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>
87 # include <afs/afsutil.h>
88 # include <WINNT\afsreg.h>
90 # include <sys/socket.h>
91 # include <sys/file.h>
93 # include <sys/stat.h>
94 # include <netinet/in.h>
95 # include <sys/time.h>
99 # include "rx_clock.h"
100 # include "rx_queue.h"
101 # include "rx_globals.h"
102 # include "rx_trace.h"
103 # include <afs/rxgen_consts.h>
106 int (*registerProgram) () = 0;
107 int (*swapNameProgram) () = 0;
109 /* Local static routines */
110 static void rxi_DestroyConnectionNoLock(register struct rx_connection *conn);
111 #ifdef RX_ENABLE_LOCKS
112 static void rxi_SetAcksInTransmitQueue(register struct rx_call *call);
115 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
117 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
118 afs_int32 rxi_start_in_error;
120 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
123 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
124 * currently allocated within rx. This number is used to allocate the
125 * memory required to return the statistics when queried.
128 static unsigned int rxi_rpc_peer_stat_cnt;
131 * rxi_rpc_process_stat_cnt counts the total number of local process stat
132 * structures currently allocated within rx. The number is used to allocate
133 * the memory required to return the statistics when queried.
136 static unsigned int rxi_rpc_process_stat_cnt;
138 #if !defined(offsetof)
139 #include <stddef.h> /* for definition of offsetof() */
142 #ifdef AFS_PTHREAD_ENV
146 * Use procedural initialization of mutexes/condition variables
150 extern pthread_mutex_t rx_stats_mutex;
151 extern pthread_mutex_t des_init_mutex;
152 extern pthread_mutex_t des_random_mutex;
153 extern pthread_mutex_t rx_clock_mutex;
154 extern pthread_mutex_t rxi_connCacheMutex;
155 extern pthread_mutex_t rx_event_mutex;
156 extern pthread_mutex_t osi_malloc_mutex;
157 extern pthread_mutex_t event_handler_mutex;
158 extern pthread_mutex_t listener_mutex;
159 extern pthread_mutex_t rx_if_init_mutex;
160 extern pthread_mutex_t rx_if_mutex;
161 extern pthread_mutex_t rxkad_client_uid_mutex;
162 extern pthread_mutex_t rxkad_random_mutex;
164 extern pthread_cond_t rx_event_handler_cond;
165 extern pthread_cond_t rx_listener_cond;
167 static pthread_mutex_t epoch_mutex;
168 static pthread_mutex_t rx_init_mutex;
169 static pthread_mutex_t rx_debug_mutex;
172 rxi_InitPthread(void)
174 assert(pthread_mutex_init(&rx_clock_mutex, (const pthread_mutexattr_t *)0)
176 assert(pthread_mutex_init(&rx_stats_mutex, (const pthread_mutexattr_t *)0)
178 assert(pthread_mutex_init
179 (&rxi_connCacheMutex, (const pthread_mutexattr_t *)0) == 0);
180 assert(pthread_mutex_init(&rx_init_mutex, (const pthread_mutexattr_t *)0)
182 assert(pthread_mutex_init(&epoch_mutex, (const pthread_mutexattr_t *)0) ==
184 assert(pthread_mutex_init(&rx_event_mutex, (const pthread_mutexattr_t *)0)
186 assert(pthread_mutex_init(&des_init_mutex, (const pthread_mutexattr_t *)0)
188 assert(pthread_mutex_init
189 (&des_random_mutex, (const pthread_mutexattr_t *)0) == 0);
190 assert(pthread_mutex_init
191 (&osi_malloc_mutex, (const pthread_mutexattr_t *)0) == 0);
192 assert(pthread_mutex_init
193 (&event_handler_mutex, (const pthread_mutexattr_t *)0) == 0);
194 assert(pthread_mutex_init(&listener_mutex, (const pthread_mutexattr_t *)0)
196 assert(pthread_mutex_init
197 (&rx_if_init_mutex, (const pthread_mutexattr_t *)0) == 0);
198 assert(pthread_mutex_init(&rx_if_mutex, (const pthread_mutexattr_t *)0) ==
200 assert(pthread_mutex_init
201 (&rxkad_client_uid_mutex, (const pthread_mutexattr_t *)0) == 0);
202 assert(pthread_mutex_init
203 (&rxkad_random_mutex, (const pthread_mutexattr_t *)0) == 0);
204 assert(pthread_mutex_init(&rx_debug_mutex, (const pthread_mutexattr_t *)0)
207 assert(pthread_cond_init
208 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
209 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
211 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
212 assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
214 rxkad_global_stats_init();
217 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
218 #define INIT_PTHREAD_LOCKS \
219 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
221 * The rx_stats_mutex mutex protects the following global variables:
226 * rxi_lowConnRefCount
227 * rxi_lowPeerRefCount
236 #define INIT_PTHREAD_LOCKS
240 /* Variables for handling the minProcs implementation. availProcs gives the
241 * number of threads available in the pool at this moment (not counting dudes
242 * executing right now). totalMin gives the total number of procs required
243 * for handling all minProcs requests. minDeficit is a dynamic variable
244 * tracking the # of procs required to satisfy all of the remaining minProcs
246 * For fine grain locking to work, the quota check and the reservation of
247 * a server thread has to come while rxi_availProcs and rxi_minDeficit
248 * are locked. To this end, the code has been modified under #ifdef
249 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
250 * same time. A new function, ReturnToServerPool() returns the allocation.
252 * A call can be on several queue's (but only one at a time). When
253 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
254 * that no one else is touching the queue. To this end, we store the address
255 * of the queue lock in the call structure (under the call lock) when we
256 * put the call on a queue, and we clear the call_queue_lock when the
257 * call is removed from a queue (once the call lock has been obtained).
258 * This allows rxi_ResetCall to safely synchronize with others wishing
259 * to manipulate the queue.
262 #ifdef RX_ENABLE_LOCKS
263 static afs_kmutex_t rx_rpc_stats;
264 void rxi_StartUnlocked();
267 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
268 ** pretty good that the next packet coming in is from the same connection
269 ** as the last packet, since we're send multiple packets in a transmit window.
271 struct rx_connection *rxLastConn = 0;
273 #ifdef RX_ENABLE_LOCKS
274 /* The locking hierarchy for rx fine grain locking is composed of these
277 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
278 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
279 * call->lock - locks call data fields.
280 * These are independent of each other:
281 * rx_freeCallQueue_lock
286 * serverQueueEntry->lock
288 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
289 * peer->lock - locks peer data fields.
290 * conn_data_lock - that more than one thread is not updating a conn data
291 * field at the same time.
299 * Do we need a lock to protect the peer field in the conn structure?
300 * conn->peer was previously a constant for all intents and so has no
301 * lock protecting this field. The multihomed client delta introduced
302 * a RX code change : change the peer field in the connection structure
303 * to that remote inetrface from which the last packet for this
304 * connection was sent out. This may become an issue if further changes
307 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
308 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
310 /* rxdb_fileID is used to identify the lock location, along with line#. */
311 static int rxdb_fileID = RXDB_FILE_RX;
312 #endif /* RX_LOCKS_DB */
313 #else /* RX_ENABLE_LOCKS */
314 #define SET_CALL_QUEUE_LOCK(C, L)
315 #define CLEAR_CALL_QUEUE_LOCK(C)
316 #endif /* RX_ENABLE_LOCKS */
317 struct rx_serverQueueEntry *rx_waitForPacket = 0;
318 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
320 /* ------------Exported Interfaces------------- */
322 /* This function allows rxkad to set the epoch to a suitably random number
323 * which rx_NewConnection will use in the future. The principle purpose is to
324 * get rxnull connections to use the same epoch as the rxkad connections do, at
325 * least once the first rxkad connection is established. This is important now
326 * that the host/port addresses aren't used in FindConnection: the uniqueness
327 * of epoch/cid matters and the start time won't do. */
329 #ifdef AFS_PTHREAD_ENV
331 * This mutex protects the following global variables:
335 #define LOCK_EPOCH assert(pthread_mutex_lock(&epoch_mutex)==0)
336 #define UNLOCK_EPOCH assert(pthread_mutex_unlock(&epoch_mutex)==0)
340 #endif /* AFS_PTHREAD_ENV */
343 rx_SetEpoch(afs_uint32 epoch)
350 /* Initialize rx. A port number may be mentioned, in which case this
351 * becomes the default port number for any service installed later.
352 * If 0 is provided for the port number, a random port will be chosen
353 * by the kernel. Whether this will ever overlap anything in
354 * /etc/services is anybody's guess... Returns 0 on success, -1 on
356 static int rxinit_status = 1;
357 #ifdef AFS_PTHREAD_ENV
359 * This mutex protects the following global variables:
363 #define LOCK_RX_INIT assert(pthread_mutex_lock(&rx_init_mutex)==0)
364 #define UNLOCK_RX_INIT assert(pthread_mutex_unlock(&rx_init_mutex)==0)
367 #define UNLOCK_RX_INIT
371 rx_InitHost(u_int host, u_int port)
378 char *htable, *ptable;
385 if (rxinit_status == 0) {
386 tmp_status = rxinit_status;
388 return tmp_status; /* Already started; return previous error code. */
394 if (afs_winsockInit() < 0)
400 * Initialize anything necessary to provide a non-premptive threading
403 rxi_InitializeThreadSupport();
406 /* Allocate and initialize a socket for client and perhaps server
409 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
410 if (rx_socket == OSI_NULLSOCKET) {
414 #ifdef RX_ENABLE_LOCKS
417 #endif /* RX_LOCKS_DB */
418 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
419 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
420 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
421 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
422 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
424 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
426 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
428 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
430 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
432 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
434 #if defined(KERNEL) && defined(AFS_HPUX110_ENV)
436 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
437 #endif /* KERNEL && AFS_HPUX110_ENV */
438 #endif /* RX_ENABLE_LOCKS */
441 rx_connDeadTime = 12;
442 rx_tranquil = 0; /* reset flag */
443 memset((char *)&rx_stats, 0, sizeof(struct rx_stats));
445 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
446 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
447 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
448 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
449 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
450 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
452 /* Malloc up a bunch of packets & buffers */
454 queue_Init(&rx_freePacketQueue);
455 rxi_NeedMorePackets = FALSE;
456 #ifdef RX_ENABLE_TSFPQ
457 rx_nPackets = 0; /* in TSFPQ version, rx_nPackets is managed by rxi_MorePackets* */
458 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
459 #else /* RX_ENABLE_TSFPQ */
460 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
461 rxi_MorePackets(rx_nPackets);
462 #endif /* RX_ENABLE_TSFPQ */
469 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
470 tv.tv_sec = clock_now.sec;
471 tv.tv_usec = clock_now.usec;
472 srand((unsigned int)tv.tv_usec);
479 #if defined(KERNEL) && !defined(UKERNEL)
480 /* Really, this should never happen in a real kernel */
483 struct sockaddr_in addr;
484 int addrlen = sizeof(addr);
485 if (getsockname((int)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
489 rx_port = addr.sin_port;
492 rx_stats.minRtt.sec = 9999999;
494 rx_SetEpoch(tv.tv_sec | 0x80000000);
496 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
497 * will provide a randomer value. */
499 MUTEX_ENTER(&rx_stats_mutex);
500 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
501 MUTEX_EXIT(&rx_stats_mutex);
502 /* *Slightly* random start time for the cid. This is just to help
503 * out with the hashing function at the peer */
504 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
505 rx_connHashTable = (struct rx_connection **)htable;
506 rx_peerHashTable = (struct rx_peer **)ptable;
508 rx_lastAckDelay.sec = 0;
509 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
510 rx_hardAckDelay.sec = 0;
511 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
512 rx_softAckDelay.sec = 0;
513 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
515 rxevent_Init(20, rxi_ReScheduleEvents);
517 /* Initialize various global queues */
518 queue_Init(&rx_idleServerQueue);
519 queue_Init(&rx_incomingCallQueue);
520 queue_Init(&rx_freeCallQueue);
522 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
523 /* Initialize our list of usable IP addresses. */
527 /* Start listener process (exact function is dependent on the
528 * implementation environment--kernel or user space) */
532 tmp_status = rxinit_status = 0;
540 return rx_InitHost(htonl(INADDR_ANY), port);
543 /* called with unincremented nRequestsRunning to see if it is OK to start
544 * a new thread in this service. Could be "no" for two reasons: over the
545 * max quota, or would prevent others from reaching their min quota.
547 #ifdef RX_ENABLE_LOCKS
548 /* This verion of QuotaOK reserves quota if it's ok while the
549 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
552 QuotaOK(register struct rx_service *aservice)
554 /* check if over max quota */
555 if (aservice->nRequestsRunning >= aservice->maxProcs) {
559 /* under min quota, we're OK */
560 /* otherwise, can use only if there are enough to allow everyone
561 * to go to their min quota after this guy starts.
563 MUTEX_ENTER(&rx_stats_mutex);
564 if ((aservice->nRequestsRunning < aservice->minProcs)
565 || (rxi_availProcs > rxi_minDeficit)) {
566 aservice->nRequestsRunning++;
567 /* just started call in minProcs pool, need fewer to maintain
569 if (aservice->nRequestsRunning <= aservice->minProcs)
572 MUTEX_EXIT(&rx_stats_mutex);
575 MUTEX_EXIT(&rx_stats_mutex);
581 ReturnToServerPool(register struct rx_service *aservice)
583 aservice->nRequestsRunning--;
584 MUTEX_ENTER(&rx_stats_mutex);
585 if (aservice->nRequestsRunning < aservice->minProcs)
588 MUTEX_EXIT(&rx_stats_mutex);
591 #else /* RX_ENABLE_LOCKS */
593 QuotaOK(register struct rx_service *aservice)
596 /* under min quota, we're OK */
597 if (aservice->nRequestsRunning < aservice->minProcs)
600 /* check if over max quota */
601 if (aservice->nRequestsRunning >= aservice->maxProcs)
604 /* otherwise, can use only if there are enough to allow everyone
605 * to go to their min quota after this guy starts.
607 if (rxi_availProcs > rxi_minDeficit)
611 #endif /* RX_ENABLE_LOCKS */
614 /* Called by rx_StartServer to start up lwp's to service calls.
615 NExistingProcs gives the number of procs already existing, and which
616 therefore needn't be created. */
618 rxi_StartServerProcs(int nExistingProcs)
620 register struct rx_service *service;
625 /* For each service, reserve N processes, where N is the "minimum"
626 * number of processes that MUST be able to execute a request in parallel,
627 * at any time, for that process. Also compute the maximum difference
628 * between any service's maximum number of processes that can run
629 * (i.e. the maximum number that ever will be run, and a guarantee
630 * that this number will run if other services aren't running), and its
631 * minimum number. The result is the extra number of processes that
632 * we need in order to provide the latter guarantee */
633 for (i = 0; i < RX_MAX_SERVICES; i++) {
635 service = rx_services[i];
636 if (service == (struct rx_service *)0)
638 nProcs += service->minProcs;
639 diff = service->maxProcs - service->minProcs;
643 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
644 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
645 for (i = 0; i < nProcs; i++) {
646 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
652 /* This routine is only required on Windows */
654 rx_StartClientThread(void)
656 #ifdef AFS_PTHREAD_ENV
658 pid = (int) pthread_self();
659 #endif /* AFS_PTHREAD_ENV */
661 #endif /* AFS_NT40_ENV */
663 /* This routine must be called if any services are exported. If the
664 * donateMe flag is set, the calling process is donated to the server
667 rx_StartServer(int donateMe)
669 register struct rx_service *service;
675 /* Start server processes, if necessary (exact function is dependent
676 * on the implementation environment--kernel or user space). DonateMe
677 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
678 * case, one less new proc will be created rx_StartServerProcs.
680 rxi_StartServerProcs(donateMe);
682 /* count up the # of threads in minProcs, and add set the min deficit to
683 * be that value, too.
685 for (i = 0; i < RX_MAX_SERVICES; i++) {
686 service = rx_services[i];
687 if (service == (struct rx_service *)0)
689 MUTEX_ENTER(&rx_stats_mutex);
690 rxi_totalMin += service->minProcs;
691 /* below works even if a thread is running, since minDeficit would
692 * still have been decremented and later re-incremented.
694 rxi_minDeficit += service->minProcs;
695 MUTEX_EXIT(&rx_stats_mutex);
698 /* Turn on reaping of idle server connections */
699 rxi_ReapConnections();
708 #ifdef AFS_PTHREAD_ENV
710 pid = (pid_t) pthread_self();
711 #else /* AFS_PTHREAD_ENV */
713 LWP_CurrentProcess(&pid);
714 #endif /* AFS_PTHREAD_ENV */
716 sprintf(name, "srv_%d", ++nProcs);
718 (*registerProgram) (pid, name);
720 #endif /* AFS_NT40_ENV */
721 rx_ServerProc(); /* Never returns */
723 #ifdef RX_ENABLE_TSFPQ
724 /* no use leaving packets around in this thread's local queue if
725 * it isn't getting donated to the server thread pool.
727 rxi_FlushLocalPacketsTSFPQ();
728 #endif /* RX_ENABLE_TSFPQ */
732 /* Create a new client connection to the specified service, using the
733 * specified security object to implement the security model for this
735 struct rx_connection *
736 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
737 register struct rx_securityClass *securityObject,
738 int serviceSecurityIndex)
742 register struct rx_connection *conn;
747 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", ntohl(shost), ntohs(sport), sservice, securityObject, serviceSecurityIndex));
749 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
750 * the case of kmem_alloc? */
751 conn = rxi_AllocConnection();
752 #ifdef RX_ENABLE_LOCKS
753 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
754 MUTEX_INIT(&conn->conn_data_lock, "conn call lock", MUTEX_DEFAULT, 0);
755 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
758 MUTEX_ENTER(&rx_connHashTable_lock);
759 cid = (rx_nextCid += RX_MAXCALLS);
760 conn->type = RX_CLIENT_CONNECTION;
762 conn->epoch = rx_epoch;
763 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
764 conn->serviceId = sservice;
765 conn->securityObject = securityObject;
766 conn->securityData = (void *) 0;
767 conn->securityIndex = serviceSecurityIndex;
768 rx_SetConnDeadTime(conn, rx_connDeadTime);
769 conn->ackRate = RX_FAST_ACK_RATE;
771 conn->specific = NULL;
772 conn->challengeEvent = NULL;
773 conn->delayedAbortEvent = NULL;
774 conn->abortCount = 0;
777 RXS_NewConnection(securityObject, conn);
779 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
781 conn->refCount++; /* no lock required since only this thread knows... */
782 conn->next = rx_connHashTable[hashindex];
783 rx_connHashTable[hashindex] = conn;
784 MUTEX_ENTER(&rx_stats_mutex);
785 rx_stats.nClientConns++;
786 MUTEX_EXIT(&rx_stats_mutex);
788 MUTEX_EXIT(&rx_connHashTable_lock);
794 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
796 /* The idea is to set the dead time to a value that allows several
797 * keepalives to be dropped without timing out the connection. */
798 conn->secondsUntilDead = MAX(seconds, 6);
799 conn->secondsUntilPing = conn->secondsUntilDead / 6;
802 int rxi_lowPeerRefCount = 0;
803 int rxi_lowConnRefCount = 0;
806 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
807 * NOTE: must not be called with rx_connHashTable_lock held.
810 rxi_CleanupConnection(struct rx_connection *conn)
812 /* Notify the service exporter, if requested, that this connection
813 * is being destroyed */
814 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
815 (*conn->service->destroyConnProc) (conn);
817 /* Notify the security module that this connection is being destroyed */
818 RXS_DestroyConnection(conn->securityObject, conn);
820 /* If this is the last connection using the rx_peer struct, set its
821 * idle time to now. rxi_ReapConnections will reap it if it's still
822 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
824 MUTEX_ENTER(&rx_peerHashTable_lock);
825 if (conn->peer->refCount < 2) {
826 conn->peer->idleWhen = clock_Sec();
827 if (conn->peer->refCount < 1) {
828 conn->peer->refCount = 1;
829 MUTEX_ENTER(&rx_stats_mutex);
830 rxi_lowPeerRefCount++;
831 MUTEX_EXIT(&rx_stats_mutex);
834 conn->peer->refCount--;
835 MUTEX_EXIT(&rx_peerHashTable_lock);
837 MUTEX_ENTER(&rx_stats_mutex);
838 if (conn->type == RX_SERVER_CONNECTION)
839 rx_stats.nServerConns--;
841 rx_stats.nClientConns--;
842 MUTEX_EXIT(&rx_stats_mutex);
845 if (conn->specific) {
847 for (i = 0; i < conn->nSpecific; i++) {
848 if (conn->specific[i] && rxi_keyCreate_destructor[i])
849 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
850 conn->specific[i] = NULL;
852 free(conn->specific);
854 conn->specific = NULL;
858 MUTEX_DESTROY(&conn->conn_call_lock);
859 MUTEX_DESTROY(&conn->conn_data_lock);
860 CV_DESTROY(&conn->conn_call_cv);
862 rxi_FreeConnection(conn);
865 /* Destroy the specified connection */
867 rxi_DestroyConnection(register struct rx_connection *conn)
869 MUTEX_ENTER(&rx_connHashTable_lock);
870 rxi_DestroyConnectionNoLock(conn);
871 /* conn should be at the head of the cleanup list */
872 if (conn == rx_connCleanup_list) {
873 rx_connCleanup_list = rx_connCleanup_list->next;
874 MUTEX_EXIT(&rx_connHashTable_lock);
875 rxi_CleanupConnection(conn);
877 #ifdef RX_ENABLE_LOCKS
879 MUTEX_EXIT(&rx_connHashTable_lock);
881 #endif /* RX_ENABLE_LOCKS */
885 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
887 register struct rx_connection **conn_ptr;
888 register int havecalls = 0;
889 struct rx_packet *packet;
896 MUTEX_ENTER(&conn->conn_data_lock);
897 if (conn->refCount > 0)
900 MUTEX_ENTER(&rx_stats_mutex);
901 rxi_lowConnRefCount++;
902 MUTEX_EXIT(&rx_stats_mutex);
905 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
906 /* Busy; wait till the last guy before proceeding */
907 MUTEX_EXIT(&conn->conn_data_lock);
912 /* If the client previously called rx_NewCall, but it is still
913 * waiting, treat this as a running call, and wait to destroy the
914 * connection later when the call completes. */
915 if ((conn->type == RX_CLIENT_CONNECTION)
916 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
917 conn->flags |= RX_CONN_DESTROY_ME;
918 MUTEX_EXIT(&conn->conn_data_lock);
922 MUTEX_EXIT(&conn->conn_data_lock);
924 /* Check for extant references to this connection */
925 for (i = 0; i < RX_MAXCALLS; i++) {
926 register struct rx_call *call = conn->call[i];
929 if (conn->type == RX_CLIENT_CONNECTION) {
930 MUTEX_ENTER(&call->lock);
931 if (call->delayedAckEvent) {
932 /* Push the final acknowledgment out now--there
933 * won't be a subsequent call to acknowledge the
934 * last reply packets */
935 rxevent_Cancel(call->delayedAckEvent, call,
936 RX_CALL_REFCOUNT_DELAY);
937 if (call->state == RX_STATE_PRECALL
938 || call->state == RX_STATE_ACTIVE) {
939 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
941 rxi_AckAll(NULL, call, 0);
944 MUTEX_EXIT(&call->lock);
948 #ifdef RX_ENABLE_LOCKS
950 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
951 MUTEX_EXIT(&conn->conn_data_lock);
953 /* Someone is accessing a packet right now. */
957 #endif /* RX_ENABLE_LOCKS */
960 /* Don't destroy the connection if there are any call
961 * structures still in use */
962 MUTEX_ENTER(&conn->conn_data_lock);
963 conn->flags |= RX_CONN_DESTROY_ME;
964 MUTEX_EXIT(&conn->conn_data_lock);
969 if (conn->delayedAbortEvent) {
970 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
971 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
973 MUTEX_ENTER(&conn->conn_data_lock);
974 rxi_SendConnectionAbort(conn, packet, 0, 1);
975 MUTEX_EXIT(&conn->conn_data_lock);
976 rxi_FreePacket(packet);
980 /* Remove from connection hash table before proceeding */
982 &rx_connHashTable[CONN_HASH
983 (peer->host, peer->port, conn->cid, conn->epoch,
985 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
986 if (*conn_ptr == conn) {
987 *conn_ptr = conn->next;
991 /* if the conn that we are destroying was the last connection, then we
992 * clear rxLastConn as well */
993 if (rxLastConn == conn)
996 /* Make sure the connection is completely reset before deleting it. */
997 /* get rid of pending events that could zap us later */
998 if (conn->challengeEvent)
999 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1000 if (conn->checkReachEvent)
1001 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1003 /* Add the connection to the list of destroyed connections that
1004 * need to be cleaned up. This is necessary to avoid deadlocks
1005 * in the routines we call to inform others that this connection is
1006 * being destroyed. */
1007 conn->next = rx_connCleanup_list;
1008 rx_connCleanup_list = conn;
1011 /* Externally available version */
1013 rx_DestroyConnection(register struct rx_connection *conn)
1018 rxi_DestroyConnection(conn);
1023 rx_GetConnection(register struct rx_connection *conn)
1028 MUTEX_ENTER(&conn->conn_data_lock);
1030 MUTEX_EXIT(&conn->conn_data_lock);
1034 /* Wait for the transmit queue to no longer be busy.
1035 * requires the call->lock to be held */
1036 static void rxi_WaitforTQBusy(struct rx_call *call) {
1037 while (call->flags & RX_CALL_TQ_BUSY) {
1038 call->flags |= RX_CALL_TQ_WAIT;
1040 #ifdef RX_ENABLE_LOCKS
1041 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1042 CV_WAIT(&call->cv_tq, &call->lock);
1043 #else /* RX_ENABLE_LOCKS */
1044 osi_rxSleep(&call->tq);
1045 #endif /* RX_ENABLE_LOCKS */
1047 if (call->tqWaiters == 0) {
1048 call->flags &= ~RX_CALL_TQ_WAIT;
1052 /* Start a new rx remote procedure call, on the specified connection.
1053 * If wait is set to 1, wait for a free call channel; otherwise return
1054 * 0. Maxtime gives the maximum number of seconds this call may take,
1055 * after rx_NewCall returns. After this time interval, a call to any
1056 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1057 * For fine grain locking, we hold the conn_call_lock in order to
1058 * to ensure that we don't get signalle after we found a call in an active
1059 * state and before we go to sleep.
1062 rx_NewCall(register struct rx_connection *conn)
1065 register struct rx_call *call;
1066 struct clock queueTime;
1070 dpf(("rx_NewCall(conn %x)\n", conn));
1073 clock_GetTime(&queueTime);
1074 MUTEX_ENTER(&conn->conn_call_lock);
1077 * Check if there are others waiting for a new call.
1078 * If so, let them go first to avoid starving them.
1079 * This is a fairly simple scheme, and might not be
1080 * a complete solution for large numbers of waiters.
1082 * makeCallWaiters keeps track of the number of
1083 * threads waiting to make calls and the
1084 * RX_CONN_MAKECALL_WAITING flag bit is used to
1085 * indicate that there are indeed calls waiting.
1086 * The flag is set when the waiter is incremented.
1087 * It is only cleared in rx_EndCall when
1088 * makeCallWaiters is 0. This prevents us from
1089 * accidently destroying the connection while it
1090 * is potentially about to be used.
1092 MUTEX_ENTER(&conn->conn_data_lock);
1093 if (conn->makeCallWaiters) {
1094 conn->flags |= RX_CONN_MAKECALL_WAITING;
1095 conn->makeCallWaiters++;
1096 MUTEX_EXIT(&conn->conn_data_lock);
1098 #ifdef RX_ENABLE_LOCKS
1099 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1103 MUTEX_ENTER(&conn->conn_data_lock);
1104 conn->makeCallWaiters--;
1106 MUTEX_EXIT(&conn->conn_data_lock);
1109 for (i = 0; i < RX_MAXCALLS; i++) {
1110 call = conn->call[i];
1112 MUTEX_ENTER(&call->lock);
1113 if (call->state == RX_STATE_DALLY) {
1114 rxi_ResetCall(call, 0);
1115 (*call->callNumber)++;
1118 MUTEX_EXIT(&call->lock);
1120 call = rxi_NewCall(conn, i);
1124 if (i < RX_MAXCALLS) {
1127 MUTEX_ENTER(&conn->conn_data_lock);
1128 conn->flags |= RX_CONN_MAKECALL_WAITING;
1129 conn->makeCallWaiters++;
1130 MUTEX_EXIT(&conn->conn_data_lock);
1132 #ifdef RX_ENABLE_LOCKS
1133 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1137 MUTEX_ENTER(&conn->conn_data_lock);
1138 conn->makeCallWaiters--;
1139 MUTEX_EXIT(&conn->conn_data_lock);
1142 * Wake up anyone else who might be giving us a chance to
1143 * run (see code above that avoids resource starvation).
1145 #ifdef RX_ENABLE_LOCKS
1146 CV_BROADCAST(&conn->conn_call_cv);
1151 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1153 /* Client is initially in send mode */
1154 call->state = RX_STATE_ACTIVE;
1155 call->error = conn->error;
1157 call->mode = RX_MODE_ERROR;
1159 call->mode = RX_MODE_SENDING;
1161 /* remember start time for call in case we have hard dead time limit */
1162 call->queueTime = queueTime;
1163 clock_GetTime(&call->startTime);
1164 hzero(call->bytesSent);
1165 hzero(call->bytesRcvd);
1167 /* Turn on busy protocol. */
1168 rxi_KeepAliveOn(call);
1170 MUTEX_EXIT(&call->lock);
1171 MUTEX_EXIT(&conn->conn_call_lock);
1174 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1175 /* Now, if TQ wasn't cleared earlier, do it now. */
1176 MUTEX_ENTER(&call->lock);
1177 rxi_WaitforTQBusy(call);
1178 if (call->flags & RX_CALL_TQ_CLEARME) {
1179 rxi_ClearTransmitQueue(call, 0);
1180 queue_Init(&call->tq);
1182 MUTEX_EXIT(&call->lock);
1183 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1185 dpf(("rx_NewCall(call %x)\n", call));
1190 rxi_HasActiveCalls(register struct rx_connection *aconn)
1193 register struct rx_call *tcall;
1197 for (i = 0; i < RX_MAXCALLS; i++) {
1198 if ((tcall = aconn->call[i])) {
1199 if ((tcall->state == RX_STATE_ACTIVE)
1200 || (tcall->state == RX_STATE_PRECALL)) {
1211 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1212 register afs_int32 * aint32s)
1215 register struct rx_call *tcall;
1219 for (i = 0; i < RX_MAXCALLS; i++) {
1220 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1221 aint32s[i] = aconn->callNumber[i] + 1;
1223 aint32s[i] = aconn->callNumber[i];
1230 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1231 register afs_int32 * aint32s)
1234 register struct rx_call *tcall;
1238 for (i = 0; i < RX_MAXCALLS; i++) {
1239 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1240 aconn->callNumber[i] = aint32s[i] - 1;
1242 aconn->callNumber[i] = aint32s[i];
1248 /* Advertise a new service. A service is named locally by a UDP port
1249 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1252 char *serviceName; Name for identification purposes (e.g. the
1253 service name might be used for probing for
1256 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1257 char *serviceName, struct rx_securityClass **securityObjects,
1258 int nSecurityObjects,
1259 afs_int32(*serviceProc) (struct rx_call * acall))
1261 osi_socket socket = OSI_NULLSOCKET;
1262 register struct rx_service *tservice;
1268 if (serviceId == 0) {
1270 "rx_NewService: service id for service %s is not non-zero.\n",
1277 "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",
1285 tservice = rxi_AllocService();
1287 for (i = 0; i < RX_MAX_SERVICES; i++) {
1288 register struct rx_service *service = rx_services[i];
1290 if (port == service->servicePort && host == service->serviceHost) {
1291 if (service->serviceId == serviceId) {
1292 /* The identical service has already been
1293 * installed; if the caller was intending to
1294 * change the security classes used by this
1295 * service, he/she loses. */
1297 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1298 serviceName, serviceId, service->serviceName);
1300 rxi_FreeService(tservice);
1303 /* Different service, same port: re-use the socket
1304 * which is bound to the same port */
1305 socket = service->socket;
1308 if (socket == OSI_NULLSOCKET) {
1309 /* If we don't already have a socket (from another
1310 * service on same port) get a new one */
1311 socket = rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
1312 if (socket == OSI_NULLSOCKET) {
1314 rxi_FreeService(tservice);
1319 service->socket = socket;
1320 service->serviceHost = host;
1321 service->servicePort = port;
1322 service->serviceId = serviceId;
1323 service->serviceName = serviceName;
1324 service->nSecurityObjects = nSecurityObjects;
1325 service->securityObjects = securityObjects;
1326 service->minProcs = 0;
1327 service->maxProcs = 1;
1328 service->idleDeadTime = 60;
1329 service->connDeadTime = rx_connDeadTime;
1330 service->executeRequestProc = serviceProc;
1331 service->checkReach = 0;
1332 rx_services[i] = service; /* not visible until now */
1338 rxi_FreeService(tservice);
1339 (osi_Msg "rx_NewService: cannot support > %d services\n",
1344 /* Set configuration options for all of a service's security objects */
1347 rx_SetSecurityConfiguration(struct rx_service *service,
1348 rx_securityConfigVariables type,
1352 for (i = 0; i<service->nSecurityObjects; i++) {
1353 if (service->securityObjects[i]) {
1354 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1362 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1363 struct rx_securityClass **securityObjects, int nSecurityObjects,
1364 afs_int32(*serviceProc) (struct rx_call * acall))
1366 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1369 /* Generic request processing loop. This routine should be called
1370 * by the implementation dependent rx_ServerProc. If socketp is
1371 * non-null, it will be set to the file descriptor that this thread
1372 * is now listening on. If socketp is null, this routine will never
1375 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1377 register struct rx_call *call;
1378 register afs_int32 code;
1379 register struct rx_service *tservice = NULL;
1386 call = rx_GetCall(threadID, tservice, socketp);
1387 if (socketp && *socketp != OSI_NULLSOCKET) {
1388 /* We are now a listener thread */
1393 /* if server is restarting( typically smooth shutdown) then do not
1394 * allow any new calls.
1397 if (rx_tranquil && (call != NULL)) {
1401 MUTEX_ENTER(&call->lock);
1403 rxi_CallError(call, RX_RESTARTING);
1404 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1406 MUTEX_EXIT(&call->lock);
1410 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1411 #ifdef RX_ENABLE_LOCKS
1413 #endif /* RX_ENABLE_LOCKS */
1414 afs_termState = AFSOP_STOP_AFS;
1415 afs_osi_Wakeup(&afs_termState);
1416 #ifdef RX_ENABLE_LOCKS
1418 #endif /* RX_ENABLE_LOCKS */
1423 tservice = call->conn->service;
1425 if (tservice->beforeProc)
1426 (*tservice->beforeProc) (call);
1428 code = call->conn->service->executeRequestProc(call);
1430 if (tservice->afterProc)
1431 (*tservice->afterProc) (call, code);
1433 rx_EndCall(call, code);
1434 MUTEX_ENTER(&rx_stats_mutex);
1436 MUTEX_EXIT(&rx_stats_mutex);
1442 rx_WakeupServerProcs(void)
1444 struct rx_serverQueueEntry *np, *tqp;
1448 MUTEX_ENTER(&rx_serverPool_lock);
1450 #ifdef RX_ENABLE_LOCKS
1451 if (rx_waitForPacket)
1452 CV_BROADCAST(&rx_waitForPacket->cv);
1453 #else /* RX_ENABLE_LOCKS */
1454 if (rx_waitForPacket)
1455 osi_rxWakeup(rx_waitForPacket);
1456 #endif /* RX_ENABLE_LOCKS */
1457 MUTEX_ENTER(&freeSQEList_lock);
1458 for (np = rx_FreeSQEList; np; np = tqp) {
1459 tqp = *(struct rx_serverQueueEntry **)np;
1460 #ifdef RX_ENABLE_LOCKS
1461 CV_BROADCAST(&np->cv);
1462 #else /* RX_ENABLE_LOCKS */
1464 #endif /* RX_ENABLE_LOCKS */
1466 MUTEX_EXIT(&freeSQEList_lock);
1467 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1468 #ifdef RX_ENABLE_LOCKS
1469 CV_BROADCAST(&np->cv);
1470 #else /* RX_ENABLE_LOCKS */
1472 #endif /* RX_ENABLE_LOCKS */
1474 MUTEX_EXIT(&rx_serverPool_lock);
1479 * One thing that seems to happen is that all the server threads get
1480 * tied up on some empty or slow call, and then a whole bunch of calls
1481 * arrive at once, using up the packet pool, so now there are more
1482 * empty calls. The most critical resources here are server threads
1483 * and the free packet pool. The "doreclaim" code seems to help in
1484 * general. I think that eventually we arrive in this state: there
1485 * are lots of pending calls which do have all their packets present,
1486 * so they won't be reclaimed, are multi-packet calls, so they won't
1487 * be scheduled until later, and thus are tying up most of the free
1488 * packet pool for a very long time.
1490 * 1. schedule multi-packet calls if all the packets are present.
1491 * Probably CPU-bound operation, useful to return packets to pool.
1492 * Do what if there is a full window, but the last packet isn't here?
1493 * 3. preserve one thread which *only* runs "best" calls, otherwise
1494 * it sleeps and waits for that type of call.
1495 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1496 * the current dataquota business is badly broken. The quota isn't adjusted
1497 * to reflect how many packets are presently queued for a running call.
1498 * So, when we schedule a queued call with a full window of packets queued
1499 * up for it, that *should* free up a window full of packets for other 2d-class
1500 * calls to be able to use from the packet pool. But it doesn't.
1502 * NB. Most of the time, this code doesn't run -- since idle server threads
1503 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1504 * as a new call arrives.
1506 /* Sleep until a call arrives. Returns a pointer to the call, ready
1507 * for an rx_Read. */
1508 #ifdef RX_ENABLE_LOCKS
1510 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1512 struct rx_serverQueueEntry *sq;
1513 register struct rx_call *call = (struct rx_call *)0;
1514 struct rx_service *service = NULL;
1517 MUTEX_ENTER(&freeSQEList_lock);
1519 if ((sq = rx_FreeSQEList)) {
1520 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1521 MUTEX_EXIT(&freeSQEList_lock);
1522 } else { /* otherwise allocate a new one and return that */
1523 MUTEX_EXIT(&freeSQEList_lock);
1524 sq = (struct rx_serverQueueEntry *)
1525 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1526 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1527 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1530 MUTEX_ENTER(&rx_serverPool_lock);
1531 if (cur_service != NULL) {
1532 ReturnToServerPool(cur_service);
1535 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1536 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1538 /* Scan for eligible incoming calls. A call is not eligible
1539 * if the maximum number of calls for its service type are
1540 * already executing */
1541 /* One thread will process calls FCFS (to prevent starvation),
1542 * while the other threads may run ahead looking for calls which
1543 * have all their input data available immediately. This helps
1544 * keep threads from blocking, waiting for data from the client. */
1545 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1546 service = tcall->conn->service;
1547 if (!QuotaOK(service)) {
1550 if (tno == rxi_fcfs_thread_num
1551 || !tcall->queue_item_header.next) {
1552 /* If we're the fcfs thread , then we'll just use
1553 * this call. If we haven't been able to find an optimal
1554 * choice, and we're at the end of the list, then use a
1555 * 2d choice if one has been identified. Otherwise... */
1556 call = (choice2 ? choice2 : tcall);
1557 service = call->conn->service;
1558 } else if (!queue_IsEmpty(&tcall->rq)) {
1559 struct rx_packet *rp;
1560 rp = queue_First(&tcall->rq, rx_packet);
1561 if (rp->header.seq == 1) {
1563 || (rp->header.flags & RX_LAST_PACKET)) {
1565 } else if (rxi_2dchoice && !choice2
1566 && !(tcall->flags & RX_CALL_CLEARED)
1567 && (tcall->rprev > rxi_HardAckRate)) {
1576 ReturnToServerPool(service);
1583 MUTEX_EXIT(&rx_serverPool_lock);
1584 MUTEX_ENTER(&call->lock);
1586 if (call->flags & RX_CALL_WAIT_PROC) {
1587 call->flags &= ~RX_CALL_WAIT_PROC;
1588 MUTEX_ENTER(&rx_stats_mutex);
1590 MUTEX_EXIT(&rx_stats_mutex);
1593 if (call->state != RX_STATE_PRECALL || call->error) {
1594 MUTEX_EXIT(&call->lock);
1595 MUTEX_ENTER(&rx_serverPool_lock);
1596 ReturnToServerPool(service);
1601 if (queue_IsEmpty(&call->rq)
1602 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1603 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1605 CLEAR_CALL_QUEUE_LOCK(call);
1608 /* If there are no eligible incoming calls, add this process
1609 * to the idle server queue, to wait for one */
1613 *socketp = OSI_NULLSOCKET;
1615 sq->socketp = socketp;
1616 queue_Append(&rx_idleServerQueue, sq);
1617 #ifndef AFS_AIX41_ENV
1618 rx_waitForPacket = sq;
1620 rx_waitingForPacket = sq;
1621 #endif /* AFS_AIX41_ENV */
1623 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1625 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1626 MUTEX_EXIT(&rx_serverPool_lock);
1627 return (struct rx_call *)0;
1630 } while (!(call = sq->newcall)
1631 && !(socketp && *socketp != OSI_NULLSOCKET));
1632 MUTEX_EXIT(&rx_serverPool_lock);
1634 MUTEX_ENTER(&call->lock);
1640 MUTEX_ENTER(&freeSQEList_lock);
1641 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1642 rx_FreeSQEList = sq;
1643 MUTEX_EXIT(&freeSQEList_lock);
1646 clock_GetTime(&call->startTime);
1647 call->state = RX_STATE_ACTIVE;
1648 call->mode = RX_MODE_RECEIVING;
1649 #ifdef RX_KERNEL_TRACE
1650 if (ICL_SETACTIVE(afs_iclSetp)) {
1651 int glockOwner = ISAFS_GLOCK();
1654 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1655 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1662 rxi_calltrace(RX_CALL_START, call);
1663 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1664 call->conn->service->servicePort, call->conn->service->serviceId,
1667 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1668 MUTEX_EXIT(&call->lock);
1670 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1675 #else /* RX_ENABLE_LOCKS */
1677 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1679 struct rx_serverQueueEntry *sq;
1680 register struct rx_call *call = (struct rx_call *)0, *choice2;
1681 struct rx_service *service = NULL;
1685 MUTEX_ENTER(&freeSQEList_lock);
1687 if ((sq = rx_FreeSQEList)) {
1688 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1689 MUTEX_EXIT(&freeSQEList_lock);
1690 } else { /* otherwise allocate a new one and return that */
1691 MUTEX_EXIT(&freeSQEList_lock);
1692 sq = (struct rx_serverQueueEntry *)
1693 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1694 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1695 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1697 MUTEX_ENTER(&sq->lock);
1699 if (cur_service != NULL) {
1700 cur_service->nRequestsRunning--;
1701 if (cur_service->nRequestsRunning < cur_service->minProcs)
1705 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1706 register struct rx_call *tcall, *ncall;
1707 /* Scan for eligible incoming calls. A call is not eligible
1708 * if the maximum number of calls for its service type are
1709 * already executing */
1710 /* One thread will process calls FCFS (to prevent starvation),
1711 * while the other threads may run ahead looking for calls which
1712 * have all their input data available immediately. This helps
1713 * keep threads from blocking, waiting for data from the client. */
1714 choice2 = (struct rx_call *)0;
1715 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1716 service = tcall->conn->service;
1717 if (QuotaOK(service)) {
1718 if (tno == rxi_fcfs_thread_num
1719 || !tcall->queue_item_header.next) {
1720 /* If we're the fcfs thread, then we'll just use
1721 * this call. If we haven't been able to find an optimal
1722 * choice, and we're at the end of the list, then use a
1723 * 2d choice if one has been identified. Otherwise... */
1724 call = (choice2 ? choice2 : tcall);
1725 service = call->conn->service;
1726 } else if (!queue_IsEmpty(&tcall->rq)) {
1727 struct rx_packet *rp;
1728 rp = queue_First(&tcall->rq, rx_packet);
1729 if (rp->header.seq == 1
1731 || (rp->header.flags & RX_LAST_PACKET))) {
1733 } else if (rxi_2dchoice && !choice2
1734 && !(tcall->flags & RX_CALL_CLEARED)
1735 && (tcall->rprev > rxi_HardAckRate)) {
1748 /* we can't schedule a call if there's no data!!! */
1749 /* send an ack if there's no data, if we're missing the
1750 * first packet, or we're missing something between first
1751 * and last -- there's a "hole" in the incoming data. */
1752 if (queue_IsEmpty(&call->rq)
1753 || queue_First(&call->rq, rx_packet)->header.seq != 1
1754 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1755 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1757 call->flags &= (~RX_CALL_WAIT_PROC);
1758 service->nRequestsRunning++;
1759 /* just started call in minProcs pool, need fewer to maintain
1761 if (service->nRequestsRunning <= service->minProcs)
1765 /* MUTEX_EXIT(&call->lock); */
1767 /* If there are no eligible incoming calls, add this process
1768 * to the idle server queue, to wait for one */
1771 *socketp = OSI_NULLSOCKET;
1773 sq->socketp = socketp;
1774 queue_Append(&rx_idleServerQueue, sq);
1778 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1780 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1781 return (struct rx_call *)0;
1784 } while (!(call = sq->newcall)
1785 && !(socketp && *socketp != OSI_NULLSOCKET));
1787 MUTEX_EXIT(&sq->lock);
1789 MUTEX_ENTER(&freeSQEList_lock);
1790 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1791 rx_FreeSQEList = sq;
1792 MUTEX_EXIT(&freeSQEList_lock);
1795 clock_GetTime(&call->startTime);
1796 call->state = RX_STATE_ACTIVE;
1797 call->mode = RX_MODE_RECEIVING;
1798 #ifdef RX_KERNEL_TRACE
1799 if (ICL_SETACTIVE(afs_iclSetp)) {
1800 int glockOwner = ISAFS_GLOCK();
1803 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1804 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1811 rxi_calltrace(RX_CALL_START, call);
1812 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1813 call->conn->service->servicePort, call->conn->service->serviceId,
1816 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1823 #endif /* RX_ENABLE_LOCKS */
1827 /* Establish a procedure to be called when a packet arrives for a
1828 * call. This routine will be called at most once after each call,
1829 * and will also be called if there is an error condition on the or
1830 * the call is complete. Used by multi rx to build a selection
1831 * function which determines which of several calls is likely to be a
1832 * good one to read from.
1833 * NOTE: the way this is currently implemented it is probably only a
1834 * good idea to (1) use it immediately after a newcall (clients only)
1835 * and (2) only use it once. Other uses currently void your warranty
1838 rx_SetArrivalProc(register struct rx_call *call,
1839 register void (*proc) (register struct rx_call * call,
1841 register int index),
1842 register void * handle, register int arg)
1844 call->arrivalProc = proc;
1845 call->arrivalProcHandle = handle;
1846 call->arrivalProcArg = arg;
1849 /* Call is finished (possibly prematurely). Return rc to the peer, if
1850 * appropriate, and return the final error code from the conversation
1854 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1856 register struct rx_connection *conn = call->conn;
1857 register struct rx_service *service;
1863 dpf(("rx_EndCall(call %x rc %d error %d abortCode %d)\n", call, rc, call->error, call->abortCode));
1866 MUTEX_ENTER(&call->lock);
1868 if (rc == 0 && call->error == 0) {
1869 call->abortCode = 0;
1870 call->abortCount = 0;
1873 call->arrivalProc = (void (*)())0;
1874 if (rc && call->error == 0) {
1875 rxi_CallError(call, rc);
1876 /* Send an abort message to the peer if this error code has
1877 * only just been set. If it was set previously, assume the
1878 * peer has already been sent the error code or will request it
1880 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1882 if (conn->type == RX_SERVER_CONNECTION) {
1883 /* Make sure reply or at least dummy reply is sent */
1884 if (call->mode == RX_MODE_RECEIVING) {
1885 rxi_WriteProc(call, 0, 0);
1887 if (call->mode == RX_MODE_SENDING) {
1888 rxi_FlushWrite(call);
1890 service = conn->service;
1891 rxi_calltrace(RX_CALL_END, call);
1892 /* Call goes to hold state until reply packets are acknowledged */
1893 if (call->tfirst + call->nSoftAcked < call->tnext) {
1894 call->state = RX_STATE_HOLD;
1896 call->state = RX_STATE_DALLY;
1897 rxi_ClearTransmitQueue(call, 0);
1898 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1899 rxevent_Cancel(call->keepAliveEvent, call,
1900 RX_CALL_REFCOUNT_ALIVE);
1902 } else { /* Client connection */
1904 /* Make sure server receives input packets, in the case where
1905 * no reply arguments are expected */
1906 if ((call->mode == RX_MODE_SENDING)
1907 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1908 (void)rxi_ReadProc(call, &dummy, 1);
1911 /* If we had an outstanding delayed ack, be nice to the server
1912 * and force-send it now.
1914 if (call->delayedAckEvent) {
1915 rxevent_Cancel(call->delayedAckEvent, call,
1916 RX_CALL_REFCOUNT_DELAY);
1917 call->delayedAckEvent = NULL;
1918 rxi_SendDelayedAck(NULL, call, NULL);
1921 /* We need to release the call lock since it's lower than the
1922 * conn_call_lock and we don't want to hold the conn_call_lock
1923 * over the rx_ReadProc call. The conn_call_lock needs to be held
1924 * here for the case where rx_NewCall is perusing the calls on
1925 * the connection structure. We don't want to signal until
1926 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1927 * have checked this call, found it active and by the time it
1928 * goes to sleep, will have missed the signal.
1930 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
1931 * there are threads waiting to use the conn object.
1933 MUTEX_EXIT(&call->lock);
1934 MUTEX_ENTER(&conn->conn_call_lock);
1935 MUTEX_ENTER(&call->lock);
1936 MUTEX_ENTER(&conn->conn_data_lock);
1937 conn->flags |= RX_CONN_BUSY;
1938 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1939 if (conn->makeCallWaiters == 0)
1940 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1941 MUTEX_EXIT(&conn->conn_data_lock);
1942 #ifdef RX_ENABLE_LOCKS
1943 CV_BROADCAST(&conn->conn_call_cv);
1948 #ifdef RX_ENABLE_LOCKS
1950 MUTEX_EXIT(&conn->conn_data_lock);
1952 #endif /* RX_ENABLE_LOCKS */
1953 call->state = RX_STATE_DALLY;
1955 error = call->error;
1957 /* currentPacket, nLeft, and NFree must be zeroed here, because
1958 * ResetCall cannot: ResetCall may be called at splnet(), in the
1959 * kernel version, and may interrupt the macros rx_Read or
1960 * rx_Write, which run at normal priority for efficiency. */
1961 if (call->currentPacket) {
1962 queue_Prepend(&call->iovq, call->currentPacket);
1963 call->currentPacket = (struct rx_packet *)0;
1966 call->nLeft = call->nFree = call->curlen = 0;
1968 /* Free any packets from the last call to ReadvProc/WritevProc */
1969 rxi_FreePackets(0, &call->iovq);
1971 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1972 MUTEX_EXIT(&call->lock);
1973 if (conn->type == RX_CLIENT_CONNECTION) {
1974 MUTEX_EXIT(&conn->conn_call_lock);
1975 conn->flags &= ~RX_CONN_BUSY;
1979 * Map errors to the local host's errno.h format.
1981 error = ntoh_syserr_conv(error);
1985 #if !defined(KERNEL)
1987 /* Call this routine when shutting down a server or client (especially
1988 * clients). This will allow Rx to gracefully garbage collect server
1989 * connections, and reduce the number of retries that a server might
1990 * make to a dead client.
1991 * This is not quite right, since some calls may still be ongoing and
1992 * we can't lock them to destroy them. */
1996 register struct rx_connection **conn_ptr, **conn_end;
2000 if (rxinit_status == 1) {
2002 return; /* Already shutdown. */
2004 rxi_DeleteCachedConnections();
2005 if (rx_connHashTable) {
2006 MUTEX_ENTER(&rx_connHashTable_lock);
2007 for (conn_ptr = &rx_connHashTable[0], conn_end =
2008 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2010 struct rx_connection *conn, *next;
2011 for (conn = *conn_ptr; conn; conn = next) {
2013 if (conn->type == RX_CLIENT_CONNECTION) {
2014 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2016 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2017 #ifdef RX_ENABLE_LOCKS
2018 rxi_DestroyConnectionNoLock(conn);
2019 #else /* RX_ENABLE_LOCKS */
2020 rxi_DestroyConnection(conn);
2021 #endif /* RX_ENABLE_LOCKS */
2025 #ifdef RX_ENABLE_LOCKS
2026 while (rx_connCleanup_list) {
2027 struct rx_connection *conn;
2028 conn = rx_connCleanup_list;
2029 rx_connCleanup_list = rx_connCleanup_list->next;
2030 MUTEX_EXIT(&rx_connHashTable_lock);
2031 rxi_CleanupConnection(conn);
2032 MUTEX_ENTER(&rx_connHashTable_lock);
2034 MUTEX_EXIT(&rx_connHashTable_lock);
2035 #endif /* RX_ENABLE_LOCKS */
2040 afs_winsockCleanup();
2048 /* if we wakeup packet waiter too often, can get in loop with two
2049 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2051 rxi_PacketsUnWait(void)
2053 if (!rx_waitingForPackets) {
2057 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2058 return; /* still over quota */
2061 rx_waitingForPackets = 0;
2062 #ifdef RX_ENABLE_LOCKS
2063 CV_BROADCAST(&rx_waitingForPackets_cv);
2065 osi_rxWakeup(&rx_waitingForPackets);
2071 /* ------------------Internal interfaces------------------------- */
2073 /* Return this process's service structure for the
2074 * specified socket and service */
2076 rxi_FindService(register osi_socket socket, register u_short serviceId)
2078 register struct rx_service **sp;
2079 for (sp = &rx_services[0]; *sp; sp++) {
2080 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2086 /* Allocate a call structure, for the indicated channel of the
2087 * supplied connection. The mode and state of the call must be set by
2088 * the caller. Returns the call with mutex locked. */
2090 rxi_NewCall(register struct rx_connection *conn, register int channel)
2092 register struct rx_call *call;
2093 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2094 register struct rx_call *cp; /* Call pointer temp */
2095 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2096 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2098 dpf(("rxi_NewCall(conn %x, channel %d)\n", conn, channel));
2100 /* Grab an existing call structure, or allocate a new one.
2101 * Existing call structures are assumed to have been left reset by
2103 MUTEX_ENTER(&rx_freeCallQueue_lock);
2105 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2107 * EXCEPT that the TQ might not yet be cleared out.
2108 * Skip over those with in-use TQs.
2111 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2112 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2118 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2119 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2120 call = queue_First(&rx_freeCallQueue, rx_call);
2121 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2123 MUTEX_ENTER(&rx_stats_mutex);
2124 rx_stats.nFreeCallStructs--;
2125 MUTEX_EXIT(&rx_stats_mutex);
2126 MUTEX_EXIT(&rx_freeCallQueue_lock);
2127 MUTEX_ENTER(&call->lock);
2128 CLEAR_CALL_QUEUE_LOCK(call);
2129 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2130 /* Now, if TQ wasn't cleared earlier, do it now. */
2131 if (call->flags & RX_CALL_TQ_CLEARME) {
2132 rxi_ClearTransmitQueue(call, 0);
2133 queue_Init(&call->tq);
2135 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2136 /* Bind the call to its connection structure */
2138 rxi_ResetCall(call, 1);
2140 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2142 MUTEX_EXIT(&rx_freeCallQueue_lock);
2143 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2144 MUTEX_ENTER(&call->lock);
2145 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2146 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2147 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2149 MUTEX_ENTER(&rx_stats_mutex);
2150 rx_stats.nCallStructs++;
2151 MUTEX_EXIT(&rx_stats_mutex);
2152 /* Initialize once-only items */
2153 queue_Init(&call->tq);
2154 queue_Init(&call->rq);
2155 queue_Init(&call->iovq);
2156 /* Bind the call to its connection structure (prereq for reset) */
2158 rxi_ResetCall(call, 1);
2160 call->channel = channel;
2161 call->callNumber = &conn->callNumber[channel];
2162 /* Note that the next expected call number is retained (in
2163 * conn->callNumber[i]), even if we reallocate the call structure
2165 conn->call[channel] = call;
2166 /* if the channel's never been used (== 0), we should start at 1, otherwise
2167 * the call number is valid from the last time this channel was used */
2168 if (*call->callNumber == 0)
2169 *call->callNumber = 1;
2174 /* A call has been inactive long enough that so we can throw away
2175 * state, including the call structure, which is placed on the call
2177 * Call is locked upon entry.
2178 * haveCTLock set if called from rxi_ReapConnections
2180 #ifdef RX_ENABLE_LOCKS
2182 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2183 #else /* RX_ENABLE_LOCKS */
2185 rxi_FreeCall(register struct rx_call *call)
2186 #endif /* RX_ENABLE_LOCKS */
2188 register int channel = call->channel;
2189 register struct rx_connection *conn = call->conn;
2192 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2193 (*call->callNumber)++;
2194 rxi_ResetCall(call, 0);
2195 call->conn->call[channel] = (struct rx_call *)0;
2197 MUTEX_ENTER(&rx_freeCallQueue_lock);
2198 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2199 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2200 /* A call may be free even though its transmit queue is still in use.
2201 * Since we search the call list from head to tail, put busy calls at
2202 * the head of the list, and idle calls at the tail.
2204 if (call->flags & RX_CALL_TQ_BUSY)
2205 queue_Prepend(&rx_freeCallQueue, call);
2207 queue_Append(&rx_freeCallQueue, call);
2208 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2209 queue_Append(&rx_freeCallQueue, call);
2210 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2211 MUTEX_ENTER(&rx_stats_mutex);
2212 rx_stats.nFreeCallStructs++;
2213 MUTEX_EXIT(&rx_stats_mutex);
2215 MUTEX_EXIT(&rx_freeCallQueue_lock);
2217 /* Destroy the connection if it was previously slated for
2218 * destruction, i.e. the Rx client code previously called
2219 * rx_DestroyConnection (client connections), or
2220 * rxi_ReapConnections called the same routine (server
2221 * connections). Only do this, however, if there are no
2222 * outstanding calls. Note that for fine grain locking, there appears
2223 * to be a deadlock in that rxi_FreeCall has a call locked and
2224 * DestroyConnectionNoLock locks each call in the conn. But note a
2225 * few lines up where we have removed this call from the conn.
2226 * If someone else destroys a connection, they either have no
2227 * call lock held or are going through this section of code.
2229 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2230 MUTEX_ENTER(&conn->conn_data_lock);
2232 MUTEX_EXIT(&conn->conn_data_lock);
2233 #ifdef RX_ENABLE_LOCKS
2235 rxi_DestroyConnectionNoLock(conn);
2237 rxi_DestroyConnection(conn);
2238 #else /* RX_ENABLE_LOCKS */
2239 rxi_DestroyConnection(conn);
2240 #endif /* RX_ENABLE_LOCKS */
2244 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2246 rxi_Alloc(register size_t size)
2250 MUTEX_ENTER(&rx_stats_mutex);
2252 rxi_Allocsize += (afs_int32)size;
2253 MUTEX_EXIT(&rx_stats_mutex);
2255 p = (char *)osi_Alloc(size);
2258 osi_Panic("rxi_Alloc error");
2264 rxi_Free(void *addr, register size_t size)
2266 MUTEX_ENTER(&rx_stats_mutex);
2268 rxi_Allocsize -= (afs_int32)size;
2269 MUTEX_EXIT(&rx_stats_mutex);
2271 osi_Free(addr, size);
2274 /* Find the peer process represented by the supplied (host,port)
2275 * combination. If there is no appropriate active peer structure, a
2276 * new one will be allocated and initialized
2277 * The origPeer, if set, is a pointer to a peer structure on which the
2278 * refcount will be be decremented. This is used to replace the peer
2279 * structure hanging off a connection structure */
2281 rxi_FindPeer(register afs_uint32 host, register u_short port,
2282 struct rx_peer *origPeer, int create)
2284 register struct rx_peer *pp;
2286 hashIndex = PEER_HASH(host, port);
2287 MUTEX_ENTER(&rx_peerHashTable_lock);
2288 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2289 if ((pp->host == host) && (pp->port == port))
2294 pp = rxi_AllocPeer(); /* This bzero's *pp */
2295 pp->host = host; /* set here or in InitPeerParams is zero */
2297 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2298 queue_Init(&pp->congestionQueue);
2299 queue_Init(&pp->rpcStats);
2300 pp->next = rx_peerHashTable[hashIndex];
2301 rx_peerHashTable[hashIndex] = pp;
2302 rxi_InitPeerParams(pp);
2303 MUTEX_ENTER(&rx_stats_mutex);
2304 rx_stats.nPeerStructs++;
2305 MUTEX_EXIT(&rx_stats_mutex);
2312 origPeer->refCount--;
2313 MUTEX_EXIT(&rx_peerHashTable_lock);
2318 /* Find the connection at (host, port) started at epoch, and with the
2319 * given connection id. Creates the server connection if necessary.
2320 * The type specifies whether a client connection or a server
2321 * connection is desired. In both cases, (host, port) specify the
2322 * peer's (host, pair) pair. Client connections are not made
2323 * automatically by this routine. The parameter socket gives the
2324 * socket descriptor on which the packet was received. This is used,
2325 * in the case of server connections, to check that *new* connections
2326 * come via a valid (port, serviceId). Finally, the securityIndex
2327 * parameter must match the existing index for the connection. If a
2328 * server connection is created, it will be created using the supplied
2329 * index, if the index is valid for this service */
2330 struct rx_connection *
2331 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2332 register u_short port, u_short serviceId, afs_uint32 cid,
2333 afs_uint32 epoch, int type, u_int securityIndex)
2335 int hashindex, flag;
2336 register struct rx_connection *conn;
2337 hashindex = CONN_HASH(host, port, cid, epoch, type);
2338 MUTEX_ENTER(&rx_connHashTable_lock);
2339 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2340 rx_connHashTable[hashindex],
2343 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2344 && (epoch == conn->epoch)) {
2345 register struct rx_peer *pp = conn->peer;
2346 if (securityIndex != conn->securityIndex) {
2347 /* this isn't supposed to happen, but someone could forge a packet
2348 * like this, and there seems to be some CM bug that makes this
2349 * happen from time to time -- in which case, the fileserver
2351 MUTEX_EXIT(&rx_connHashTable_lock);
2352 return (struct rx_connection *)0;
2354 if (pp->host == host && pp->port == port)
2356 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2358 /* So what happens when it's a callback connection? */
2359 if ( /*type == RX_CLIENT_CONNECTION && */
2360 (conn->epoch & 0x80000000))
2364 /* the connection rxLastConn that was used the last time is not the
2365 ** one we are looking for now. Hence, start searching in the hash */
2367 conn = rx_connHashTable[hashindex];
2372 struct rx_service *service;
2373 if (type == RX_CLIENT_CONNECTION) {
2374 MUTEX_EXIT(&rx_connHashTable_lock);
2375 return (struct rx_connection *)0;
2377 service = rxi_FindService(socket, serviceId);
2378 if (!service || (securityIndex >= service->nSecurityObjects)
2379 || (service->securityObjects[securityIndex] == 0)) {
2380 MUTEX_EXIT(&rx_connHashTable_lock);
2381 return (struct rx_connection *)0;
2383 conn = rxi_AllocConnection(); /* This bzero's the connection */
2384 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2385 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2386 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2387 conn->next = rx_connHashTable[hashindex];
2388 rx_connHashTable[hashindex] = conn;
2389 conn->peer = rxi_FindPeer(host, port, 0, 1);
2390 conn->type = RX_SERVER_CONNECTION;
2391 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2392 conn->epoch = epoch;
2393 conn->cid = cid & RX_CIDMASK;
2394 /* conn->serial = conn->lastSerial = 0; */
2395 /* conn->timeout = 0; */
2396 conn->ackRate = RX_FAST_ACK_RATE;
2397 conn->service = service;
2398 conn->serviceId = serviceId;
2399 conn->securityIndex = securityIndex;
2400 conn->securityObject = service->securityObjects[securityIndex];
2401 conn->nSpecific = 0;
2402 conn->specific = NULL;
2403 rx_SetConnDeadTime(conn, service->connDeadTime);
2404 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2405 /* Notify security object of the new connection */
2406 RXS_NewConnection(conn->securityObject, conn);
2407 /* XXXX Connection timeout? */
2408 if (service->newConnProc)
2409 (*service->newConnProc) (conn);
2410 MUTEX_ENTER(&rx_stats_mutex);
2411 rx_stats.nServerConns++;
2412 MUTEX_EXIT(&rx_stats_mutex);
2415 MUTEX_ENTER(&conn->conn_data_lock);
2417 MUTEX_EXIT(&conn->conn_data_lock);
2419 rxLastConn = conn; /* store this connection as the last conn used */
2420 MUTEX_EXIT(&rx_connHashTable_lock);
2424 /* There are two packet tracing routines available for testing and monitoring
2425 * Rx. One is called just after every packet is received and the other is
2426 * called just before every packet is sent. Received packets, have had their
2427 * headers decoded, and packets to be sent have not yet had their headers
2428 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2429 * containing the network address. Both can be modified. The return value, if
2430 * non-zero, indicates that the packet should be dropped. */
2432 int (*rx_justReceived) () = 0;
2433 int (*rx_almostSent) () = 0;
2435 /* A packet has been received off the interface. Np is the packet, socket is
2436 * the socket number it was received from (useful in determining which service
2437 * this packet corresponds to), and (host, port) reflect the host,port of the
2438 * sender. This call returns the packet to the caller if it is finished with
2439 * it, rather than de-allocating it, just as a small performance hack */
2442 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2443 afs_uint32 host, u_short port, int *tnop,
2444 struct rx_call **newcallp)
2446 register struct rx_call *call;
2447 register struct rx_connection *conn;
2449 afs_uint32 currentCallNumber;
2455 struct rx_packet *tnp;
2458 /* We don't print out the packet until now because (1) the time may not be
2459 * accurate enough until now in the lwp implementation (rx_Listener only gets
2460 * the time after the packet is read) and (2) from a protocol point of view,
2461 * this is the first time the packet has been seen */
2462 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2463 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2464 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2465 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2466 np->header.epoch, np->header.cid, np->header.callNumber,
2467 np->header.seq, np->header.flags, np));
2470 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2471 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2474 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2475 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2478 /* If an input tracer function is defined, call it with the packet and
2479 * network address. Note this function may modify its arguments. */
2480 if (rx_justReceived) {
2481 struct sockaddr_in addr;
2483 addr.sin_family = AF_INET;
2484 addr.sin_port = port;
2485 addr.sin_addr.s_addr = host;
2486 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2487 addr.sin_len = sizeof(addr);
2488 #endif /* AFS_OSF_ENV */
2489 drop = (*rx_justReceived) (np, &addr);
2490 /* drop packet if return value is non-zero */
2493 port = addr.sin_port; /* in case fcn changed addr */
2494 host = addr.sin_addr.s_addr;
2498 /* If packet was not sent by the client, then *we* must be the client */
2499 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2500 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2502 /* Find the connection (or fabricate one, if we're the server & if
2503 * necessary) associated with this packet */
2505 rxi_FindConnection(socket, host, port, np->header.serviceId,
2506 np->header.cid, np->header.epoch, type,
2507 np->header.securityIndex);
2510 /* If no connection found or fabricated, just ignore the packet.
2511 * (An argument could be made for sending an abort packet for
2516 MUTEX_ENTER(&conn->conn_data_lock);
2517 if (conn->maxSerial < np->header.serial)
2518 conn->maxSerial = np->header.serial;
2519 MUTEX_EXIT(&conn->conn_data_lock);
2521 /* If the connection is in an error state, send an abort packet and ignore
2522 * the incoming packet */
2524 /* Don't respond to an abort packet--we don't want loops! */
2525 MUTEX_ENTER(&conn->conn_data_lock);
2526 if (np->header.type != RX_PACKET_TYPE_ABORT)
2527 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2529 MUTEX_EXIT(&conn->conn_data_lock);
2533 /* Check for connection-only requests (i.e. not call specific). */
2534 if (np->header.callNumber == 0) {
2535 switch (np->header.type) {
2536 case RX_PACKET_TYPE_ABORT: {
2537 /* What if the supplied error is zero? */
2538 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2539 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2540 rxi_ConnectionError(conn, errcode);
2541 MUTEX_ENTER(&conn->conn_data_lock);
2543 MUTEX_EXIT(&conn->conn_data_lock);
2546 case RX_PACKET_TYPE_CHALLENGE:
2547 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2548 MUTEX_ENTER(&conn->conn_data_lock);
2550 MUTEX_EXIT(&conn->conn_data_lock);
2552 case RX_PACKET_TYPE_RESPONSE:
2553 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2554 MUTEX_ENTER(&conn->conn_data_lock);
2556 MUTEX_EXIT(&conn->conn_data_lock);
2558 case RX_PACKET_TYPE_PARAMS:
2559 case RX_PACKET_TYPE_PARAMS + 1:
2560 case RX_PACKET_TYPE_PARAMS + 2:
2561 /* ignore these packet types for now */
2562 MUTEX_ENTER(&conn->conn_data_lock);
2564 MUTEX_EXIT(&conn->conn_data_lock);
2569 /* Should not reach here, unless the peer is broken: send an
2571 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2572 MUTEX_ENTER(&conn->conn_data_lock);
2573 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2575 MUTEX_EXIT(&conn->conn_data_lock);
2580 channel = np->header.cid & RX_CHANNELMASK;
2581 call = conn->call[channel];
2582 #ifdef RX_ENABLE_LOCKS
2584 MUTEX_ENTER(&call->lock);
2585 /* Test to see if call struct is still attached to conn. */
2586 if (call != conn->call[channel]) {
2588 MUTEX_EXIT(&call->lock);
2589 if (type == RX_SERVER_CONNECTION) {
2590 call = conn->call[channel];
2591 /* If we started with no call attached and there is one now,
2592 * another thread is also running this routine and has gotten
2593 * the connection channel. We should drop this packet in the tests
2594 * below. If there was a call on this connection and it's now
2595 * gone, then we'll be making a new call below.
2596 * If there was previously a call and it's now different then
2597 * the old call was freed and another thread running this routine
2598 * has created a call on this channel. One of these two threads
2599 * has a packet for the old call and the code below handles those
2603 MUTEX_ENTER(&call->lock);
2605 /* This packet can't be for this call. If the new call address is
2606 * 0 then no call is running on this channel. If there is a call
2607 * then, since this is a client connection we're getting data for
2608 * it must be for the previous call.
2610 MUTEX_ENTER(&rx_stats_mutex);
2611 rx_stats.spuriousPacketsRead++;
2612 MUTEX_EXIT(&rx_stats_mutex);
2613 MUTEX_ENTER(&conn->conn_data_lock);
2615 MUTEX_EXIT(&conn->conn_data_lock);
2620 currentCallNumber = conn->callNumber[channel];
2622 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2623 if (np->header.callNumber < currentCallNumber) {
2624 MUTEX_ENTER(&rx_stats_mutex);
2625 rx_stats.spuriousPacketsRead++;
2626 MUTEX_EXIT(&rx_stats_mutex);
2627 #ifdef RX_ENABLE_LOCKS
2629 MUTEX_EXIT(&call->lock);
2631 MUTEX_ENTER(&conn->conn_data_lock);
2633 MUTEX_EXIT(&conn->conn_data_lock);
2637 MUTEX_ENTER(&conn->conn_call_lock);
2638 call = rxi_NewCall(conn, channel);
2639 MUTEX_EXIT(&conn->conn_call_lock);
2640 *call->callNumber = np->header.callNumber;
2641 if (np->header.callNumber == 0)
2642 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %lx resend %d.%0.3d len %d", np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port), np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq, np->header.flags, (unsigned long)np, np->retryTime.sec, np->retryTime.usec / 1000, np->length));
2644 call->state = RX_STATE_PRECALL;
2645 clock_GetTime(&call->queueTime);
2646 hzero(call->bytesSent);
2647 hzero(call->bytesRcvd);
2649 * If the number of queued calls exceeds the overload
2650 * threshold then abort this call.
2652 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2653 struct rx_packet *tp;
2655 rxi_CallError(call, rx_BusyError);
2656 tp = rxi_SendCallAbort(call, np, 1, 0);
2657 MUTEX_EXIT(&call->lock);
2658 MUTEX_ENTER(&conn->conn_data_lock);
2660 MUTEX_EXIT(&conn->conn_data_lock);
2661 MUTEX_ENTER(&rx_stats_mutex);
2663 MUTEX_EXIT(&rx_stats_mutex);
2666 rxi_KeepAliveOn(call);
2667 } else if (np->header.callNumber != currentCallNumber) {
2668 /* Wait until the transmit queue is idle before deciding
2669 * whether to reset the current call. Chances are that the
2670 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2673 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2674 while ((call->state == RX_STATE_ACTIVE)
2675 && (call->flags & RX_CALL_TQ_BUSY)) {
2676 call->flags |= RX_CALL_TQ_WAIT;
2678 #ifdef RX_ENABLE_LOCKS
2679 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2680 CV_WAIT(&call->cv_tq, &call->lock);
2681 #else /* RX_ENABLE_LOCKS */
2682 osi_rxSleep(&call->tq);
2683 #endif /* RX_ENABLE_LOCKS */
2685 if (call->tqWaiters == 0)
2686 call->flags &= ~RX_CALL_TQ_WAIT;
2688 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2689 /* If the new call cannot be taken right now send a busy and set
2690 * the error condition in this call, so that it terminates as
2691 * quickly as possible */
2692 if (call->state == RX_STATE_ACTIVE) {
2693 struct rx_packet *tp;
2695 rxi_CallError(call, RX_CALL_DEAD);
2696 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2698 MUTEX_EXIT(&call->lock);
2699 MUTEX_ENTER(&conn->conn_data_lock);
2701 MUTEX_EXIT(&conn->conn_data_lock);
2704 rxi_ResetCall(call, 0);
2705 *call->callNumber = np->header.callNumber;
2706 if (np->header.callNumber == 0)
2707 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %lx resend %d.%0.3d len %d", np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port), np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq, np->header.flags, (unsigned long)np, np->retryTime.sec, np->retryTime.usec / 1000, np->length));
2709 call->state = RX_STATE_PRECALL;
2710 clock_GetTime(&call->queueTime);
2711 hzero(call->bytesSent);
2712 hzero(call->bytesRcvd);
2714 * If the number of queued calls exceeds the overload
2715 * threshold then abort this call.
2717 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2718 struct rx_packet *tp;
2720 rxi_CallError(call, rx_BusyError);
2721 tp = rxi_SendCallAbort(call, np, 1, 0);
2722 MUTEX_EXIT(&call->lock);
2723 MUTEX_ENTER(&conn->conn_data_lock);
2725 MUTEX_EXIT(&conn->conn_data_lock);
2726 MUTEX_ENTER(&rx_stats_mutex);
2728 MUTEX_EXIT(&rx_stats_mutex);
2731 rxi_KeepAliveOn(call);
2733 /* Continuing call; do nothing here. */
2735 } else { /* we're the client */
2736 /* Ignore all incoming acknowledgements for calls in DALLY state */
2737 if (call && (call->state == RX_STATE_DALLY)
2738 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2739 MUTEX_ENTER(&rx_stats_mutex);
2740 rx_stats.ignorePacketDally++;
2741 MUTEX_EXIT(&rx_stats_mutex);
2742 #ifdef RX_ENABLE_LOCKS
2744 MUTEX_EXIT(&call->lock);
2747 MUTEX_ENTER(&conn->conn_data_lock);
2749 MUTEX_EXIT(&conn->conn_data_lock);
2753 /* Ignore anything that's not relevant to the current call. If there
2754 * isn't a current call, then no packet is relevant. */
2755 if (!call || (np->header.callNumber != currentCallNumber)) {
2756 MUTEX_ENTER(&rx_stats_mutex);
2757 rx_stats.spuriousPacketsRead++;
2758 MUTEX_EXIT(&rx_stats_mutex);
2759 #ifdef RX_ENABLE_LOCKS
2761 MUTEX_EXIT(&call->lock);
2764 MUTEX_ENTER(&conn->conn_data_lock);
2766 MUTEX_EXIT(&conn->conn_data_lock);
2769 /* If the service security object index stamped in the packet does not
2770 * match the connection's security index, ignore the packet */
2771 if (np->header.securityIndex != conn->securityIndex) {
2772 #ifdef RX_ENABLE_LOCKS
2773 MUTEX_EXIT(&call->lock);
2775 MUTEX_ENTER(&conn->conn_data_lock);
2777 MUTEX_EXIT(&conn->conn_data_lock);
2781 /* If we're receiving the response, then all transmit packets are
2782 * implicitly acknowledged. Get rid of them. */
2783 if (np->header.type == RX_PACKET_TYPE_DATA) {
2784 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2785 /* XXX Hack. Because we must release the global rx lock when
2786 * sending packets (osi_NetSend) we drop all acks while we're
2787 * traversing the tq in rxi_Start sending packets out because
2788 * packets may move to the freePacketQueue as result of being here!
2789 * So we drop these packets until we're safely out of the
2790 * traversing. Really ugly!
2791 * For fine grain RX locking, we set the acked field in the
2792 * packets and let rxi_Start remove them from the transmit queue.
2794 if (call->flags & RX_CALL_TQ_BUSY) {
2795 #ifdef RX_ENABLE_LOCKS
2796 rxi_SetAcksInTransmitQueue(call);
2799 return np; /* xmitting; drop packet */
2802 rxi_ClearTransmitQueue(call, 0);
2804 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2805 rxi_ClearTransmitQueue(call, 0);
2806 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2808 if (np->header.type == RX_PACKET_TYPE_ACK) {
2809 /* now check to see if this is an ack packet acknowledging that the
2810 * server actually *lost* some hard-acked data. If this happens we
2811 * ignore this packet, as it may indicate that the server restarted in
2812 * the middle of a call. It is also possible that this is an old ack
2813 * packet. We don't abort the connection in this case, because this
2814 * *might* just be an old ack packet. The right way to detect a server
2815 * restart in the midst of a call is to notice that the server epoch
2817 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2818 * XXX unacknowledged. I think that this is off-by-one, but
2819 * XXX I don't dare change it just yet, since it will
2820 * XXX interact badly with the server-restart detection
2821 * XXX code in receiveackpacket. */
2822 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2823 MUTEX_ENTER(&rx_stats_mutex);
2824 rx_stats.spuriousPacketsRead++;
2825 MUTEX_EXIT(&rx_stats_mutex);
2826 MUTEX_EXIT(&call->lock);
2827 MUTEX_ENTER(&conn->conn_data_lock);
2829 MUTEX_EXIT(&conn->conn_data_lock);
2833 } /* else not a data packet */
2836 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2837 /* Set remote user defined status from packet */
2838 call->remoteStatus = np->header.userStatus;
2840 /* Note the gap between the expected next packet and the actual
2841 * packet that arrived, when the new packet has a smaller serial number
2842 * than expected. Rioses frequently reorder packets all by themselves,
2843 * so this will be quite important with very large window sizes.
2844 * Skew is checked against 0 here to avoid any dependence on the type of
2845 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2847 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2848 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2849 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2851 MUTEX_ENTER(&conn->conn_data_lock);
2852 skew = conn->lastSerial - np->header.serial;
2853 conn->lastSerial = np->header.serial;
2854 MUTEX_EXIT(&conn->conn_data_lock);
2856 register struct rx_peer *peer;
2858 if (skew > peer->inPacketSkew) {
2859 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2861 peer->inPacketSkew = skew;
2865 /* Now do packet type-specific processing */
2866 switch (np->header.type) {
2867 case RX_PACKET_TYPE_DATA:
2868 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2871 case RX_PACKET_TYPE_ACK:
2872 /* Respond immediately to ack packets requesting acknowledgement
2874 if (np->header.flags & RX_REQUEST_ACK) {
2876 (void)rxi_SendCallAbort(call, 0, 1, 0);
2878 (void)rxi_SendAck(call, 0, np->header.serial,
2879 RX_ACK_PING_RESPONSE, 1);
2881 np = rxi_ReceiveAckPacket(call, np, 1);
2883 case RX_PACKET_TYPE_ABORT: {
2884 /* An abort packet: reset the call, passing the error up to the user. */
2885 /* What if error is zero? */
2886 /* What if the error is -1? the application will treat it as a timeout. */
2887 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
2888 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
2889 rxi_CallError(call, errdata);
2890 MUTEX_EXIT(&call->lock);
2891 MUTEX_ENTER(&conn->conn_data_lock);
2893 MUTEX_EXIT(&conn->conn_data_lock);
2894 return np; /* xmitting; drop packet */
2896 case RX_PACKET_TYPE_BUSY:
2899 case RX_PACKET_TYPE_ACKALL:
2900 /* All packets acknowledged, so we can drop all packets previously
2901 * readied for sending */
2902 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2903 /* XXX Hack. We because we can't release the global rx lock when
2904 * sending packets (osi_NetSend) we drop all ack pkts while we're
2905 * traversing the tq in rxi_Start sending packets out because
2906 * packets may move to the freePacketQueue as result of being
2907 * here! So we drop these packets until we're safely out of the
2908 * traversing. Really ugly!
2909 * For fine grain RX locking, we set the acked field in the packets
2910 * and let rxi_Start remove the packets from the transmit queue.
2912 if (call->flags & RX_CALL_TQ_BUSY) {
2913 #ifdef RX_ENABLE_LOCKS
2914 rxi_SetAcksInTransmitQueue(call);
2916 #else /* RX_ENABLE_LOCKS */
2917 MUTEX_EXIT(&call->lock);
2918 MUTEX_ENTER(&conn->conn_data_lock);
2920 MUTEX_EXIT(&conn->conn_data_lock);
2921 return np; /* xmitting; drop packet */
2922 #endif /* RX_ENABLE_LOCKS */
2924 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2925 rxi_ClearTransmitQueue(call, 0);
2928 /* Should not reach here, unless the peer is broken: send an abort
2930 rxi_CallError(call, RX_PROTOCOL_ERROR);
2931 np = rxi_SendCallAbort(call, np, 1, 0);
2934 /* Note when this last legitimate packet was received, for keep-alive
2935 * processing. Note, we delay getting the time until now in the hope that
2936 * the packet will be delivered to the user before any get time is required
2937 * (if not, then the time won't actually be re-evaluated here). */
2938 call->lastReceiveTime = clock_Sec();
2939 MUTEX_EXIT(&call->lock);
2940 MUTEX_ENTER(&conn->conn_data_lock);
2942 MUTEX_EXIT(&conn->conn_data_lock);
2946 /* return true if this is an "interesting" connection from the point of view
2947 of someone trying to debug the system */
2949 rxi_IsConnInteresting(struct rx_connection *aconn)
2952 register struct rx_call *tcall;
2954 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2956 for (i = 0; i < RX_MAXCALLS; i++) {
2957 tcall = aconn->call[i];
2959 if ((tcall->state == RX_STATE_PRECALL)
2960 || (tcall->state == RX_STATE_ACTIVE))
2962 if ((tcall->mode == RX_MODE_SENDING)
2963 || (tcall->mode == RX_MODE_RECEIVING))
2971 /* if this is one of the last few packets AND it wouldn't be used by the
2972 receiving call to immediately satisfy a read request, then drop it on
2973 the floor, since accepting it might prevent a lock-holding thread from
2974 making progress in its reading. If a call has been cleared while in
2975 the precall state then ignore all subsequent packets until the call
2976 is assigned to a thread. */
2979 TooLow(struct rx_packet *ap, struct rx_call *acall)
2982 MUTEX_ENTER(&rx_stats_mutex);
2983 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
2984 && (acall->state == RX_STATE_PRECALL))
2985 || ((rx_nFreePackets < rxi_dataQuota + 2)
2986 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
2987 && (acall->flags & RX_CALL_READER_WAIT)))) {
2990 MUTEX_EXIT(&rx_stats_mutex);
2996 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
2997 struct rx_call *acall)
2999 struct rx_call *call = acall;
3003 MUTEX_ENTER(&conn->conn_data_lock);
3004 conn->checkReachEvent = NULL;
3005 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3008 MUTEX_EXIT(&conn->conn_data_lock);
3012 MUTEX_ENTER(&conn->conn_call_lock);
3013 MUTEX_ENTER(&conn->conn_data_lock);
3014 for (i = 0; i < RX_MAXCALLS; i++) {
3015 struct rx_call *tc = conn->call[i];
3016 if (tc && tc->state == RX_STATE_PRECALL) {
3022 /* Indicate that rxi_CheckReachEvent is no longer running by
3023 * clearing the flag. Must be atomic under conn_data_lock to
3024 * avoid a new call slipping by: rxi_CheckConnReach holds
3025 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3027 conn->flags &= ~RX_CONN_ATTACHWAIT;
3028 MUTEX_EXIT(&conn->conn_data_lock);
3029 MUTEX_EXIT(&conn->conn_call_lock);
3034 MUTEX_ENTER(&call->lock);
3035 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3037 MUTEX_EXIT(&call->lock);
3039 clock_GetTime(&when);
3040 when.sec += RX_CHECKREACH_TIMEOUT;
3041 MUTEX_ENTER(&conn->conn_data_lock);
3042 if (!conn->checkReachEvent) {
3044 conn->checkReachEvent =
3045 rxevent_Post(&when, rxi_CheckReachEvent, conn, NULL);
3047 MUTEX_EXIT(&conn->conn_data_lock);
3053 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3055 struct rx_service *service = conn->service;
3056 struct rx_peer *peer = conn->peer;
3057 afs_uint32 now, lastReach;
3059 if (service->checkReach == 0)
3063 MUTEX_ENTER(&peer->peer_lock);
3064 lastReach = peer->lastReachTime;
3065 MUTEX_EXIT(&peer->peer_lock);
3066 if (now - lastReach < RX_CHECKREACH_TTL)
3069 MUTEX_ENTER(&conn->conn_data_lock);
3070 if (conn->flags & RX_CONN_ATTACHWAIT) {
3071 MUTEX_EXIT(&conn->conn_data_lock);
3074 conn->flags |= RX_CONN_ATTACHWAIT;
3075 MUTEX_EXIT(&conn->conn_data_lock);
3076 if (!conn->checkReachEvent)
3077 rxi_CheckReachEvent(NULL, conn, call);
3082 /* try to attach call, if authentication is complete */
3084 TryAttach(register struct rx_call *acall, register osi_socket socket,
3085 register int *tnop, register struct rx_call **newcallp,
3088 struct rx_connection *conn = acall->conn;
3090 if (conn->type == RX_SERVER_CONNECTION
3091 && acall->state == RX_STATE_PRECALL) {
3092 /* Don't attach until we have any req'd. authentication. */
3093 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3094 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3095 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3096 /* Note: this does not necessarily succeed; there
3097 * may not any proc available
3100 rxi_ChallengeOn(acall->conn);
3105 /* A data packet has been received off the interface. This packet is
3106 * appropriate to the call (the call is in the right state, etc.). This
3107 * routine can return a packet to the caller, for re-use */
3110 rxi_ReceiveDataPacket(register struct rx_call *call,
3111 register struct rx_packet *np, int istack,
3112 osi_socket socket, afs_uint32 host, u_short port,
3113 int *tnop, struct rx_call **newcallp)
3115 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3119 afs_uint32 seq, serial, flags;
3121 struct rx_packet *tnp;
3123 MUTEX_ENTER(&rx_stats_mutex);
3124 rx_stats.dataPacketsRead++;
3125 MUTEX_EXIT(&rx_stats_mutex);
3128 /* If there are no packet buffers, drop this new packet, unless we can find
3129 * packet buffers from inactive calls */
3131 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3132 MUTEX_ENTER(&rx_freePktQ_lock);
3133 rxi_NeedMorePackets = TRUE;
3134 MUTEX_EXIT(&rx_freePktQ_lock);
3135 MUTEX_ENTER(&rx_stats_mutex);
3136 rx_stats.noPacketBuffersOnRead++;
3137 MUTEX_EXIT(&rx_stats_mutex);
3138 call->rprev = np->header.serial;
3139 rxi_calltrace(RX_TRACE_DROP, call);
3140 dpf(("packet %x dropped on receipt - quota problems", np));
3142 rxi_ClearReceiveQueue(call);
3143 clock_GetTime(&when);
3144 clock_Add(&when, &rx_softAckDelay);
3145 if (!call->delayedAckEvent
3146 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3147 rxevent_Cancel(call->delayedAckEvent, call,
3148 RX_CALL_REFCOUNT_DELAY);
3149 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3150 call->delayedAckEvent =
3151 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3153 /* we've damaged this call already, might as well do it in. */
3159 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3160 * packet is one of several packets transmitted as a single
3161 * datagram. Do not send any soft or hard acks until all packets
3162 * in a jumbogram have been processed. Send negative acks right away.
3164 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3165 /* tnp is non-null when there are more packets in the
3166 * current jumbo gram */
3173 seq = np->header.seq;
3174 serial = np->header.serial;
3175 flags = np->header.flags;
3177 /* If the call is in an error state, send an abort message */
3179 return rxi_SendCallAbort(call, np, istack, 0);
3181 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3182 * AFS 3.5 jumbogram. */
3183 if (flags & RX_JUMBO_PACKET) {
3184 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3189 if (np->header.spare != 0) {
3190 MUTEX_ENTER(&call->conn->conn_data_lock);
3191 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3192 MUTEX_EXIT(&call->conn->conn_data_lock);
3195 /* The usual case is that this is the expected next packet */
3196 if (seq == call->rnext) {
3198 /* Check to make sure it is not a duplicate of one already queued */
3199 if (queue_IsNotEmpty(&call->rq)
3200 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3201 MUTEX_ENTER(&rx_stats_mutex);
3202 rx_stats.dupPacketsRead++;
3203 MUTEX_EXIT(&rx_stats_mutex);
3204 dpf(("packet %x dropped on receipt - duplicate", np));
3205 rxevent_Cancel(call->delayedAckEvent, call,
3206 RX_CALL_REFCOUNT_DELAY);
3207 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3213 /* It's the next packet. Stick it on the receive queue
3214 * for this call. Set newPackets to make sure we wake
3215 * the reader once all packets have been processed */
3216 queue_Prepend(&call->rq, np);
3218 np = NULL; /* We can't use this anymore */
3221 /* If an ack is requested then set a flag to make sure we
3222 * send an acknowledgement for this packet */
3223 if (flags & RX_REQUEST_ACK) {
3224 ackNeeded = RX_ACK_REQUESTED;
3227 /* Keep track of whether we have received the last packet */
3228 if (flags & RX_LAST_PACKET) {
3229 call->flags |= RX_CALL_HAVE_LAST;
3233 /* Check whether we have all of the packets for this call */
3234 if (call->flags & RX_CALL_HAVE_LAST) {
3235 afs_uint32 tseq; /* temporary sequence number */
3236 struct rx_packet *tp; /* Temporary packet pointer */
3237 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3239 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3240 if (tseq != tp->header.seq)
3242 if (tp->header.flags & RX_LAST_PACKET) {
3243 call->flags |= RX_CALL_RECEIVE_DONE;
3250 /* Provide asynchronous notification for those who want it
3251 * (e.g. multi rx) */
3252 if (call->arrivalProc) {
3253 (*call->arrivalProc) (call, call->arrivalProcHandle,
3254 call->arrivalProcArg);
3255 call->arrivalProc = (void (*)())0;
3258 /* Update last packet received */
3261 /* If there is no server process serving this call, grab
3262 * one, if available. We only need to do this once. If a
3263 * server thread is available, this thread becomes a server
3264 * thread and the server thread becomes a listener thread. */
3266 TryAttach(call, socket, tnop, newcallp, 0);
3269 /* This is not the expected next packet. */
3271 /* Determine whether this is a new or old packet, and if it's
3272 * a new one, whether it fits into the current receive window.
3273 * Also figure out whether the packet was delivered in sequence.
3274 * We use the prev variable to determine whether the new packet
3275 * is the successor of its immediate predecessor in the
3276 * receive queue, and the missing flag to determine whether
3277 * any of this packets predecessors are missing. */
3279 afs_uint32 prev; /* "Previous packet" sequence number */
3280 struct rx_packet *tp; /* Temporary packet pointer */
3281 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3282 int missing; /* Are any predecessors missing? */
3284 /* If the new packet's sequence number has been sent to the
3285 * application already, then this is a duplicate */
3286 if (seq < call->rnext) {
3287 MUTEX_ENTER(&rx_stats_mutex);
3288 rx_stats.dupPacketsRead++;
3289 MUTEX_EXIT(&rx_stats_mutex);
3290 rxevent_Cancel(call->delayedAckEvent, call,
3291 RX_CALL_REFCOUNT_DELAY);
3292 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3298 /* If the sequence number is greater than what can be
3299 * accomodated by the current window, then send a negative
3300 * acknowledge and drop the packet */
3301 if ((call->rnext + call->rwind) <= seq) {
3302 rxevent_Cancel(call->delayedAckEvent, call,
3303 RX_CALL_REFCOUNT_DELAY);
3304 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3311 /* Look for the packet in the queue of old received packets */
3312 for (prev = call->rnext - 1, missing =
3313 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3314 /*Check for duplicate packet */
3315 if (seq == tp->header.seq) {
3316 MUTEX_ENTER(&rx_stats_mutex);
3317 rx_stats.dupPacketsRead++;
3318 MUTEX_EXIT(&rx_stats_mutex);
3319 rxevent_Cancel(call->delayedAckEvent, call,
3320 RX_CALL_REFCOUNT_DELAY);
3321 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3327 /* If we find a higher sequence packet, break out and
3328 * insert the new packet here. */
3329 if (seq < tp->header.seq)
3331 /* Check for missing packet */
3332 if (tp->header.seq != prev + 1) {
3336 prev = tp->header.seq;
3339 /* Keep track of whether we have received the last packet. */
3340 if (flags & RX_LAST_PACKET) {
3341 call->flags |= RX_CALL_HAVE_LAST;
3344 /* It's within the window: add it to the the receive queue.
3345 * tp is left by the previous loop either pointing at the
3346 * packet before which to insert the new packet, or at the
3347 * queue head if the queue is empty or the packet should be
3349 queue_InsertBefore(tp, np);
3353 /* Check whether we have all of the packets for this call */
3354 if ((call->flags & RX_CALL_HAVE_LAST)
3355 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3356 afs_uint32 tseq; /* temporary sequence number */
3359 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3360 if (tseq != tp->header.seq)
3362 if (tp->header.flags & RX_LAST_PACKET) {
3363 call->flags |= RX_CALL_RECEIVE_DONE;
3370 /* We need to send an ack of the packet is out of sequence,
3371 * or if an ack was requested by the peer. */
3372 if (seq != prev + 1 || missing) {
3373 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3374 } else if (flags & RX_REQUEST_ACK) {
3375 ackNeeded = RX_ACK_REQUESTED;
3378 /* Acknowledge the last packet for each call */
3379 if (flags & RX_LAST_PACKET) {
3390 * If the receiver is waiting for an iovec, fill the iovec
3391 * using the data from the receive queue */
3392 if (call->flags & RX_CALL_IOVEC_WAIT) {
3393 didHardAck = rxi_FillReadVec(call, serial);
3394 /* the call may have been aborted */
3403 /* Wakeup the reader if any */
3404 if ((call->flags & RX_CALL_READER_WAIT)
3405 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3406 || (call->iovNext >= call->iovMax)
3407 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3408 call->flags &= ~RX_CALL_READER_WAIT;
3409 #ifdef RX_ENABLE_LOCKS
3410 CV_BROADCAST(&call->cv_rq);
3412 osi_rxWakeup(&call->rq);
3418 * Send an ack when requested by the peer, or once every
3419 * rxi_SoftAckRate packets until the last packet has been
3420 * received. Always send a soft ack for the last packet in
3421 * the server's reply. */
3423 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3424 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3425 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3426 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3427 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3428 } else if (call->nSoftAcks) {
3429 clock_GetTime(&when);
3430 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3431 clock_Add(&when, &rx_lastAckDelay);
3433 clock_Add(&when, &rx_softAckDelay);
3435 if (!call->delayedAckEvent
3436 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3437 rxevent_Cancel(call->delayedAckEvent, call,
3438 RX_CALL_REFCOUNT_DELAY);
3439 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3440 call->delayedAckEvent =
3441 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3443 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3444 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3451 static void rxi_ComputeRate();
3455 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3457 struct rx_peer *peer = conn->peer;
3459 MUTEX_ENTER(&peer->peer_lock);
3460 peer->lastReachTime = clock_Sec();
3461 MUTEX_EXIT(&peer->peer_lock);
3463 MUTEX_ENTER(&conn->conn_data_lock);
3464 if (conn->flags & RX_CONN_ATTACHWAIT) {
3467 conn->flags &= ~RX_CONN_ATTACHWAIT;
3468 MUTEX_EXIT(&conn->conn_data_lock);
3470 for (i = 0; i < RX_MAXCALLS; i++) {
3471 struct rx_call *call = conn->call[i];
3474 MUTEX_ENTER(&call->lock);
3475 /* tnop can be null if newcallp is null */
3476 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3478 MUTEX_EXIT(&call->lock);
3482 MUTEX_EXIT(&conn->conn_data_lock);
3486 rx_ack_reason(int reason)
3489 case RX_ACK_REQUESTED:
3491 case RX_ACK_DUPLICATE:
3493 case RX_ACK_OUT_OF_SEQUENCE:
3495 case RX_ACK_EXCEEDS_WINDOW:
3497 case RX_ACK_NOSPACE:
3501 case RX_ACK_PING_RESPONSE:
3513 /* rxi_ComputePeerNetStats
3515 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3516 * estimates (like RTT and throughput) based on ack packets. Caller
3517 * must ensure that the packet in question is the right one (i.e.
3518 * serial number matches).
3521 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3522 struct rx_ackPacket *ap, struct rx_packet *np)
3524 struct rx_peer *peer = call->conn->peer;
3526 /* Use RTT if not delayed by client. */
3527 if (ap->reason != RX_ACK_DELAY)
3528 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3530 rxi_ComputeRate(peer, call, p, np, ap->reason);
3534 /* The real smarts of the whole thing. */
3536 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3539 struct rx_ackPacket *ap;
3541 register struct rx_packet *tp;
3542 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3543 register struct rx_connection *conn = call->conn;
3544 struct rx_peer *peer = conn->peer;
3547 /* because there are CM's that are bogus, sending weird values for this. */
3548 afs_uint32 skew = 0;
3553 int newAckCount = 0;
3554 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3555 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3557 MUTEX_ENTER(&rx_stats_mutex);
3558 rx_stats.ackPacketsRead++;
3559 MUTEX_EXIT(&rx_stats_mutex);
3560 ap = (struct rx_ackPacket *)rx_DataOf(np);
3561 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3563 return np; /* truncated ack packet */
3565 /* depends on ack packet struct */
3566 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3567 first = ntohl(ap->firstPacket);
3568 serial = ntohl(ap->serial);
3569 /* temporarily disabled -- needs to degrade over time
3570 * skew = ntohs(ap->maxSkew); */
3572 /* Ignore ack packets received out of order */
3573 if (first < call->tfirst) {
3577 if (np->header.flags & RX_SLOW_START_OK) {
3578 call->flags |= RX_CALL_SLOW_START_OK;
3581 if (ap->reason == RX_ACK_PING_RESPONSE)
3582 rxi_UpdatePeerReach(conn, call);
3586 if (rxdebug_active) {
3590 len = _snprintf(msg, sizeof(msg),
3591 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3592 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3593 ntohl(ap->serial), ntohl(ap->previousPacket),
3594 (unsigned int)np->header.seq, (unsigned int)skew,
3595 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3599 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3600 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3604 OutputDebugString(msg);
3606 #else /* AFS_NT40_ENV */
3609 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3610 ap->reason, ntohl(ap->previousPacket),
3611 (unsigned int)np->header.seq, (unsigned int)serial,
3612 (unsigned int)skew, ntohl(ap->firstPacket));
3615 for (offset = 0; offset < nAcks; offset++)
3616 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3621 #endif /* AFS_NT40_ENV */
3624 /* Update the outgoing packet skew value to the latest value of
3625 * the peer's incoming packet skew value. The ack packet, of
3626 * course, could arrive out of order, but that won't affect things
3628 MUTEX_ENTER(&peer->peer_lock);
3629 peer->outPacketSkew = skew;
3631 /* Check for packets that no longer need to be transmitted, and
3632 * discard them. This only applies to packets positively
3633 * acknowledged as having been sent to the peer's upper level.
3634 * All other packets must be retained. So only packets with
3635 * sequence numbers < ap->firstPacket are candidates. */
3636 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3637 if (tp->header.seq >= first)
3639 call->tfirst = tp->header.seq + 1;
3641 && (tp->header.serial == serial || tp->firstSerial == serial))
3642 rxi_ComputePeerNetStats(call, tp, ap, np);
3643 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3646 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3647 /* XXX Hack. Because we have to release the global rx lock when sending
3648 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3649 * in rxi_Start sending packets out because packets may move to the
3650 * freePacketQueue as result of being here! So we drop these packets until
3651 * we're safely out of the traversing. Really ugly!
3652 * To make it even uglier, if we're using fine grain locking, we can
3653 * set the ack bits in the packets and have rxi_Start remove the packets
3654 * when it's done transmitting.
3656 if (call->flags & RX_CALL_TQ_BUSY) {
3657 #ifdef RX_ENABLE_LOCKS
3658 tp->flags |= RX_PKTFLAG_ACKED;
3659 call->flags |= RX_CALL_TQ_SOME_ACKED;
3660 #else /* RX_ENABLE_LOCKS */
3662 #endif /* RX_ENABLE_LOCKS */
3664 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3667 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3672 /* Give rate detector a chance to respond to ping requests */
3673 if (ap->reason == RX_ACK_PING_RESPONSE) {
3674 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3678 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3680 /* Now go through explicit acks/nacks and record the results in
3681 * the waiting packets. These are packets that can't be released
3682 * yet, even with a positive acknowledge. This positive
3683 * acknowledge only means the packet has been received by the
3684 * peer, not that it will be retained long enough to be sent to
3685 * the peer's upper level. In addition, reset the transmit timers
3686 * of any missing packets (those packets that must be missing
3687 * because this packet was out of sequence) */
3689 call->nSoftAcked = 0;
3690 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3691 /* Update round trip time if the ack was stimulated on receipt
3693 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3694 #ifdef RX_ENABLE_LOCKS
3695 if (tp->header.seq >= first)
3696 #endif /* RX_ENABLE_LOCKS */
3697 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3699 && (tp->header.serial == serial || tp->firstSerial == serial))
3700 rxi_ComputePeerNetStats(call, tp, ap, np);
3702 /* Set the acknowledge flag per packet based on the
3703 * information in the ack packet. An acknowlegded packet can
3704 * be downgraded when the server has discarded a packet it
3705 * soacked previously, or when an ack packet is received
3706 * out of sequence. */
3707 if (tp->header.seq < first) {
3708 /* Implicit ack information */
3709 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3712 tp->flags |= RX_PKTFLAG_ACKED;
3713 } else if (tp->header.seq < first + nAcks) {
3714 /* Explicit ack information: set it in the packet appropriately */
3715 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3716 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3718 tp->flags |= RX_PKTFLAG_ACKED;
3725 } else /* RX_ACK_TYPE_NACK */ {
3726 tp->flags &= ~RX_PKTFLAG_ACKED;
3730 tp->flags &= ~RX_PKTFLAG_ACKED;
3734 /* If packet isn't yet acked, and it has been transmitted at least
3735 * once, reset retransmit time using latest timeout
3736 * ie, this should readjust the retransmit timer for all outstanding
3737 * packets... So we don't just retransmit when we should know better*/
3739 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3740 tp->retryTime = tp->timeSent;
3741 clock_Add(&tp->retryTime, &peer->timeout);
3742 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3743 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3747 /* If the window has been extended by this acknowledge packet,
3748 * then wakeup a sender waiting in alloc for window space, or try
3749 * sending packets now, if he's been sitting on packets due to
3750 * lack of window space */
3751 if (call->tnext < (call->tfirst + call->twind)) {
3752 #ifdef RX_ENABLE_LOCKS
3753 CV_SIGNAL(&call->cv_twind);
3755 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3756 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3757 osi_rxWakeup(&call->twind);
3760 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3761 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3765 /* if the ack packet has a receivelen field hanging off it,
3766 * update our state */
3767 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3770 /* If the ack packet has a "recommended" size that is less than
3771 * what I am using now, reduce my size to match */
3772 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3773 (int)sizeof(afs_int32), &tSize);
3774 tSize = (afs_uint32) ntohl(tSize);
3775 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3777 /* Get the maximum packet size to send to this peer */
3778 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
3780 tSize = (afs_uint32) ntohl(tSize);
3781 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3782 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3784 /* sanity check - peer might have restarted with different params.
3785 * If peer says "send less", dammit, send less... Peer should never
3786 * be unable to accept packets of the size that prior AFS versions would
3787 * send without asking. */
3788 if (peer->maxMTU != tSize) {
3789 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
3791 peer->maxMTU = tSize;
3792 peer->MTU = MIN(tSize, peer->MTU);
3793 call->MTU = MIN(call->MTU, tSize);
3796 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3799 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3800 (int)sizeof(afs_int32), &tSize);
3801 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3802 if (tSize < call->twind) { /* smaller than our send */
3803 call->twind = tSize; /* window, we must send less... */
3804 call->ssthresh = MIN(call->twind, call->ssthresh);
3807 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3808 * network MTU confused with the loopback MTU. Calculate the
3809 * maximum MTU here for use in the slow start code below.
3811 maxMTU = peer->maxMTU;
3812 /* Did peer restart with older RX version? */
3813 if (peer->maxDgramPackets > 1) {
3814 peer->maxDgramPackets = 1;
3816 } else if (np->length >=
3817 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3820 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3821 sizeof(afs_int32), &tSize);
3822 tSize = (afs_uint32) ntohl(tSize);
3824 * As of AFS 3.5 we set the send window to match the receive window.
3826 if (tSize < call->twind) {
3827 call->twind = tSize;
3828 call->ssthresh = MIN(call->twind, call->ssthresh);
3829 } else if (tSize > call->twind) {
3830 call->twind = tSize;
3834 * As of AFS 3.5, a jumbogram is more than one fixed size
3835 * packet transmitted in a single UDP datagram. If the remote
3836 * MTU is smaller than our local MTU then never send a datagram
3837 * larger than the natural MTU.
3840 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3841 sizeof(afs_int32), &tSize);
3842 maxDgramPackets = (afs_uint32) ntohl(tSize);
3843 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3844 maxDgramPackets = MIN(maxDgramPackets, peer->ifDgramPackets);
3845 if (peer->natMTU < peer->ifMTU)
3846 maxDgramPackets = MIN(maxDgramPackets, rxi_AdjustDgramPackets(1, peer->natMTU));
3847 if (maxDgramPackets > 1) {
3848 peer->maxDgramPackets = maxDgramPackets;
3849 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3851 peer->maxDgramPackets = 1;
3852 call->MTU = peer->natMTU;
3854 } else if (peer->maxDgramPackets > 1) {
3855 /* Restarted with lower version of RX */
3856 peer->maxDgramPackets = 1;
3858 } else if (peer->maxDgramPackets > 1
3859 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3860 /* Restarted with lower version of RX */
3861 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3862 peer->natMTU = OLD_MAX_PACKET_SIZE;
3863 peer->MTU = OLD_MAX_PACKET_SIZE;
3864 peer->maxDgramPackets = 1;
3865 peer->nDgramPackets = 1;
3867 call->MTU = OLD_MAX_PACKET_SIZE;
3872 * Calculate how many datagrams were successfully received after
3873 * the first missing packet and adjust the negative ack counter
3878 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3879 if (call->nNacks < nNacked) {
3880 call->nNacks = nNacked;
3889 if (call->flags & RX_CALL_FAST_RECOVER) {
3891 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3893 call->flags &= ~RX_CALL_FAST_RECOVER;
3894 call->cwind = call->nextCwind;
3895 call->nextCwind = 0;
3898 call->nCwindAcks = 0;
3899 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3900 /* Three negative acks in a row trigger congestion recovery */
3901 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3902 MUTEX_EXIT(&peer->peer_lock);
3903 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3904 /* someone else is waiting to start recovery */
3907 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3908 rxi_WaitforTQBusy(call);
3909 MUTEX_ENTER(&peer->peer_lock);
3910 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3911 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3912 call->flags |= RX_CALL_FAST_RECOVER;
3913 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3915 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3916 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3917 call->nextCwind = call->ssthresh;
3920 peer->MTU = call->MTU;
3921 peer->cwind = call->nextCwind;
3922 peer->nDgramPackets = call->nDgramPackets;
3924 call->congestSeq = peer->congestSeq;
3925 /* Reset the resend times on the packets that were nacked
3926 * so we will retransmit as soon as the window permits*/
3927 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3929 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3930 clock_Zero(&tp->retryTime);
3932 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3937 /* If cwind is smaller than ssthresh, then increase
3938 * the window one packet for each ack we receive (exponential
3940 * If cwind is greater than or equal to ssthresh then increase
3941 * the congestion window by one packet for each cwind acks we
3942 * receive (linear growth). */
3943 if (call->cwind < call->ssthresh) {
3945 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3946 call->nCwindAcks = 0;
3948 call->nCwindAcks += newAckCount;
3949 if (call->nCwindAcks >= call->cwind) {
3950 call->nCwindAcks = 0;
3951 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3955 * If we have received several acknowledgements in a row then
3956 * it is time to increase the size of our datagrams
3958 if ((int)call->nAcks > rx_nDgramThreshold) {
3959 if (peer->maxDgramPackets > 1) {
3960 if (call->nDgramPackets < peer->maxDgramPackets) {
3961 call->nDgramPackets++;
3963 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3964 } else if (call->MTU < peer->maxMTU) {
3965 call->MTU += peer->natMTU;
3966 call->MTU = MIN(call->MTU, peer->maxMTU);
3972 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3974 /* Servers need to hold the call until all response packets have
3975 * been acknowledged. Soft acks are good enough since clients
3976 * are not allowed to clear their receive queues. */
3977 if (call->state == RX_STATE_HOLD
3978 && call->tfirst + call->nSoftAcked >= call->tnext) {
3979 call->state = RX_STATE_DALLY;
3980 rxi_ClearTransmitQueue(call, 0);
3981 } else if (!queue_IsEmpty(&call->tq)) {
3982 rxi_Start(0, call, 0, istack);
3987 /* Received a response to a challenge packet */
3989 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
3990 register struct rx_packet *np, int istack)
3994 /* Ignore the packet if we're the client */
3995 if (conn->type == RX_CLIENT_CONNECTION)
3998 /* If already authenticated, ignore the packet (it's probably a retry) */
3999 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4002 /* Otherwise, have the security object evaluate the response packet */
4003 error = RXS_CheckResponse(conn->securityObject, conn, np);
4005 /* If the response is invalid, reset the connection, sending
4006 * an abort to the peer */
4010 rxi_ConnectionError(conn, error);
4011 MUTEX_ENTER(&conn->conn_data_lock);
4012 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4013 MUTEX_EXIT(&conn->conn_data_lock);
4016 /* If the response is valid, any calls waiting to attach
4017 * servers can now do so */
4020 for (i = 0; i < RX_MAXCALLS; i++) {
4021 struct rx_call *call = conn->call[i];
4023 MUTEX_ENTER(&call->lock);
4024 if (call->state == RX_STATE_PRECALL)
4025 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4026 /* tnop can be null if newcallp is null */
4027 MUTEX_EXIT(&call->lock);
4031 /* Update the peer reachability information, just in case
4032 * some calls went into attach-wait while we were waiting
4033 * for authentication..
4035 rxi_UpdatePeerReach(conn, NULL);
4040 /* A client has received an authentication challenge: the security
4041 * object is asked to cough up a respectable response packet to send
4042 * back to the server. The server is responsible for retrying the
4043 * challenge if it fails to get a response. */
4046 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
4047 register struct rx_packet *np, int istack)
4051 /* Ignore the challenge if we're the server */
4052 if (conn->type == RX_SERVER_CONNECTION)
4055 /* Ignore the challenge if the connection is otherwise idle; someone's
4056 * trying to use us as an oracle. */
4057 if (!rxi_HasActiveCalls(conn))
4060 /* Send the security object the challenge packet. It is expected to fill
4061 * in the response. */
4062 error = RXS_GetResponse(conn->securityObject, conn, np);
4064 /* If the security object is unable to return a valid response, reset the
4065 * connection and send an abort to the peer. Otherwise send the response
4066 * packet to the peer connection. */
4068 rxi_ConnectionError(conn, error);
4069 MUTEX_ENTER(&conn->conn_data_lock);
4070 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4071 MUTEX_EXIT(&conn->conn_data_lock);
4073 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4074 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4080 /* Find an available server process to service the current request in
4081 * the given call structure. If one isn't available, queue up this
4082 * call so it eventually gets one */
4084 rxi_AttachServerProc(register struct rx_call *call,
4085 register osi_socket socket, register int *tnop,
4086 register struct rx_call **newcallp)
4088 register struct rx_serverQueueEntry *sq;
4089 register struct rx_service *service = call->conn->service;
4090 register int haveQuota = 0;
4092 /* May already be attached */
4093 if (call->state == RX_STATE_ACTIVE)
4096 MUTEX_ENTER(&rx_serverPool_lock);
4098 haveQuota = QuotaOK(service);
4099 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4100 /* If there are no processes available to service this call,
4101 * put the call on the incoming call queue (unless it's
4102 * already on the queue).
4104 #ifdef RX_ENABLE_LOCKS
4106 ReturnToServerPool(service);
4107 #endif /* RX_ENABLE_LOCKS */
4109 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4110 call->flags |= RX_CALL_WAIT_PROC;
4111 MUTEX_ENTER(&rx_stats_mutex);
4114 MUTEX_EXIT(&rx_stats_mutex);
4115 rxi_calltrace(RX_CALL_ARRIVAL, call);
4116 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4117 queue_Append(&rx_incomingCallQueue, call);
4120 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4122 /* If hot threads are enabled, and both newcallp and sq->socketp
4123 * are non-null, then this thread will process the call, and the
4124 * idle server thread will start listening on this threads socket.
4127 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4130 *sq->socketp = socket;
4131 clock_GetTime(&call->startTime);
4132 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4136 if (call->flags & RX_CALL_WAIT_PROC) {
4137 /* Conservative: I don't think this should happen */
4138 call->flags &= ~RX_CALL_WAIT_PROC;
4139 if (queue_IsOnQueue(call)) {
4141 MUTEX_ENTER(&rx_stats_mutex);
4143 MUTEX_EXIT(&rx_stats_mutex);
4146 call->state = RX_STATE_ACTIVE;
4147 call->mode = RX_MODE_RECEIVING;
4148 #ifdef RX_KERNEL_TRACE
4150 int glockOwner = ISAFS_GLOCK();
4153 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4154 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4160 if (call->flags & RX_CALL_CLEARED) {
4161 /* send an ack now to start the packet flow up again */
4162 call->flags &= ~RX_CALL_CLEARED;
4163 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4165 #ifdef RX_ENABLE_LOCKS
4168 service->nRequestsRunning++;
4169 if (service->nRequestsRunning <= service->minProcs)
4175 MUTEX_EXIT(&rx_serverPool_lock);
4178 /* Delay the sending of an acknowledge event for a short while, while
4179 * a new call is being prepared (in the case of a client) or a reply
4180 * is being prepared (in the case of a server). Rather than sending
4181 * an ack packet, an ACKALL packet is sent. */
4183 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4185 #ifdef RX_ENABLE_LOCKS
4187 MUTEX_ENTER(&call->lock);
4188 call->delayedAckEvent = NULL;
4189 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4191 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4192 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4194 MUTEX_EXIT(&call->lock);
4195 #else /* RX_ENABLE_LOCKS */
4197 call->delayedAckEvent = NULL;
4198 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4199 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4200 #endif /* RX_ENABLE_LOCKS */
4204 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4207 #ifdef RX_ENABLE_LOCKS
4209 MUTEX_ENTER(&call->lock);
4210 if (event == call->delayedAckEvent)
4211 call->delayedAckEvent = NULL;
4212 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4214 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4216 MUTEX_EXIT(&call->lock);
4217 #else /* RX_ENABLE_LOCKS */
4219 call->delayedAckEvent = NULL;
4220 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4221 #endif /* RX_ENABLE_LOCKS */
4225 #ifdef RX_ENABLE_LOCKS
4226 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4227 * clearing them out.
4230 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4232 register struct rx_packet *p, *tp;
4235 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4236 p->flags |= RX_PKTFLAG_ACKED;
4240 call->flags |= RX_CALL_TQ_CLEARME;
4241 call->flags |= RX_CALL_TQ_SOME_ACKED;
4244 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4245 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4246 call->tfirst = call->tnext;
4247 call->nSoftAcked = 0;
4249 if (call->flags & RX_CALL_FAST_RECOVER) {
4250 call->flags &= ~RX_CALL_FAST_RECOVER;
4251 call->cwind = call->nextCwind;
4252 call->nextCwind = 0;
4255 CV_SIGNAL(&call->cv_twind);
4257 #endif /* RX_ENABLE_LOCKS */
4259 /* Clear out the transmit queue for the current call (all packets have
4260 * been received by peer) */
4262 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4264 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4265 register struct rx_packet *p, *tp;
4267 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4269 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4270 p->flags |= RX_PKTFLAG_ACKED;
4274 call->flags |= RX_CALL_TQ_CLEARME;
4275 call->flags |= RX_CALL_TQ_SOME_ACKED;
4278 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4279 rxi_FreePackets(0, &call->tq);
4280 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4281 call->flags &= ~RX_CALL_TQ_CLEARME;
4283 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4285 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4286 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4287 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4288 call->nSoftAcked = 0;
4290 if (call->flags & RX_CALL_FAST_RECOVER) {
4291 call->flags &= ~RX_CALL_FAST_RECOVER;
4292 call->cwind = call->nextCwind;
4294 #ifdef RX_ENABLE_LOCKS
4295 CV_SIGNAL(&call->cv_twind);
4297 osi_rxWakeup(&call->twind);
4302 rxi_ClearReceiveQueue(register struct rx_call *call)
4304 if (queue_IsNotEmpty(&call->rq)) {
4305 rx_packetReclaims += rxi_FreePackets(0, &call->rq);
4306 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4308 if (call->state == RX_STATE_PRECALL) {
4309 call->flags |= RX_CALL_CLEARED;
4313 /* Send an abort packet for the specified call */
4315 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4316 int istack, int force)
4324 /* Clients should never delay abort messages */
4325 if (rx_IsClientConn(call->conn))
4328 if (call->abortCode != call->error) {
4329 call->abortCode = call->error;
4330 call->abortCount = 0;
4333 if (force || rxi_callAbortThreshhold == 0
4334 || call->abortCount < rxi_callAbortThreshhold) {
4335 if (call->delayedAbortEvent) {
4336 rxevent_Cancel(call->delayedAbortEvent, call,
4337 RX_CALL_REFCOUNT_ABORT);
4339 error = htonl(call->error);
4342 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4343 (char *)&error, sizeof(error), istack);
4344 } else if (!call->delayedAbortEvent) {
4345 clock_GetTime(&when);
4346 clock_Addmsec(&when, rxi_callAbortDelay);
4347 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4348 call->delayedAbortEvent =
4349 rxevent_Post(&when, rxi_SendDelayedCallAbort, call, 0);
4354 /* Send an abort packet for the specified connection. Packet is an
4355 * optional pointer to a packet that can be used to send the abort.
4356 * Once the number of abort messages reaches the threshhold, an
4357 * event is scheduled to send the abort. Setting the force flag
4358 * overrides sending delayed abort messages.
4360 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4361 * to send the abort packet.
4364 rxi_SendConnectionAbort(register struct rx_connection *conn,
4365 struct rx_packet *packet, int istack, int force)
4373 /* Clients should never delay abort messages */
4374 if (rx_IsClientConn(conn))
4377 if (force || rxi_connAbortThreshhold == 0
4378 || conn->abortCount < rxi_connAbortThreshhold) {
4379 if (conn->delayedAbortEvent) {
4380 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4382 error = htonl(conn->error);
4384 MUTEX_EXIT(&conn->conn_data_lock);
4386 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4387 RX_PACKET_TYPE_ABORT, (char *)&error,
4388 sizeof(error), istack);
4389 MUTEX_ENTER(&conn->conn_data_lock);
4390 } else if (!conn->delayedAbortEvent) {
4391 clock_GetTime(&when);
4392 clock_Addmsec(&when, rxi_connAbortDelay);
4393 conn->delayedAbortEvent =
4394 rxevent_Post(&when, rxi_SendDelayedConnAbort, conn, 0);
4399 /* Associate an error all of the calls owned by a connection. Called
4400 * with error non-zero. This is only for really fatal things, like
4401 * bad authentication responses. The connection itself is set in
4402 * error at this point, so that future packets received will be
4405 rxi_ConnectionError(register struct rx_connection *conn,
4406 register afs_int32 error)
4411 dpf(("rxi_ConnectionError conn %x error %d", conn, error));
4413 MUTEX_ENTER(&conn->conn_data_lock);
4414 if (conn->challengeEvent)
4415 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4416 if (conn->checkReachEvent) {
4417 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4418 conn->checkReachEvent = 0;
4419 conn->flags &= ~RX_CONN_ATTACHWAIT;
4422 MUTEX_EXIT(&conn->conn_data_lock);
4423 for (i = 0; i < RX_MAXCALLS; i++) {
4424 struct rx_call *call = conn->call[i];
4426 MUTEX_ENTER(&call->lock);
4427 rxi_CallError(call, error);
4428 MUTEX_EXIT(&call->lock);
4431 conn->error = error;
4432 MUTEX_ENTER(&rx_stats_mutex);
4433 rx_stats.fatalErrors++;
4434 MUTEX_EXIT(&rx_stats_mutex);
4439 rxi_CallError(register struct rx_call *call, afs_int32 error)
4441 dpf(("rxi_CallError call %x error %d call->error %d", call, error, call->error));
4443 error = call->error;
4445 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4446 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4447 rxi_ResetCall(call, 0);
4450 rxi_ResetCall(call, 0);
4452 call->error = error;
4453 call->mode = RX_MODE_ERROR;
4456 /* Reset various fields in a call structure, and wakeup waiting
4457 * processes. Some fields aren't changed: state & mode are not
4458 * touched (these must be set by the caller), and bufptr, nLeft, and
4459 * nFree are not reset, since these fields are manipulated by
4460 * unprotected macros, and may only be reset by non-interrupting code.
4463 /* this code requires that call->conn be set properly as a pre-condition. */
4464 #endif /* ADAPT_WINDOW */
4467 rxi_ResetCall(register struct rx_call *call, register int newcall)
4470 register struct rx_peer *peer;
4471 struct rx_packet *packet;
4473 dpf(("rxi_ResetCall(call %x, newcall %d)\n", call, newcall));
4475 /* Notify anyone who is waiting for asynchronous packet arrival */
4476 if (call->arrivalProc) {
4477 (*call->arrivalProc) (call, call->arrivalProcHandle,
4478 call->arrivalProcArg);
4479 call->arrivalProc = (void (*)())0;
4482 if (call->delayedAbortEvent) {
4483 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4484 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4486 rxi_SendCallAbort(call, packet, 0, 1);
4487 rxi_FreePacket(packet);
4492 * Update the peer with the congestion information in this call
4493 * so other calls on this connection can pick up where this call
4494 * left off. If the congestion sequence numbers don't match then
4495 * another call experienced a retransmission.
4497 peer = call->conn->peer;
4498 MUTEX_ENTER(&peer->peer_lock);
4500 if (call->congestSeq == peer->congestSeq) {
4501 peer->cwind = MAX(peer->cwind, call->cwind);
4502 peer->MTU = MAX(peer->MTU, call->MTU);
4503 peer->nDgramPackets =
4504 MAX(peer->nDgramPackets, call->nDgramPackets);
4507 call->abortCode = 0;
4508 call->abortCount = 0;
4510 if (peer->maxDgramPackets > 1) {
4511 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4513 call->MTU = peer->MTU;
4515 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4516 call->ssthresh = rx_maxSendWindow;
4517 call->nDgramPackets = peer->nDgramPackets;
4518 call->congestSeq = peer->congestSeq;
4519 MUTEX_EXIT(&peer->peer_lock);
4521 flags = call->flags;
4522 rxi_ClearReceiveQueue(call);
4523 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4524 if (flags & RX_CALL_TQ_BUSY) {
4525 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4526 call->flags |= (flags & RX_CALL_TQ_WAIT);
4528 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4530 rxi_ClearTransmitQueue(call, 0);
4531 queue_Init(&call->tq);
4532 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4533 dpf(("rcall %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4536 while (call->tqWaiters) {
4537 #ifdef RX_ENABLE_LOCKS
4538 CV_BROADCAST(&call->cv_tq);
4539 #else /* RX_ENABLE_LOCKS */
4540 osi_rxWakeup(&call->tq);
4541 #endif /* RX_ENABLE_LOCKS */
4545 queue_Init(&call->rq);
4547 call->rwind = rx_initReceiveWindow;
4548 call->twind = rx_initSendWindow;
4549 call->nSoftAcked = 0;
4550 call->nextCwind = 0;
4553 call->nCwindAcks = 0;
4554 call->nSoftAcks = 0;
4555 call->nHardAcks = 0;
4557 call->tfirst = call->rnext = call->tnext = 1;
4559 call->lastAcked = 0;
4560 call->localStatus = call->remoteStatus = 0;
4562 if (flags & RX_CALL_READER_WAIT) {
4563 #ifdef RX_ENABLE_LOCKS
4564 CV_BROADCAST(&call->cv_rq);
4566 osi_rxWakeup(&call->rq);
4569 if (flags & RX_CALL_WAIT_PACKETS) {
4570 MUTEX_ENTER(&rx_freePktQ_lock);
4571 rxi_PacketsUnWait(); /* XXX */
4572 MUTEX_EXIT(&rx_freePktQ_lock);
4574 #ifdef RX_ENABLE_LOCKS
4575 CV_SIGNAL(&call->cv_twind);
4577 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4578 osi_rxWakeup(&call->twind);
4581 #ifdef RX_ENABLE_LOCKS
4582 /* The following ensures that we don't mess with any queue while some
4583 * other thread might also be doing so. The call_queue_lock field is
4584 * is only modified under the call lock. If the call is in the process
4585 * of being removed from a queue, the call is not locked until the
4586 * the queue lock is dropped and only then is the call_queue_lock field
4587 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4588 * Note that any other routine which removes a call from a queue has to
4589 * obtain the queue lock before examing the queue and removing the call.
4591 if (call->call_queue_lock) {
4592 MUTEX_ENTER(call->call_queue_lock);
4593 if (queue_IsOnQueue(call)) {
4595 if (flags & RX_CALL_WAIT_PROC) {
4596 MUTEX_ENTER(&rx_stats_mutex);
4598 MUTEX_EXIT(&rx_stats_mutex);
4601 MUTEX_EXIT(call->call_queue_lock);
4602 CLEAR_CALL_QUEUE_LOCK(call);
4604 #else /* RX_ENABLE_LOCKS */
4605 if (queue_IsOnQueue(call)) {
4607 if (flags & RX_CALL_WAIT_PROC)
4610 #endif /* RX_ENABLE_LOCKS */
4612 rxi_KeepAliveOff(call);
4613 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4616 /* Send an acknowledge for the indicated packet (seq,serial) of the
4617 * indicated call, for the indicated reason (reason). This
4618 * acknowledge will specifically acknowledge receiving the packet, and
4619 * will also specify which other packets for this call have been
4620 * received. This routine returns the packet that was used to the
4621 * caller. The caller is responsible for freeing it or re-using it.
4622 * This acknowledgement also returns the highest sequence number
4623 * actually read out by the higher level to the sender; the sender
4624 * promises to keep around packets that have not been read by the
4625 * higher level yet (unless, of course, the sender decides to abort
4626 * the call altogether). Any of p, seq, serial, pflags, or reason may
4627 * be set to zero without ill effect. That is, if they are zero, they
4628 * will not convey any information.
4629 * NOW there is a trailer field, after the ack where it will safely be
4630 * ignored by mundanes, which indicates the maximum size packet this
4631 * host can swallow. */
4633 register struct rx_packet *optionalPacket; use to send ack (or null)
4634 int seq; Sequence number of the packet we are acking
4635 int serial; Serial number of the packet
4636 int pflags; Flags field from packet header
4637 int reason; Reason an acknowledge was prompted
4641 rxi_SendAck(register struct rx_call *call,
4642 register struct rx_packet *optionalPacket, int serial, int reason,
4645 struct rx_ackPacket *ap;
4646 register struct rx_packet *rqp;
4647 register struct rx_packet *nxp; /* For queue_Scan */
4648 register struct rx_packet *p;
4651 #ifdef RX_ENABLE_TSFPQ
4652 struct rx_ts_info_t * rx_ts_info;
4656 * Open the receive window once a thread starts reading packets
4658 if (call->rnext > 1) {
4659 call->rwind = rx_maxReceiveWindow;
4662 call->nHardAcks = 0;
4663 call->nSoftAcks = 0;
4664 if (call->rnext > call->lastAcked)
4665 call->lastAcked = call->rnext;
4669 rx_computelen(p, p->length); /* reset length, you never know */
4670 } /* where that's been... */
4671 #ifdef RX_ENABLE_TSFPQ
4673 RX_TS_INFO_GET(rx_ts_info);
4674 if ((p = rx_ts_info->local_special_packet)) {
4675 rx_computelen(p, p->length);
4676 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4677 rx_ts_info->local_special_packet = p;
4678 } else { /* We won't send the ack, but don't panic. */
4679 return optionalPacket;
4683 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4684 /* We won't send the ack, but don't panic. */
4685 return optionalPacket;
4690 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4693 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4694 #ifndef RX_ENABLE_TSFPQ
4695 if (!optionalPacket)
4698 return optionalPacket;
4700 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4701 if (rx_Contiguous(p) < templ) {
4702 #ifndef RX_ENABLE_TSFPQ
4703 if (!optionalPacket)
4706 return optionalPacket;
4711 /* MTUXXX failing to send an ack is very serious. We should */
4712 /* try as hard as possible to send even a partial ack; it's */
4713 /* better than nothing. */
4714 ap = (struct rx_ackPacket *)rx_DataOf(p);
4715 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4716 ap->reason = reason;
4718 /* The skew computation used to be bogus, I think it's better now. */
4719 /* We should start paying attention to skew. XXX */
4720 ap->serial = htonl(serial);
4721 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4723 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4724 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4726 /* No fear of running out of ack packet here because there can only be at most
4727 * one window full of unacknowledged packets. The window size must be constrained
4728 * to be less than the maximum ack size, of course. Also, an ack should always
4729 * fit into a single packet -- it should not ever be fragmented. */
4730 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4731 if (!rqp || !call->rq.next
4732 || (rqp->header.seq > (call->rnext + call->rwind))) {
4733 #ifndef RX_ENABLE_TSFPQ
4734 if (!optionalPacket)
4737 rxi_CallError(call, RX_CALL_DEAD);
4738 return optionalPacket;
4741 while (rqp->header.seq > call->rnext + offset)
4742 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4743 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4745 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4746 #ifndef RX_ENABLE_TSFPQ
4747 if (!optionalPacket)
4750 rxi_CallError(call, RX_CALL_DEAD);
4751 return optionalPacket;
4756 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4758 /* these are new for AFS 3.3 */
4759 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4760 templ = htonl(templ);
4761 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4762 templ = htonl(call->conn->peer->ifMTU);
4763 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4764 sizeof(afs_int32), &templ);
4766 /* new for AFS 3.4 */
4767 templ = htonl(call->rwind);
4768 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4769 sizeof(afs_int32), &templ);
4771 /* new for AFS 3.5 */
4772 templ = htonl(call->conn->peer->ifDgramPackets);
4773 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4774 sizeof(afs_int32), &templ);
4776 p->header.serviceId = call->conn->serviceId;
4777 p->header.cid = (call->conn->cid | call->channel);
4778 p->header.callNumber = *call->callNumber;
4780 p->header.securityIndex = call->conn->securityIndex;
4781 p->header.epoch = call->conn->epoch;
4782 p->header.type = RX_PACKET_TYPE_ACK;
4783 p->header.flags = RX_SLOW_START_OK;
4784 if (reason == RX_ACK_PING) {
4785 p->header.flags |= RX_REQUEST_ACK;
4787 clock_GetTime(&call->pingRequestTime);
4790 if (call->conn->type == RX_CLIENT_CONNECTION)
4791 p->header.flags |= RX_CLIENT_INITIATED;
4795 if (rxdebug_active) {
4799 len = _snprintf(msg, sizeof(msg),
4800 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4801 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4802 ntohl(ap->serial), ntohl(ap->previousPacket),
4803 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
4804 ap->nAcks, ntohs(ap->bufferSpace) );
4808 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
4809 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4813 OutputDebugString(msg);
4815 #else /* AFS_NT40_ENV */
4817 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
4818 ap->reason, ntohl(ap->previousPacket),
4819 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4821 for (offset = 0; offset < ap->nAcks; offset++)
4822 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4827 #endif /* AFS_NT40_ENV */
4830 register int i, nbytes = p->length;
4832 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4833 if (nbytes <= p->wirevec[i].iov_len) {
4834 register int savelen, saven;
4836 savelen = p->wirevec[i].iov_len;
4838 p->wirevec[i].iov_len = nbytes;
4840 rxi_Send(call, p, istack);
4841 p->wirevec[i].iov_len = savelen;
4845 nbytes -= p->wirevec[i].iov_len;
4848 MUTEX_ENTER(&rx_stats_mutex);
4849 rx_stats.ackPacketsSent++;
4850 MUTEX_EXIT(&rx_stats_mutex);
4851 #ifndef RX_ENABLE_TSFPQ
4852 if (!optionalPacket)
4855 return optionalPacket; /* Return packet for re-use by caller */
4858 /* Send all of the packets in the list in single datagram */
4860 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4861 int istack, int moreFlag, struct clock *now,
4862 struct clock *retryTime, int resending)
4867 struct rx_connection *conn = call->conn;
4868 struct rx_peer *peer = conn->peer;
4870 MUTEX_ENTER(&peer->peer_lock);
4873 peer->reSends += len;
4874 MUTEX_ENTER(&rx_stats_mutex);
4875 rx_stats.dataPacketsSent += len;
4876 MUTEX_EXIT(&rx_stats_mutex);
4877 MUTEX_EXIT(&peer->peer_lock);
4879 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4883 /* Set the packet flags and schedule the resend events */
4884 /* Only request an ack for the last packet in the list */
4885 for (i = 0; i < len; i++) {
4886 list[i]->retryTime = *retryTime;
4887 if (list[i]->header.serial) {
4888 /* Exponentially backoff retry times */
4889 if (list[i]->backoff < MAXBACKOFF) {
4890 /* so it can't stay == 0 */
4891 list[i]->backoff = (list[i]->backoff << 1) + 1;
4894 clock_Addmsec(&(list[i]->retryTime),
4895 ((afs_uint32) list[i]->backoff) << 8);
4898 /* Wait a little extra for the ack on the last packet */
4899 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4900 clock_Addmsec(&(list[i]->retryTime), 400);
4903 /* Record the time sent */
4904 list[i]->timeSent = *now;
4906 /* Ask for an ack on retransmitted packets, on every other packet
4907 * if the peer doesn't support slow start. Ask for an ack on every
4908 * packet until the congestion window reaches the ack rate. */
4909 if (list[i]->header.serial) {
4911 MUTEX_ENTER(&rx_stats_mutex);
4912 rx_stats.dataPacketsReSent++;
4913 MUTEX_EXIT(&rx_stats_mutex);
4915 /* improved RTO calculation- not Karn */
4916 list[i]->firstSent = *now;
4917 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4918 || (!(call->flags & RX_CALL_SLOW_START_OK)
4919 && (list[i]->header.seq & 1)))) {
4924 MUTEX_ENTER(&peer->peer_lock);
4928 MUTEX_ENTER(&rx_stats_mutex);
4929 rx_stats.dataPacketsSent++;
4930 MUTEX_EXIT(&rx_stats_mutex);
4931 MUTEX_EXIT(&peer->peer_lock);
4933 /* Tag this packet as not being the last in this group,
4934 * for the receiver's benefit */
4935 if (i < len - 1 || moreFlag) {
4936 list[i]->header.flags |= RX_MORE_PACKETS;
4939 /* Install the new retransmit time for the packet, and
4940 * record the time sent */
4941 list[i]->timeSent = *now;
4945 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4948 /* Since we're about to send a data packet to the peer, it's
4949 * safe to nuke any scheduled end-of-packets ack */
4950 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4952 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4953 MUTEX_EXIT(&call->lock);
4955 rxi_SendPacketList(call, conn, list, len, istack);
4957 rxi_SendPacket(call, conn, list[0], istack);
4959 MUTEX_ENTER(&call->lock);
4960 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4962 /* Update last send time for this call (for keep-alive
4963 * processing), and for the connection (so that we can discover
4964 * idle connections) */
4965 conn->lastSendTime = call->lastSendTime = clock_Sec();
4968 /* When sending packets we need to follow these rules:
4969 * 1. Never send more than maxDgramPackets in a jumbogram.
4970 * 2. Never send a packet with more than two iovecs in a jumbogram.
4971 * 3. Never send a retransmitted packet in a jumbogram.
4972 * 4. Never send more than cwind/4 packets in a jumbogram
4973 * We always keep the last list we should have sent so we
4974 * can set the RX_MORE_PACKETS flags correctly.
4977 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4978 int istack, struct clock *now, struct clock *retryTime,
4981 int i, cnt, lastCnt = 0;
4982 struct rx_packet **listP, **lastP = 0;
4983 struct rx_peer *peer = call->conn->peer;
4984 int morePackets = 0;
4986 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4987 /* Does the current packet force us to flush the current list? */
4989 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4990 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4992 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
4994 /* If the call enters an error state stop sending, or if
4995 * we entered congestion recovery mode, stop sending */
4996 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5004 /* Add the current packet to the list if it hasn't been acked.
5005 * Otherwise adjust the list pointer to skip the current packet. */
5006 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5008 /* Do we need to flush the list? */
5009 if (cnt >= (int)peer->maxDgramPackets
5010 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5011 || list[i]->header.serial
5012 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5014 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5015 retryTime, resending);
5016 /* If the call enters an error state stop sending, or if
5017 * we entered congestion recovery mode, stop sending */
5019 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5024 listP = &list[i + 1];
5029 osi_Panic("rxi_SendList error");
5031 listP = &list[i + 1];
5035 /* Send the whole list when the call is in receive mode, when
5036 * the call is in eof mode, when we are in fast recovery mode,
5037 * and when we have the last packet */
5038 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5039 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5040 || (call->flags & RX_CALL_FAST_RECOVER)) {
5041 /* Check for the case where the current list contains
5042 * an acked packet. Since we always send retransmissions
5043 * in a separate packet, we only need to check the first
5044 * packet in the list */
5045 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5049 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5050 retryTime, resending);
5051 /* If the call enters an error state stop sending, or if
5052 * we entered congestion recovery mode, stop sending */
5053 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5057 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5060 } else if (lastCnt > 0) {
5061 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5066 #ifdef RX_ENABLE_LOCKS
5067 /* Call rxi_Start, below, but with the call lock held. */
5069 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
5070 void *arg1, int istack)
5072 MUTEX_ENTER(&call->lock);
5073 rxi_Start(event, call, arg1, istack);
5074 MUTEX_EXIT(&call->lock);
5076 #endif /* RX_ENABLE_LOCKS */
5078 /* This routine is called when new packets are readied for
5079 * transmission and when retransmission may be necessary, or when the
5080 * transmission window or burst count are favourable. This should be
5081 * better optimized for new packets, the usual case, now that we've
5082 * got rid of queues of send packets. XXXXXXXXXXX */
5084 rxi_Start(struct rxevent *event, register struct rx_call *call,
5085 void *arg1, int istack)
5087 struct rx_packet *p;
5088 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
5089 struct rx_peer *peer = call->conn->peer;
5090 struct clock now, retryTime;
5094 struct rx_packet **xmitList;
5097 /* If rxi_Start is being called as a result of a resend event,
5098 * then make sure that the event pointer is removed from the call
5099 * structure, since there is no longer a per-call retransmission
5101 if (event && event == call->resendEvent) {
5102 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5103 call->resendEvent = NULL;
5105 if (queue_IsEmpty(&call->tq)) {
5109 /* Timeouts trigger congestion recovery */
5110 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5111 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5112 /* someone else is waiting to start recovery */
5115 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5116 rxi_WaitforTQBusy(call);
5117 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5118 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5119 call->flags |= RX_CALL_FAST_RECOVER;
5120 if (peer->maxDgramPackets > 1) {
5121 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5123 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5125 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5126 call->nDgramPackets = 1;
5128 call->nextCwind = 1;
5131 MUTEX_ENTER(&peer->peer_lock);
5132 peer->MTU = call->MTU;
5133 peer->cwind = call->cwind;
5134 peer->nDgramPackets = 1;
5136 call->congestSeq = peer->congestSeq;
5137 MUTEX_EXIT(&peer->peer_lock);
5138 /* Clear retry times on packets. Otherwise, it's possible for
5139 * some packets in the queue to force resends at rates faster
5140 * than recovery rates.
5142 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5143 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5144 clock_Zero(&p->retryTime);
5149 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5150 MUTEX_ENTER(&rx_stats_mutex);
5151 rx_tq_debug.rxi_start_in_error++;
5152 MUTEX_EXIT(&rx_stats_mutex);
5157 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5158 /* Get clock to compute the re-transmit time for any packets
5159 * in this burst. Note, if we back off, it's reasonable to
5160 * back off all of the packets in the same manner, even if
5161 * some of them have been retransmitted more times than more
5162 * recent additions */
5163 clock_GetTime(&now);
5164 retryTime = now; /* initialize before use */
5165 MUTEX_ENTER(&peer->peer_lock);
5166 clock_Add(&retryTime, &peer->timeout);
5167 MUTEX_EXIT(&peer->peer_lock);
5169 /* Send (or resend) any packets that need it, subject to
5170 * window restrictions and congestion burst control
5171 * restrictions. Ask for an ack on the last packet sent in
5172 * this burst. For now, we're relying upon the window being
5173 * considerably bigger than the largest number of packets that
5174 * are typically sent at once by one initial call to
5175 * rxi_Start. This is probably bogus (perhaps we should ask
5176 * for an ack when we're half way through the current
5177 * window?). Also, for non file transfer applications, this
5178 * may end up asking for an ack for every packet. Bogus. XXXX
5181 * But check whether we're here recursively, and let the other guy
5184 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5185 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5186 call->flags |= RX_CALL_TQ_BUSY;
5188 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5190 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5191 call->flags &= ~RX_CALL_NEED_START;
5192 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5194 maxXmitPackets = MIN(call->twind, call->cwind);
5195 xmitList = (struct rx_packet **)
5196 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5197 if (xmitList == NULL)
5198 osi_Panic("rxi_Start, failed to allocate xmit list");
5199 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5200 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5201 /* We shouldn't be sending packets if a thread is waiting
5202 * to initiate congestion recovery */
5206 && (call->flags & RX_CALL_FAST_RECOVER)) {
5207 /* Only send one packet during fast recovery */
5210 if ((p->flags & RX_PKTFLAG_FREE)
5211 || (!queue_IsEnd(&call->tq, nxp)
5212 && (nxp->flags & RX_PKTFLAG_FREE))
5213 || (p == (struct rx_packet *)&rx_freePacketQueue)
5214 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5215 osi_Panic("rxi_Start: xmit queue clobbered");
5217 if (p->flags & RX_PKTFLAG_ACKED) {
5218 MUTEX_ENTER(&rx_stats_mutex);
5219 rx_stats.ignoreAckedPacket++;
5220 MUTEX_EXIT(&rx_stats_mutex);
5221 continue; /* Ignore this packet if it has been acknowledged */
5224 /* Turn off all flags except these ones, which are the same
5225 * on each transmission */
5226 p->header.flags &= RX_PRESET_FLAGS;
5228 if (p->header.seq >=
5229 call->tfirst + MIN((int)call->twind,
5230 (int)(call->nSoftAcked +
5232 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5233 /* Note: if we're waiting for more window space, we can
5234 * still send retransmits; hence we don't return here, but
5235 * break out to schedule a retransmit event */
5236 dpf(("call %d waiting for window",
5237 *(call->callNumber)));
5241 /* Transmit the packet if it needs to be sent. */
5242 if (!clock_Lt(&now, &p->retryTime)) {
5243 if (nXmitPackets == maxXmitPackets) {
5244 rxi_SendXmitList(call, xmitList, nXmitPackets,
5245 istack, &now, &retryTime,
5247 osi_Free(xmitList, maxXmitPackets *
5248 sizeof(struct rx_packet *));
5251 xmitList[nXmitPackets++] = p;
5255 /* xmitList now hold pointers to all of the packets that are
5256 * ready to send. Now we loop to send the packets */
5257 if (nXmitPackets > 0) {
5258 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5259 &now, &retryTime, resending);
5262 maxXmitPackets * sizeof(struct rx_packet *));
5264 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5266 * TQ references no longer protected by this flag; they must remain
5267 * protected by the global lock.
5269 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5270 call->flags &= ~RX_CALL_TQ_BUSY;
5271 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5272 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5273 #ifdef RX_ENABLE_LOCKS
5274 osirx_AssertMine(&call->lock, "rxi_Start start");
5275 CV_BROADCAST(&call->cv_tq);
5276 #else /* RX_ENABLE_LOCKS */
5277 osi_rxWakeup(&call->tq);
5278 #endif /* RX_ENABLE_LOCKS */
5283 /* We went into the error state while sending packets. Now is
5284 * the time to reset the call. This will also inform the using
5285 * process that the call is in an error state.
5287 MUTEX_ENTER(&rx_stats_mutex);
5288 rx_tq_debug.rxi_start_aborted++;
5289 MUTEX_EXIT(&rx_stats_mutex);
5290 call->flags &= ~RX_CALL_TQ_BUSY;
5291 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5292 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5293 #ifdef RX_ENABLE_LOCKS
5294 osirx_AssertMine(&call->lock, "rxi_Start middle");
5295 CV_BROADCAST(&call->cv_tq);
5296 #else /* RX_ENABLE_LOCKS */
5297 osi_rxWakeup(&call->tq);
5298 #endif /* RX_ENABLE_LOCKS */
5300 rxi_CallError(call, call->error);
5303 #ifdef RX_ENABLE_LOCKS
5304 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5305 register int missing;
5306 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5307 /* Some packets have received acks. If they all have, we can clear
5308 * the transmit queue.
5311 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5312 if (p->header.seq < call->tfirst
5313 && (p->flags & RX_PKTFLAG_ACKED)) {
5320 call->flags |= RX_CALL_TQ_CLEARME;
5322 #endif /* RX_ENABLE_LOCKS */
5323 /* Don't bother doing retransmits if the TQ is cleared. */
5324 if (call->flags & RX_CALL_TQ_CLEARME) {
5325 rxi_ClearTransmitQueue(call, 1);
5327 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5330 /* Always post a resend event, if there is anything in the
5331 * queue, and resend is possible. There should be at least
5332 * one unacknowledged packet in the queue ... otherwise none
5333 * of these packets should be on the queue in the first place.
5335 if (call->resendEvent) {
5336 /* Cancel the existing event and post a new one */
5337 rxevent_Cancel(call->resendEvent, call,
5338 RX_CALL_REFCOUNT_RESEND);
5341 /* The retry time is the retry time on the first unacknowledged
5342 * packet inside the current window */
5344 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5345 /* Don't set timers for packets outside the window */
5346 if (p->header.seq >= call->tfirst + call->twind) {
5350 if (!(p->flags & RX_PKTFLAG_ACKED)
5351 && !clock_IsZero(&p->retryTime)) {
5353 retryTime = p->retryTime;
5358 /* Post a new event to re-run rxi_Start when retries may be needed */
5359 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5360 #ifdef RX_ENABLE_LOCKS
5361 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5363 rxevent_Post2(&retryTime, rxi_StartUnlocked,
5364 (void *)call, 0, istack);
5365 #else /* RX_ENABLE_LOCKS */
5367 rxevent_Post2(&retryTime, rxi_Start, (void *)call,
5369 #endif /* RX_ENABLE_LOCKS */
5372 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5373 } while (call->flags & RX_CALL_NEED_START);
5375 * TQ references no longer protected by this flag; they must remain
5376 * protected by the global lock.
5378 call->flags &= ~RX_CALL_TQ_BUSY;
5379 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5380 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5381 #ifdef RX_ENABLE_LOCKS
5382 osirx_AssertMine(&call->lock, "rxi_Start end");
5383 CV_BROADCAST(&call->cv_tq);
5384 #else /* RX_ENABLE_LOCKS */
5385 osi_rxWakeup(&call->tq);
5386 #endif /* RX_ENABLE_LOCKS */
5389 call->flags |= RX_CALL_NEED_START;
5391 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5393 if (call->resendEvent) {
5394 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5399 /* Also adjusts the keep alive parameters for the call, to reflect
5400 * that we have just sent a packet (so keep alives aren't sent
5403 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5406 register struct rx_connection *conn = call->conn;
5408 /* Stamp each packet with the user supplied status */
5409 p->header.userStatus = call->localStatus;
5411 /* Allow the security object controlling this call's security to
5412 * make any last-minute changes to the packet */
5413 RXS_SendPacket(conn->securityObject, call, p);
5415 /* Since we're about to send SOME sort of packet to the peer, it's
5416 * safe to nuke any scheduled end-of-packets ack */
5417 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5419 /* Actually send the packet, filling in more connection-specific fields */
5420 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5421 MUTEX_EXIT(&call->lock);
5422 rxi_SendPacket(call, conn, p, istack);
5423 MUTEX_ENTER(&call->lock);
5424 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5426 /* Update last send time for this call (for keep-alive
5427 * processing), and for the connection (so that we can discover
5428 * idle connections) */
5429 conn->lastSendTime = call->lastSendTime = clock_Sec();
5433 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5434 * that things are fine. Also called periodically to guarantee that nothing
5435 * falls through the cracks (e.g. (error + dally) connections have keepalive
5436 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5438 * haveCTLock Set if calling from rxi_ReapConnections
5440 #ifdef RX_ENABLE_LOCKS
5442 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5443 #else /* RX_ENABLE_LOCKS */
5445 rxi_CheckCall(register struct rx_call *call)
5446 #endif /* RX_ENABLE_LOCKS */
5448 register struct rx_connection *conn = call->conn;
5450 afs_uint32 deadTime;
5452 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5453 if (call->flags & RX_CALL_TQ_BUSY) {
5454 /* Call is active and will be reset by rxi_Start if it's
5455 * in an error state.
5460 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5462 (((afs_uint32) conn->secondsUntilDead << 10) +
5463 ((afs_uint32) conn->peer->rtt >> 3) +
5464 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5466 /* These are computed to the second (+- 1 second). But that's
5467 * good enough for these values, which should be a significant
5468 * number of seconds. */
5469 if (now > (call->lastReceiveTime + deadTime)) {
5470 if (call->state == RX_STATE_ACTIVE) {
5471 rxi_CallError(call, RX_CALL_DEAD);
5474 #ifdef RX_ENABLE_LOCKS
5475 /* Cancel pending events */
5476 rxevent_Cancel(call->delayedAckEvent, call,
5477 RX_CALL_REFCOUNT_DELAY);
5478 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5479 rxevent_Cancel(call->keepAliveEvent, call,
5480 RX_CALL_REFCOUNT_ALIVE);
5481 if (call->refCount == 0) {
5482 rxi_FreeCall(call, haveCTLock);
5486 #else /* RX_ENABLE_LOCKS */
5489 #endif /* RX_ENABLE_LOCKS */
5491 /* Non-active calls are destroyed if they are not responding
5492 * to pings; active calls are simply flagged in error, so the
5493 * attached process can die reasonably gracefully. */
5495 /* see if we have a non-activity timeout */
5496 if (call->startWait && conn->idleDeadTime
5497 && ((call->startWait + conn->idleDeadTime) < now)) {
5498 if (call->state == RX_STATE_ACTIVE) {
5499 rxi_CallError(call, RX_CALL_TIMEOUT);
5503 /* see if we have a hard timeout */
5504 if (conn->hardDeadTime
5505 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5506 if (call->state == RX_STATE_ACTIVE)
5507 rxi_CallError(call, RX_CALL_TIMEOUT);
5514 /* When a call is in progress, this routine is called occasionally to
5515 * make sure that some traffic has arrived (or been sent to) the peer.
5516 * If nothing has arrived in a reasonable amount of time, the call is
5517 * declared dead; if nothing has been sent for a while, we send a
5518 * keep-alive packet (if we're actually trying to keep the call alive)
5521 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5524 struct rx_connection *conn;
5527 MUTEX_ENTER(&call->lock);
5528 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5529 if (event == call->keepAliveEvent)
5530 call->keepAliveEvent = NULL;
5533 #ifdef RX_ENABLE_LOCKS
5534 if (rxi_CheckCall(call, 0)) {
5535 MUTEX_EXIT(&call->lock);
5538 #else /* RX_ENABLE_LOCKS */
5539 if (rxi_CheckCall(call))
5541 #endif /* RX_ENABLE_LOCKS */
5543 /* Don't try to keep alive dallying calls */
5544 if (call->state == RX_STATE_DALLY) {
5545 MUTEX_EXIT(&call->lock);
5550 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5551 /* Don't try to send keepalives if there is unacknowledged data */
5552 /* the rexmit code should be good enough, this little hack
5553 * doesn't quite work XXX */
5554 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5556 rxi_ScheduleKeepAliveEvent(call);
5557 MUTEX_EXIT(&call->lock);
5562 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5564 if (!call->keepAliveEvent) {
5566 clock_GetTime(&when);
5567 when.sec += call->conn->secondsUntilPing;
5568 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5569 call->keepAliveEvent =
5570 rxevent_Post(&when, rxi_KeepAliveEvent, call, 0);
5574 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5576 rxi_KeepAliveOn(register struct rx_call *call)
5578 /* Pretend last packet received was received now--i.e. if another
5579 * packet isn't received within the keep alive time, then the call
5580 * will die; Initialize last send time to the current time--even
5581 * if a packet hasn't been sent yet. This will guarantee that a
5582 * keep-alive is sent within the ping time */
5583 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5584 rxi_ScheduleKeepAliveEvent(call);
5587 /* This routine is called to send connection abort messages
5588 * that have been delayed to throttle looping clients. */
5590 rxi_SendDelayedConnAbort(struct rxevent *event,
5591 register struct rx_connection *conn, char *dummy)
5594 struct rx_packet *packet;
5596 MUTEX_ENTER(&conn->conn_data_lock);
5597 conn->delayedAbortEvent = NULL;
5598 error = htonl(conn->error);
5600 MUTEX_EXIT(&conn->conn_data_lock);
5601 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5604 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5605 RX_PACKET_TYPE_ABORT, (char *)&error,
5607 rxi_FreePacket(packet);
5611 /* This routine is called to send call abort messages
5612 * that have been delayed to throttle looping clients. */
5614 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5618 struct rx_packet *packet;
5620 MUTEX_ENTER(&call->lock);
5621 call->delayedAbortEvent = NULL;
5622 error = htonl(call->error);
5624 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5627 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5628 (char *)&error, sizeof(error), 0);
5629 rxi_FreePacket(packet);
5631 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5632 MUTEX_EXIT(&call->lock);
5635 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5636 * seconds) to ask the client to authenticate itself. The routine
5637 * issues a challenge to the client, which is obtained from the
5638 * security object associated with the connection */
5640 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5641 void *arg1, int tries)
5643 conn->challengeEvent = NULL;
5644 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5645 register struct rx_packet *packet;
5649 /* We've failed to authenticate for too long.
5650 * Reset any calls waiting for authentication;
5651 * they are all in RX_STATE_PRECALL.
5655 MUTEX_ENTER(&conn->conn_call_lock);
5656 for (i = 0; i < RX_MAXCALLS; i++) {
5657 struct rx_call *call = conn->call[i];
5659 MUTEX_ENTER(&call->lock);
5660 if (call->state == RX_STATE_PRECALL) {
5661 rxi_CallError(call, RX_CALL_DEAD);
5662 rxi_SendCallAbort(call, NULL, 0, 0);
5664 MUTEX_EXIT(&call->lock);
5667 MUTEX_EXIT(&conn->conn_call_lock);
5671 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5673 /* If there's no packet available, do this later. */
5674 RXS_GetChallenge(conn->securityObject, conn, packet);
5675 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5676 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5677 rxi_FreePacket(packet);
5679 clock_GetTime(&when);
5680 when.sec += RX_CHALLENGE_TIMEOUT;
5681 conn->challengeEvent =
5682 rxevent_Post2(&when, rxi_ChallengeEvent, conn, 0,
5687 /* Call this routine to start requesting the client to authenticate
5688 * itself. This will continue until authentication is established,
5689 * the call times out, or an invalid response is returned. The
5690 * security object associated with the connection is asked to create
5691 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5692 * defined earlier. */
5694 rxi_ChallengeOn(register struct rx_connection *conn)
5696 if (!conn->challengeEvent) {
5697 RXS_CreateChallenge(conn->securityObject, conn);
5698 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5703 /* Compute round trip time of the packet provided, in *rttp.
5706 /* rxi_ComputeRoundTripTime is called with peer locked. */
5707 /* sentp and/or peer may be null */
5709 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5710 register struct clock *sentp,
5711 register struct rx_peer *peer)
5713 struct clock thisRtt, *rttp = &thisRtt;
5715 register int rtt_timeout;
5717 clock_GetTime(rttp);
5719 if (clock_Lt(rttp, sentp)) {
5721 return; /* somebody set the clock back, don't count this time. */
5723 clock_Sub(rttp, sentp);
5724 MUTEX_ENTER(&rx_stats_mutex);
5725 if (clock_Lt(rttp, &rx_stats.minRtt))
5726 rx_stats.minRtt = *rttp;
5727 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5728 if (rttp->sec > 60) {
5729 MUTEX_EXIT(&rx_stats_mutex);
5730 return; /* somebody set the clock ahead */
5732 rx_stats.maxRtt = *rttp;
5734 clock_Add(&rx_stats.totalRtt, rttp);
5735 rx_stats.nRttSamples++;
5736 MUTEX_EXIT(&rx_stats_mutex);
5738 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5740 /* Apply VanJacobson round-trip estimations */
5745 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5746 * srtt is stored as fixed point with 3 bits after the binary
5747 * point (i.e., scaled by 8). The following magic is
5748 * equivalent to the smoothing algorithm in rfc793 with an
5749 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5750 * srtt*8 = srtt*8 + rtt - srtt
5751 * srtt = srtt + rtt/8 - srtt/8
5754 delta = MSEC(rttp) - (peer->rtt >> 3);
5758 * We accumulate a smoothed rtt variance (actually, a smoothed
5759 * mean difference), then set the retransmit timer to smoothed
5760 * rtt + 4 times the smoothed variance (was 2x in van's original
5761 * paper, but 4x works better for me, and apparently for him as
5763 * rttvar is stored as
5764 * fixed point with 2 bits after the binary point (scaled by
5765 * 4). The following is equivalent to rfc793 smoothing with
5766 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5767 * replaces rfc793's wired-in beta.
5768 * dev*4 = dev*4 + (|actual - expected| - dev)
5774 delta -= (peer->rtt_dev >> 2);
5775 peer->rtt_dev += delta;
5777 /* I don't have a stored RTT so I start with this value. Since I'm
5778 * probably just starting a call, and will be pushing more data down
5779 * this, I expect congestion to increase rapidly. So I fudge a
5780 * little, and I set deviance to half the rtt. In practice,
5781 * deviance tends to approach something a little less than
5782 * half the smoothed rtt. */
5783 peer->rtt = (MSEC(rttp) << 3) + 8;
5784 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5786 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5787 * the other of these connections is usually in a user process, and can
5788 * be switched and/or swapped out. So on fast, reliable networks, the
5789 * timeout would otherwise be too short.
5791 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5792 clock_Zero(&(peer->timeout));
5793 clock_Addmsec(&(peer->timeout), rtt_timeout);
5795 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)));
5799 /* Find all server connections that have not been active for a long time, and
5802 rxi_ReapConnections(void)
5805 clock_GetTime(&now);
5807 /* Find server connection structures that haven't been used for
5808 * greater than rx_idleConnectionTime */
5810 struct rx_connection **conn_ptr, **conn_end;
5811 int i, havecalls = 0;
5812 MUTEX_ENTER(&rx_connHashTable_lock);
5813 for (conn_ptr = &rx_connHashTable[0], conn_end =
5814 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5816 struct rx_connection *conn, *next;
5817 struct rx_call *call;
5821 for (conn = *conn_ptr; conn; conn = next) {
5822 /* XXX -- Shouldn't the connection be locked? */
5825 for (i = 0; i < RX_MAXCALLS; i++) {
5826 call = conn->call[i];
5829 MUTEX_ENTER(&call->lock);
5830 #ifdef RX_ENABLE_LOCKS
5831 result = rxi_CheckCall(call, 1);
5832 #else /* RX_ENABLE_LOCKS */
5833 result = rxi_CheckCall(call);
5834 #endif /* RX_ENABLE_LOCKS */
5835 MUTEX_EXIT(&call->lock);
5837 /* If CheckCall freed the call, it might
5838 * have destroyed the connection as well,
5839 * which screws up the linked lists.
5845 if (conn->type == RX_SERVER_CONNECTION) {
5846 /* This only actually destroys the connection if
5847 * there are no outstanding calls */
5848 MUTEX_ENTER(&conn->conn_data_lock);
5849 if (!havecalls && !conn->refCount
5850 && ((conn->lastSendTime + rx_idleConnectionTime) <
5852 conn->refCount++; /* it will be decr in rx_DestroyConn */
5853 MUTEX_EXIT(&conn->conn_data_lock);
5854 #ifdef RX_ENABLE_LOCKS
5855 rxi_DestroyConnectionNoLock(conn);
5856 #else /* RX_ENABLE_LOCKS */
5857 rxi_DestroyConnection(conn);
5858 #endif /* RX_ENABLE_LOCKS */
5860 #ifdef RX_ENABLE_LOCKS
5862 MUTEX_EXIT(&conn->conn_data_lock);
5864 #endif /* RX_ENABLE_LOCKS */
5868 #ifdef RX_ENABLE_LOCKS
5869 while (rx_connCleanup_list) {
5870 struct rx_connection *conn;
5871 conn = rx_connCleanup_list;
5872 rx_connCleanup_list = rx_connCleanup_list->next;
5873 MUTEX_EXIT(&rx_connHashTable_lock);
5874 rxi_CleanupConnection(conn);
5875 MUTEX_ENTER(&rx_connHashTable_lock);
5877 MUTEX_EXIT(&rx_connHashTable_lock);
5878 #endif /* RX_ENABLE_LOCKS */
5881 /* Find any peer structures that haven't been used (haven't had an
5882 * associated connection) for greater than rx_idlePeerTime */
5884 struct rx_peer **peer_ptr, **peer_end;
5886 MUTEX_ENTER(&rx_rpc_stats);
5887 MUTEX_ENTER(&rx_peerHashTable_lock);
5888 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5889 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5891 struct rx_peer *peer, *next, *prev;
5892 for (prev = peer = *peer_ptr; peer; peer = next) {
5894 code = MUTEX_TRYENTER(&peer->peer_lock);
5895 if ((code) && (peer->refCount == 0)
5896 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5897 rx_interface_stat_p rpc_stat, nrpc_stat;
5899 MUTEX_EXIT(&peer->peer_lock);
5900 MUTEX_DESTROY(&peer->peer_lock);
5902 (&peer->rpcStats, rpc_stat, nrpc_stat,
5903 rx_interface_stat)) {
5904 unsigned int num_funcs;
5907 queue_Remove(&rpc_stat->queue_header);
5908 queue_Remove(&rpc_stat->all_peers);
5909 num_funcs = rpc_stat->stats[0].func_total;
5911 sizeof(rx_interface_stat_t) +
5912 rpc_stat->stats[0].func_total *
5913 sizeof(rx_function_entry_v1_t);
5915 rxi_Free(rpc_stat, space);
5916 rxi_rpc_peer_stat_cnt -= num_funcs;
5919 MUTEX_ENTER(&rx_stats_mutex);
5920 rx_stats.nPeerStructs--;
5921 MUTEX_EXIT(&rx_stats_mutex);
5922 if (peer == *peer_ptr) {
5929 MUTEX_EXIT(&peer->peer_lock);
5935 MUTEX_EXIT(&rx_peerHashTable_lock);
5936 MUTEX_EXIT(&rx_rpc_stats);
5939 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5940 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5941 * GC, just below. Really, we shouldn't have to keep moving packets from
5942 * one place to another, but instead ought to always know if we can
5943 * afford to hold onto a packet in its particular use. */
5944 MUTEX_ENTER(&rx_freePktQ_lock);
5945 if (rx_waitingForPackets) {
5946 rx_waitingForPackets = 0;
5947 #ifdef RX_ENABLE_LOCKS
5948 CV_BROADCAST(&rx_waitingForPackets_cv);
5950 osi_rxWakeup(&rx_waitingForPackets);
5953 MUTEX_EXIT(&rx_freePktQ_lock);
5955 now.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5956 rxevent_Post(&now, rxi_ReapConnections, 0, 0);
5960 /* rxs_Release - This isn't strictly necessary but, since the macro name from
5961 * rx.h is sort of strange this is better. This is called with a security
5962 * object before it is discarded. Each connection using a security object has
5963 * its own refcount to the object so it won't actually be freed until the last
5964 * connection is destroyed.
5966 * This is the only rxs module call. A hold could also be written but no one
5970 rxs_Release(struct rx_securityClass *aobj)
5972 return RXS_Close(aobj);
5976 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
5977 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
5978 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
5979 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
5981 /* Adjust our estimate of the transmission rate to this peer, given
5982 * that the packet p was just acked. We can adjust peer->timeout and
5983 * call->twind. Pragmatically, this is called
5984 * only with packets of maximal length.
5985 * Called with peer and call locked.
5989 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
5990 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
5992 afs_int32 xferSize, xferMs;
5993 register afs_int32 minTime;
5996 /* Count down packets */
5997 if (peer->rateFlag > 0)
5999 /* Do nothing until we're enabled */
6000 if (peer->rateFlag != 0)
6005 /* Count only when the ack seems legitimate */
6006 switch (ackReason) {
6007 case RX_ACK_REQUESTED:
6009 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6013 case RX_ACK_PING_RESPONSE:
6014 if (p) /* want the response to ping-request, not data send */
6016 clock_GetTime(&newTO);
6017 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6018 clock_Sub(&newTO, &call->pingRequestTime);
6019 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6023 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6030 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));
6032 /* Track only packets that are big enough. */
6033 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6037 /* absorb RTT data (in milliseconds) for these big packets */
6038 if (peer->smRtt == 0) {
6039 peer->smRtt = xferMs;
6041 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6046 if (peer->countDown) {
6050 peer->countDown = 10; /* recalculate only every so often */
6052 /* In practice, we can measure only the RTT for full packets,
6053 * because of the way Rx acks the data that it receives. (If it's
6054 * smaller than a full packet, it often gets implicitly acked
6055 * either by the call response (from a server) or by the next call
6056 * (from a client), and either case confuses transmission times
6057 * with processing times.) Therefore, replace the above
6058 * more-sophisticated processing with a simpler version, where the
6059 * smoothed RTT is kept for full-size packets, and the time to
6060 * transmit a windowful of full-size packets is simply RTT *
6061 * windowSize. Again, we take two steps:
6062 - ensure the timeout is large enough for a single packet's RTT;
6063 - ensure that the window is small enough to fit in the desired timeout.*/
6065 /* First, the timeout check. */
6066 minTime = peer->smRtt;
6067 /* Get a reasonable estimate for a timeout period */
6069 newTO.sec = minTime / 1000;
6070 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6072 /* Increase the timeout period so that we can always do at least
6073 * one packet exchange */
6074 if (clock_Gt(&newTO, &peer->timeout)) {
6076 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));
6078 peer->timeout = newTO;
6081 /* Now, get an estimate for the transmit window size. */
6082 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6083 /* Now, convert to the number of full packets that could fit in a
6084 * reasonable fraction of that interval */
6085 minTime /= (peer->smRtt << 1);
6086 xferSize = minTime; /* (make a copy) */
6088 /* Now clamp the size to reasonable bounds. */
6091 else if (minTime > rx_Window)
6092 minTime = rx_Window;
6093 /* if (minTime != peer->maxWindow) {
6094 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6095 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6096 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6098 peer->maxWindow = minTime;
6099 elide... call->twind = minTime;
6103 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6104 * Discern this by calculating the timeout necessary for rx_Window
6106 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6107 /* calculate estimate for transmission interval in milliseconds */
6108 minTime = rx_Window * peer->smRtt;
6109 if (minTime < 1000) {
6110 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
6111 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6112 peer->timeout.usec, peer->smRtt, peer->packetSize));
6114 newTO.sec = 0; /* cut back on timeout by half a second */
6115 newTO.usec = 500000;
6116 clock_Sub(&peer->timeout, &newTO);
6121 } /* end of rxi_ComputeRate */
6122 #endif /* ADAPT_WINDOW */
6130 #define TRACE_OPTION_DEBUGLOG 4
6138 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6139 0, KEY_QUERY_VALUE, &parmKey);
6140 if (code != ERROR_SUCCESS)
6143 dummyLen = sizeof(TraceOption);
6144 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6145 (BYTE *) &TraceOption, &dummyLen);
6146 if (code == ERROR_SUCCESS) {
6147 rxdebug_active = (TraceOption & TRACE_OPTION_DEBUGLOG) ? 1 : 0;
6149 RegCloseKey (parmKey);
6150 #endif /* AFS_NT40_ENV */
6155 rx_DebugOnOff(int on)
6157 rxdebug_active = on;
6159 #endif /* AFS_NT40_ENV */
6162 /* Don't call this debugging routine directly; use dpf */
6164 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
6165 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
6173 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6176 len = _snprintf(msg, sizeof(msg)-2,
6177 tformat, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10,
6178 a11, a12, a13, a14, a15);
6180 if (msg[len-1] != '\n') {
6184 OutputDebugString(msg);
6189 clock_GetTime(&now);
6190 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
6191 (unsigned int)now.usec / 1000);
6192 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
6199 * This function is used to process the rx_stats structure that is local
6200 * to a process as well as an rx_stats structure received from a remote
6201 * process (via rxdebug). Therefore, it needs to do minimal version
6205 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
6206 afs_int32 freePackets, char version)
6210 if (size != sizeof(struct rx_stats)) {
6212 "Unexpected size of stats structure: was %d, expected %d\n",
6213 size, sizeof(struct rx_stats));
6216 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6219 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6220 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6221 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6222 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6223 s->specialPktAllocFailures);
6225 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6226 s->receivePktAllocFailures, s->sendPktAllocFailures,
6227 s->specialPktAllocFailures);
6231 " greedy %d, " "bogusReads %d (last from host %x), "
6232 "noPackets %d, " "noBuffers %d, " "selects %d, "
6233 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6234 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6235 s->selects, s->sendSelects);
6237 fprintf(file, " packets read: ");
6238 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6239 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6241 fprintf(file, "\n");
6244 " other read counters: data %d, " "ack %d, " "dup %d "
6245 "spurious %d " "dally %d\n", s->dataPacketsRead,
6246 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6247 s->ignorePacketDally);
6249 fprintf(file, " packets sent: ");
6250 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6251 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6253 fprintf(file, "\n");
6256 " other send counters: ack %d, " "data %d (not resends), "
6257 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6258 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6259 s->dataPacketsPushed, s->ignoreAckedPacket);
6262 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6263 s->netSendFailures, (int)s->fatalErrors);
6265 if (s->nRttSamples) {
6266 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6267 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6269 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6270 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6274 " %d server connections, " "%d client connections, "
6275 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6276 s->nServerConns, s->nClientConns, s->nPeerStructs,
6277 s->nCallStructs, s->nFreeCallStructs);
6279 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6280 fprintf(file, " %d clock updates\n", clock_nUpdates);
6285 /* for backward compatibility */
6287 rx_PrintStats(FILE * file)
6289 MUTEX_ENTER(&rx_stats_mutex);
6290 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6292 MUTEX_EXIT(&rx_stats_mutex);
6296 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6298 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6299 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6300 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6303 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6304 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6305 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6308 " Packet size %d, " "max in packet skew %d, "
6309 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6310 (int)peer->outPacketSkew);
6313 #ifdef AFS_PTHREAD_ENV
6315 * This mutex protects the following static variables:
6319 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6320 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6322 #define LOCK_RX_DEBUG
6323 #define UNLOCK_RX_DEBUG
6324 #endif /* AFS_PTHREAD_ENV */
6327 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6328 u_char type, void *inputData, size_t inputLength,
6329 void *outputData, size_t outputLength)
6331 static afs_int32 counter = 100;
6332 time_t waitTime, waitCount, startTime, endTime;
6333 struct rx_header theader;
6335 register afs_int32 code;
6336 struct timeval tv_now, tv_wake, tv_delta;
6337 struct sockaddr_in taddr, faddr;
6342 startTime = time(0);
6348 tp = &tbuffer[sizeof(struct rx_header)];
6349 taddr.sin_family = AF_INET;
6350 taddr.sin_port = remotePort;
6351 taddr.sin_addr.s_addr = remoteAddr;
6352 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6353 taddr.sin_len = sizeof(struct sockaddr_in);
6356 memset(&theader, 0, sizeof(theader));
6357 theader.epoch = htonl(999);
6359 theader.callNumber = htonl(counter);
6362 theader.type = type;
6363 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6364 theader.serviceId = 0;
6366 memcpy(tbuffer, &theader, sizeof(theader));
6367 memcpy(tp, inputData, inputLength);
6369 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6370 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6372 /* see if there's a packet available */
6373 gettimeofday(&tv_wake,0);
6374 tv_wake.tv_sec += waitTime;
6377 FD_SET(socket, &imask);
6378 tv_delta.tv_sec = tv_wake.tv_sec;
6379 tv_delta.tv_usec = tv_wake.tv_usec;
6380 gettimeofday(&tv_now, 0);
6382 if (tv_delta.tv_usec < tv_now.tv_usec) {
6384 tv_delta.tv_usec += 1000000;
6387 tv_delta.tv_usec -= tv_now.tv_usec;
6389 if (tv_delta.tv_sec < tv_now.tv_sec) {
6393 tv_delta.tv_sec -= tv_now.tv_sec;
6395 code = select(socket + 1, &imask, 0, 0, &tv_delta);
6396 if (code == 1 && FD_ISSET(socket, &imask)) {
6397 /* now receive a packet */
6398 faddrLen = sizeof(struct sockaddr_in);
6400 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6401 (struct sockaddr *)&faddr, &faddrLen);
6404 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6405 if (counter == ntohl(theader.callNumber))
6413 /* see if we've timed out */
6421 code -= sizeof(struct rx_header);
6422 if (code > outputLength)
6423 code = outputLength;
6424 memcpy(outputData, tp, code);
6429 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6430 afs_uint16 remotePort, struct rx_debugStats * stat,
6431 afs_uint32 * supportedValues)
6433 struct rx_debugIn in;
6436 *supportedValues = 0;
6437 in.type = htonl(RX_DEBUGI_GETSTATS);
6440 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6441 &in, sizeof(in), stat, sizeof(*stat));
6444 * If the call was successful, fixup the version and indicate
6445 * what contents of the stat structure are valid.
6446 * Also do net to host conversion of fields here.
6450 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6451 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6453 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6454 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6456 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6457 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6459 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6460 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6462 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6463 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6465 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6466 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6468 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6469 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6471 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6472 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6475 stat->nFreePackets = ntohl(stat->nFreePackets);
6476 stat->packetReclaims = ntohl(stat->packetReclaims);
6477 stat->callsExecuted = ntohl(stat->callsExecuted);
6478 stat->nWaiting = ntohl(stat->nWaiting);
6479 stat->idleThreads = ntohl(stat->idleThreads);
6486 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6487 afs_uint16 remotePort, struct rx_stats * stat,
6488 afs_uint32 * supportedValues)
6490 struct rx_debugIn in;
6491 afs_int32 *lp = (afs_int32 *) stat;
6496 * supportedValues is currently unused, but added to allow future
6497 * versioning of this function.
6500 *supportedValues = 0;
6501 in.type = htonl(RX_DEBUGI_RXSTATS);
6503 memset(stat, 0, sizeof(*stat));
6505 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6506 &in, sizeof(in), stat, sizeof(*stat));
6511 * Do net to host conversion here
6514 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6523 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6524 afs_uint16 remotePort, size_t version_length,
6528 return MakeDebugCall(socket, remoteAddr, remotePort,
6529 RX_PACKET_TYPE_VERSION, a, 1, version,
6534 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6535 afs_uint16 remotePort, afs_int32 * nextConnection,
6536 int allConnections, afs_uint32 debugSupportedValues,
6537 struct rx_debugConn * conn,
6538 afs_uint32 * supportedValues)
6540 struct rx_debugIn in;
6545 * supportedValues is currently unused, but added to allow future
6546 * versioning of this function.
6549 *supportedValues = 0;
6550 if (allConnections) {
6551 in.type = htonl(RX_DEBUGI_GETALLCONN);
6553 in.type = htonl(RX_DEBUGI_GETCONN);
6555 in.index = htonl(*nextConnection);
6556 memset(conn, 0, sizeof(*conn));
6558 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6559 &in, sizeof(in), conn, sizeof(*conn));
6562 *nextConnection += 1;
6565 * Convert old connection format to new structure.
6568 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6569 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6570 #define MOVEvL(a) (conn->a = vL->a)
6572 /* any old or unrecognized version... */
6573 for (i = 0; i < RX_MAXCALLS; i++) {
6574 MOVEvL(callState[i]);
6575 MOVEvL(callMode[i]);
6576 MOVEvL(callFlags[i]);
6577 MOVEvL(callOther[i]);
6579 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6580 MOVEvL(secStats.type);
6581 MOVEvL(secStats.level);
6582 MOVEvL(secStats.flags);
6583 MOVEvL(secStats.expires);
6584 MOVEvL(secStats.packetsReceived);
6585 MOVEvL(secStats.packetsSent);
6586 MOVEvL(secStats.bytesReceived);
6587 MOVEvL(secStats.bytesSent);
6592 * Do net to host conversion here
6594 * I don't convert host or port since we are most likely
6595 * going to want these in NBO.
6597 conn->cid = ntohl(conn->cid);
6598 conn->serial = ntohl(conn->serial);
6599 for (i = 0; i < RX_MAXCALLS; i++) {
6600 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6602 conn->error = ntohl(conn->error);
6603 conn->secStats.flags = ntohl(conn->secStats.flags);
6604 conn->secStats.expires = ntohl(conn->secStats.expires);
6605 conn->secStats.packetsReceived =
6606 ntohl(conn->secStats.packetsReceived);
6607 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6608 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6609 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6610 conn->epoch = ntohl(conn->epoch);
6611 conn->natMTU = ntohl(conn->natMTU);
6618 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6619 afs_uint16 remotePort, afs_int32 * nextPeer,
6620 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6621 afs_uint32 * supportedValues)
6623 struct rx_debugIn in;
6627 * supportedValues is currently unused, but added to allow future
6628 * versioning of this function.
6631 *supportedValues = 0;
6632 in.type = htonl(RX_DEBUGI_GETPEER);
6633 in.index = htonl(*nextPeer);
6634 memset(peer, 0, sizeof(*peer));
6636 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6637 &in, sizeof(in), peer, sizeof(*peer));
6643 * Do net to host conversion here
6645 * I don't convert host or port since we are most likely
6646 * going to want these in NBO.
6648 peer->ifMTU = ntohs(peer->ifMTU);
6649 peer->idleWhen = ntohl(peer->idleWhen);
6650 peer->refCount = ntohs(peer->refCount);
6651 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6652 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6653 peer->rtt = ntohl(peer->rtt);
6654 peer->rtt_dev = ntohl(peer->rtt_dev);
6655 peer->timeout.sec = ntohl(peer->timeout.sec);
6656 peer->timeout.usec = ntohl(peer->timeout.usec);
6657 peer->nSent = ntohl(peer->nSent);
6658 peer->reSends = ntohl(peer->reSends);
6659 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6660 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6661 peer->rateFlag = ntohl(peer->rateFlag);
6662 peer->natMTU = ntohs(peer->natMTU);
6663 peer->maxMTU = ntohs(peer->maxMTU);
6664 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6665 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6666 peer->MTU = ntohs(peer->MTU);
6667 peer->cwind = ntohs(peer->cwind);
6668 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6669 peer->congestSeq = ntohs(peer->congestSeq);
6670 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6671 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6672 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6673 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6678 #endif /* RXDEBUG */
6683 struct rx_serverQueueEntry *np;
6686 register struct rx_call *call;
6687 register struct rx_serverQueueEntry *sq;
6691 if (rxinit_status == 1) {
6693 return; /* Already shutdown. */
6697 #ifndef AFS_PTHREAD_ENV
6698 FD_ZERO(&rx_selectMask);
6699 #endif /* AFS_PTHREAD_ENV */
6700 rxi_dataQuota = RX_MAX_QUOTA;
6701 #ifndef AFS_PTHREAD_ENV
6703 #endif /* AFS_PTHREAD_ENV */
6706 #ifndef AFS_PTHREAD_ENV
6707 #ifndef AFS_USE_GETTIMEOFDAY
6709 #endif /* AFS_USE_GETTIMEOFDAY */
6710 #endif /* AFS_PTHREAD_ENV */
6712 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6713 call = queue_First(&rx_freeCallQueue, rx_call);
6715 rxi_Free(call, sizeof(struct rx_call));
6718 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6719 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6725 struct rx_peer **peer_ptr, **peer_end;
6726 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6727 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6729 struct rx_peer *peer, *next;
6730 for (peer = *peer_ptr; peer; peer = next) {
6731 rx_interface_stat_p rpc_stat, nrpc_stat;
6734 (&peer->rpcStats, rpc_stat, nrpc_stat,
6735 rx_interface_stat)) {
6736 unsigned int num_funcs;
6739 queue_Remove(&rpc_stat->queue_header);
6740 queue_Remove(&rpc_stat->all_peers);
6741 num_funcs = rpc_stat->stats[0].func_total;
6743 sizeof(rx_interface_stat_t) +
6744 rpc_stat->stats[0].func_total *
6745 sizeof(rx_function_entry_v1_t);
6747 rxi_Free(rpc_stat, space);
6748 MUTEX_ENTER(&rx_rpc_stats);
6749 rxi_rpc_peer_stat_cnt -= num_funcs;
6750 MUTEX_EXIT(&rx_rpc_stats);
6754 MUTEX_ENTER(&rx_stats_mutex);
6755 rx_stats.nPeerStructs--;
6756 MUTEX_EXIT(&rx_stats_mutex);
6760 for (i = 0; i < RX_MAX_SERVICES; i++) {
6762 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6764 for (i = 0; i < rx_hashTableSize; i++) {
6765 register struct rx_connection *tc, *ntc;
6766 MUTEX_ENTER(&rx_connHashTable_lock);
6767 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6769 for (j = 0; j < RX_MAXCALLS; j++) {
6771 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6774 rxi_Free(tc, sizeof(*tc));
6776 MUTEX_EXIT(&rx_connHashTable_lock);
6779 MUTEX_ENTER(&freeSQEList_lock);
6781 while ((np = rx_FreeSQEList)) {
6782 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6783 MUTEX_DESTROY(&np->lock);
6784 rxi_Free(np, sizeof(*np));
6787 MUTEX_EXIT(&freeSQEList_lock);
6788 MUTEX_DESTROY(&freeSQEList_lock);
6789 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6790 MUTEX_DESTROY(&rx_connHashTable_lock);
6791 MUTEX_DESTROY(&rx_peerHashTable_lock);
6792 MUTEX_DESTROY(&rx_serverPool_lock);
6794 osi_Free(rx_connHashTable,
6795 rx_hashTableSize * sizeof(struct rx_connection *));
6796 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6798 UNPIN(rx_connHashTable,
6799 rx_hashTableSize * sizeof(struct rx_connection *));
6800 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6802 rxi_FreeAllPackets();
6804 MUTEX_ENTER(&rx_stats_mutex);
6805 rxi_dataQuota = RX_MAX_QUOTA;
6806 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6807 MUTEX_EXIT(&rx_stats_mutex);
6813 #ifdef RX_ENABLE_LOCKS
6815 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6817 if (!MUTEX_ISMINE(lockaddr))
6818 osi_Panic("Lock not held: %s", msg);
6820 #endif /* RX_ENABLE_LOCKS */
6825 * Routines to implement connection specific data.
6829 rx_KeyCreate(rx_destructor_t rtn)
6832 MUTEX_ENTER(&rxi_keyCreate_lock);
6833 key = rxi_keyCreate_counter++;
6834 rxi_keyCreate_destructor = (rx_destructor_t *)
6835 realloc((void *)rxi_keyCreate_destructor,
6836 (key + 1) * sizeof(rx_destructor_t));
6837 rxi_keyCreate_destructor[key] = rtn;
6838 MUTEX_EXIT(&rxi_keyCreate_lock);
6843 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6846 MUTEX_ENTER(&conn->conn_data_lock);
6847 if (!conn->specific) {
6848 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6849 for (i = 0; i < key; i++)
6850 conn->specific[i] = NULL;
6851 conn->nSpecific = key + 1;
6852 conn->specific[key] = ptr;
6853 } else if (key >= conn->nSpecific) {
6854 conn->specific = (void **)
6855 realloc(conn->specific, (key + 1) * sizeof(void *));
6856 for (i = conn->nSpecific; i < key; i++)
6857 conn->specific[i] = NULL;
6858 conn->nSpecific = key + 1;
6859 conn->specific[key] = ptr;
6861 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6862 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6863 conn->specific[key] = ptr;
6865 MUTEX_EXIT(&conn->conn_data_lock);
6869 rx_GetSpecific(struct rx_connection *conn, int key)
6872 MUTEX_ENTER(&conn->conn_data_lock);
6873 if (key >= conn->nSpecific)
6876 ptr = conn->specific[key];
6877 MUTEX_EXIT(&conn->conn_data_lock);
6881 #endif /* !KERNEL */
6884 * processStats is a queue used to store the statistics for the local
6885 * process. Its contents are similar to the contents of the rpcStats
6886 * queue on a rx_peer structure, but the actual data stored within
6887 * this queue contains totals across the lifetime of the process (assuming
6888 * the stats have not been reset) - unlike the per peer structures
6889 * which can come and go based upon the peer lifetime.
6892 static struct rx_queue processStats = { &processStats, &processStats };
6895 * peerStats is a queue used to store the statistics for all peer structs.
6896 * Its contents are the union of all the peer rpcStats queues.
6899 static struct rx_queue peerStats = { &peerStats, &peerStats };
6902 * rxi_monitor_processStats is used to turn process wide stat collection
6906 static int rxi_monitor_processStats = 0;
6909 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6912 static int rxi_monitor_peerStats = 0;
6915 * rxi_AddRpcStat - given all of the information for a particular rpc
6916 * call, create (if needed) and update the stat totals for the rpc.
6920 * IN stats - the queue of stats that will be updated with the new value
6922 * IN rxInterface - a unique number that identifies the rpc interface
6924 * IN currentFunc - the index of the function being invoked
6926 * IN totalFunc - the total number of functions in this interface
6928 * IN queueTime - the amount of time this function waited for a thread
6930 * IN execTime - the amount of time this function invocation took to execute
6932 * IN bytesSent - the number bytes sent by this invocation
6934 * IN bytesRcvd - the number bytes received by this invocation
6936 * IN isServer - if true, this invocation was made to a server
6938 * IN remoteHost - the ip address of the remote host
6940 * IN remotePort - the port of the remote host
6942 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6944 * INOUT counter - if a new stats structure is allocated, the counter will
6945 * be updated with the new number of allocated stat structures
6953 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6954 afs_uint32 currentFunc, afs_uint32 totalFunc,
6955 struct clock *queueTime, struct clock *execTime,
6956 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6957 afs_uint32 remoteHost, afs_uint32 remotePort,
6958 int addToPeerList, unsigned int *counter)
6961 rx_interface_stat_p rpc_stat, nrpc_stat;
6964 * See if there's already a structure for this interface
6967 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6968 if ((rpc_stat->stats[0].interfaceId == rxInterface)
6969 && (rpc_stat->stats[0].remote_is_server == isServer))
6974 * Didn't find a match so allocate a new structure and add it to the
6978 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
6979 || (rpc_stat->stats[0].interfaceId != rxInterface)
6980 || (rpc_stat->stats[0].remote_is_server != isServer)) {
6985 sizeof(rx_interface_stat_t) +
6986 totalFunc * sizeof(rx_function_entry_v1_t);
6988 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
6989 if (rpc_stat == NULL) {
6993 *counter += totalFunc;
6994 for (i = 0; i < totalFunc; i++) {
6995 rpc_stat->stats[i].remote_peer = remoteHost;
6996 rpc_stat->stats[i].remote_port = remotePort;
6997 rpc_stat->stats[i].remote_is_server = isServer;
6998 rpc_stat->stats[i].interfaceId = rxInterface;
6999 rpc_stat->stats[i].func_total = totalFunc;
7000 rpc_stat->stats[i].func_index = i;
7001 hzero(rpc_stat->stats[i].invocations);
7002 hzero(rpc_stat->stats[i].bytes_sent);
7003 hzero(rpc_stat->stats[i].bytes_rcvd);
7004 rpc_stat->stats[i].queue_time_sum.sec = 0;
7005 rpc_stat->stats[i].queue_time_sum.usec = 0;
7006 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7007 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7008 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7009 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7010 rpc_stat->stats[i].queue_time_max.sec = 0;
7011 rpc_stat->stats[i].queue_time_max.usec = 0;
7012 rpc_stat->stats[i].execution_time_sum.sec = 0;
7013 rpc_stat->stats[i].execution_time_sum.usec = 0;
7014 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7015 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7016 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7017 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7018 rpc_stat->stats[i].execution_time_max.sec = 0;
7019 rpc_stat->stats[i].execution_time_max.usec = 0;
7021 queue_Prepend(stats, rpc_stat);
7022 if (addToPeerList) {
7023 queue_Prepend(&peerStats, &rpc_stat->all_peers);
7028 * Increment the stats for this function
7031 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
7032 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
7033 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
7034 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
7035 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7036 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7037 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7039 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7040 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7042 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7043 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7045 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7046 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7048 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7049 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7057 * rx_IncrementTimeAndCount - increment the times and count for a particular
7062 * IN peer - the peer who invoked the rpc
7064 * IN rxInterface - a unique number that identifies the rpc interface
7066 * IN currentFunc - the index of the function being invoked
7068 * IN totalFunc - the total number of functions in this interface
7070 * IN queueTime - the amount of time this function waited for a thread
7072 * IN execTime - the amount of time this function invocation took to execute
7074 * IN bytesSent - the number bytes sent by this invocation
7076 * IN bytesRcvd - the number bytes received by this invocation
7078 * IN isServer - if true, this invocation was made to a server
7086 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7087 afs_uint32 currentFunc, afs_uint32 totalFunc,
7088 struct clock *queueTime, struct clock *execTime,
7089 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7093 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
7096 MUTEX_ENTER(&rx_rpc_stats);
7097 MUTEX_ENTER(&peer->peer_lock);
7099 if (rxi_monitor_peerStats) {
7100 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7101 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7102 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7105 if (rxi_monitor_processStats) {
7106 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7107 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7108 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7111 MUTEX_EXIT(&peer->peer_lock);
7112 MUTEX_EXIT(&rx_rpc_stats);
7117 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7121 * IN callerVersion - the rpc stat version of the caller.
7123 * IN count - the number of entries to marshall.
7125 * IN stats - pointer to stats to be marshalled.
7127 * OUT ptr - Where to store the marshalled data.
7134 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7135 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7141 * We only support the first version
7143 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7144 *(ptr++) = stats->remote_peer;
7145 *(ptr++) = stats->remote_port;
7146 *(ptr++) = stats->remote_is_server;
7147 *(ptr++) = stats->interfaceId;
7148 *(ptr++) = stats->func_total;
7149 *(ptr++) = stats->func_index;
7150 *(ptr++) = hgethi(stats->invocations);
7151 *(ptr++) = hgetlo(stats->invocations);
7152 *(ptr++) = hgethi(stats->bytes_sent);
7153 *(ptr++) = hgetlo(stats->bytes_sent);
7154 *(ptr++) = hgethi(stats->bytes_rcvd);
7155 *(ptr++) = hgetlo(stats->bytes_rcvd);
7156 *(ptr++) = stats->queue_time_sum.sec;
7157 *(ptr++) = stats->queue_time_sum.usec;
7158 *(ptr++) = stats->queue_time_sum_sqr.sec;
7159 *(ptr++) = stats->queue_time_sum_sqr.usec;
7160 *(ptr++) = stats->queue_time_min.sec;
7161 *(ptr++) = stats->queue_time_min.usec;
7162 *(ptr++) = stats->queue_time_max.sec;
7163 *(ptr++) = stats->queue_time_max.usec;
7164 *(ptr++) = stats->execution_time_sum.sec;
7165 *(ptr++) = stats->execution_time_sum.usec;
7166 *(ptr++) = stats->execution_time_sum_sqr.sec;
7167 *(ptr++) = stats->execution_time_sum_sqr.usec;
7168 *(ptr++) = stats->execution_time_min.sec;
7169 *(ptr++) = stats->execution_time_min.usec;
7170 *(ptr++) = stats->execution_time_max.sec;
7171 *(ptr++) = stats->execution_time_max.usec;
7177 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7182 * IN callerVersion - the rpc stat version of the caller
7184 * OUT myVersion - the rpc stat version of this function
7186 * OUT clock_sec - local time seconds
7188 * OUT clock_usec - local time microseconds
7190 * OUT allocSize - the number of bytes allocated to contain stats
7192 * OUT statCount - the number stats retrieved from this process.
7194 * OUT stats - the actual stats retrieved from this process.
7198 * Returns void. If successful, stats will != NULL.
7202 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7203 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7204 size_t * allocSize, afs_uint32 * statCount,
7205 afs_uint32 ** stats)
7215 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7218 * Check to see if stats are enabled
7221 MUTEX_ENTER(&rx_rpc_stats);
7222 if (!rxi_monitor_processStats) {
7223 MUTEX_EXIT(&rx_rpc_stats);
7227 clock_GetTime(&now);
7228 *clock_sec = now.sec;
7229 *clock_usec = now.usec;
7232 * Allocate the space based upon the caller version
7234 * If the client is at an older version than we are,
7235 * we return the statistic data in the older data format, but
7236 * we still return our version number so the client knows we
7237 * are maintaining more data than it can retrieve.
7240 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7241 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7242 *statCount = rxi_rpc_process_stat_cnt;
7245 * This can't happen yet, but in the future version changes
7246 * can be handled by adding additional code here
7250 if (space > (size_t) 0) {
7252 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7255 rx_interface_stat_p rpc_stat, nrpc_stat;
7259 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7261 * Copy the data based upon the caller version
7263 rx_MarshallProcessRPCStats(callerVersion,
7264 rpc_stat->stats[0].func_total,
7265 rpc_stat->stats, &ptr);
7271 MUTEX_EXIT(&rx_rpc_stats);
7276 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7280 * IN callerVersion - the rpc stat version of the caller
7282 * OUT myVersion - the rpc stat version of this function
7284 * OUT clock_sec - local time seconds
7286 * OUT clock_usec - local time microseconds
7288 * OUT allocSize - the number of bytes allocated to contain stats
7290 * OUT statCount - the number of stats retrieved from the individual
7293 * OUT stats - the actual stats retrieved from the individual peer structures.
7297 * Returns void. If successful, stats will != NULL.
7301 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7302 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7303 size_t * allocSize, afs_uint32 * statCount,
7304 afs_uint32 ** stats)
7314 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7317 * Check to see if stats are enabled
7320 MUTEX_ENTER(&rx_rpc_stats);
7321 if (!rxi_monitor_peerStats) {
7322 MUTEX_EXIT(&rx_rpc_stats);
7326 clock_GetTime(&now);
7327 *clock_sec = now.sec;
7328 *clock_usec = now.usec;
7331 * Allocate the space based upon the caller version
7333 * If the client is at an older version than we are,
7334 * we return the statistic data in the older data format, but
7335 * we still return our version number so the client knows we
7336 * are maintaining more data than it can retrieve.
7339 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7340 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7341 *statCount = rxi_rpc_peer_stat_cnt;
7344 * This can't happen yet, but in the future version changes
7345 * can be handled by adding additional code here
7349 if (space > (size_t) 0) {
7351 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7354 rx_interface_stat_p rpc_stat, nrpc_stat;
7358 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7360 * We have to fix the offset of rpc_stat since we are
7361 * keeping this structure on two rx_queues. The rx_queue
7362 * package assumes that the rx_queue member is the first
7363 * member of the structure. That is, rx_queue assumes that
7364 * any one item is only on one queue at a time. We are
7365 * breaking that assumption and so we have to do a little
7366 * math to fix our pointers.
7369 fix_offset = (char *)rpc_stat;
7370 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7371 rpc_stat = (rx_interface_stat_p) fix_offset;
7374 * Copy the data based upon the caller version
7376 rx_MarshallProcessRPCStats(callerVersion,
7377 rpc_stat->stats[0].func_total,
7378 rpc_stat->stats, &ptr);
7384 MUTEX_EXIT(&rx_rpc_stats);
7389 * rx_FreeRPCStats - free memory allocated by
7390 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7394 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7395 * rx_RetrievePeerRPCStats
7397 * IN allocSize - the number of bytes in stats.
7405 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7407 rxi_Free(stats, allocSize);
7411 * rx_queryProcessRPCStats - see if process rpc stat collection is
7412 * currently enabled.
7418 * Returns 0 if stats are not enabled != 0 otherwise
7422 rx_queryProcessRPCStats(void)
7425 MUTEX_ENTER(&rx_rpc_stats);
7426 rc = rxi_monitor_processStats;
7427 MUTEX_EXIT(&rx_rpc_stats);
7432 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7438 * Returns 0 if stats are not enabled != 0 otherwise
7442 rx_queryPeerRPCStats(void)
7445 MUTEX_ENTER(&rx_rpc_stats);
7446 rc = rxi_monitor_peerStats;
7447 MUTEX_EXIT(&rx_rpc_stats);
7452 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7462 rx_enableProcessRPCStats(void)
7464 MUTEX_ENTER(&rx_rpc_stats);
7465 rx_enable_stats = 1;
7466 rxi_monitor_processStats = 1;
7467 MUTEX_EXIT(&rx_rpc_stats);
7471 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7481 rx_enablePeerRPCStats(void)
7483 MUTEX_ENTER(&rx_rpc_stats);
7484 rx_enable_stats = 1;
7485 rxi_monitor_peerStats = 1;
7486 MUTEX_EXIT(&rx_rpc_stats);
7490 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7500 rx_disableProcessRPCStats(void)
7502 rx_interface_stat_p rpc_stat, nrpc_stat;
7505 MUTEX_ENTER(&rx_rpc_stats);
7508 * Turn off process statistics and if peer stats is also off, turn
7512 rxi_monitor_processStats = 0;
7513 if (rxi_monitor_peerStats == 0) {
7514 rx_enable_stats = 0;
7517 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7518 unsigned int num_funcs = 0;
7521 queue_Remove(rpc_stat);
7522 num_funcs = rpc_stat->stats[0].func_total;
7524 sizeof(rx_interface_stat_t) +
7525 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7527 rxi_Free(rpc_stat, space);
7528 rxi_rpc_process_stat_cnt -= num_funcs;
7530 MUTEX_EXIT(&rx_rpc_stats);
7534 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7544 rx_disablePeerRPCStats(void)
7546 struct rx_peer **peer_ptr, **peer_end;
7549 MUTEX_ENTER(&rx_rpc_stats);
7552 * Turn off peer statistics and if process stats is also off, turn
7556 rxi_monitor_peerStats = 0;
7557 if (rxi_monitor_processStats == 0) {
7558 rx_enable_stats = 0;
7561 MUTEX_ENTER(&rx_peerHashTable_lock);
7562 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7563 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7565 struct rx_peer *peer, *next, *prev;
7566 for (prev = peer = *peer_ptr; peer; peer = next) {
7568 code = MUTEX_TRYENTER(&peer->peer_lock);
7570 rx_interface_stat_p rpc_stat, nrpc_stat;
7573 (&peer->rpcStats, rpc_stat, nrpc_stat,
7574 rx_interface_stat)) {
7575 unsigned int num_funcs = 0;
7578 queue_Remove(&rpc_stat->queue_header);
7579 queue_Remove(&rpc_stat->all_peers);
7580 num_funcs = rpc_stat->stats[0].func_total;
7582 sizeof(rx_interface_stat_t) +
7583 rpc_stat->stats[0].func_total *
7584 sizeof(rx_function_entry_v1_t);
7586 rxi_Free(rpc_stat, space);
7587 rxi_rpc_peer_stat_cnt -= num_funcs;
7589 MUTEX_EXIT(&peer->peer_lock);
7590 if (prev == *peer_ptr) {
7600 MUTEX_EXIT(&rx_peerHashTable_lock);
7601 MUTEX_EXIT(&rx_rpc_stats);
7605 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7610 * IN clearFlag - flag indicating which stats to clear
7618 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7620 rx_interface_stat_p rpc_stat, nrpc_stat;
7622 MUTEX_ENTER(&rx_rpc_stats);
7624 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7625 unsigned int num_funcs = 0, i;
7626 num_funcs = rpc_stat->stats[0].func_total;
7627 for (i = 0; i < num_funcs; i++) {
7628 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7629 hzero(rpc_stat->stats[i].invocations);
7631 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7632 hzero(rpc_stat->stats[i].bytes_sent);
7634 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7635 hzero(rpc_stat->stats[i].bytes_rcvd);
7637 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7638 rpc_stat->stats[i].queue_time_sum.sec = 0;
7639 rpc_stat->stats[i].queue_time_sum.usec = 0;
7641 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7642 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7643 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7645 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7646 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7647 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7649 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7650 rpc_stat->stats[i].queue_time_max.sec = 0;
7651 rpc_stat->stats[i].queue_time_max.usec = 0;
7653 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7654 rpc_stat->stats[i].execution_time_sum.sec = 0;
7655 rpc_stat->stats[i].execution_time_sum.usec = 0;
7657 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7658 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7659 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7661 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7662 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7663 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7665 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7666 rpc_stat->stats[i].execution_time_max.sec = 0;
7667 rpc_stat->stats[i].execution_time_max.usec = 0;
7672 MUTEX_EXIT(&rx_rpc_stats);
7676 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7681 * IN clearFlag - flag indicating which stats to clear
7689 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7691 rx_interface_stat_p rpc_stat, nrpc_stat;
7693 MUTEX_ENTER(&rx_rpc_stats);
7695 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7696 unsigned int num_funcs = 0, i;
7699 * We have to fix the offset of rpc_stat since we are
7700 * keeping this structure on two rx_queues. The rx_queue
7701 * package assumes that the rx_queue member is the first
7702 * member of the structure. That is, rx_queue assumes that
7703 * any one item is only on one queue at a time. We are
7704 * breaking that assumption and so we have to do a little
7705 * math to fix our pointers.
7708 fix_offset = (char *)rpc_stat;
7709 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7710 rpc_stat = (rx_interface_stat_p) fix_offset;
7712 num_funcs = rpc_stat->stats[0].func_total;
7713 for (i = 0; i < num_funcs; i++) {
7714 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7715 hzero(rpc_stat->stats[i].invocations);
7717 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7718 hzero(rpc_stat->stats[i].bytes_sent);
7720 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7721 hzero(rpc_stat->stats[i].bytes_rcvd);
7723 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7724 rpc_stat->stats[i].queue_time_sum.sec = 0;
7725 rpc_stat->stats[i].queue_time_sum.usec = 0;
7727 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7728 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7729 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7731 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7732 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7733 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7735 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7736 rpc_stat->stats[i].queue_time_max.sec = 0;
7737 rpc_stat->stats[i].queue_time_max.usec = 0;
7739 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7740 rpc_stat->stats[i].execution_time_sum.sec = 0;
7741 rpc_stat->stats[i].execution_time_sum.usec = 0;
7743 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7744 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7745 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7747 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7748 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7749 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7751 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7752 rpc_stat->stats[i].execution_time_max.sec = 0;
7753 rpc_stat->stats[i].execution_time_max.usec = 0;
7758 MUTEX_EXIT(&rx_rpc_stats);
7762 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7763 * is authorized to enable/disable/clear RX statistics.
7765 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7768 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7770 rxi_rxstat_userok = proc;
7774 rx_RxStatUserOk(struct rx_call *call)
7776 if (!rxi_rxstat_userok)
7778 return rxi_rxstat_userok(call);
7783 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
7784 * function in the MSVC runtime DLL (msvcrt.dll).
7786 * Note: the system serializes calls to this function.
7789 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
7790 DWORD reason, /* reason function is being called */
7791 LPVOID reserved) /* reserved for future use */
7794 case DLL_PROCESS_ATTACH:
7795 /* library is being attached to a process */
7799 case DLL_PROCESS_DETACH: