i] = NULL;
6611 conn->nSpecific = key + 1;
6612 conn->specific[key] = ptr;
6613 } else if (key >= conn->nSpecific) {
6614 conn->specific = (void **)
6615 realloc(conn->specific, (key + 1) * sizeof(void *));
6616 for (i = conn->nSpecific; i < key; i++)
6617 conn->specific[i] = NULL;
6618 conn->nSpecific = key + 1;
6619 conn->specific[key] = ptr;
6621 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6622 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6623 conn->specific[key] = ptr;
6625 MUTEX_EXIT(&conn->conn_data_lock);
6629 rx_GetSpecific(struct rx_connection *conn, int key)
6632 MUTEX_ENTER(&conn->conn_data_lock);
6633 if (key >= conn->nSpecific)
6636 ptr = conn->specific[key];
6637 MUTEX_EXIT(&conn->conn_data_lock);
6641 #endif /* !KERNEL */
6644 * processStats is a queue used to store the statistics for the local
6645 * process. Its contents are similar to the contents of the rpcStats
6646 * queue on a rx_peer structure, but the actual data stored within
6647 * this queue contains totals across the lifetime of the process (assuming
6648 * the stats have not been reset) - unlike the per peer structures
6649 * which can come and go based upon the peer lifetime.
6652 static struct rx_queue processStats = { &processStats, &processStats };
6655 * peerStats is a queue used to store the statistics for all peer structs.
6656 * Its contents are the union of all the peer rpcStats queues.
6659 static struct rx_queue peerStats = { &peerStats, &peerStats };
6662 * rxi_monitor_processStats is used to turn process wide stat collection
6666 static int rxi_monitor_processStats = 0;
6669 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6672 static int rxi_monitor_peerStats = 0;
6675 * rxi_AddRpcStat - given all of the information for a particular rpc
6676 * call, create (if needed) and update the stat totals for the rpc.
6680 * IN stats - the queue of stats that will be updated with the new value
6682 * IN rxInterface - a unique number that identifies the rpc interface
6684 * IN currentFunc - the index of the function being invoked
6686 * IN totalFunc - the total number of functions in this interface
6688 * IN queueTime - the amount of time this function waited for a thread
6690 * IN execTime - the amount of time this function invocation took to execute
6692 * IN bytesSent - the number bytes sent by this invocation
6694 * IN bytesRcvd - the number bytes received by this invocation
6696 * IN isServer - if true, this invocation was made to a server
6698 * IN remoteHost - the ip address of the remote host
6700 * IN remotePort - the port of the remote host
6702 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6704 * INOUT counter - if a new stats structure is allocated, the counter will
6705 * be updated with the new number of allocated stat structures
6713 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6714 afs_uint32 currentFunc, afs_uint32 totalFunc,
6715 struct clock *queueTime, struct clock *execTime,
6716 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6717 afs_uint32 remoteHost, afs_uint32 remotePort,
6718 int addToPeerList, unsigned int *counter)
6721 rx_interface_stat_p rpc_stat, nrpc_stat;
6724 * See if there's already a structure for this interface
6727 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6728 if ((rpc_stat->stats[0].interfaceId == rxInterface)
6729 && (rpc_stat->stats[0].remote_is_server == isServer))
6734 * Didn't find a match so allocate a new structure and add it to the
6738 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
6739 || (rpc_stat->stats[0].interfaceId != rxInterface)
6740 || (rpc_stat->stats[0].remote_is_server != isServer)) {
6745 sizeof(rx_interface_stat_t) +
6746 totalFunc * sizeof(rx_function_entry_v1_t);
6748 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
6749 if (rpc_stat == NULL) {
6753 *counter += totalFunc;
6754 for (i = 0; i < totalFunc; i++) {
6755 rpc_stat->stats[i].remote_peer = remoteHost;
6756 rpc_stat->stats[i].remote_port = remotePort;
6757 rpc_stat->stats[i].remote_is_server = isServer;
6758 rpc_stat->stats[i].interfaceId = rxInterface;
6759 rpc_stat->stats[i].func_total = totalFunc;
6760 rpc_stat->stats[i].func_index = i;
6761 hzero(rpc_stat->stats[i].invocations);
6762 hzero(rpc_stat->stats[i].bytes_sent);
6763 hzero(rpc_stat->stats[i].bytes_rcvd);
6764 rpc_stat->stats[i].queue_time_sum.sec = 0;
6765 rpc_stat->stats[i].queue_time_sum.usec = 0;
6766 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
6767 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
6768 rpc_stat->stats[i].queue_time_min.sec = 9999999;
6769 rpc_stat->stats[i].queue_time_min.usec = 9999999;
6770 rpc_stat->stats[i].queue_time_max.sec = 0;
6771 rpc_stat->stats[i].queue_time_max.usec = 0;
6772 rpc_stat->stats[i].execution_time_sum.sec = 0;
6773 rpc_stat->stats[i].execution_time_sum.usec = 0;
6774 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
6775 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
6776 rpc_stat->stats[i].execution_time_min.sec = 9999999;
6777 rpc_stat->stats[i].execution_time_min.usec = 9999999;
6778 rpc_stat->stats[i].execution_time_max.sec = 0;
6779 rpc_stat->stats[i].execution_time_max.usec = 0;
6781 queue_Prepend(stats, rpc_stat);
6782 if (addToPeerList) {
6783 queue_Prepend(&peerStats, &rpc_stat->all_peers);
6788 * Increment the stats for this function
6791 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
6792 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
6793 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
6794 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
6795 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
6796 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
6797 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
6799 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
6800 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
6802 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
6803 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
6805 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
6806 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
6808 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
6809 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
6817 * rx_IncrementTimeAndCount - increment the times and count for a particular
6822 * IN peer - the peer who invoked the rpc
6824 * IN rxInterface - a unique number that identifies the rpc interface
6826 * IN currentFunc - the index of the function being invoked
6828 * IN totalFunc - the total number of functions in this interface
6830 * IN queueTime - the amount of time this function waited for a thread
6832 * IN execTime - the amount of time this function invocation took to execute
6834 * IN bytesSent - the number bytes sent by this invocation
6836 * IN bytesRcvd - the number bytes received by this invocation
6838 * IN isServer - if true, this invocation was made to a server
6846 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
6847 afs_uint32 currentFunc, afs_uint32 totalFunc,
6848 struct clock *queueTime, struct clock *execTime,
6849 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
6853 MUTEX_ENTER(&rx_rpc_stats);
6854 MUTEX_ENTER(&peer->peer_lock);
6856 if (rxi_monitor_peerStats) {
6857 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
6858 queueTime, execTime, bytesSent, bytesRcvd, isServer,
6859 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
6862 if (rxi_monitor_processStats) {
6863 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
6864 queueTime, execTime, bytesSent, bytesRcvd, isServer,
6865 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
6868 MUTEX_EXIT(&peer->peer_lock);
6869 MUTEX_EXIT(&rx_rpc_stats);
6874 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
6878 * IN callerVersion - the rpc stat version of the caller.
6880 * IN count - the number of entries to marshall.
6882 * IN stats - pointer to stats to be marshalled.
6884 * OUT ptr - Where to store the marshalled data.
6891 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
6892 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
6898 * We only support the first version
6900 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
6901 *(ptr++) = stats->remote_peer;
6902 *(ptr++) = stats->remote_port;
6903 *(ptr++) = stats->remote_is_server;
6904 *(ptr++) = stats->interfaceId;
6905 *(ptr++) = stats->func_total;
6906 *(ptr++) = stats->func_index;
6907 *(ptr++) = hgethi(stats->invocations);
6908 *(ptr++) = hgetlo(stats->invocations);
6909 *(ptr++) = hgethi(stats->bytes_sent);
6910 *(ptr++) = hgetlo(stats->bytes_sent);
6911 *(ptr++) = hgethi(stats->bytes_rcvd);
6912 *(ptr++) = hgetlo(stats->bytes_rcvd);
6913 *(ptr++) = stats->queue_time_sum.sec;
6914 *(ptr++) = stats->queue_time_sum.usec;
6915 *(ptr++) = stats->queue_time_sum_sqr.sec;
6916 *(ptr++) = stats->queue_time_sum_sqr.usec;
6917 *(ptr++) = stats->queue_time_min.sec;
6918 *(ptr++) = stats->queue_time_min.usec;
6919 *(ptr++) = stats->queue_time_max.sec;
6920 *(ptr++) = stats->queue_time_max.usec;
6921 *(ptr++) = stats->execution_time_sum.sec;
6922 *(ptr++) = stats->execution_time_sum.usec;
6923 *(ptr++) = stats->execution_time_sum_sqr.sec;
6924 *(ptr++) = stats->execution_time_sum_sqr.usec;
6925 *(ptr++) = stats->execution_time_min.sec;
6926 *(ptr++) = stats->execution_time_min.usec;
6927 *(ptr++) = stats->execution_time_max.sec;
6928 *(ptr++) = stats->execution_time_max.usec;
6934 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
6939 * IN callerVersion - the rpc stat version of the caller
6941 * OUT myVersion - the rpc stat version of this function
6943 * OUT clock_sec - local time seconds
6945 * OUT clock_usec - local time microseconds
6947 * OUT allocSize - the number of bytes allocated to contain stats
6949 * OUT statCount - the number stats retrieved from this process.
6951 * OUT stats - the actual stats retrieved from this process.
6955 * Returns void. If successful, stats will != NULL.
6959 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
6960 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
6961 size_t * allocSize, afs_uint32 * statCount,
6962 afs_uint32 ** stats)
6972 *myVersion = RX_STATS_RETRIEVAL_VERSION;
6975 * Check to see if stats are enabled
6978 MUTEX_ENTER(&rx_rpc_stats);
6979 if (!rxi_monitor_processStats) {
6980 MUTEX_EXIT(&rx_rpc_stats);
6984 clock_GetTime(&now);
6985 *clock_sec = now.sec;
6986 *clock_usec = now.usec;
6989 * Allocate the space based upon the caller version
6991 * If the client is at an older version than we are,
6992 * we return the statistic data in the older data format, but
6993 * we still return our version number so the client knows we
6994 * are maintaining more data than it can retrieve.
6997 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
6998 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
6999 *statCount = rxi_rpc_process_stat_cnt;
7002 * This can't happen yet, but in the future version changes
7003 * can be handled by adding additional code here
7007 if (space > (size_t) 0) {
7009 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7012 rx_interface_stat_p rpc_stat, nrpc_stat;
7016 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7018 * Copy the data based upon the caller version
7020 rx_MarshallProcessRPCStats(callerVersion,
7021 rpc_stat->stats[0].func_total,
7022 rpc_stat->stats, &ptr);
7028 MUTEX_EXIT(&rx_rpc_stats);
7033 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7037 * IN callerVersion - the rpc stat version of the caller
7039 * OUT myVersion - the rpc stat version of this function
7041 * OUT clock_sec - local time seconds
7043 * OUT clock_usec - local time microseconds
7045 * OUT allocSize - the number of bytes allocated to contain stats
7047 * OUT statCount - the number of stats retrieved from the individual
7050 * OUT stats - the actual stats retrieved from the individual peer structures.
7054 * Returns void. If successful, stats will != NULL.
7058 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7059 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7060 size_t * allocSize, afs_uint32 * statCount,
7061 afs_uint32 ** stats)
7071 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7074 * Check to see if stats are enabled
7077 MUTEX_ENTER(&rx_rpc_stats);
7078 if (!rxi_monitor_peerStats) {
7079 MUTEX_EXIT(&rx_rpc_stats);
7083 clock_GetTime(&now);
7084 *clock_sec = now.sec;
7085 *clock_usec = now.usec;
7088 * Allocate the space based upon the caller version
7090 * If the client is at an older version than we are,
7091 * we return the statistic data in the older data format, but
7092 * we still return our version number so the client knows we
7093 * are maintaining more data than it can retrieve.
7096 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7097 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7098 *statCount = rxi_rpc_peer_stat_cnt;
7101 * This can't happen yet, but in the future version changes
7102 * can be handled by adding additional code here
7106 if (space > (size_t) 0) {
7108 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7111 rx_interface_stat_p rpc_stat, nrpc_stat;
7115 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7117 * We have to fix the offset of rpc_stat since we are
7118 * keeping this structure on two rx_queues. The rx_queue
7119 * package assumes that the rx_queue member is the first
7120 * member of the structure. That is, rx_queue assumes that
7121 * any one item is only on one queue at a time. We are
7122 * breaking that assumption and so we have to do a little
7123 * math to fix our pointers.
7126 fix_offset = (char *)rpc_stat;
7127 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7128 rpc_stat = (rx_interface_stat_p) fix_offset;
7131 * Copy the data based upon the caller version
7133 rx_MarshallProcessRPCStats(callerVersion,
7134 rpc_stat->stats[0].func_total,
7135 rpc_stat->stats, &ptr);
7141 MUTEX_EXIT(&rx_rpc_stats);
7146 * rx_FreeRPCStats - free memory allocated by
7147 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7151 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7152 * rx_RetrievePeerRPCStats
7154 * IN allocSize - the number of bytes in stats.
7162 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7164 rxi_Free(stats, allocSize);
7168 * rx_queryProcessRPCStats - see if process rpc stat collection is
7169 * currently enabled.
7175 * Returns 0 if stats are not enabled != 0 otherwise
7179 rx_queryProcessRPCStats(void)
7182 MUTEX_ENTER(&rx_rpc_stats);
7183 rc = rxi_monitor_processStats;
7184 MUTEX_EXIT(&rx_rpc_stats);
7189 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7195 * Returns 0 if stats are not enabled != 0 otherwise
7199 rx_queryPeerRPCStats(void)
7202 MUTEX_ENTER(&rx_rpc_stats);
7203 rc = rxi_monitor_peerStats;
7204 MUTEX_EXIT(&rx_rpc_stats);
7209 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7219 rx_enableProcessRPCStats(void)
7221 MUTEX_ENTER(&rx_rpc_stats);
7222 rx_enable_stats = 1;
7223 rxi_monitor_processStats = 1;
7224 MUTEX_EXIT(&rx_rpc_stats);
7228 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7238 rx_enablePeerRPCStats(void)
7240 MUTEX_ENTER(&rx_rpc_stats);
7241 rx_enable_stats = 1;
7242 rxi_monitor_peerStats = 1;
7243 MUTEX_EXIT(&rx_rpc_stats);
7247 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7257 rx_disableProcessRPCStats(void)
7259 rx_interface_stat_p rpc_stat, nrpc_stat;
7262 MUTEX_ENTER(&rx_rpc_stats);
7265 * Turn off process statistics and if peer stats is also off, turn
7269 rxi_monitor_processStats = 0;
7270 if (rxi_monitor_peerStats == 0) {
7271 rx_enable_stats = 0;
7274 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7275 unsigned int num_funcs = 0;
7278 queue_Remove(rpc_stat);
7279 num_funcs = rpc_stat->stats[0].func_total;
7281 sizeof(rx_interface_stat_t) +
7282 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7284 rxi_Free(rpc_stat, space);
7285 rxi_rpc_process_stat_cnt -= num_funcs;
7287 MUTEX_EXIT(&rx_rpc_stats);
7291 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7301 rx_disablePeerRPCStats(void)
7303 struct rx_peer **peer_ptr, **peer_end;
7306 MUTEX_ENTER(&rx_rpc_stats);
7309 * Turn off peer statistics and if process stats is also off, turn
7313 rxi_monitor_peerStats = 0;
7314 if (rxi_monitor_processStats == 0) {
7315 rx_enable_stats = 0;
7318 MUTEX_ENTER(&rx_peerHashTable_lock);
7319 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7320 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7322 struct rx_peer *peer, *next, *prev;
7323 for (prev = peer = *peer_ptr; peer; peer = next) {
7325 code = MUTEX_TRYENTER(&peer->peer_lock);
7327 rx_interface_stat_p rpc_stat, nrpc_stat;
7330 (&peer->rpcStats, rpc_stat, nrpc_stat,
7331 rx_interface_stat)) {
7332 unsigned int num_funcs = 0;
7335 queue_Remove(&rpc_stat->queue_header);
7336 queue_Remove(&rpc_stat->all_peers);
7337 num_funcs = rpc_stat->stats[0].func_total;
7339 sizeof(rx_interface_stat_t) +
7340 rpc_stat->stats[0].func_total *
7341 sizeof(rx_function_entry_v1_t);
7343 rxi_Free(rpc_stat, space);
7344 rxi_rpc_peer_stat_cnt -= num_funcs;
7346 MUTEX_EXIT(&peer->peer_lock);
7347 if (prev == *peer_ptr) {
7357 MUTEX_EXIT(&rx_peerHashTable_lock);
7358 MUTEX_EXIT(&rx_rpc_stats);
7362 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7367 * IN clearFlag - flag indicating which stats to clear
7375 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7377 rx_interface_stat_p rpc_stat, nrpc_stat;
7379 MUTEX_ENTER(&rx_rpc_stats);
7381 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7382 unsigned int num_funcs = 0, i;
7383 num_funcs = rpc_stat->stats[0].func_total;
7384 for (i = 0; i < num_funcs; i++) {
7385 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7386 hzero(rpc_stat->stats[i].invocations);
7388 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7389 hzero(rpc_stat->stats[i].bytes_sent);
7391 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7392 hzero(rpc_stat->stats[i].bytes_rcvd);
7394 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7395 rpc_stat->stats[i].queue_time_sum.sec = 0;
7396 rpc_stat->stats[i].queue_time_sum.usec = 0;
7398 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7399 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7400 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7402 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7403 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7404 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7406 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7407 rpc_stat->stats[i].queue_time_max.sec = 0;
7408 rpc_stat->stats[i].queue_time_max.usec = 0;
7410 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7411 rpc_stat->stats[i].execution_time_sum.sec = 0;
7412 rpc_stat->stats[i].execution_time_sum.usec = 0;
7414 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7415 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7416 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7418 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7419 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7420 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7422 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7423 rpc_stat->stats[i].execution_time_max.sec = 0;
7424 rpc_stat->stats[i].execution_time_max.usec = 0;
7429 MUTEX_EXIT(&rx_rpc_stats);
7433 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7438 * IN clearFlag - flag indicating which stats to clear
7446 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7448 rx_interface_stat_p rpc_stat, nrpc_stat;
7450 MUTEX_ENTER(&rx_rpc_stats);
7452 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7453 unsigned int num_funcs = 0, i;
7456 * We have to fix the offset of rpc_stat since we are
7457 * keeping this structure on two rx_queues. The rx_queue
7458 * package assumes that the rx_queue member is the first
7459 * member of the structure. That is, rx_queue assumes that
7460 * any one item is only on one queue at a time. We are
7461 * breaking that assumption and so we have to do a little
7462 * math to fix our pointers.
7465 fix_offset = (char *)rpc_stat;
7466 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7467 rpc_stat = (rx_interface_stat_p) fix_offset;
7469 num_funcs = rpc_stat->stats[0].func_total;
7470 for (i = 0; i < num_funcs; i++) {
7471 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7472 hzero(rpc_stat->stats[i].invocations);
7474 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7475 hzero(rpc_stat->stats[i].bytes_sent);
7477 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7478 hzero(rpc_stat->stats[i].bytes_rcvd);
7480 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7481 rpc_stat->stats[i].queue_time_sum.sec = 0;
7482 rpc_stat->stats[i].queue_time_sum.usec = 0;
7484 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7485 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7486 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7488 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7489 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7490 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7492 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7493 rpc_stat->stats[i].queue_time_max.sec = 0;
7494 rpc_stat->stats[i].queue_time_max.usec = 0;
7496 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7497 rpc_stat->stats[i].execution_time_sum.sec = 0;
7498 rpc_stat->stats[i].execution_time_sum.usec = 0;
7500 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7501 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7502 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7504 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7505 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7506 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7508 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7509 rpc_stat->stats[i].execution_time_max.sec = 0;
7510 rpc_stat->stats[i].execution_time_max.usec = 0;
7515 MUTEX_EXIT(&rx_rpc_stats);
7519 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7520 * is authorized to enable/disable/clear RX statistics.
7522 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7525 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7527 rxi_rxstat_userok = proc;
7531 rx_RxStatUserOk(struct rx_call *call)
7533 if (!rxi_rxstat_userok)
7535 return rxi_rxstat_userok(call);