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 /* This doesn't work in all compilers with void (they're buggy), so fake it
768 conn->securityData = (VOID *) 0;
769 conn->securityIndex = serviceSecurityIndex;
770 rx_SetConnDeadTime(conn, rx_connDeadTime);
771 conn->ackRate = RX_FAST_ACK_RATE;
773 conn->specific = NULL;
774 conn->challengeEvent = NULL;
775 conn->delayedAbortEvent = NULL;
776 conn->abortCount = 0;
779 RXS_NewConnection(securityObject, conn);
781 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
783 conn->refCount++; /* no lock required since only this thread knows... */
784 conn->next = rx_connHashTable[hashindex];
785 rx_connHashTable[hashindex] = conn;
786 MUTEX_ENTER(&rx_stats_mutex);
787 rx_stats.nClientConns++;
788 MUTEX_EXIT(&rx_stats_mutex);
790 MUTEX_EXIT(&rx_connHashTable_lock);
796 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
798 /* The idea is to set the dead time to a value that allows several
799 * keepalives to be dropped without timing out the connection. */
800 conn->secondsUntilDead = MAX(seconds, 6);
801 conn->secondsUntilPing = conn->secondsUntilDead / 6;
804 int rxi_lowPeerRefCount = 0;
805 int rxi_lowConnRefCount = 0;
808 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
809 * NOTE: must not be called with rx_connHashTable_lock held.
812 rxi_CleanupConnection(struct rx_connection *conn)
814 /* Notify the service exporter, if requested, that this connection
815 * is being destroyed */
816 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
817 (*conn->service->destroyConnProc) (conn);
819 /* Notify the security module that this connection is being destroyed */
820 RXS_DestroyConnection(conn->securityObject, conn);
822 /* If this is the last connection using the rx_peer struct, set its
823 * idle time to now. rxi_ReapConnections will reap it if it's still
824 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
826 MUTEX_ENTER(&rx_peerHashTable_lock);
827 if (conn->peer->refCount < 2) {
828 conn->peer->idleWhen = clock_Sec();
829 if (conn->peer->refCount < 1) {
830 conn->peer->refCount = 1;
831 MUTEX_ENTER(&rx_stats_mutex);
832 rxi_lowPeerRefCount++;
833 MUTEX_EXIT(&rx_stats_mutex);
836 conn->peer->refCount--;
837 MUTEX_EXIT(&rx_peerHashTable_lock);
839 MUTEX_ENTER(&rx_stats_mutex);
840 if (conn->type == RX_SERVER_CONNECTION)
841 rx_stats.nServerConns--;
843 rx_stats.nClientConns--;
844 MUTEX_EXIT(&rx_stats_mutex);
847 if (conn->specific) {
849 for (i = 0; i < conn->nSpecific; i++) {
850 if (conn->specific[i] && rxi_keyCreate_destructor[i])
851 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
852 conn->specific[i] = NULL;
854 free(conn->specific);
856 conn->specific = NULL;
860 MUTEX_DESTROY(&conn->conn_call_lock);
861 MUTEX_DESTROY(&conn->conn_data_lock);
862 CV_DESTROY(&conn->conn_call_cv);
864 rxi_FreeConnection(conn);
867 /* Destroy the specified connection */
869 rxi_DestroyConnection(register struct rx_connection *conn)
871 MUTEX_ENTER(&rx_connHashTable_lock);
872 rxi_DestroyConnectionNoLock(conn);
873 /* conn should be at the head of the cleanup list */
874 if (conn == rx_connCleanup_list) {
875 rx_connCleanup_list = rx_connCleanup_list->next;
876 MUTEX_EXIT(&rx_connHashTable_lock);
877 rxi_CleanupConnection(conn);
879 #ifdef RX_ENABLE_LOCKS
881 MUTEX_EXIT(&rx_connHashTable_lock);
883 #endif /* RX_ENABLE_LOCKS */
887 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
889 register struct rx_connection **conn_ptr;
890 register int havecalls = 0;
891 struct rx_packet *packet;
898 MUTEX_ENTER(&conn->conn_data_lock);
899 if (conn->refCount > 0)
902 MUTEX_ENTER(&rx_stats_mutex);
903 rxi_lowConnRefCount++;
904 MUTEX_EXIT(&rx_stats_mutex);
907 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
908 /* Busy; wait till the last guy before proceeding */
909 MUTEX_EXIT(&conn->conn_data_lock);
914 /* If the client previously called rx_NewCall, but it is still
915 * waiting, treat this as a running call, and wait to destroy the
916 * connection later when the call completes. */
917 if ((conn->type == RX_CLIENT_CONNECTION)
918 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
919 conn->flags |= RX_CONN_DESTROY_ME;
920 MUTEX_EXIT(&conn->conn_data_lock);
924 MUTEX_EXIT(&conn->conn_data_lock);
926 /* Check for extant references to this connection */
927 for (i = 0; i < RX_MAXCALLS; i++) {
928 register struct rx_call *call = conn->call[i];
931 if (conn->type == RX_CLIENT_CONNECTION) {
932 MUTEX_ENTER(&call->lock);
933 if (call->delayedAckEvent) {
934 /* Push the final acknowledgment out now--there
935 * won't be a subsequent call to acknowledge the
936 * last reply packets */
937 rxevent_Cancel(call->delayedAckEvent, call,
938 RX_CALL_REFCOUNT_DELAY);
939 if (call->state == RX_STATE_PRECALL
940 || call->state == RX_STATE_ACTIVE) {
941 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
943 rxi_AckAll(NULL, call, 0);
946 MUTEX_EXIT(&call->lock);
950 #ifdef RX_ENABLE_LOCKS
952 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
953 MUTEX_EXIT(&conn->conn_data_lock);
955 /* Someone is accessing a packet right now. */
959 #endif /* RX_ENABLE_LOCKS */
962 /* Don't destroy the connection if there are any call
963 * structures still in use */
964 MUTEX_ENTER(&conn->conn_data_lock);
965 conn->flags |= RX_CONN_DESTROY_ME;
966 MUTEX_EXIT(&conn->conn_data_lock);
971 if (conn->delayedAbortEvent) {
972 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
973 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
975 MUTEX_ENTER(&conn->conn_data_lock);
976 rxi_SendConnectionAbort(conn, packet, 0, 1);
977 MUTEX_EXIT(&conn->conn_data_lock);
978 rxi_FreePacket(packet);
982 /* Remove from connection hash table before proceeding */
984 &rx_connHashTable[CONN_HASH
985 (peer->host, peer->port, conn->cid, conn->epoch,
987 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
988 if (*conn_ptr == conn) {
989 *conn_ptr = conn->next;
993 /* if the conn that we are destroying was the last connection, then we
994 * clear rxLastConn as well */
995 if (rxLastConn == conn)
998 /* Make sure the connection is completely reset before deleting it. */
999 /* get rid of pending events that could zap us later */
1000 if (conn->challengeEvent)
1001 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1002 if (conn->checkReachEvent)
1003 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1005 /* Add the connection to the list of destroyed connections that
1006 * need to be cleaned up. This is necessary to avoid deadlocks
1007 * in the routines we call to inform others that this connection is
1008 * being destroyed. */
1009 conn->next = rx_connCleanup_list;
1010 rx_connCleanup_list = conn;
1013 /* Externally available version */
1015 rx_DestroyConnection(register struct rx_connection *conn)
1020 rxi_DestroyConnection(conn);
1025 rx_GetConnection(register struct rx_connection *conn)
1030 MUTEX_ENTER(&conn->conn_data_lock);
1032 MUTEX_EXIT(&conn->conn_data_lock);
1036 /* Wait for the transmit queue to no longer be busy.
1037 * requires the call->lock to be held */
1038 static void rxi_WaitforTQBusy(struct rx_call *call) {
1039 while (call->flags & RX_CALL_TQ_BUSY) {
1040 call->flags |= RX_CALL_TQ_WAIT;
1042 #ifdef RX_ENABLE_LOCKS
1043 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1044 CV_WAIT(&call->cv_tq, &call->lock);
1045 #else /* RX_ENABLE_LOCKS */
1046 osi_rxSleep(&call->tq);
1047 #endif /* RX_ENABLE_LOCKS */
1049 if (call->tqWaiters == 0) {
1050 call->flags &= ~RX_CALL_TQ_WAIT;
1054 /* Start a new rx remote procedure call, on the specified connection.
1055 * If wait is set to 1, wait for a free call channel; otherwise return
1056 * 0. Maxtime gives the maximum number of seconds this call may take,
1057 * after rx_NewCall returns. After this time interval, a call to any
1058 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1059 * For fine grain locking, we hold the conn_call_lock in order to
1060 * to ensure that we don't get signalle after we found a call in an active
1061 * state and before we go to sleep.
1064 rx_NewCall(register struct rx_connection *conn)
1067 register struct rx_call *call;
1068 struct clock queueTime;
1072 dpf(("rx_NewCall(conn %x)\n", conn));
1075 clock_GetTime(&queueTime);
1076 MUTEX_ENTER(&conn->conn_call_lock);
1079 * Check if there are others waiting for a new call.
1080 * If so, let them go first to avoid starving them.
1081 * This is a fairly simple scheme, and might not be
1082 * a complete solution for large numbers of waiters.
1084 * makeCallWaiters keeps track of the number of
1085 * threads waiting to make calls and the
1086 * RX_CONN_MAKECALL_WAITING flag bit is used to
1087 * indicate that there are indeed calls waiting.
1088 * The flag is set when the waiter is incremented.
1089 * It is only cleared in rx_EndCall when
1090 * makeCallWaiters is 0. This prevents us from
1091 * accidently destroying the connection while it
1092 * is potentially about to be used.
1094 MUTEX_ENTER(&conn->conn_data_lock);
1095 if (conn->makeCallWaiters) {
1096 conn->flags |= RX_CONN_MAKECALL_WAITING;
1097 conn->makeCallWaiters++;
1098 MUTEX_EXIT(&conn->conn_data_lock);
1100 #ifdef RX_ENABLE_LOCKS
1101 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1105 MUTEX_ENTER(&conn->conn_data_lock);
1106 conn->makeCallWaiters--;
1108 MUTEX_EXIT(&conn->conn_data_lock);
1111 for (i = 0; i < RX_MAXCALLS; i++) {
1112 call = conn->call[i];
1114 MUTEX_ENTER(&call->lock);
1115 if (call->state == RX_STATE_DALLY) {
1116 rxi_ResetCall(call, 0);
1117 (*call->callNumber)++;
1120 MUTEX_EXIT(&call->lock);
1122 call = rxi_NewCall(conn, i);
1126 if (i < RX_MAXCALLS) {
1129 MUTEX_ENTER(&conn->conn_data_lock);
1130 conn->flags |= RX_CONN_MAKECALL_WAITING;
1131 conn->makeCallWaiters++;
1132 MUTEX_EXIT(&conn->conn_data_lock);
1134 #ifdef RX_ENABLE_LOCKS
1135 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1139 MUTEX_ENTER(&conn->conn_data_lock);
1140 conn->makeCallWaiters--;
1141 MUTEX_EXIT(&conn->conn_data_lock);
1144 * Wake up anyone else who might be giving us a chance to
1145 * run (see code above that avoids resource starvation).
1147 #ifdef RX_ENABLE_LOCKS
1148 CV_BROADCAST(&conn->conn_call_cv);
1153 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1155 /* Client is initially in send mode */
1156 call->state = RX_STATE_ACTIVE;
1157 call->error = conn->error;
1159 call->mode = RX_MODE_ERROR;
1161 call->mode = RX_MODE_SENDING;
1163 /* remember start time for call in case we have hard dead time limit */
1164 call->queueTime = queueTime;
1165 clock_GetTime(&call->startTime);
1166 hzero(call->bytesSent);
1167 hzero(call->bytesRcvd);
1169 /* Turn on busy protocol. */
1170 rxi_KeepAliveOn(call);
1172 MUTEX_EXIT(&call->lock);
1173 MUTEX_EXIT(&conn->conn_call_lock);
1176 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1177 /* Now, if TQ wasn't cleared earlier, do it now. */
1178 MUTEX_ENTER(&call->lock);
1179 rxi_WaitforTQBusy(call);
1180 if (call->flags & RX_CALL_TQ_CLEARME) {
1181 rxi_ClearTransmitQueue(call, 0);
1182 queue_Init(&call->tq);
1184 MUTEX_EXIT(&call->lock);
1185 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1187 dpf(("rx_NewCall(call %x)\n", call));
1192 rxi_HasActiveCalls(register struct rx_connection *aconn)
1195 register struct rx_call *tcall;
1199 for (i = 0; i < RX_MAXCALLS; i++) {
1200 if ((tcall = aconn->call[i])) {
1201 if ((tcall->state == RX_STATE_ACTIVE)
1202 || (tcall->state == RX_STATE_PRECALL)) {
1213 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1214 register afs_int32 * aint32s)
1217 register struct rx_call *tcall;
1221 for (i = 0; i < RX_MAXCALLS; i++) {
1222 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1223 aint32s[i] = aconn->callNumber[i] + 1;
1225 aint32s[i] = aconn->callNumber[i];
1232 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1233 register afs_int32 * aint32s)
1236 register struct rx_call *tcall;
1240 for (i = 0; i < RX_MAXCALLS; i++) {
1241 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1242 aconn->callNumber[i] = aint32s[i] - 1;
1244 aconn->callNumber[i] = aint32s[i];
1250 /* Advertise a new service. A service is named locally by a UDP port
1251 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1254 char *serviceName; Name for identification purposes (e.g. the
1255 service name might be used for probing for
1258 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1259 char *serviceName, struct rx_securityClass **securityObjects,
1260 int nSecurityObjects,
1261 afs_int32(*serviceProc) (struct rx_call * acall))
1263 osi_socket socket = OSI_NULLSOCKET;
1264 register struct rx_service *tservice;
1270 if (serviceId == 0) {
1272 "rx_NewService: service id for service %s is not non-zero.\n",
1279 "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",
1287 tservice = rxi_AllocService();
1289 for (i = 0; i < RX_MAX_SERVICES; i++) {
1290 register struct rx_service *service = rx_services[i];
1292 if (port == service->servicePort && host == service->serviceHost) {
1293 if (service->serviceId == serviceId) {
1294 /* The identical service has already been
1295 * installed; if the caller was intending to
1296 * change the security classes used by this
1297 * service, he/she loses. */
1299 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1300 serviceName, serviceId, service->serviceName);
1302 rxi_FreeService(tservice);
1305 /* Different service, same port: re-use the socket
1306 * which is bound to the same port */
1307 socket = service->socket;
1310 if (socket == OSI_NULLSOCKET) {
1311 /* If we don't already have a socket (from another
1312 * service on same port) get a new one */
1313 socket = rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
1314 if (socket == OSI_NULLSOCKET) {
1316 rxi_FreeService(tservice);
1321 service->socket = socket;
1322 service->serviceHost = host;
1323 service->servicePort = port;
1324 service->serviceId = serviceId;
1325 service->serviceName = serviceName;
1326 service->nSecurityObjects = nSecurityObjects;
1327 service->securityObjects = securityObjects;
1328 service->minProcs = 0;
1329 service->maxProcs = 1;
1330 service->idleDeadTime = 60;
1331 service->connDeadTime = rx_connDeadTime;
1332 service->executeRequestProc = serviceProc;
1333 service->checkReach = 0;
1334 rx_services[i] = service; /* not visible until now */
1340 rxi_FreeService(tservice);
1341 (osi_Msg "rx_NewService: cannot support > %d services\n",
1346 /* Set configuration options for all of a service's security objects */
1349 rx_SetSecurityConfiguration(struct rx_service *service,
1350 rx_securityConfigVariables type,
1354 for (i = 0; i<service->nSecurityObjects; i++) {
1355 if (service->securityObjects[i]) {
1356 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1364 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1365 struct rx_securityClass **securityObjects, int nSecurityObjects,
1366 afs_int32(*serviceProc) (struct rx_call * acall))
1368 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1371 /* Generic request processing loop. This routine should be called
1372 * by the implementation dependent rx_ServerProc. If socketp is
1373 * non-null, it will be set to the file descriptor that this thread
1374 * is now listening on. If socketp is null, this routine will never
1377 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1379 register struct rx_call *call;
1380 register afs_int32 code;
1381 register struct rx_service *tservice = NULL;
1388 call = rx_GetCall(threadID, tservice, socketp);
1389 if (socketp && *socketp != OSI_NULLSOCKET) {
1390 /* We are now a listener thread */
1395 /* if server is restarting( typically smooth shutdown) then do not
1396 * allow any new calls.
1399 if (rx_tranquil && (call != NULL)) {
1403 MUTEX_ENTER(&call->lock);
1405 rxi_CallError(call, RX_RESTARTING);
1406 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1408 MUTEX_EXIT(&call->lock);
1412 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1413 #ifdef RX_ENABLE_LOCKS
1415 #endif /* RX_ENABLE_LOCKS */
1416 afs_termState = AFSOP_STOP_AFS;
1417 afs_osi_Wakeup(&afs_termState);
1418 #ifdef RX_ENABLE_LOCKS
1420 #endif /* RX_ENABLE_LOCKS */
1425 tservice = call->conn->service;
1427 if (tservice->beforeProc)
1428 (*tservice->beforeProc) (call);
1430 code = call->conn->service->executeRequestProc(call);
1432 if (tservice->afterProc)
1433 (*tservice->afterProc) (call, code);
1435 rx_EndCall(call, code);
1436 MUTEX_ENTER(&rx_stats_mutex);
1438 MUTEX_EXIT(&rx_stats_mutex);
1444 rx_WakeupServerProcs(void)
1446 struct rx_serverQueueEntry *np, *tqp;
1450 MUTEX_ENTER(&rx_serverPool_lock);
1452 #ifdef RX_ENABLE_LOCKS
1453 if (rx_waitForPacket)
1454 CV_BROADCAST(&rx_waitForPacket->cv);
1455 #else /* RX_ENABLE_LOCKS */
1456 if (rx_waitForPacket)
1457 osi_rxWakeup(rx_waitForPacket);
1458 #endif /* RX_ENABLE_LOCKS */
1459 MUTEX_ENTER(&freeSQEList_lock);
1460 for (np = rx_FreeSQEList; np; np = tqp) {
1461 tqp = *(struct rx_serverQueueEntry **)np;
1462 #ifdef RX_ENABLE_LOCKS
1463 CV_BROADCAST(&np->cv);
1464 #else /* RX_ENABLE_LOCKS */
1466 #endif /* RX_ENABLE_LOCKS */
1468 MUTEX_EXIT(&freeSQEList_lock);
1469 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1470 #ifdef RX_ENABLE_LOCKS
1471 CV_BROADCAST(&np->cv);
1472 #else /* RX_ENABLE_LOCKS */
1474 #endif /* RX_ENABLE_LOCKS */
1476 MUTEX_EXIT(&rx_serverPool_lock);
1481 * One thing that seems to happen is that all the server threads get
1482 * tied up on some empty or slow call, and then a whole bunch of calls
1483 * arrive at once, using up the packet pool, so now there are more
1484 * empty calls. The most critical resources here are server threads
1485 * and the free packet pool. The "doreclaim" code seems to help in
1486 * general. I think that eventually we arrive in this state: there
1487 * are lots of pending calls which do have all their packets present,
1488 * so they won't be reclaimed, are multi-packet calls, so they won't
1489 * be scheduled until later, and thus are tying up most of the free
1490 * packet pool for a very long time.
1492 * 1. schedule multi-packet calls if all the packets are present.
1493 * Probably CPU-bound operation, useful to return packets to pool.
1494 * Do what if there is a full window, but the last packet isn't here?
1495 * 3. preserve one thread which *only* runs "best" calls, otherwise
1496 * it sleeps and waits for that type of call.
1497 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1498 * the current dataquota business is badly broken. The quota isn't adjusted
1499 * to reflect how many packets are presently queued for a running call.
1500 * So, when we schedule a queued call with a full window of packets queued
1501 * up for it, that *should* free up a window full of packets for other 2d-class
1502 * calls to be able to use from the packet pool. But it doesn't.
1504 * NB. Most of the time, this code doesn't run -- since idle server threads
1505 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1506 * as a new call arrives.
1508 /* Sleep until a call arrives. Returns a pointer to the call, ready
1509 * for an rx_Read. */
1510 #ifdef RX_ENABLE_LOCKS
1512 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1514 struct rx_serverQueueEntry *sq;
1515 register struct rx_call *call = (struct rx_call *)0;
1516 struct rx_service *service = NULL;
1519 MUTEX_ENTER(&freeSQEList_lock);
1521 if ((sq = rx_FreeSQEList)) {
1522 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1523 MUTEX_EXIT(&freeSQEList_lock);
1524 } else { /* otherwise allocate a new one and return that */
1525 MUTEX_EXIT(&freeSQEList_lock);
1526 sq = (struct rx_serverQueueEntry *)
1527 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1528 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1529 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1532 MUTEX_ENTER(&rx_serverPool_lock);
1533 if (cur_service != NULL) {
1534 ReturnToServerPool(cur_service);
1537 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1538 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1540 /* Scan for eligible incoming calls. A call is not eligible
1541 * if the maximum number of calls for its service type are
1542 * already executing */
1543 /* One thread will process calls FCFS (to prevent starvation),
1544 * while the other threads may run ahead looking for calls which
1545 * have all their input data available immediately. This helps
1546 * keep threads from blocking, waiting for data from the client. */
1547 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1548 service = tcall->conn->service;
1549 if (!QuotaOK(service)) {
1552 if (tno == rxi_fcfs_thread_num
1553 || !tcall->queue_item_header.next) {
1554 /* If we're the fcfs thread , then we'll just use
1555 * this call. If we haven't been able to find an optimal
1556 * choice, and we're at the end of the list, then use a
1557 * 2d choice if one has been identified. Otherwise... */
1558 call = (choice2 ? choice2 : tcall);
1559 service = call->conn->service;
1560 } else if (!queue_IsEmpty(&tcall->rq)) {
1561 struct rx_packet *rp;
1562 rp = queue_First(&tcall->rq, rx_packet);
1563 if (rp->header.seq == 1) {
1565 || (rp->header.flags & RX_LAST_PACKET)) {
1567 } else if (rxi_2dchoice && !choice2
1568 && !(tcall->flags & RX_CALL_CLEARED)
1569 && (tcall->rprev > rxi_HardAckRate)) {
1578 ReturnToServerPool(service);
1585 MUTEX_EXIT(&rx_serverPool_lock);
1586 MUTEX_ENTER(&call->lock);
1588 if (call->flags & RX_CALL_WAIT_PROC) {
1589 call->flags &= ~RX_CALL_WAIT_PROC;
1590 MUTEX_ENTER(&rx_stats_mutex);
1592 MUTEX_EXIT(&rx_stats_mutex);
1595 if (call->state != RX_STATE_PRECALL || call->error) {
1596 MUTEX_EXIT(&call->lock);
1597 MUTEX_ENTER(&rx_serverPool_lock);
1598 ReturnToServerPool(service);
1603 if (queue_IsEmpty(&call->rq)
1604 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1605 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1607 CLEAR_CALL_QUEUE_LOCK(call);
1610 /* If there are no eligible incoming calls, add this process
1611 * to the idle server queue, to wait for one */
1615 *socketp = OSI_NULLSOCKET;
1617 sq->socketp = socketp;
1618 queue_Append(&rx_idleServerQueue, sq);
1619 #ifndef AFS_AIX41_ENV
1620 rx_waitForPacket = sq;
1622 rx_waitingForPacket = sq;
1623 #endif /* AFS_AIX41_ENV */
1625 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1627 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1628 MUTEX_EXIT(&rx_serverPool_lock);
1629 return (struct rx_call *)0;
1632 } while (!(call = sq->newcall)
1633 && !(socketp && *socketp != OSI_NULLSOCKET));
1634 MUTEX_EXIT(&rx_serverPool_lock);
1636 MUTEX_ENTER(&call->lock);
1642 MUTEX_ENTER(&freeSQEList_lock);
1643 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1644 rx_FreeSQEList = sq;
1645 MUTEX_EXIT(&freeSQEList_lock);
1648 clock_GetTime(&call->startTime);
1649 call->state = RX_STATE_ACTIVE;
1650 call->mode = RX_MODE_RECEIVING;
1651 #ifdef RX_KERNEL_TRACE
1652 if (ICL_SETACTIVE(afs_iclSetp)) {
1653 int glockOwner = ISAFS_GLOCK();
1656 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1657 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1664 rxi_calltrace(RX_CALL_START, call);
1665 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1666 call->conn->service->servicePort, call->conn->service->serviceId,
1669 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1670 MUTEX_EXIT(&call->lock);
1672 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1677 #else /* RX_ENABLE_LOCKS */
1679 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1681 struct rx_serverQueueEntry *sq;
1682 register struct rx_call *call = (struct rx_call *)0, *choice2;
1683 struct rx_service *service = NULL;
1687 MUTEX_ENTER(&freeSQEList_lock);
1689 if ((sq = rx_FreeSQEList)) {
1690 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1691 MUTEX_EXIT(&freeSQEList_lock);
1692 } else { /* otherwise allocate a new one and return that */
1693 MUTEX_EXIT(&freeSQEList_lock);
1694 sq = (struct rx_serverQueueEntry *)
1695 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1696 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1697 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1699 MUTEX_ENTER(&sq->lock);
1701 if (cur_service != NULL) {
1702 cur_service->nRequestsRunning--;
1703 if (cur_service->nRequestsRunning < cur_service->minProcs)
1707 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1708 register struct rx_call *tcall, *ncall;
1709 /* Scan for eligible incoming calls. A call is not eligible
1710 * if the maximum number of calls for its service type are
1711 * already executing */
1712 /* One thread will process calls FCFS (to prevent starvation),
1713 * while the other threads may run ahead looking for calls which
1714 * have all their input data available immediately. This helps
1715 * keep threads from blocking, waiting for data from the client. */
1716 choice2 = (struct rx_call *)0;
1717 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1718 service = tcall->conn->service;
1719 if (QuotaOK(service)) {
1720 if (tno == rxi_fcfs_thread_num
1721 || !tcall->queue_item_header.next) {
1722 /* If we're the fcfs thread, then we'll just use
1723 * this call. If we haven't been able to find an optimal
1724 * choice, and we're at the end of the list, then use a
1725 * 2d choice if one has been identified. Otherwise... */
1726 call = (choice2 ? choice2 : tcall);
1727 service = call->conn->service;
1728 } else if (!queue_IsEmpty(&tcall->rq)) {
1729 struct rx_packet *rp;
1730 rp = queue_First(&tcall->rq, rx_packet);
1731 if (rp->header.seq == 1
1733 || (rp->header.flags & RX_LAST_PACKET))) {
1735 } else if (rxi_2dchoice && !choice2
1736 && !(tcall->flags & RX_CALL_CLEARED)
1737 && (tcall->rprev > rxi_HardAckRate)) {
1750 /* we can't schedule a call if there's no data!!! */
1751 /* send an ack if there's no data, if we're missing the
1752 * first packet, or we're missing something between first
1753 * and last -- there's a "hole" in the incoming data. */
1754 if (queue_IsEmpty(&call->rq)
1755 || queue_First(&call->rq, rx_packet)->header.seq != 1
1756 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1757 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1759 call->flags &= (~RX_CALL_WAIT_PROC);
1760 service->nRequestsRunning++;
1761 /* just started call in minProcs pool, need fewer to maintain
1763 if (service->nRequestsRunning <= service->minProcs)
1767 /* MUTEX_EXIT(&call->lock); */
1769 /* If there are no eligible incoming calls, add this process
1770 * to the idle server queue, to wait for one */
1773 *socketp = OSI_NULLSOCKET;
1775 sq->socketp = socketp;
1776 queue_Append(&rx_idleServerQueue, sq);
1780 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1782 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1783 return (struct rx_call *)0;
1786 } while (!(call = sq->newcall)
1787 && !(socketp && *socketp != OSI_NULLSOCKET));
1789 MUTEX_EXIT(&sq->lock);
1791 MUTEX_ENTER(&freeSQEList_lock);
1792 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1793 rx_FreeSQEList = sq;
1794 MUTEX_EXIT(&freeSQEList_lock);
1797 clock_GetTime(&call->startTime);
1798 call->state = RX_STATE_ACTIVE;
1799 call->mode = RX_MODE_RECEIVING;
1800 #ifdef RX_KERNEL_TRACE
1801 if (ICL_SETACTIVE(afs_iclSetp)) {
1802 int glockOwner = ISAFS_GLOCK();
1805 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1806 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1813 rxi_calltrace(RX_CALL_START, call);
1814 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1815 call->conn->service->servicePort, call->conn->service->serviceId,
1818 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1825 #endif /* RX_ENABLE_LOCKS */
1829 /* Establish a procedure to be called when a packet arrives for a
1830 * call. This routine will be called at most once after each call,
1831 * and will also be called if there is an error condition on the or
1832 * the call is complete. Used by multi rx to build a selection
1833 * function which determines which of several calls is likely to be a
1834 * good one to read from.
1835 * NOTE: the way this is currently implemented it is probably only a
1836 * good idea to (1) use it immediately after a newcall (clients only)
1837 * and (2) only use it once. Other uses currently void your warranty
1840 rx_SetArrivalProc(register struct rx_call *call,
1841 register void (*proc) (register struct rx_call * call,
1843 register int index),
1844 register VOID * handle, register int arg)
1846 call->arrivalProc = proc;
1847 call->arrivalProcHandle = handle;
1848 call->arrivalProcArg = arg;
1851 /* Call is finished (possibly prematurely). Return rc to the peer, if
1852 * appropriate, and return the final error code from the conversation
1856 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1858 register struct rx_connection *conn = call->conn;
1859 register struct rx_service *service;
1865 dpf(("rx_EndCall(call %x rc %d error %d abortCode %d)\n", call, rc, call->error, call->abortCode));
1868 MUTEX_ENTER(&call->lock);
1870 if (rc == 0 && call->error == 0) {
1871 call->abortCode = 0;
1872 call->abortCount = 0;
1875 call->arrivalProc = (void (*)())0;
1876 if (rc && call->error == 0) {
1877 rxi_CallError(call, rc);
1878 /* Send an abort message to the peer if this error code has
1879 * only just been set. If it was set previously, assume the
1880 * peer has already been sent the error code or will request it
1882 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1884 if (conn->type == RX_SERVER_CONNECTION) {
1885 /* Make sure reply or at least dummy reply is sent */
1886 if (call->mode == RX_MODE_RECEIVING) {
1887 rxi_WriteProc(call, 0, 0);
1889 if (call->mode == RX_MODE_SENDING) {
1890 rxi_FlushWrite(call);
1892 service = conn->service;
1893 rxi_calltrace(RX_CALL_END, call);
1894 /* Call goes to hold state until reply packets are acknowledged */
1895 if (call->tfirst + call->nSoftAcked < call->tnext) {
1896 call->state = RX_STATE_HOLD;
1898 call->state = RX_STATE_DALLY;
1899 rxi_ClearTransmitQueue(call, 0);
1900 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1901 rxevent_Cancel(call->keepAliveEvent, call,
1902 RX_CALL_REFCOUNT_ALIVE);
1904 } else { /* Client connection */
1906 /* Make sure server receives input packets, in the case where
1907 * no reply arguments are expected */
1908 if ((call->mode == RX_MODE_SENDING)
1909 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1910 (void)rxi_ReadProc(call, &dummy, 1);
1913 /* If we had an outstanding delayed ack, be nice to the server
1914 * and force-send it now.
1916 if (call->delayedAckEvent) {
1917 rxevent_Cancel(call->delayedAckEvent, call,
1918 RX_CALL_REFCOUNT_DELAY);
1919 call->delayedAckEvent = NULL;
1920 rxi_SendDelayedAck(NULL, call, NULL);
1923 /* We need to release the call lock since it's lower than the
1924 * conn_call_lock and we don't want to hold the conn_call_lock
1925 * over the rx_ReadProc call. The conn_call_lock needs to be held
1926 * here for the case where rx_NewCall is perusing the calls on
1927 * the connection structure. We don't want to signal until
1928 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1929 * have checked this call, found it active and by the time it
1930 * goes to sleep, will have missed the signal.
1932 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
1933 * there are threads waiting to use the conn object.
1935 MUTEX_EXIT(&call->lock);
1936 MUTEX_ENTER(&conn->conn_call_lock);
1937 MUTEX_ENTER(&call->lock);
1938 MUTEX_ENTER(&conn->conn_data_lock);
1939 conn->flags |= RX_CONN_BUSY;
1940 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1941 if (conn->makeCallWaiters == 0)
1942 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1943 MUTEX_EXIT(&conn->conn_data_lock);
1944 #ifdef RX_ENABLE_LOCKS
1945 CV_BROADCAST(&conn->conn_call_cv);
1950 #ifdef RX_ENABLE_LOCKS
1952 MUTEX_EXIT(&conn->conn_data_lock);
1954 #endif /* RX_ENABLE_LOCKS */
1955 call->state = RX_STATE_DALLY;
1957 error = call->error;
1959 /* currentPacket, nLeft, and NFree must be zeroed here, because
1960 * ResetCall cannot: ResetCall may be called at splnet(), in the
1961 * kernel version, and may interrupt the macros rx_Read or
1962 * rx_Write, which run at normal priority for efficiency. */
1963 if (call->currentPacket) {
1964 queue_Prepend(&call->iovq, call->currentPacket);
1965 call->currentPacket = (struct rx_packet *)0;
1968 call->nLeft = call->nFree = call->curlen = 0;
1970 /* Free any packets from the last call to ReadvProc/WritevProc */
1971 rxi_FreePackets(0, &call->iovq);
1973 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1974 MUTEX_EXIT(&call->lock);
1975 if (conn->type == RX_CLIENT_CONNECTION) {
1976 MUTEX_EXIT(&conn->conn_call_lock);
1977 conn->flags &= ~RX_CONN_BUSY;
1981 * Map errors to the local host's errno.h format.
1983 error = ntoh_syserr_conv(error);
1987 #if !defined(KERNEL)
1989 /* Call this routine when shutting down a server or client (especially
1990 * clients). This will allow Rx to gracefully garbage collect server
1991 * connections, and reduce the number of retries that a server might
1992 * make to a dead client.
1993 * This is not quite right, since some calls may still be ongoing and
1994 * we can't lock them to destroy them. */
1998 register struct rx_connection **conn_ptr, **conn_end;
2002 if (rxinit_status == 1) {
2004 return; /* Already shutdown. */
2006 rxi_DeleteCachedConnections();
2007 if (rx_connHashTable) {
2008 MUTEX_ENTER(&rx_connHashTable_lock);
2009 for (conn_ptr = &rx_connHashTable[0], conn_end =
2010 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2012 struct rx_connection *conn, *next;
2013 for (conn = *conn_ptr; conn; conn = next) {
2015 if (conn->type == RX_CLIENT_CONNECTION) {
2016 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2018 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2019 #ifdef RX_ENABLE_LOCKS
2020 rxi_DestroyConnectionNoLock(conn);
2021 #else /* RX_ENABLE_LOCKS */
2022 rxi_DestroyConnection(conn);
2023 #endif /* RX_ENABLE_LOCKS */
2027 #ifdef RX_ENABLE_LOCKS
2028 while (rx_connCleanup_list) {
2029 struct rx_connection *conn;
2030 conn = rx_connCleanup_list;
2031 rx_connCleanup_list = rx_connCleanup_list->next;
2032 MUTEX_EXIT(&rx_connHashTable_lock);
2033 rxi_CleanupConnection(conn);
2034 MUTEX_ENTER(&rx_connHashTable_lock);
2036 MUTEX_EXIT(&rx_connHashTable_lock);
2037 #endif /* RX_ENABLE_LOCKS */
2042 afs_winsockCleanup();
2050 /* if we wakeup packet waiter too often, can get in loop with two
2051 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2053 rxi_PacketsUnWait(void)
2055 if (!rx_waitingForPackets) {
2059 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2060 return; /* still over quota */
2063 rx_waitingForPackets = 0;
2064 #ifdef RX_ENABLE_LOCKS
2065 CV_BROADCAST(&rx_waitingForPackets_cv);
2067 osi_rxWakeup(&rx_waitingForPackets);
2073 /* ------------------Internal interfaces------------------------- */
2075 /* Return this process's service structure for the
2076 * specified socket and service */
2078 rxi_FindService(register osi_socket socket, register u_short serviceId)
2080 register struct rx_service **sp;
2081 for (sp = &rx_services[0]; *sp; sp++) {
2082 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2088 /* Allocate a call structure, for the indicated channel of the
2089 * supplied connection. The mode and state of the call must be set by
2090 * the caller. Returns the call with mutex locked. */
2092 rxi_NewCall(register struct rx_connection *conn, register int channel)
2094 register struct rx_call *call;
2095 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2096 register struct rx_call *cp; /* Call pointer temp */
2097 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2098 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2100 dpf(("rxi_NewCall(conn %x, channel %d)\n", conn, channel));
2102 /* Grab an existing call structure, or allocate a new one.
2103 * Existing call structures are assumed to have been left reset by
2105 MUTEX_ENTER(&rx_freeCallQueue_lock);
2107 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2109 * EXCEPT that the TQ might not yet be cleared out.
2110 * Skip over those with in-use TQs.
2113 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2114 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2120 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2121 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2122 call = queue_First(&rx_freeCallQueue, rx_call);
2123 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2125 MUTEX_ENTER(&rx_stats_mutex);
2126 rx_stats.nFreeCallStructs--;
2127 MUTEX_EXIT(&rx_stats_mutex);
2128 MUTEX_EXIT(&rx_freeCallQueue_lock);
2129 MUTEX_ENTER(&call->lock);
2130 CLEAR_CALL_QUEUE_LOCK(call);
2131 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2132 /* Now, if TQ wasn't cleared earlier, do it now. */
2133 if (call->flags & RX_CALL_TQ_CLEARME) {
2134 rxi_ClearTransmitQueue(call, 0);
2135 queue_Init(&call->tq);
2137 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2138 /* Bind the call to its connection structure */
2140 rxi_ResetCall(call, 1);
2142 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2144 MUTEX_EXIT(&rx_freeCallQueue_lock);
2145 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2146 MUTEX_ENTER(&call->lock);
2147 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2148 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2149 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2151 MUTEX_ENTER(&rx_stats_mutex);
2152 rx_stats.nCallStructs++;
2153 MUTEX_EXIT(&rx_stats_mutex);
2154 /* Initialize once-only items */
2155 queue_Init(&call->tq);
2156 queue_Init(&call->rq);
2157 queue_Init(&call->iovq);
2158 /* Bind the call to its connection structure (prereq for reset) */
2160 rxi_ResetCall(call, 1);
2162 call->channel = channel;
2163 call->callNumber = &conn->callNumber[channel];
2164 /* Note that the next expected call number is retained (in
2165 * conn->callNumber[i]), even if we reallocate the call structure
2167 conn->call[channel] = call;
2168 /* if the channel's never been used (== 0), we should start at 1, otherwise
2169 * the call number is valid from the last time this channel was used */
2170 if (*call->callNumber == 0)
2171 *call->callNumber = 1;
2176 /* A call has been inactive long enough that so we can throw away
2177 * state, including the call structure, which is placed on the call
2179 * Call is locked upon entry.
2180 * haveCTLock set if called from rxi_ReapConnections
2182 #ifdef RX_ENABLE_LOCKS
2184 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2185 #else /* RX_ENABLE_LOCKS */
2187 rxi_FreeCall(register struct rx_call *call)
2188 #endif /* RX_ENABLE_LOCKS */
2190 register int channel = call->channel;
2191 register struct rx_connection *conn = call->conn;
2194 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2195 (*call->callNumber)++;
2196 rxi_ResetCall(call, 0);
2197 call->conn->call[channel] = (struct rx_call *)0;
2199 MUTEX_ENTER(&rx_freeCallQueue_lock);
2200 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2201 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2202 /* A call may be free even though its transmit queue is still in use.
2203 * Since we search the call list from head to tail, put busy calls at
2204 * the head of the list, and idle calls at the tail.
2206 if (call->flags & RX_CALL_TQ_BUSY)
2207 queue_Prepend(&rx_freeCallQueue, call);
2209 queue_Append(&rx_freeCallQueue, call);
2210 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2211 queue_Append(&rx_freeCallQueue, call);
2212 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2213 MUTEX_ENTER(&rx_stats_mutex);
2214 rx_stats.nFreeCallStructs++;
2215 MUTEX_EXIT(&rx_stats_mutex);
2217 MUTEX_EXIT(&rx_freeCallQueue_lock);
2219 /* Destroy the connection if it was previously slated for
2220 * destruction, i.e. the Rx client code previously called
2221 * rx_DestroyConnection (client connections), or
2222 * rxi_ReapConnections called the same routine (server
2223 * connections). Only do this, however, if there are no
2224 * outstanding calls. Note that for fine grain locking, there appears
2225 * to be a deadlock in that rxi_FreeCall has a call locked and
2226 * DestroyConnectionNoLock locks each call in the conn. But note a
2227 * few lines up where we have removed this call from the conn.
2228 * If someone else destroys a connection, they either have no
2229 * call lock held or are going through this section of code.
2231 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2232 MUTEX_ENTER(&conn->conn_data_lock);
2234 MUTEX_EXIT(&conn->conn_data_lock);
2235 #ifdef RX_ENABLE_LOCKS
2237 rxi_DestroyConnectionNoLock(conn);
2239 rxi_DestroyConnection(conn);
2240 #else /* RX_ENABLE_LOCKS */
2241 rxi_DestroyConnection(conn);
2242 #endif /* RX_ENABLE_LOCKS */
2246 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2248 rxi_Alloc(register size_t size)
2252 MUTEX_ENTER(&rx_stats_mutex);
2254 rxi_Allocsize += (afs_int32)size;
2255 MUTEX_EXIT(&rx_stats_mutex);
2257 p = (char *)osi_Alloc(size);
2260 osi_Panic("rxi_Alloc error");
2266 rxi_Free(void *addr, register size_t size)
2268 MUTEX_ENTER(&rx_stats_mutex);
2270 rxi_Allocsize -= (afs_int32)size;
2271 MUTEX_EXIT(&rx_stats_mutex);
2273 osi_Free(addr, size);
2276 /* Find the peer process represented by the supplied (host,port)
2277 * combination. If there is no appropriate active peer structure, a
2278 * new one will be allocated and initialized
2279 * The origPeer, if set, is a pointer to a peer structure on which the
2280 * refcount will be be decremented. This is used to replace the peer
2281 * structure hanging off a connection structure */
2283 rxi_FindPeer(register afs_uint32 host, register u_short port,
2284 struct rx_peer *origPeer, int create)
2286 register struct rx_peer *pp;
2288 hashIndex = PEER_HASH(host, port);
2289 MUTEX_ENTER(&rx_peerHashTable_lock);
2290 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2291 if ((pp->host == host) && (pp->port == port))
2296 pp = rxi_AllocPeer(); /* This bzero's *pp */
2297 pp->host = host; /* set here or in InitPeerParams is zero */
2299 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2300 queue_Init(&pp->congestionQueue);
2301 queue_Init(&pp->rpcStats);
2302 pp->next = rx_peerHashTable[hashIndex];
2303 rx_peerHashTable[hashIndex] = pp;
2304 rxi_InitPeerParams(pp);
2305 MUTEX_ENTER(&rx_stats_mutex);
2306 rx_stats.nPeerStructs++;
2307 MUTEX_EXIT(&rx_stats_mutex);
2314 origPeer->refCount--;
2315 MUTEX_EXIT(&rx_peerHashTable_lock);
2320 /* Find the connection at (host, port) started at epoch, and with the
2321 * given connection id. Creates the server connection if necessary.
2322 * The type specifies whether a client connection or a server
2323 * connection is desired. In both cases, (host, port) specify the
2324 * peer's (host, pair) pair. Client connections are not made
2325 * automatically by this routine. The parameter socket gives the
2326 * socket descriptor on which the packet was received. This is used,
2327 * in the case of server connections, to check that *new* connections
2328 * come via a valid (port, serviceId). Finally, the securityIndex
2329 * parameter must match the existing index for the connection. If a
2330 * server connection is created, it will be created using the supplied
2331 * index, if the index is valid for this service */
2332 struct rx_connection *
2333 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2334 register u_short port, u_short serviceId, afs_uint32 cid,
2335 afs_uint32 epoch, int type, u_int securityIndex)
2337 int hashindex, flag;
2338 register struct rx_connection *conn;
2339 hashindex = CONN_HASH(host, port, cid, epoch, type);
2340 MUTEX_ENTER(&rx_connHashTable_lock);
2341 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2342 rx_connHashTable[hashindex],
2345 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2346 && (epoch == conn->epoch)) {
2347 register struct rx_peer *pp = conn->peer;
2348 if (securityIndex != conn->securityIndex) {
2349 /* this isn't supposed to happen, but someone could forge a packet
2350 * like this, and there seems to be some CM bug that makes this
2351 * happen from time to time -- in which case, the fileserver
2353 MUTEX_EXIT(&rx_connHashTable_lock);
2354 return (struct rx_connection *)0;
2356 if (pp->host == host && pp->port == port)
2358 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2360 /* So what happens when it's a callback connection? */
2361 if ( /*type == RX_CLIENT_CONNECTION && */
2362 (conn->epoch & 0x80000000))
2366 /* the connection rxLastConn that was used the last time is not the
2367 ** one we are looking for now. Hence, start searching in the hash */
2369 conn = rx_connHashTable[hashindex];
2374 struct rx_service *service;
2375 if (type == RX_CLIENT_CONNECTION) {
2376 MUTEX_EXIT(&rx_connHashTable_lock);
2377 return (struct rx_connection *)0;
2379 service = rxi_FindService(socket, serviceId);
2380 if (!service || (securityIndex >= service->nSecurityObjects)
2381 || (service->securityObjects[securityIndex] == 0)) {
2382 MUTEX_EXIT(&rx_connHashTable_lock);
2383 return (struct rx_connection *)0;
2385 conn = rxi_AllocConnection(); /* This bzero's the connection */
2386 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2387 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2388 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2389 conn->next = rx_connHashTable[hashindex];
2390 rx_connHashTable[hashindex] = conn;
2391 conn->peer = rxi_FindPeer(host, port, 0, 1);
2392 conn->type = RX_SERVER_CONNECTION;
2393 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2394 conn->epoch = epoch;
2395 conn->cid = cid & RX_CIDMASK;
2396 /* conn->serial = conn->lastSerial = 0; */
2397 /* conn->timeout = 0; */
2398 conn->ackRate = RX_FAST_ACK_RATE;
2399 conn->service = service;
2400 conn->serviceId = serviceId;
2401 conn->securityIndex = securityIndex;
2402 conn->securityObject = service->securityObjects[securityIndex];
2403 conn->nSpecific = 0;
2404 conn->specific = NULL;
2405 rx_SetConnDeadTime(conn, service->connDeadTime);
2406 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2407 /* Notify security object of the new connection */
2408 RXS_NewConnection(conn->securityObject, conn);
2409 /* XXXX Connection timeout? */
2410 if (service->newConnProc)
2411 (*service->newConnProc) (conn);
2412 MUTEX_ENTER(&rx_stats_mutex);
2413 rx_stats.nServerConns++;
2414 MUTEX_EXIT(&rx_stats_mutex);
2417 MUTEX_ENTER(&conn->conn_data_lock);
2419 MUTEX_EXIT(&conn->conn_data_lock);
2421 rxLastConn = conn; /* store this connection as the last conn used */
2422 MUTEX_EXIT(&rx_connHashTable_lock);
2426 /* There are two packet tracing routines available for testing and monitoring
2427 * Rx. One is called just after every packet is received and the other is
2428 * called just before every packet is sent. Received packets, have had their
2429 * headers decoded, and packets to be sent have not yet had their headers
2430 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2431 * containing the network address. Both can be modified. The return value, if
2432 * non-zero, indicates that the packet should be dropped. */
2434 int (*rx_justReceived) () = 0;
2435 int (*rx_almostSent) () = 0;
2437 /* A packet has been received off the interface. Np is the packet, socket is
2438 * the socket number it was received from (useful in determining which service
2439 * this packet corresponds to), and (host, port) reflect the host,port of the
2440 * sender. This call returns the packet to the caller if it is finished with
2441 * it, rather than de-allocating it, just as a small performance hack */
2444 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2445 afs_uint32 host, u_short port, int *tnop,
2446 struct rx_call **newcallp)
2448 register struct rx_call *call;
2449 register struct rx_connection *conn;
2451 afs_uint32 currentCallNumber;
2457 struct rx_packet *tnp;
2460 /* We don't print out the packet until now because (1) the time may not be
2461 * accurate enough until now in the lwp implementation (rx_Listener only gets
2462 * the time after the packet is read) and (2) from a protocol point of view,
2463 * this is the first time the packet has been seen */
2464 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2465 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2466 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2467 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2468 np->header.epoch, np->header.cid, np->header.callNumber,
2469 np->header.seq, np->header.flags, np));
2472 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2473 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2476 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2477 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2480 /* If an input tracer function is defined, call it with the packet and
2481 * network address. Note this function may modify its arguments. */
2482 if (rx_justReceived) {
2483 struct sockaddr_in addr;
2485 addr.sin_family = AF_INET;
2486 addr.sin_port = port;
2487 addr.sin_addr.s_addr = host;
2488 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2489 addr.sin_len = sizeof(addr);
2490 #endif /* AFS_OSF_ENV */
2491 drop = (*rx_justReceived) (np, &addr);
2492 /* drop packet if return value is non-zero */
2495 port = addr.sin_port; /* in case fcn changed addr */
2496 host = addr.sin_addr.s_addr;
2500 /* If packet was not sent by the client, then *we* must be the client */
2501 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2502 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2504 /* Find the connection (or fabricate one, if we're the server & if
2505 * necessary) associated with this packet */
2507 rxi_FindConnection(socket, host, port, np->header.serviceId,
2508 np->header.cid, np->header.epoch, type,
2509 np->header.securityIndex);
2512 /* If no connection found or fabricated, just ignore the packet.
2513 * (An argument could be made for sending an abort packet for
2518 MUTEX_ENTER(&conn->conn_data_lock);
2519 if (conn->maxSerial < np->header.serial)
2520 conn->maxSerial = np->header.serial;
2521 MUTEX_EXIT(&conn->conn_data_lock);
2523 /* If the connection is in an error state, send an abort packet and ignore
2524 * the incoming packet */
2526 /* Don't respond to an abort packet--we don't want loops! */
2527 MUTEX_ENTER(&conn->conn_data_lock);
2528 if (np->header.type != RX_PACKET_TYPE_ABORT)
2529 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2531 MUTEX_EXIT(&conn->conn_data_lock);
2535 /* Check for connection-only requests (i.e. not call specific). */
2536 if (np->header.callNumber == 0) {
2537 switch (np->header.type) {
2538 case RX_PACKET_TYPE_ABORT: {
2539 /* What if the supplied error is zero? */
2540 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2541 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2542 rxi_ConnectionError(conn, errcode);
2543 MUTEX_ENTER(&conn->conn_data_lock);
2545 MUTEX_EXIT(&conn->conn_data_lock);
2548 case RX_PACKET_TYPE_CHALLENGE:
2549 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2550 MUTEX_ENTER(&conn->conn_data_lock);
2552 MUTEX_EXIT(&conn->conn_data_lock);
2554 case RX_PACKET_TYPE_RESPONSE:
2555 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2556 MUTEX_ENTER(&conn->conn_data_lock);
2558 MUTEX_EXIT(&conn->conn_data_lock);
2560 case RX_PACKET_TYPE_PARAMS:
2561 case RX_PACKET_TYPE_PARAMS + 1:
2562 case RX_PACKET_TYPE_PARAMS + 2:
2563 /* ignore these packet types for now */
2564 MUTEX_ENTER(&conn->conn_data_lock);
2566 MUTEX_EXIT(&conn->conn_data_lock);
2571 /* Should not reach here, unless the peer is broken: send an
2573 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2574 MUTEX_ENTER(&conn->conn_data_lock);
2575 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2577 MUTEX_EXIT(&conn->conn_data_lock);
2582 channel = np->header.cid & RX_CHANNELMASK;
2583 call = conn->call[channel];
2584 #ifdef RX_ENABLE_LOCKS
2586 MUTEX_ENTER(&call->lock);
2587 /* Test to see if call struct is still attached to conn. */
2588 if (call != conn->call[channel]) {
2590 MUTEX_EXIT(&call->lock);
2591 if (type == RX_SERVER_CONNECTION) {
2592 call = conn->call[channel];
2593 /* If we started with no call attached and there is one now,
2594 * another thread is also running this routine and has gotten
2595 * the connection channel. We should drop this packet in the tests
2596 * below. If there was a call on this connection and it's now
2597 * gone, then we'll be making a new call below.
2598 * If there was previously a call and it's now different then
2599 * the old call was freed and another thread running this routine
2600 * has created a call on this channel. One of these two threads
2601 * has a packet for the old call and the code below handles those
2605 MUTEX_ENTER(&call->lock);
2607 /* This packet can't be for this call. If the new call address is
2608 * 0 then no call is running on this channel. If there is a call
2609 * then, since this is a client connection we're getting data for
2610 * it must be for the previous call.
2612 MUTEX_ENTER(&rx_stats_mutex);
2613 rx_stats.spuriousPacketsRead++;
2614 MUTEX_EXIT(&rx_stats_mutex);
2615 MUTEX_ENTER(&conn->conn_data_lock);
2617 MUTEX_EXIT(&conn->conn_data_lock);
2622 currentCallNumber = conn->callNumber[channel];
2624 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2625 if (np->header.callNumber < currentCallNumber) {
2626 MUTEX_ENTER(&rx_stats_mutex);
2627 rx_stats.spuriousPacketsRead++;
2628 MUTEX_EXIT(&rx_stats_mutex);
2629 #ifdef RX_ENABLE_LOCKS
2631 MUTEX_EXIT(&call->lock);
2633 MUTEX_ENTER(&conn->conn_data_lock);
2635 MUTEX_EXIT(&conn->conn_data_lock);
2639 MUTEX_ENTER(&conn->conn_call_lock);
2640 call = rxi_NewCall(conn, channel);
2641 MUTEX_EXIT(&conn->conn_call_lock);
2642 *call->callNumber = np->header.callNumber;
2643 if (np->header.callNumber == 0)
2644 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));
2646 call->state = RX_STATE_PRECALL;
2647 clock_GetTime(&call->queueTime);
2648 hzero(call->bytesSent);
2649 hzero(call->bytesRcvd);
2651 * If the number of queued calls exceeds the overload
2652 * threshold then abort this call.
2654 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2655 struct rx_packet *tp;
2657 rxi_CallError(call, rx_BusyError);
2658 tp = rxi_SendCallAbort(call, np, 1, 0);
2659 MUTEX_EXIT(&call->lock);
2660 MUTEX_ENTER(&conn->conn_data_lock);
2662 MUTEX_EXIT(&conn->conn_data_lock);
2663 MUTEX_ENTER(&rx_stats_mutex);
2665 MUTEX_EXIT(&rx_stats_mutex);
2668 rxi_KeepAliveOn(call);
2669 } else if (np->header.callNumber != currentCallNumber) {
2670 /* Wait until the transmit queue is idle before deciding
2671 * whether to reset the current call. Chances are that the
2672 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2675 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2676 while ((call->state == RX_STATE_ACTIVE)
2677 && (call->flags & RX_CALL_TQ_BUSY)) {
2678 call->flags |= RX_CALL_TQ_WAIT;
2680 #ifdef RX_ENABLE_LOCKS
2681 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2682 CV_WAIT(&call->cv_tq, &call->lock);
2683 #else /* RX_ENABLE_LOCKS */
2684 osi_rxSleep(&call->tq);
2685 #endif /* RX_ENABLE_LOCKS */
2687 if (call->tqWaiters == 0)
2688 call->flags &= ~RX_CALL_TQ_WAIT;
2690 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2691 /* If the new call cannot be taken right now send a busy and set
2692 * the error condition in this call, so that it terminates as
2693 * quickly as possible */
2694 if (call->state == RX_STATE_ACTIVE) {
2695 struct rx_packet *tp;
2697 rxi_CallError(call, RX_CALL_DEAD);
2698 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2700 MUTEX_EXIT(&call->lock);
2701 MUTEX_ENTER(&conn->conn_data_lock);
2703 MUTEX_EXIT(&conn->conn_data_lock);
2706 rxi_ResetCall(call, 0);
2707 *call->callNumber = np->header.callNumber;
2708 if (np->header.callNumber == 0)
2709 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));
2711 call->state = RX_STATE_PRECALL;
2712 clock_GetTime(&call->queueTime);
2713 hzero(call->bytesSent);
2714 hzero(call->bytesRcvd);
2716 * If the number of queued calls exceeds the overload
2717 * threshold then abort this call.
2719 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2720 struct rx_packet *tp;
2722 rxi_CallError(call, rx_BusyError);
2723 tp = rxi_SendCallAbort(call, np, 1, 0);
2724 MUTEX_EXIT(&call->lock);
2725 MUTEX_ENTER(&conn->conn_data_lock);
2727 MUTEX_EXIT(&conn->conn_data_lock);
2728 MUTEX_ENTER(&rx_stats_mutex);
2730 MUTEX_EXIT(&rx_stats_mutex);
2733 rxi_KeepAliveOn(call);
2735 /* Continuing call; do nothing here. */
2737 } else { /* we're the client */
2738 /* Ignore all incoming acknowledgements for calls in DALLY state */
2739 if (call && (call->state == RX_STATE_DALLY)
2740 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2741 MUTEX_ENTER(&rx_stats_mutex);
2742 rx_stats.ignorePacketDally++;
2743 MUTEX_EXIT(&rx_stats_mutex);
2744 #ifdef RX_ENABLE_LOCKS
2746 MUTEX_EXIT(&call->lock);
2749 MUTEX_ENTER(&conn->conn_data_lock);
2751 MUTEX_EXIT(&conn->conn_data_lock);
2755 /* Ignore anything that's not relevant to the current call. If there
2756 * isn't a current call, then no packet is relevant. */
2757 if (!call || (np->header.callNumber != currentCallNumber)) {
2758 MUTEX_ENTER(&rx_stats_mutex);
2759 rx_stats.spuriousPacketsRead++;
2760 MUTEX_EXIT(&rx_stats_mutex);
2761 #ifdef RX_ENABLE_LOCKS
2763 MUTEX_EXIT(&call->lock);
2766 MUTEX_ENTER(&conn->conn_data_lock);
2768 MUTEX_EXIT(&conn->conn_data_lock);
2771 /* If the service security object index stamped in the packet does not
2772 * match the connection's security index, ignore the packet */
2773 if (np->header.securityIndex != conn->securityIndex) {
2774 #ifdef RX_ENABLE_LOCKS
2775 MUTEX_EXIT(&call->lock);
2777 MUTEX_ENTER(&conn->conn_data_lock);
2779 MUTEX_EXIT(&conn->conn_data_lock);
2783 /* If we're receiving the response, then all transmit packets are
2784 * implicitly acknowledged. Get rid of them. */
2785 if (np->header.type == RX_PACKET_TYPE_DATA) {
2786 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2787 /* XXX Hack. Because we must release the global rx lock when
2788 * sending packets (osi_NetSend) we drop all acks while we're
2789 * traversing the tq in rxi_Start sending packets out because
2790 * packets may move to the freePacketQueue as result of being here!
2791 * So we drop these packets until we're safely out of the
2792 * traversing. Really ugly!
2793 * For fine grain RX locking, we set the acked field in the
2794 * packets and let rxi_Start remove them from the transmit queue.
2796 if (call->flags & RX_CALL_TQ_BUSY) {
2797 #ifdef RX_ENABLE_LOCKS
2798 rxi_SetAcksInTransmitQueue(call);
2801 return np; /* xmitting; drop packet */
2804 rxi_ClearTransmitQueue(call, 0);
2806 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2807 rxi_ClearTransmitQueue(call, 0);
2808 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2810 if (np->header.type == RX_PACKET_TYPE_ACK) {
2811 /* now check to see if this is an ack packet acknowledging that the
2812 * server actually *lost* some hard-acked data. If this happens we
2813 * ignore this packet, as it may indicate that the server restarted in
2814 * the middle of a call. It is also possible that this is an old ack
2815 * packet. We don't abort the connection in this case, because this
2816 * *might* just be an old ack packet. The right way to detect a server
2817 * restart in the midst of a call is to notice that the server epoch
2819 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2820 * XXX unacknowledged. I think that this is off-by-one, but
2821 * XXX I don't dare change it just yet, since it will
2822 * XXX interact badly with the server-restart detection
2823 * XXX code in receiveackpacket. */
2824 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2825 MUTEX_ENTER(&rx_stats_mutex);
2826 rx_stats.spuriousPacketsRead++;
2827 MUTEX_EXIT(&rx_stats_mutex);
2828 MUTEX_EXIT(&call->lock);
2829 MUTEX_ENTER(&conn->conn_data_lock);
2831 MUTEX_EXIT(&conn->conn_data_lock);
2835 } /* else not a data packet */
2838 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2839 /* Set remote user defined status from packet */
2840 call->remoteStatus = np->header.userStatus;
2842 /* Note the gap between the expected next packet and the actual
2843 * packet that arrived, when the new packet has a smaller serial number
2844 * than expected. Rioses frequently reorder packets all by themselves,
2845 * so this will be quite important with very large window sizes.
2846 * Skew is checked against 0 here to avoid any dependence on the type of
2847 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2849 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2850 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2851 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2853 MUTEX_ENTER(&conn->conn_data_lock);
2854 skew = conn->lastSerial - np->header.serial;
2855 conn->lastSerial = np->header.serial;
2856 MUTEX_EXIT(&conn->conn_data_lock);
2858 register struct rx_peer *peer;
2860 if (skew > peer->inPacketSkew) {
2861 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2863 peer->inPacketSkew = skew;
2867 /* Now do packet type-specific processing */
2868 switch (np->header.type) {
2869 case RX_PACKET_TYPE_DATA:
2870 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2873 case RX_PACKET_TYPE_ACK:
2874 /* Respond immediately to ack packets requesting acknowledgement
2876 if (np->header.flags & RX_REQUEST_ACK) {
2878 (void)rxi_SendCallAbort(call, 0, 1, 0);
2880 (void)rxi_SendAck(call, 0, np->header.serial,
2881 RX_ACK_PING_RESPONSE, 1);
2883 np = rxi_ReceiveAckPacket(call, np, 1);
2885 case RX_PACKET_TYPE_ABORT: {
2886 /* An abort packet: reset the call, passing the error up to the user. */
2887 /* What if error is zero? */
2888 /* What if the error is -1? the application will treat it as a timeout. */
2889 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
2890 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
2891 rxi_CallError(call, errdata);
2892 MUTEX_EXIT(&call->lock);
2893 MUTEX_ENTER(&conn->conn_data_lock);
2895 MUTEX_EXIT(&conn->conn_data_lock);
2896 return np; /* xmitting; drop packet */
2898 case RX_PACKET_TYPE_BUSY:
2901 case RX_PACKET_TYPE_ACKALL:
2902 /* All packets acknowledged, so we can drop all packets previously
2903 * readied for sending */
2904 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2905 /* XXX Hack. We because we can't release the global rx lock when
2906 * sending packets (osi_NetSend) we drop all ack pkts while we're
2907 * traversing the tq in rxi_Start sending packets out because
2908 * packets may move to the freePacketQueue as result of being
2909 * here! So we drop these packets until we're safely out of the
2910 * traversing. Really ugly!
2911 * For fine grain RX locking, we set the acked field in the packets
2912 * and let rxi_Start remove the packets from the transmit queue.
2914 if (call->flags & RX_CALL_TQ_BUSY) {
2915 #ifdef RX_ENABLE_LOCKS
2916 rxi_SetAcksInTransmitQueue(call);
2918 #else /* RX_ENABLE_LOCKS */
2919 MUTEX_EXIT(&call->lock);
2920 MUTEX_ENTER(&conn->conn_data_lock);
2922 MUTEX_EXIT(&conn->conn_data_lock);
2923 return np; /* xmitting; drop packet */
2924 #endif /* RX_ENABLE_LOCKS */
2926 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2927 rxi_ClearTransmitQueue(call, 0);
2930 /* Should not reach here, unless the peer is broken: send an abort
2932 rxi_CallError(call, RX_PROTOCOL_ERROR);
2933 np = rxi_SendCallAbort(call, np, 1, 0);
2936 /* Note when this last legitimate packet was received, for keep-alive
2937 * processing. Note, we delay getting the time until now in the hope that
2938 * the packet will be delivered to the user before any get time is required
2939 * (if not, then the time won't actually be re-evaluated here). */
2940 call->lastReceiveTime = clock_Sec();
2941 MUTEX_EXIT(&call->lock);
2942 MUTEX_ENTER(&conn->conn_data_lock);
2944 MUTEX_EXIT(&conn->conn_data_lock);
2948 /* return true if this is an "interesting" connection from the point of view
2949 of someone trying to debug the system */
2951 rxi_IsConnInteresting(struct rx_connection *aconn)
2954 register struct rx_call *tcall;
2956 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2958 for (i = 0; i < RX_MAXCALLS; i++) {
2959 tcall = aconn->call[i];
2961 if ((tcall->state == RX_STATE_PRECALL)
2962 || (tcall->state == RX_STATE_ACTIVE))
2964 if ((tcall->mode == RX_MODE_SENDING)
2965 || (tcall->mode == RX_MODE_RECEIVING))
2973 /* if this is one of the last few packets AND it wouldn't be used by the
2974 receiving call to immediately satisfy a read request, then drop it on
2975 the floor, since accepting it might prevent a lock-holding thread from
2976 making progress in its reading. If a call has been cleared while in
2977 the precall state then ignore all subsequent packets until the call
2978 is assigned to a thread. */
2981 TooLow(struct rx_packet *ap, struct rx_call *acall)
2984 MUTEX_ENTER(&rx_stats_mutex);
2985 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
2986 && (acall->state == RX_STATE_PRECALL))
2987 || ((rx_nFreePackets < rxi_dataQuota + 2)
2988 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
2989 && (acall->flags & RX_CALL_READER_WAIT)))) {
2992 MUTEX_EXIT(&rx_stats_mutex);
2998 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
2999 struct rx_call *acall)
3001 struct rx_call *call = acall;
3005 MUTEX_ENTER(&conn->conn_data_lock);
3006 conn->checkReachEvent = NULL;
3007 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3010 MUTEX_EXIT(&conn->conn_data_lock);
3014 MUTEX_ENTER(&conn->conn_call_lock);
3015 MUTEX_ENTER(&conn->conn_data_lock);
3016 for (i = 0; i < RX_MAXCALLS; i++) {
3017 struct rx_call *tc = conn->call[i];
3018 if (tc && tc->state == RX_STATE_PRECALL) {
3024 /* Indicate that rxi_CheckReachEvent is no longer running by
3025 * clearing the flag. Must be atomic under conn_data_lock to
3026 * avoid a new call slipping by: rxi_CheckConnReach holds
3027 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3029 conn->flags &= ~RX_CONN_ATTACHWAIT;
3030 MUTEX_EXIT(&conn->conn_data_lock);
3031 MUTEX_EXIT(&conn->conn_call_lock);
3036 MUTEX_ENTER(&call->lock);
3037 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3039 MUTEX_EXIT(&call->lock);
3041 clock_GetTime(&when);
3042 when.sec += RX_CHECKREACH_TIMEOUT;
3043 MUTEX_ENTER(&conn->conn_data_lock);
3044 if (!conn->checkReachEvent) {
3046 conn->checkReachEvent =
3047 rxevent_Post(&when, rxi_CheckReachEvent, conn, NULL);
3049 MUTEX_EXIT(&conn->conn_data_lock);
3055 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3057 struct rx_service *service = conn->service;
3058 struct rx_peer *peer = conn->peer;
3059 afs_uint32 now, lastReach;
3061 if (service->checkReach == 0)
3065 MUTEX_ENTER(&peer->peer_lock);
3066 lastReach = peer->lastReachTime;
3067 MUTEX_EXIT(&peer->peer_lock);
3068 if (now - lastReach < RX_CHECKREACH_TTL)
3071 MUTEX_ENTER(&conn->conn_data_lock);
3072 if (conn->flags & RX_CONN_ATTACHWAIT) {
3073 MUTEX_EXIT(&conn->conn_data_lock);
3076 conn->flags |= RX_CONN_ATTACHWAIT;
3077 MUTEX_EXIT(&conn->conn_data_lock);
3078 if (!conn->checkReachEvent)
3079 rxi_CheckReachEvent(NULL, conn, call);
3084 /* try to attach call, if authentication is complete */
3086 TryAttach(register struct rx_call *acall, register osi_socket socket,
3087 register int *tnop, register struct rx_call **newcallp,
3090 struct rx_connection *conn = acall->conn;
3092 if (conn->type == RX_SERVER_CONNECTION
3093 && acall->state == RX_STATE_PRECALL) {
3094 /* Don't attach until we have any req'd. authentication. */
3095 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3096 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3097 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3098 /* Note: this does not necessarily succeed; there
3099 * may not any proc available
3102 rxi_ChallengeOn(acall->conn);
3107 /* A data packet has been received off the interface. This packet is
3108 * appropriate to the call (the call is in the right state, etc.). This
3109 * routine can return a packet to the caller, for re-use */
3112 rxi_ReceiveDataPacket(register struct rx_call *call,
3113 register struct rx_packet *np, int istack,
3114 osi_socket socket, afs_uint32 host, u_short port,
3115 int *tnop, struct rx_call **newcallp)
3117 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3121 afs_uint32 seq, serial, flags;
3123 struct rx_packet *tnp;
3125 MUTEX_ENTER(&rx_stats_mutex);
3126 rx_stats.dataPacketsRead++;
3127 MUTEX_EXIT(&rx_stats_mutex);
3130 /* If there are no packet buffers, drop this new packet, unless we can find
3131 * packet buffers from inactive calls */
3133 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3134 MUTEX_ENTER(&rx_freePktQ_lock);
3135 rxi_NeedMorePackets = TRUE;
3136 MUTEX_EXIT(&rx_freePktQ_lock);
3137 MUTEX_ENTER(&rx_stats_mutex);
3138 rx_stats.noPacketBuffersOnRead++;
3139 MUTEX_EXIT(&rx_stats_mutex);
3140 call->rprev = np->header.serial;
3141 rxi_calltrace(RX_TRACE_DROP, call);
3142 dpf(("packet %x dropped on receipt - quota problems", np));
3144 rxi_ClearReceiveQueue(call);
3145 clock_GetTime(&when);
3146 clock_Add(&when, &rx_softAckDelay);
3147 if (!call->delayedAckEvent
3148 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3149 rxevent_Cancel(call->delayedAckEvent, call,
3150 RX_CALL_REFCOUNT_DELAY);
3151 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3152 call->delayedAckEvent =
3153 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3155 /* we've damaged this call already, might as well do it in. */
3161 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3162 * packet is one of several packets transmitted as a single
3163 * datagram. Do not send any soft or hard acks until all packets
3164 * in a jumbogram have been processed. Send negative acks right away.
3166 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3167 /* tnp is non-null when there are more packets in the
3168 * current jumbo gram */
3175 seq = np->header.seq;
3176 serial = np->header.serial;
3177 flags = np->header.flags;
3179 /* If the call is in an error state, send an abort message */
3181 return rxi_SendCallAbort(call, np, istack, 0);
3183 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3184 * AFS 3.5 jumbogram. */
3185 if (flags & RX_JUMBO_PACKET) {
3186 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3191 if (np->header.spare != 0) {
3192 MUTEX_ENTER(&call->conn->conn_data_lock);
3193 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3194 MUTEX_EXIT(&call->conn->conn_data_lock);
3197 /* The usual case is that this is the expected next packet */
3198 if (seq == call->rnext) {
3200 /* Check to make sure it is not a duplicate of one already queued */
3201 if (queue_IsNotEmpty(&call->rq)
3202 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3203 MUTEX_ENTER(&rx_stats_mutex);
3204 rx_stats.dupPacketsRead++;
3205 MUTEX_EXIT(&rx_stats_mutex);
3206 dpf(("packet %x dropped on receipt - duplicate", np));
3207 rxevent_Cancel(call->delayedAckEvent, call,
3208 RX_CALL_REFCOUNT_DELAY);
3209 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3215 /* It's the next packet. Stick it on the receive queue
3216 * for this call. Set newPackets to make sure we wake
3217 * the reader once all packets have been processed */
3218 queue_Prepend(&call->rq, np);
3220 np = NULL; /* We can't use this anymore */
3223 /* If an ack is requested then set a flag to make sure we
3224 * send an acknowledgement for this packet */
3225 if (flags & RX_REQUEST_ACK) {
3226 ackNeeded = RX_ACK_REQUESTED;
3229 /* Keep track of whether we have received the last packet */
3230 if (flags & RX_LAST_PACKET) {
3231 call->flags |= RX_CALL_HAVE_LAST;
3235 /* Check whether we have all of the packets for this call */
3236 if (call->flags & RX_CALL_HAVE_LAST) {
3237 afs_uint32 tseq; /* temporary sequence number */
3238 struct rx_packet *tp; /* Temporary packet pointer */
3239 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3241 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3242 if (tseq != tp->header.seq)
3244 if (tp->header.flags & RX_LAST_PACKET) {
3245 call->flags |= RX_CALL_RECEIVE_DONE;
3252 /* Provide asynchronous notification for those who want it
3253 * (e.g. multi rx) */
3254 if (call->arrivalProc) {
3255 (*call->arrivalProc) (call, call->arrivalProcHandle,
3256 call->arrivalProcArg);
3257 call->arrivalProc = (void (*)())0;
3260 /* Update last packet received */
3263 /* If there is no server process serving this call, grab
3264 * one, if available. We only need to do this once. If a
3265 * server thread is available, this thread becomes a server
3266 * thread and the server thread becomes a listener thread. */
3268 TryAttach(call, socket, tnop, newcallp, 0);
3271 /* This is not the expected next packet. */
3273 /* Determine whether this is a new or old packet, and if it's
3274 * a new one, whether it fits into the current receive window.
3275 * Also figure out whether the packet was delivered in sequence.
3276 * We use the prev variable to determine whether the new packet
3277 * is the successor of its immediate predecessor in the
3278 * receive queue, and the missing flag to determine whether
3279 * any of this packets predecessors are missing. */
3281 afs_uint32 prev; /* "Previous packet" sequence number */
3282 struct rx_packet *tp; /* Temporary packet pointer */
3283 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3284 int missing; /* Are any predecessors missing? */
3286 /* If the new packet's sequence number has been sent to the
3287 * application already, then this is a duplicate */
3288 if (seq < call->rnext) {
3289 MUTEX_ENTER(&rx_stats_mutex);
3290 rx_stats.dupPacketsRead++;
3291 MUTEX_EXIT(&rx_stats_mutex);
3292 rxevent_Cancel(call->delayedAckEvent, call,
3293 RX_CALL_REFCOUNT_DELAY);
3294 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3300 /* If the sequence number is greater than what can be
3301 * accomodated by the current window, then send a negative
3302 * acknowledge and drop the packet */
3303 if ((call->rnext + call->rwind) <= seq) {
3304 rxevent_Cancel(call->delayedAckEvent, call,
3305 RX_CALL_REFCOUNT_DELAY);
3306 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3313 /* Look for the packet in the queue of old received packets */
3314 for (prev = call->rnext - 1, missing =
3315 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3316 /*Check for duplicate packet */
3317 if (seq == tp->header.seq) {
3318 MUTEX_ENTER(&rx_stats_mutex);
3319 rx_stats.dupPacketsRead++;
3320 MUTEX_EXIT(&rx_stats_mutex);
3321 rxevent_Cancel(call->delayedAckEvent, call,
3322 RX_CALL_REFCOUNT_DELAY);
3323 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3329 /* If we find a higher sequence packet, break out and
3330 * insert the new packet here. */
3331 if (seq < tp->header.seq)
3333 /* Check for missing packet */
3334 if (tp->header.seq != prev + 1) {
3338 prev = tp->header.seq;
3341 /* Keep track of whether we have received the last packet. */
3342 if (flags & RX_LAST_PACKET) {
3343 call->flags |= RX_CALL_HAVE_LAST;
3346 /* It's within the window: add it to the the receive queue.
3347 * tp is left by the previous loop either pointing at the
3348 * packet before which to insert the new packet, or at the
3349 * queue head if the queue is empty or the packet should be
3351 queue_InsertBefore(tp, np);
3355 /* Check whether we have all of the packets for this call */
3356 if ((call->flags & RX_CALL_HAVE_LAST)
3357 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3358 afs_uint32 tseq; /* temporary sequence number */
3361 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3362 if (tseq != tp->header.seq)
3364 if (tp->header.flags & RX_LAST_PACKET) {
3365 call->flags |= RX_CALL_RECEIVE_DONE;
3372 /* We need to send an ack of the packet is out of sequence,
3373 * or if an ack was requested by the peer. */
3374 if (seq != prev + 1 || missing) {
3375 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3376 } else if (flags & RX_REQUEST_ACK) {
3377 ackNeeded = RX_ACK_REQUESTED;
3380 /* Acknowledge the last packet for each call */
3381 if (flags & RX_LAST_PACKET) {
3392 * If the receiver is waiting for an iovec, fill the iovec
3393 * using the data from the receive queue */
3394 if (call->flags & RX_CALL_IOVEC_WAIT) {
3395 didHardAck = rxi_FillReadVec(call, serial);
3396 /* the call may have been aborted */
3405 /* Wakeup the reader if any */
3406 if ((call->flags & RX_CALL_READER_WAIT)
3407 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3408 || (call->iovNext >= call->iovMax)
3409 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3410 call->flags &= ~RX_CALL_READER_WAIT;
3411 #ifdef RX_ENABLE_LOCKS
3412 CV_BROADCAST(&call->cv_rq);
3414 osi_rxWakeup(&call->rq);
3420 * Send an ack when requested by the peer, or once every
3421 * rxi_SoftAckRate packets until the last packet has been
3422 * received. Always send a soft ack for the last packet in
3423 * the server's reply. */
3425 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3426 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3427 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3428 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3429 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3430 } else if (call->nSoftAcks) {
3431 clock_GetTime(&when);
3432 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3433 clock_Add(&when, &rx_lastAckDelay);
3435 clock_Add(&when, &rx_softAckDelay);
3437 if (!call->delayedAckEvent
3438 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3439 rxevent_Cancel(call->delayedAckEvent, call,
3440 RX_CALL_REFCOUNT_DELAY);
3441 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3442 call->delayedAckEvent =
3443 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3445 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3446 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3453 static void rxi_ComputeRate();
3457 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3459 struct rx_peer *peer = conn->peer;
3461 MUTEX_ENTER(&peer->peer_lock);
3462 peer->lastReachTime = clock_Sec();
3463 MUTEX_EXIT(&peer->peer_lock);
3465 MUTEX_ENTER(&conn->conn_data_lock);
3466 if (conn->flags & RX_CONN_ATTACHWAIT) {
3469 conn->flags &= ~RX_CONN_ATTACHWAIT;
3470 MUTEX_EXIT(&conn->conn_data_lock);
3472 for (i = 0; i < RX_MAXCALLS; i++) {
3473 struct rx_call *call = conn->call[i];
3476 MUTEX_ENTER(&call->lock);
3477 /* tnop can be null if newcallp is null */
3478 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3480 MUTEX_EXIT(&call->lock);
3484 MUTEX_EXIT(&conn->conn_data_lock);
3488 rx_ack_reason(int reason)
3491 case RX_ACK_REQUESTED:
3493 case RX_ACK_DUPLICATE:
3495 case RX_ACK_OUT_OF_SEQUENCE:
3497 case RX_ACK_EXCEEDS_WINDOW:
3499 case RX_ACK_NOSPACE:
3503 case RX_ACK_PING_RESPONSE:
3515 /* rxi_ComputePeerNetStats
3517 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3518 * estimates (like RTT and throughput) based on ack packets. Caller
3519 * must ensure that the packet in question is the right one (i.e.
3520 * serial number matches).
3523 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3524 struct rx_ackPacket *ap, struct rx_packet *np)
3526 struct rx_peer *peer = call->conn->peer;
3528 /* Use RTT if not delayed by client. */
3529 if (ap->reason != RX_ACK_DELAY)
3530 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3532 rxi_ComputeRate(peer, call, p, np, ap->reason);
3536 /* The real smarts of the whole thing. */
3538 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3541 struct rx_ackPacket *ap;
3543 register struct rx_packet *tp;
3544 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3545 register struct rx_connection *conn = call->conn;
3546 struct rx_peer *peer = conn->peer;
3549 /* because there are CM's that are bogus, sending weird values for this. */
3550 afs_uint32 skew = 0;
3555 int newAckCount = 0;
3556 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3557 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3559 MUTEX_ENTER(&rx_stats_mutex);
3560 rx_stats.ackPacketsRead++;
3561 MUTEX_EXIT(&rx_stats_mutex);
3562 ap = (struct rx_ackPacket *)rx_DataOf(np);
3563 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3565 return np; /* truncated ack packet */
3567 /* depends on ack packet struct */
3568 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3569 first = ntohl(ap->firstPacket);
3570 serial = ntohl(ap->serial);
3571 /* temporarily disabled -- needs to degrade over time
3572 * skew = ntohs(ap->maxSkew); */
3574 /* Ignore ack packets received out of order */
3575 if (first < call->tfirst) {
3579 if (np->header.flags & RX_SLOW_START_OK) {
3580 call->flags |= RX_CALL_SLOW_START_OK;
3583 if (ap->reason == RX_ACK_PING_RESPONSE)
3584 rxi_UpdatePeerReach(conn, call);
3588 if (rxdebug_active) {
3592 len = _snprintf(msg, sizeof(msg),
3593 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3594 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3595 ntohl(ap->serial), ntohl(ap->previousPacket),
3596 (unsigned int)np->header.seq, (unsigned int)skew,
3597 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3601 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3602 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3606 OutputDebugString(msg);
3608 #else /* AFS_NT40_ENV */
3611 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3612 ap->reason, ntohl(ap->previousPacket),
3613 (unsigned int)np->header.seq, (unsigned int)serial,
3614 (unsigned int)skew, ntohl(ap->firstPacket));
3617 for (offset = 0; offset < nAcks; offset++)
3618 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3623 #endif /* AFS_NT40_ENV */
3626 /* Update the outgoing packet skew value to the latest value of
3627 * the peer's incoming packet skew value. The ack packet, of
3628 * course, could arrive out of order, but that won't affect things
3630 MUTEX_ENTER(&peer->peer_lock);
3631 peer->outPacketSkew = skew;
3633 /* Check for packets that no longer need to be transmitted, and
3634 * discard them. This only applies to packets positively
3635 * acknowledged as having been sent to the peer's upper level.
3636 * All other packets must be retained. So only packets with
3637 * sequence numbers < ap->firstPacket are candidates. */
3638 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3639 if (tp->header.seq >= first)
3641 call->tfirst = tp->header.seq + 1;
3643 && (tp->header.serial == serial || tp->firstSerial == serial))
3644 rxi_ComputePeerNetStats(call, tp, ap, np);
3645 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3648 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3649 /* XXX Hack. Because we have to release the global rx lock when sending
3650 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3651 * in rxi_Start sending packets out because packets may move to the
3652 * freePacketQueue as result of being here! So we drop these packets until
3653 * we're safely out of the traversing. Really ugly!
3654 * To make it even uglier, if we're using fine grain locking, we can
3655 * set the ack bits in the packets and have rxi_Start remove the packets
3656 * when it's done transmitting.
3658 if (call->flags & RX_CALL_TQ_BUSY) {
3659 #ifdef RX_ENABLE_LOCKS
3660 tp->flags |= RX_PKTFLAG_ACKED;
3661 call->flags |= RX_CALL_TQ_SOME_ACKED;
3662 #else /* RX_ENABLE_LOCKS */
3664 #endif /* RX_ENABLE_LOCKS */
3666 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3669 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3674 /* Give rate detector a chance to respond to ping requests */
3675 if (ap->reason == RX_ACK_PING_RESPONSE) {
3676 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3680 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3682 /* Now go through explicit acks/nacks and record the results in
3683 * the waiting packets. These are packets that can't be released
3684 * yet, even with a positive acknowledge. This positive
3685 * acknowledge only means the packet has been received by the
3686 * peer, not that it will be retained long enough to be sent to
3687 * the peer's upper level. In addition, reset the transmit timers
3688 * of any missing packets (those packets that must be missing
3689 * because this packet was out of sequence) */
3691 call->nSoftAcked = 0;
3692 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3693 /* Update round trip time if the ack was stimulated on receipt
3695 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3696 #ifdef RX_ENABLE_LOCKS
3697 if (tp->header.seq >= first)
3698 #endif /* RX_ENABLE_LOCKS */
3699 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3701 && (tp->header.serial == serial || tp->firstSerial == serial))
3702 rxi_ComputePeerNetStats(call, tp, ap, np);
3704 /* Set the acknowledge flag per packet based on the
3705 * information in the ack packet. An acknowlegded packet can
3706 * be downgraded when the server has discarded a packet it
3707 * soacked previously, or when an ack packet is received
3708 * out of sequence. */
3709 if (tp->header.seq < first) {
3710 /* Implicit ack information */
3711 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3714 tp->flags |= RX_PKTFLAG_ACKED;
3715 } else if (tp->header.seq < first + nAcks) {
3716 /* Explicit ack information: set it in the packet appropriately */
3717 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3718 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3720 tp->flags |= RX_PKTFLAG_ACKED;
3727 } else /* RX_ACK_TYPE_NACK */ {
3728 tp->flags &= ~RX_PKTFLAG_ACKED;
3732 tp->flags &= ~RX_PKTFLAG_ACKED;
3736 /* If packet isn't yet acked, and it has been transmitted at least
3737 * once, reset retransmit time using latest timeout
3738 * ie, this should readjust the retransmit timer for all outstanding
3739 * packets... So we don't just retransmit when we should know better*/
3741 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3742 tp->retryTime = tp->timeSent;
3743 clock_Add(&tp->retryTime, &peer->timeout);
3744 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3745 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3749 /* If the window has been extended by this acknowledge packet,
3750 * then wakeup a sender waiting in alloc for window space, or try
3751 * sending packets now, if he's been sitting on packets due to
3752 * lack of window space */
3753 if (call->tnext < (call->tfirst + call->twind)) {
3754 #ifdef RX_ENABLE_LOCKS
3755 CV_SIGNAL(&call->cv_twind);
3757 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3758 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3759 osi_rxWakeup(&call->twind);
3762 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3763 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3767 /* if the ack packet has a receivelen field hanging off it,
3768 * update our state */
3769 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3772 /* If the ack packet has a "recommended" size that is less than
3773 * what I am using now, reduce my size to match */
3774 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3775 (int)sizeof(afs_int32), &tSize);
3776 tSize = (afs_uint32) ntohl(tSize);
3777 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3779 /* Get the maximum packet size to send to this peer */
3780 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
3782 tSize = (afs_uint32) ntohl(tSize);
3783 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3784 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3786 /* sanity check - peer might have restarted with different params.
3787 * If peer says "send less", dammit, send less... Peer should never
3788 * be unable to accept packets of the size that prior AFS versions would
3789 * send without asking. */
3790 if (peer->maxMTU != tSize) {
3791 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
3793 peer->maxMTU = tSize;
3794 peer->MTU = MIN(tSize, peer->MTU);
3795 call->MTU = MIN(call->MTU, tSize);
3798 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3801 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3802 (int)sizeof(afs_int32), &tSize);
3803 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3804 if (tSize < call->twind) { /* smaller than our send */
3805 call->twind = tSize; /* window, we must send less... */
3806 call->ssthresh = MIN(call->twind, call->ssthresh);
3809 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3810 * network MTU confused with the loopback MTU. Calculate the
3811 * maximum MTU here for use in the slow start code below.
3813 maxMTU = peer->maxMTU;
3814 /* Did peer restart with older RX version? */
3815 if (peer->maxDgramPackets > 1) {
3816 peer->maxDgramPackets = 1;
3818 } else if (np->length >=
3819 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3822 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3823 sizeof(afs_int32), &tSize);
3824 tSize = (afs_uint32) ntohl(tSize);
3826 * As of AFS 3.5 we set the send window to match the receive window.
3828 if (tSize < call->twind) {
3829 call->twind = tSize;
3830 call->ssthresh = MIN(call->twind, call->ssthresh);
3831 } else if (tSize > call->twind) {
3832 call->twind = tSize;
3836 * As of AFS 3.5, a jumbogram is more than one fixed size
3837 * packet transmitted in a single UDP datagram. If the remote
3838 * MTU is smaller than our local MTU then never send a datagram
3839 * larger than the natural MTU.
3842 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3843 sizeof(afs_int32), &tSize);
3844 maxDgramPackets = (afs_uint32) ntohl(tSize);
3845 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3846 maxDgramPackets = MIN(maxDgramPackets, peer->ifDgramPackets);
3847 if (peer->natMTU < peer->ifMTU)
3848 maxDgramPackets = MIN(maxDgramPackets, rxi_AdjustDgramPackets(1, peer->natMTU));
3849 if (maxDgramPackets > 1) {
3850 peer->maxDgramPackets = maxDgramPackets;
3851 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3853 peer->maxDgramPackets = 1;
3854 call->MTU = peer->natMTU;
3856 } else if (peer->maxDgramPackets > 1) {
3857 /* Restarted with lower version of RX */
3858 peer->maxDgramPackets = 1;
3860 } else if (peer->maxDgramPackets > 1
3861 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3862 /* Restarted with lower version of RX */
3863 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3864 peer->natMTU = OLD_MAX_PACKET_SIZE;
3865 peer->MTU = OLD_MAX_PACKET_SIZE;
3866 peer->maxDgramPackets = 1;
3867 peer->nDgramPackets = 1;
3869 call->MTU = OLD_MAX_PACKET_SIZE;
3874 * Calculate how many datagrams were successfully received after
3875 * the first missing packet and adjust the negative ack counter
3880 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3881 if (call->nNacks < nNacked) {
3882 call->nNacks = nNacked;
3891 if (call->flags & RX_CALL_FAST_RECOVER) {
3893 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3895 call->flags &= ~RX_CALL_FAST_RECOVER;
3896 call->cwind = call->nextCwind;
3897 call->nextCwind = 0;
3900 call->nCwindAcks = 0;
3901 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3902 /* Three negative acks in a row trigger congestion recovery */
3903 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3904 MUTEX_EXIT(&peer->peer_lock);
3905 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3906 /* someone else is waiting to start recovery */
3909 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3910 rxi_WaitforTQBusy(call);
3911 MUTEX_ENTER(&peer->peer_lock);
3912 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3913 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3914 call->flags |= RX_CALL_FAST_RECOVER;
3915 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3917 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3918 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3919 call->nextCwind = call->ssthresh;
3922 peer->MTU = call->MTU;
3923 peer->cwind = call->nextCwind;
3924 peer->nDgramPackets = call->nDgramPackets;
3926 call->congestSeq = peer->congestSeq;
3927 /* Reset the resend times on the packets that were nacked
3928 * so we will retransmit as soon as the window permits*/
3929 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3931 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3932 clock_Zero(&tp->retryTime);
3934 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3939 /* If cwind is smaller than ssthresh, then increase
3940 * the window one packet for each ack we receive (exponential
3942 * If cwind is greater than or equal to ssthresh then increase
3943 * the congestion window by one packet for each cwind acks we
3944 * receive (linear growth). */
3945 if (call->cwind < call->ssthresh) {
3947 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3948 call->nCwindAcks = 0;
3950 call->nCwindAcks += newAckCount;
3951 if (call->nCwindAcks >= call->cwind) {
3952 call->nCwindAcks = 0;
3953 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3957 * If we have received several acknowledgements in a row then
3958 * it is time to increase the size of our datagrams
3960 if ((int)call->nAcks > rx_nDgramThreshold) {
3961 if (peer->maxDgramPackets > 1) {
3962 if (call->nDgramPackets < peer->maxDgramPackets) {
3963 call->nDgramPackets++;
3965 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3966 } else if (call->MTU < peer->maxMTU) {
3967 call->MTU += peer->natMTU;
3968 call->MTU = MIN(call->MTU, peer->maxMTU);
3974 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3976 /* Servers need to hold the call until all response packets have
3977 * been acknowledged. Soft acks are good enough since clients
3978 * are not allowed to clear their receive queues. */
3979 if (call->state == RX_STATE_HOLD
3980 && call->tfirst + call->nSoftAcked >= call->tnext) {
3981 call->state = RX_STATE_DALLY;
3982 rxi_ClearTransmitQueue(call, 0);
3983 } else if (!queue_IsEmpty(&call->tq)) {
3984 rxi_Start(0, call, 0, istack);
3989 /* Received a response to a challenge packet */
3991 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
3992 register struct rx_packet *np, int istack)
3996 /* Ignore the packet if we're the client */
3997 if (conn->type == RX_CLIENT_CONNECTION)
4000 /* If already authenticated, ignore the packet (it's probably a retry) */
4001 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4004 /* Otherwise, have the security object evaluate the response packet */
4005 error = RXS_CheckResponse(conn->securityObject, conn, np);
4007 /* If the response is invalid, reset the connection, sending
4008 * an abort to the peer */
4012 rxi_ConnectionError(conn, error);
4013 MUTEX_ENTER(&conn->conn_data_lock);
4014 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4015 MUTEX_EXIT(&conn->conn_data_lock);
4018 /* If the response is valid, any calls waiting to attach
4019 * servers can now do so */
4022 for (i = 0; i < RX_MAXCALLS; i++) {
4023 struct rx_call *call = conn->call[i];
4025 MUTEX_ENTER(&call->lock);
4026 if (call->state == RX_STATE_PRECALL)
4027 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4028 /* tnop can be null if newcallp is null */
4029 MUTEX_EXIT(&call->lock);
4033 /* Update the peer reachability information, just in case
4034 * some calls went into attach-wait while we were waiting
4035 * for authentication..
4037 rxi_UpdatePeerReach(conn, NULL);
4042 /* A client has received an authentication challenge: the security
4043 * object is asked to cough up a respectable response packet to send
4044 * back to the server. The server is responsible for retrying the
4045 * challenge if it fails to get a response. */
4048 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
4049 register struct rx_packet *np, int istack)
4053 /* Ignore the challenge if we're the server */
4054 if (conn->type == RX_SERVER_CONNECTION)
4057 /* Ignore the challenge if the connection is otherwise idle; someone's
4058 * trying to use us as an oracle. */
4059 if (!rxi_HasActiveCalls(conn))
4062 /* Send the security object the challenge packet. It is expected to fill
4063 * in the response. */
4064 error = RXS_GetResponse(conn->securityObject, conn, np);
4066 /* If the security object is unable to return a valid response, reset the
4067 * connection and send an abort to the peer. Otherwise send the response
4068 * packet to the peer connection. */
4070 rxi_ConnectionError(conn, error);
4071 MUTEX_ENTER(&conn->conn_data_lock);
4072 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4073 MUTEX_EXIT(&conn->conn_data_lock);
4075 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4076 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4082 /* Find an available server process to service the current request in
4083 * the given call structure. If one isn't available, queue up this
4084 * call so it eventually gets one */
4086 rxi_AttachServerProc(register struct rx_call *call,
4087 register osi_socket socket, register int *tnop,
4088 register struct rx_call **newcallp)
4090 register struct rx_serverQueueEntry *sq;
4091 register struct rx_service *service = call->conn->service;
4092 register int haveQuota = 0;
4094 /* May already be attached */
4095 if (call->state == RX_STATE_ACTIVE)
4098 MUTEX_ENTER(&rx_serverPool_lock);
4100 haveQuota = QuotaOK(service);
4101 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4102 /* If there are no processes available to service this call,
4103 * put the call on the incoming call queue (unless it's
4104 * already on the queue).
4106 #ifdef RX_ENABLE_LOCKS
4108 ReturnToServerPool(service);
4109 #endif /* RX_ENABLE_LOCKS */
4111 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4112 call->flags |= RX_CALL_WAIT_PROC;
4113 MUTEX_ENTER(&rx_stats_mutex);
4116 MUTEX_EXIT(&rx_stats_mutex);
4117 rxi_calltrace(RX_CALL_ARRIVAL, call);
4118 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4119 queue_Append(&rx_incomingCallQueue, call);
4122 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4124 /* If hot threads are enabled, and both newcallp and sq->socketp
4125 * are non-null, then this thread will process the call, and the
4126 * idle server thread will start listening on this threads socket.
4129 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4132 *sq->socketp = socket;
4133 clock_GetTime(&call->startTime);
4134 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4138 if (call->flags & RX_CALL_WAIT_PROC) {
4139 /* Conservative: I don't think this should happen */
4140 call->flags &= ~RX_CALL_WAIT_PROC;
4141 if (queue_IsOnQueue(call)) {
4143 MUTEX_ENTER(&rx_stats_mutex);
4145 MUTEX_EXIT(&rx_stats_mutex);
4148 call->state = RX_STATE_ACTIVE;
4149 call->mode = RX_MODE_RECEIVING;
4150 #ifdef RX_KERNEL_TRACE
4152 int glockOwner = ISAFS_GLOCK();
4155 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4156 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4162 if (call->flags & RX_CALL_CLEARED) {
4163 /* send an ack now to start the packet flow up again */
4164 call->flags &= ~RX_CALL_CLEARED;
4165 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4167 #ifdef RX_ENABLE_LOCKS
4170 service->nRequestsRunning++;
4171 if (service->nRequestsRunning <= service->minProcs)
4177 MUTEX_EXIT(&rx_serverPool_lock);
4180 /* Delay the sending of an acknowledge event for a short while, while
4181 * a new call is being prepared (in the case of a client) or a reply
4182 * is being prepared (in the case of a server). Rather than sending
4183 * an ack packet, an ACKALL packet is sent. */
4185 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4187 #ifdef RX_ENABLE_LOCKS
4189 MUTEX_ENTER(&call->lock);
4190 call->delayedAckEvent = NULL;
4191 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4193 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4194 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4196 MUTEX_EXIT(&call->lock);
4197 #else /* RX_ENABLE_LOCKS */
4199 call->delayedAckEvent = NULL;
4200 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4201 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4202 #endif /* RX_ENABLE_LOCKS */
4206 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4209 #ifdef RX_ENABLE_LOCKS
4211 MUTEX_ENTER(&call->lock);
4212 if (event == call->delayedAckEvent)
4213 call->delayedAckEvent = NULL;
4214 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4216 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4218 MUTEX_EXIT(&call->lock);
4219 #else /* RX_ENABLE_LOCKS */
4221 call->delayedAckEvent = NULL;
4222 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4223 #endif /* RX_ENABLE_LOCKS */
4227 #ifdef RX_ENABLE_LOCKS
4228 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4229 * clearing them out.
4232 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4234 register struct rx_packet *p, *tp;
4237 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4238 p->flags |= RX_PKTFLAG_ACKED;
4242 call->flags |= RX_CALL_TQ_CLEARME;
4243 call->flags |= RX_CALL_TQ_SOME_ACKED;
4246 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4247 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4248 call->tfirst = call->tnext;
4249 call->nSoftAcked = 0;
4251 if (call->flags & RX_CALL_FAST_RECOVER) {
4252 call->flags &= ~RX_CALL_FAST_RECOVER;
4253 call->cwind = call->nextCwind;
4254 call->nextCwind = 0;
4257 CV_SIGNAL(&call->cv_twind);
4259 #endif /* RX_ENABLE_LOCKS */
4261 /* Clear out the transmit queue for the current call (all packets have
4262 * been received by peer) */
4264 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4266 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4267 register struct rx_packet *p, *tp;
4269 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4271 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4272 p->flags |= RX_PKTFLAG_ACKED;
4276 call->flags |= RX_CALL_TQ_CLEARME;
4277 call->flags |= RX_CALL_TQ_SOME_ACKED;
4280 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4281 rxi_FreePackets(0, &call->tq);
4282 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4283 call->flags &= ~RX_CALL_TQ_CLEARME;
4285 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4287 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4288 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4289 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4290 call->nSoftAcked = 0;
4292 if (call->flags & RX_CALL_FAST_RECOVER) {
4293 call->flags &= ~RX_CALL_FAST_RECOVER;
4294 call->cwind = call->nextCwind;
4296 #ifdef RX_ENABLE_LOCKS
4297 CV_SIGNAL(&call->cv_twind);
4299 osi_rxWakeup(&call->twind);
4304 rxi_ClearReceiveQueue(register struct rx_call *call)
4306 if (queue_IsNotEmpty(&call->rq)) {
4307 rx_packetReclaims += rxi_FreePackets(0, &call->rq);
4308 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4310 if (call->state == RX_STATE_PRECALL) {
4311 call->flags |= RX_CALL_CLEARED;
4315 /* Send an abort packet for the specified call */
4317 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4318 int istack, int force)
4326 /* Clients should never delay abort messages */
4327 if (rx_IsClientConn(call->conn))
4330 if (call->abortCode != call->error) {
4331 call->abortCode = call->error;
4332 call->abortCount = 0;
4335 if (force || rxi_callAbortThreshhold == 0
4336 || call->abortCount < rxi_callAbortThreshhold) {
4337 if (call->delayedAbortEvent) {
4338 rxevent_Cancel(call->delayedAbortEvent, call,
4339 RX_CALL_REFCOUNT_ABORT);
4341 error = htonl(call->error);
4344 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4345 (char *)&error, sizeof(error), istack);
4346 } else if (!call->delayedAbortEvent) {
4347 clock_GetTime(&when);
4348 clock_Addmsec(&when, rxi_callAbortDelay);
4349 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4350 call->delayedAbortEvent =
4351 rxevent_Post(&when, rxi_SendDelayedCallAbort, call, 0);
4356 /* Send an abort packet for the specified connection. Packet is an
4357 * optional pointer to a packet that can be used to send the abort.
4358 * Once the number of abort messages reaches the threshhold, an
4359 * event is scheduled to send the abort. Setting the force flag
4360 * overrides sending delayed abort messages.
4362 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4363 * to send the abort packet.
4366 rxi_SendConnectionAbort(register struct rx_connection *conn,
4367 struct rx_packet *packet, int istack, int force)
4375 /* Clients should never delay abort messages */
4376 if (rx_IsClientConn(conn))
4379 if (force || rxi_connAbortThreshhold == 0
4380 || conn->abortCount < rxi_connAbortThreshhold) {
4381 if (conn->delayedAbortEvent) {
4382 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4384 error = htonl(conn->error);
4386 MUTEX_EXIT(&conn->conn_data_lock);
4388 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4389 RX_PACKET_TYPE_ABORT, (char *)&error,
4390 sizeof(error), istack);
4391 MUTEX_ENTER(&conn->conn_data_lock);
4392 } else if (!conn->delayedAbortEvent) {
4393 clock_GetTime(&when);
4394 clock_Addmsec(&when, rxi_connAbortDelay);
4395 conn->delayedAbortEvent =
4396 rxevent_Post(&when, rxi_SendDelayedConnAbort, conn, 0);
4401 /* Associate an error all of the calls owned by a connection. Called
4402 * with error non-zero. This is only for really fatal things, like
4403 * bad authentication responses. The connection itself is set in
4404 * error at this point, so that future packets received will be
4407 rxi_ConnectionError(register struct rx_connection *conn,
4408 register afs_int32 error)
4413 dpf(("rxi_ConnectionError conn %x error %d", conn, error));
4415 MUTEX_ENTER(&conn->conn_data_lock);
4416 if (conn->challengeEvent)
4417 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4418 if (conn->checkReachEvent) {
4419 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4420 conn->checkReachEvent = 0;
4421 conn->flags &= ~RX_CONN_ATTACHWAIT;
4424 MUTEX_EXIT(&conn->conn_data_lock);
4425 for (i = 0; i < RX_MAXCALLS; i++) {
4426 struct rx_call *call = conn->call[i];
4428 MUTEX_ENTER(&call->lock);
4429 rxi_CallError(call, error);
4430 MUTEX_EXIT(&call->lock);
4433 conn->error = error;
4434 MUTEX_ENTER(&rx_stats_mutex);
4435 rx_stats.fatalErrors++;
4436 MUTEX_EXIT(&rx_stats_mutex);
4441 rxi_CallError(register struct rx_call *call, afs_int32 error)
4443 dpf(("rxi_CallError call %x error %d call->error %d", call, error, call->error));
4445 error = call->error;
4447 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4448 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4449 rxi_ResetCall(call, 0);
4452 rxi_ResetCall(call, 0);
4454 call->error = error;
4455 call->mode = RX_MODE_ERROR;
4458 /* Reset various fields in a call structure, and wakeup waiting
4459 * processes. Some fields aren't changed: state & mode are not
4460 * touched (these must be set by the caller), and bufptr, nLeft, and
4461 * nFree are not reset, since these fields are manipulated by
4462 * unprotected macros, and may only be reset by non-interrupting code.
4465 /* this code requires that call->conn be set properly as a pre-condition. */
4466 #endif /* ADAPT_WINDOW */
4469 rxi_ResetCall(register struct rx_call *call, register int newcall)
4472 register struct rx_peer *peer;
4473 struct rx_packet *packet;
4475 dpf(("rxi_ResetCall(call %x, newcall %d)\n", call, newcall));
4477 /* Notify anyone who is waiting for asynchronous packet arrival */
4478 if (call->arrivalProc) {
4479 (*call->arrivalProc) (call, call->arrivalProcHandle,
4480 call->arrivalProcArg);
4481 call->arrivalProc = (void (*)())0;
4484 if (call->delayedAbortEvent) {
4485 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4486 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4488 rxi_SendCallAbort(call, packet, 0, 1);
4489 rxi_FreePacket(packet);
4494 * Update the peer with the congestion information in this call
4495 * so other calls on this connection can pick up where this call
4496 * left off. If the congestion sequence numbers don't match then
4497 * another call experienced a retransmission.
4499 peer = call->conn->peer;
4500 MUTEX_ENTER(&peer->peer_lock);
4502 if (call->congestSeq == peer->congestSeq) {
4503 peer->cwind = MAX(peer->cwind, call->cwind);
4504 peer->MTU = MAX(peer->MTU, call->MTU);
4505 peer->nDgramPackets =
4506 MAX(peer->nDgramPackets, call->nDgramPackets);
4509 call->abortCode = 0;
4510 call->abortCount = 0;
4512 if (peer->maxDgramPackets > 1) {
4513 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4515 call->MTU = peer->MTU;
4517 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4518 call->ssthresh = rx_maxSendWindow;
4519 call->nDgramPackets = peer->nDgramPackets;
4520 call->congestSeq = peer->congestSeq;
4521 MUTEX_EXIT(&peer->peer_lock);
4523 flags = call->flags;
4524 rxi_ClearReceiveQueue(call);
4525 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4526 if (flags & RX_CALL_TQ_BUSY) {
4527 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4528 call->flags |= (flags & RX_CALL_TQ_WAIT);
4530 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4532 rxi_ClearTransmitQueue(call, 0);
4533 queue_Init(&call->tq);
4534 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4535 dpf(("rcall %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4538 while (call->tqWaiters) {
4539 #ifdef RX_ENABLE_LOCKS
4540 CV_BROADCAST(&call->cv_tq);
4541 #else /* RX_ENABLE_LOCKS */
4542 osi_rxWakeup(&call->tq);
4543 #endif /* RX_ENABLE_LOCKS */
4547 queue_Init(&call->rq);
4549 call->rwind = rx_initReceiveWindow;
4550 call->twind = rx_initSendWindow;
4551 call->nSoftAcked = 0;
4552 call->nextCwind = 0;
4555 call->nCwindAcks = 0;
4556 call->nSoftAcks = 0;
4557 call->nHardAcks = 0;
4559 call->tfirst = call->rnext = call->tnext = 1;
4561 call->lastAcked = 0;
4562 call->localStatus = call->remoteStatus = 0;
4564 if (flags & RX_CALL_READER_WAIT) {
4565 #ifdef RX_ENABLE_LOCKS
4566 CV_BROADCAST(&call->cv_rq);
4568 osi_rxWakeup(&call->rq);
4571 if (flags & RX_CALL_WAIT_PACKETS) {
4572 MUTEX_ENTER(&rx_freePktQ_lock);
4573 rxi_PacketsUnWait(); /* XXX */
4574 MUTEX_EXIT(&rx_freePktQ_lock);
4576 #ifdef RX_ENABLE_LOCKS
4577 CV_SIGNAL(&call->cv_twind);
4579 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4580 osi_rxWakeup(&call->twind);
4583 #ifdef RX_ENABLE_LOCKS
4584 /* The following ensures that we don't mess with any queue while some
4585 * other thread might also be doing so. The call_queue_lock field is
4586 * is only modified under the call lock. If the call is in the process
4587 * of being removed from a queue, the call is not locked until the
4588 * the queue lock is dropped and only then is the call_queue_lock field
4589 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4590 * Note that any other routine which removes a call from a queue has to
4591 * obtain the queue lock before examing the queue and removing the call.
4593 if (call->call_queue_lock) {
4594 MUTEX_ENTER(call->call_queue_lock);
4595 if (queue_IsOnQueue(call)) {
4597 if (flags & RX_CALL_WAIT_PROC) {
4598 MUTEX_ENTER(&rx_stats_mutex);
4600 MUTEX_EXIT(&rx_stats_mutex);
4603 MUTEX_EXIT(call->call_queue_lock);
4604 CLEAR_CALL_QUEUE_LOCK(call);
4606 #else /* RX_ENABLE_LOCKS */
4607 if (queue_IsOnQueue(call)) {
4609 if (flags & RX_CALL_WAIT_PROC)
4612 #endif /* RX_ENABLE_LOCKS */
4614 rxi_KeepAliveOff(call);
4615 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4618 /* Send an acknowledge for the indicated packet (seq,serial) of the
4619 * indicated call, for the indicated reason (reason). This
4620 * acknowledge will specifically acknowledge receiving the packet, and
4621 * will also specify which other packets for this call have been
4622 * received. This routine returns the packet that was used to the
4623 * caller. The caller is responsible for freeing it or re-using it.
4624 * This acknowledgement also returns the highest sequence number
4625 * actually read out by the higher level to the sender; the sender
4626 * promises to keep around packets that have not been read by the
4627 * higher level yet (unless, of course, the sender decides to abort
4628 * the call altogether). Any of p, seq, serial, pflags, or reason may
4629 * be set to zero without ill effect. That is, if they are zero, they
4630 * will not convey any information.
4631 * NOW there is a trailer field, after the ack where it will safely be
4632 * ignored by mundanes, which indicates the maximum size packet this
4633 * host can swallow. */
4635 register struct rx_packet *optionalPacket; use to send ack (or null)
4636 int seq; Sequence number of the packet we are acking
4637 int serial; Serial number of the packet
4638 int pflags; Flags field from packet header
4639 int reason; Reason an acknowledge was prompted
4643 rxi_SendAck(register struct rx_call *call,
4644 register struct rx_packet *optionalPacket, int serial, int reason,
4647 struct rx_ackPacket *ap;
4648 register struct rx_packet *rqp;
4649 register struct rx_packet *nxp; /* For queue_Scan */
4650 register struct rx_packet *p;
4653 #ifdef RX_ENABLE_TSFPQ
4654 struct rx_ts_info_t * rx_ts_info;
4658 * Open the receive window once a thread starts reading packets
4660 if (call->rnext > 1) {
4661 call->rwind = rx_maxReceiveWindow;
4664 call->nHardAcks = 0;
4665 call->nSoftAcks = 0;
4666 if (call->rnext > call->lastAcked)
4667 call->lastAcked = call->rnext;
4671 rx_computelen(p, p->length); /* reset length, you never know */
4672 } /* where that's been... */
4673 #ifdef RX_ENABLE_TSFPQ
4675 RX_TS_INFO_GET(rx_ts_info);
4676 if ((p = rx_ts_info->local_special_packet)) {
4677 rx_computelen(p, p->length);
4678 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4679 rx_ts_info->local_special_packet = p;
4680 } else { /* We won't send the ack, but don't panic. */
4681 return optionalPacket;
4685 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4686 /* We won't send the ack, but don't panic. */
4687 return optionalPacket;
4692 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4695 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4696 #ifndef RX_ENABLE_TSFPQ
4697 if (!optionalPacket)
4700 return optionalPacket;
4702 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4703 if (rx_Contiguous(p) < templ) {
4704 #ifndef RX_ENABLE_TSFPQ
4705 if (!optionalPacket)
4708 return optionalPacket;
4713 /* MTUXXX failing to send an ack is very serious. We should */
4714 /* try as hard as possible to send even a partial ack; it's */
4715 /* better than nothing. */
4716 ap = (struct rx_ackPacket *)rx_DataOf(p);
4717 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4718 ap->reason = reason;
4720 /* The skew computation used to be bogus, I think it's better now. */
4721 /* We should start paying attention to skew. XXX */
4722 ap->serial = htonl(serial);
4723 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4725 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4726 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4728 /* No fear of running out of ack packet here because there can only be at most
4729 * one window full of unacknowledged packets. The window size must be constrained
4730 * to be less than the maximum ack size, of course. Also, an ack should always
4731 * fit into a single packet -- it should not ever be fragmented. */
4732 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4733 if (!rqp || !call->rq.next
4734 || (rqp->header.seq > (call->rnext + call->rwind))) {
4735 #ifndef RX_ENABLE_TSFPQ
4736 if (!optionalPacket)
4739 rxi_CallError(call, RX_CALL_DEAD);
4740 return optionalPacket;
4743 while (rqp->header.seq > call->rnext + offset)
4744 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4745 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4747 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4748 #ifndef RX_ENABLE_TSFPQ
4749 if (!optionalPacket)
4752 rxi_CallError(call, RX_CALL_DEAD);
4753 return optionalPacket;
4758 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4760 /* these are new for AFS 3.3 */
4761 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4762 templ = htonl(templ);
4763 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4764 templ = htonl(call->conn->peer->ifMTU);
4765 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4766 sizeof(afs_int32), &templ);
4768 /* new for AFS 3.4 */
4769 templ = htonl(call->rwind);
4770 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4771 sizeof(afs_int32), &templ);
4773 /* new for AFS 3.5 */
4774 templ = htonl(call->conn->peer->ifDgramPackets);
4775 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4776 sizeof(afs_int32), &templ);
4778 p->header.serviceId = call->conn->serviceId;
4779 p->header.cid = (call->conn->cid | call->channel);
4780 p->header.callNumber = *call->callNumber;
4782 p->header.securityIndex = call->conn->securityIndex;
4783 p->header.epoch = call->conn->epoch;
4784 p->header.type = RX_PACKET_TYPE_ACK;
4785 p->header.flags = RX_SLOW_START_OK;
4786 if (reason == RX_ACK_PING) {
4787 p->header.flags |= RX_REQUEST_ACK;
4789 clock_GetTime(&call->pingRequestTime);
4792 if (call->conn->type == RX_CLIENT_CONNECTION)
4793 p->header.flags |= RX_CLIENT_INITIATED;
4797 if (rxdebug_active) {
4801 len = _snprintf(msg, sizeof(msg),
4802 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4803 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4804 ntohl(ap->serial), ntohl(ap->previousPacket),
4805 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
4806 ap->nAcks, ntohs(ap->bufferSpace) );
4810 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
4811 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4815 OutputDebugString(msg);
4817 #else /* AFS_NT40_ENV */
4819 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
4820 ap->reason, ntohl(ap->previousPacket),
4821 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4823 for (offset = 0; offset < ap->nAcks; offset++)
4824 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4829 #endif /* AFS_NT40_ENV */
4832 register int i, nbytes = p->length;
4834 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4835 if (nbytes <= p->wirevec[i].iov_len) {
4836 register int savelen, saven;
4838 savelen = p->wirevec[i].iov_len;
4840 p->wirevec[i].iov_len = nbytes;
4842 rxi_Send(call, p, istack);
4843 p->wirevec[i].iov_len = savelen;
4847 nbytes -= p->wirevec[i].iov_len;
4850 MUTEX_ENTER(&rx_stats_mutex);
4851 rx_stats.ackPacketsSent++;
4852 MUTEX_EXIT(&rx_stats_mutex);
4853 #ifndef RX_ENABLE_TSFPQ
4854 if (!optionalPacket)
4857 return optionalPacket; /* Return packet for re-use by caller */
4860 /* Send all of the packets in the list in single datagram */
4862 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4863 int istack, int moreFlag, struct clock *now,
4864 struct clock *retryTime, int resending)
4869 struct rx_connection *conn = call->conn;
4870 struct rx_peer *peer = conn->peer;
4872 MUTEX_ENTER(&peer->peer_lock);
4875 peer->reSends += len;
4876 MUTEX_ENTER(&rx_stats_mutex);
4877 rx_stats.dataPacketsSent += len;
4878 MUTEX_EXIT(&rx_stats_mutex);
4879 MUTEX_EXIT(&peer->peer_lock);
4881 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4885 /* Set the packet flags and schedule the resend events */
4886 /* Only request an ack for the last packet in the list */
4887 for (i = 0; i < len; i++) {
4888 list[i]->retryTime = *retryTime;
4889 if (list[i]->header.serial) {
4890 /* Exponentially backoff retry times */
4891 if (list[i]->backoff < MAXBACKOFF) {
4892 /* so it can't stay == 0 */
4893 list[i]->backoff = (list[i]->backoff << 1) + 1;
4896 clock_Addmsec(&(list[i]->retryTime),
4897 ((afs_uint32) list[i]->backoff) << 8);
4900 /* Wait a little extra for the ack on the last packet */
4901 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4902 clock_Addmsec(&(list[i]->retryTime), 400);
4905 /* Record the time sent */
4906 list[i]->timeSent = *now;
4908 /* Ask for an ack on retransmitted packets, on every other packet
4909 * if the peer doesn't support slow start. Ask for an ack on every
4910 * packet until the congestion window reaches the ack rate. */
4911 if (list[i]->header.serial) {
4913 MUTEX_ENTER(&rx_stats_mutex);
4914 rx_stats.dataPacketsReSent++;
4915 MUTEX_EXIT(&rx_stats_mutex);
4917 /* improved RTO calculation- not Karn */
4918 list[i]->firstSent = *now;
4919 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4920 || (!(call->flags & RX_CALL_SLOW_START_OK)
4921 && (list[i]->header.seq & 1)))) {
4926 MUTEX_ENTER(&peer->peer_lock);
4930 MUTEX_ENTER(&rx_stats_mutex);
4931 rx_stats.dataPacketsSent++;
4932 MUTEX_EXIT(&rx_stats_mutex);
4933 MUTEX_EXIT(&peer->peer_lock);
4935 /* Tag this packet as not being the last in this group,
4936 * for the receiver's benefit */
4937 if (i < len - 1 || moreFlag) {
4938 list[i]->header.flags |= RX_MORE_PACKETS;
4941 /* Install the new retransmit time for the packet, and
4942 * record the time sent */
4943 list[i]->timeSent = *now;
4947 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4950 /* Since we're about to send a data packet to the peer, it's
4951 * safe to nuke any scheduled end-of-packets ack */
4952 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4954 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4955 MUTEX_EXIT(&call->lock);
4957 rxi_SendPacketList(call, conn, list, len, istack);
4959 rxi_SendPacket(call, conn, list[0], istack);
4961 MUTEX_ENTER(&call->lock);
4962 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4964 /* Update last send time for this call (for keep-alive
4965 * processing), and for the connection (so that we can discover
4966 * idle connections) */
4967 conn->lastSendTime = call->lastSendTime = clock_Sec();
4970 /* When sending packets we need to follow these rules:
4971 * 1. Never send more than maxDgramPackets in a jumbogram.
4972 * 2. Never send a packet with more than two iovecs in a jumbogram.
4973 * 3. Never send a retransmitted packet in a jumbogram.
4974 * 4. Never send more than cwind/4 packets in a jumbogram
4975 * We always keep the last list we should have sent so we
4976 * can set the RX_MORE_PACKETS flags correctly.
4979 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4980 int istack, struct clock *now, struct clock *retryTime,
4983 int i, cnt, lastCnt = 0;
4984 struct rx_packet **listP, **lastP = 0;
4985 struct rx_peer *peer = call->conn->peer;
4986 int morePackets = 0;
4988 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4989 /* Does the current packet force us to flush the current list? */
4991 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4992 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4994 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
4996 /* If the call enters an error state stop sending, or if
4997 * we entered congestion recovery mode, stop sending */
4998 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5006 /* Add the current packet to the list if it hasn't been acked.
5007 * Otherwise adjust the list pointer to skip the current packet. */
5008 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5010 /* Do we need to flush the list? */
5011 if (cnt >= (int)peer->maxDgramPackets
5012 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5013 || list[i]->header.serial
5014 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5016 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5017 retryTime, resending);
5018 /* If the call enters an error state stop sending, or if
5019 * we entered congestion recovery mode, stop sending */
5021 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5026 listP = &list[i + 1];
5031 osi_Panic("rxi_SendList error");
5033 listP = &list[i + 1];
5037 /* Send the whole list when the call is in receive mode, when
5038 * the call is in eof mode, when we are in fast recovery mode,
5039 * and when we have the last packet */
5040 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5041 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5042 || (call->flags & RX_CALL_FAST_RECOVER)) {
5043 /* Check for the case where the current list contains
5044 * an acked packet. Since we always send retransmissions
5045 * in a separate packet, we only need to check the first
5046 * packet in the list */
5047 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5051 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5052 retryTime, resending);
5053 /* If the call enters an error state stop sending, or if
5054 * we entered congestion recovery mode, stop sending */
5055 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5059 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5062 } else if (lastCnt > 0) {
5063 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5068 #ifdef RX_ENABLE_LOCKS
5069 /* Call rxi_Start, below, but with the call lock held. */
5071 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
5072 void *arg1, int istack)
5074 MUTEX_ENTER(&call->lock);
5075 rxi_Start(event, call, arg1, istack);
5076 MUTEX_EXIT(&call->lock);
5078 #endif /* RX_ENABLE_LOCKS */
5080 /* This routine is called when new packets are readied for
5081 * transmission and when retransmission may be necessary, or when the
5082 * transmission window or burst count are favourable. This should be
5083 * better optimized for new packets, the usual case, now that we've
5084 * got rid of queues of send packets. XXXXXXXXXXX */
5086 rxi_Start(struct rxevent *event, register struct rx_call *call,
5087 void *arg1, int istack)
5089 struct rx_packet *p;
5090 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
5091 struct rx_peer *peer = call->conn->peer;
5092 struct clock now, retryTime;
5096 struct rx_packet **xmitList;
5099 /* If rxi_Start is being called as a result of a resend event,
5100 * then make sure that the event pointer is removed from the call
5101 * structure, since there is no longer a per-call retransmission
5103 if (event && event == call->resendEvent) {
5104 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5105 call->resendEvent = NULL;
5107 if (queue_IsEmpty(&call->tq)) {
5111 /* Timeouts trigger congestion recovery */
5112 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5113 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5114 /* someone else is waiting to start recovery */
5117 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5118 rxi_WaitforTQBusy(call);
5119 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5120 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5121 call->flags |= RX_CALL_FAST_RECOVER;
5122 if (peer->maxDgramPackets > 1) {
5123 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5125 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5127 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5128 call->nDgramPackets = 1;
5130 call->nextCwind = 1;
5133 MUTEX_ENTER(&peer->peer_lock);
5134 peer->MTU = call->MTU;
5135 peer->cwind = call->cwind;
5136 peer->nDgramPackets = 1;
5138 call->congestSeq = peer->congestSeq;
5139 MUTEX_EXIT(&peer->peer_lock);
5140 /* Clear retry times on packets. Otherwise, it's possible for
5141 * some packets in the queue to force resends at rates faster
5142 * than recovery rates.
5144 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5145 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5146 clock_Zero(&p->retryTime);
5151 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5152 MUTEX_ENTER(&rx_stats_mutex);
5153 rx_tq_debug.rxi_start_in_error++;
5154 MUTEX_EXIT(&rx_stats_mutex);
5159 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5160 /* Get clock to compute the re-transmit time for any packets
5161 * in this burst. Note, if we back off, it's reasonable to
5162 * back off all of the packets in the same manner, even if
5163 * some of them have been retransmitted more times than more
5164 * recent additions */
5165 clock_GetTime(&now);
5166 retryTime = now; /* initialize before use */
5167 MUTEX_ENTER(&peer->peer_lock);
5168 clock_Add(&retryTime, &peer->timeout);
5169 MUTEX_EXIT(&peer->peer_lock);
5171 /* Send (or resend) any packets that need it, subject to
5172 * window restrictions and congestion burst control
5173 * restrictions. Ask for an ack on the last packet sent in
5174 * this burst. For now, we're relying upon the window being
5175 * considerably bigger than the largest number of packets that
5176 * are typically sent at once by one initial call to
5177 * rxi_Start. This is probably bogus (perhaps we should ask
5178 * for an ack when we're half way through the current
5179 * window?). Also, for non file transfer applications, this
5180 * may end up asking for an ack for every packet. Bogus. XXXX
5183 * But check whether we're here recursively, and let the other guy
5186 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5187 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5188 call->flags |= RX_CALL_TQ_BUSY;
5190 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5192 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5193 call->flags &= ~RX_CALL_NEED_START;
5194 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5196 maxXmitPackets = MIN(call->twind, call->cwind);
5197 xmitList = (struct rx_packet **)
5198 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5199 if (xmitList == NULL)
5200 osi_Panic("rxi_Start, failed to allocate xmit list");
5201 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5202 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5203 /* We shouldn't be sending packets if a thread is waiting
5204 * to initiate congestion recovery */
5208 && (call->flags & RX_CALL_FAST_RECOVER)) {
5209 /* Only send one packet during fast recovery */
5212 if ((p->flags & RX_PKTFLAG_FREE)
5213 || (!queue_IsEnd(&call->tq, nxp)
5214 && (nxp->flags & RX_PKTFLAG_FREE))
5215 || (p == (struct rx_packet *)&rx_freePacketQueue)
5216 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5217 osi_Panic("rxi_Start: xmit queue clobbered");
5219 if (p->flags & RX_PKTFLAG_ACKED) {
5220 MUTEX_ENTER(&rx_stats_mutex);
5221 rx_stats.ignoreAckedPacket++;
5222 MUTEX_EXIT(&rx_stats_mutex);
5223 continue; /* Ignore this packet if it has been acknowledged */
5226 /* Turn off all flags except these ones, which are the same
5227 * on each transmission */
5228 p->header.flags &= RX_PRESET_FLAGS;
5230 if (p->header.seq >=
5231 call->tfirst + MIN((int)call->twind,
5232 (int)(call->nSoftAcked +
5234 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5235 /* Note: if we're waiting for more window space, we can
5236 * still send retransmits; hence we don't return here, but
5237 * break out to schedule a retransmit event */
5238 dpf(("call %d waiting for window",
5239 *(call->callNumber)));
5243 /* Transmit the packet if it needs to be sent. */
5244 if (!clock_Lt(&now, &p->retryTime)) {
5245 if (nXmitPackets == maxXmitPackets) {
5246 rxi_SendXmitList(call, xmitList, nXmitPackets,
5247 istack, &now, &retryTime,
5249 osi_Free(xmitList, maxXmitPackets *
5250 sizeof(struct rx_packet *));
5253 xmitList[nXmitPackets++] = p;
5257 /* xmitList now hold pointers to all of the packets that are
5258 * ready to send. Now we loop to send the packets */
5259 if (nXmitPackets > 0) {
5260 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5261 &now, &retryTime, resending);
5264 maxXmitPackets * sizeof(struct rx_packet *));
5266 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5268 * TQ references no longer protected by this flag; they must remain
5269 * protected by the global lock.
5271 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5272 call->flags &= ~RX_CALL_TQ_BUSY;
5273 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5274 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5275 #ifdef RX_ENABLE_LOCKS
5276 osirx_AssertMine(&call->lock, "rxi_Start start");
5277 CV_BROADCAST(&call->cv_tq);
5278 #else /* RX_ENABLE_LOCKS */
5279 osi_rxWakeup(&call->tq);
5280 #endif /* RX_ENABLE_LOCKS */
5285 /* We went into the error state while sending packets. Now is
5286 * the time to reset the call. This will also inform the using
5287 * process that the call is in an error state.
5289 MUTEX_ENTER(&rx_stats_mutex);
5290 rx_tq_debug.rxi_start_aborted++;
5291 MUTEX_EXIT(&rx_stats_mutex);
5292 call->flags &= ~RX_CALL_TQ_BUSY;
5293 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5294 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5295 #ifdef RX_ENABLE_LOCKS
5296 osirx_AssertMine(&call->lock, "rxi_Start middle");
5297 CV_BROADCAST(&call->cv_tq);
5298 #else /* RX_ENABLE_LOCKS */
5299 osi_rxWakeup(&call->tq);
5300 #endif /* RX_ENABLE_LOCKS */
5302 rxi_CallError(call, call->error);
5305 #ifdef RX_ENABLE_LOCKS
5306 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5307 register int missing;
5308 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5309 /* Some packets have received acks. If they all have, we can clear
5310 * the transmit queue.
5313 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5314 if (p->header.seq < call->tfirst
5315 && (p->flags & RX_PKTFLAG_ACKED)) {
5322 call->flags |= RX_CALL_TQ_CLEARME;
5324 #endif /* RX_ENABLE_LOCKS */
5325 /* Don't bother doing retransmits if the TQ is cleared. */
5326 if (call->flags & RX_CALL_TQ_CLEARME) {
5327 rxi_ClearTransmitQueue(call, 1);
5329 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5332 /* Always post a resend event, if there is anything in the
5333 * queue, and resend is possible. There should be at least
5334 * one unacknowledged packet in the queue ... otherwise none
5335 * of these packets should be on the queue in the first place.
5337 if (call->resendEvent) {
5338 /* Cancel the existing event and post a new one */
5339 rxevent_Cancel(call->resendEvent, call,
5340 RX_CALL_REFCOUNT_RESEND);
5343 /* The retry time is the retry time on the first unacknowledged
5344 * packet inside the current window */
5346 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5347 /* Don't set timers for packets outside the window */
5348 if (p->header.seq >= call->tfirst + call->twind) {
5352 if (!(p->flags & RX_PKTFLAG_ACKED)
5353 && !clock_IsZero(&p->retryTime)) {
5355 retryTime = p->retryTime;
5360 /* Post a new event to re-run rxi_Start when retries may be needed */
5361 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5362 #ifdef RX_ENABLE_LOCKS
5363 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5365 rxevent_Post2(&retryTime, rxi_StartUnlocked,
5366 (void *)call, 0, istack);
5367 #else /* RX_ENABLE_LOCKS */
5369 rxevent_Post2(&retryTime, rxi_Start, (void *)call,
5371 #endif /* RX_ENABLE_LOCKS */
5374 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5375 } while (call->flags & RX_CALL_NEED_START);
5377 * TQ references no longer protected by this flag; they must remain
5378 * protected by the global lock.
5380 call->flags &= ~RX_CALL_TQ_BUSY;
5381 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5382 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5383 #ifdef RX_ENABLE_LOCKS
5384 osirx_AssertMine(&call->lock, "rxi_Start end");
5385 CV_BROADCAST(&call->cv_tq);
5386 #else /* RX_ENABLE_LOCKS */
5387 osi_rxWakeup(&call->tq);
5388 #endif /* RX_ENABLE_LOCKS */
5391 call->flags |= RX_CALL_NEED_START;
5393 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5395 if (call->resendEvent) {
5396 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5401 /* Also adjusts the keep alive parameters for the call, to reflect
5402 * that we have just sent a packet (so keep alives aren't sent
5405 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5408 register struct rx_connection *conn = call->conn;
5410 /* Stamp each packet with the user supplied status */
5411 p->header.userStatus = call->localStatus;
5413 /* Allow the security object controlling this call's security to
5414 * make any last-minute changes to the packet */
5415 RXS_SendPacket(conn->securityObject, call, p);
5417 /* Since we're about to send SOME sort of packet to the peer, it's
5418 * safe to nuke any scheduled end-of-packets ack */
5419 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5421 /* Actually send the packet, filling in more connection-specific fields */
5422 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5423 MUTEX_EXIT(&call->lock);
5424 rxi_SendPacket(call, conn, p, istack);
5425 MUTEX_ENTER(&call->lock);
5426 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5428 /* Update last send time for this call (for keep-alive
5429 * processing), and for the connection (so that we can discover
5430 * idle connections) */
5431 conn->lastSendTime = call->lastSendTime = clock_Sec();
5435 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5436 * that things are fine. Also called periodically to guarantee that nothing
5437 * falls through the cracks (e.g. (error + dally) connections have keepalive
5438 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5440 * haveCTLock Set if calling from rxi_ReapConnections
5442 #ifdef RX_ENABLE_LOCKS
5444 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5445 #else /* RX_ENABLE_LOCKS */
5447 rxi_CheckCall(register struct rx_call *call)
5448 #endif /* RX_ENABLE_LOCKS */
5450 register struct rx_connection *conn = call->conn;
5452 afs_uint32 deadTime;
5454 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5455 if (call->flags & RX_CALL_TQ_BUSY) {
5456 /* Call is active and will be reset by rxi_Start if it's
5457 * in an error state.
5462 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5464 (((afs_uint32) conn->secondsUntilDead << 10) +
5465 ((afs_uint32) conn->peer->rtt >> 3) +
5466 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5468 /* These are computed to the second (+- 1 second). But that's
5469 * good enough for these values, which should be a significant
5470 * number of seconds. */
5471 if (now > (call->lastReceiveTime + deadTime)) {
5472 if (call->state == RX_STATE_ACTIVE) {
5473 rxi_CallError(call, RX_CALL_DEAD);
5476 #ifdef RX_ENABLE_LOCKS
5477 /* Cancel pending events */
5478 rxevent_Cancel(call->delayedAckEvent, call,
5479 RX_CALL_REFCOUNT_DELAY);
5480 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5481 rxevent_Cancel(call->keepAliveEvent, call,
5482 RX_CALL_REFCOUNT_ALIVE);
5483 if (call->refCount == 0) {
5484 rxi_FreeCall(call, haveCTLock);
5488 #else /* RX_ENABLE_LOCKS */
5491 #endif /* RX_ENABLE_LOCKS */
5493 /* Non-active calls are destroyed if they are not responding
5494 * to pings; active calls are simply flagged in error, so the
5495 * attached process can die reasonably gracefully. */
5497 /* see if we have a non-activity timeout */
5498 if (call->startWait && conn->idleDeadTime
5499 && ((call->startWait + conn->idleDeadTime) < now)) {
5500 if (call->state == RX_STATE_ACTIVE) {
5501 rxi_CallError(call, RX_CALL_TIMEOUT);
5505 /* see if we have a hard timeout */
5506 if (conn->hardDeadTime
5507 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5508 if (call->state == RX_STATE_ACTIVE)
5509 rxi_CallError(call, RX_CALL_TIMEOUT);
5516 /* When a call is in progress, this routine is called occasionally to
5517 * make sure that some traffic has arrived (or been sent to) the peer.
5518 * If nothing has arrived in a reasonable amount of time, the call is
5519 * declared dead; if nothing has been sent for a while, we send a
5520 * keep-alive packet (if we're actually trying to keep the call alive)
5523 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5526 struct rx_connection *conn;
5529 MUTEX_ENTER(&call->lock);
5530 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5531 if (event == call->keepAliveEvent)
5532 call->keepAliveEvent = NULL;
5535 #ifdef RX_ENABLE_LOCKS
5536 if (rxi_CheckCall(call, 0)) {
5537 MUTEX_EXIT(&call->lock);
5540 #else /* RX_ENABLE_LOCKS */
5541 if (rxi_CheckCall(call))
5543 #endif /* RX_ENABLE_LOCKS */
5545 /* Don't try to keep alive dallying calls */
5546 if (call->state == RX_STATE_DALLY) {
5547 MUTEX_EXIT(&call->lock);
5552 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5553 /* Don't try to send keepalives if there is unacknowledged data */
5554 /* the rexmit code should be good enough, this little hack
5555 * doesn't quite work XXX */
5556 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5558 rxi_ScheduleKeepAliveEvent(call);
5559 MUTEX_EXIT(&call->lock);
5564 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5566 if (!call->keepAliveEvent) {
5568 clock_GetTime(&when);
5569 when.sec += call->conn->secondsUntilPing;
5570 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5571 call->keepAliveEvent =
5572 rxevent_Post(&when, rxi_KeepAliveEvent, call, 0);
5576 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5578 rxi_KeepAliveOn(register struct rx_call *call)
5580 /* Pretend last packet received was received now--i.e. if another
5581 * packet isn't received within the keep alive time, then the call
5582 * will die; Initialize last send time to the current time--even
5583 * if a packet hasn't been sent yet. This will guarantee that a
5584 * keep-alive is sent within the ping time */
5585 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5586 rxi_ScheduleKeepAliveEvent(call);
5589 /* This routine is called to send connection abort messages
5590 * that have been delayed to throttle looping clients. */
5592 rxi_SendDelayedConnAbort(struct rxevent *event,
5593 register struct rx_connection *conn, char *dummy)
5596 struct rx_packet *packet;
5598 MUTEX_ENTER(&conn->conn_data_lock);
5599 conn->delayedAbortEvent = NULL;
5600 error = htonl(conn->error);
5602 MUTEX_EXIT(&conn->conn_data_lock);
5603 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5606 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5607 RX_PACKET_TYPE_ABORT, (char *)&error,
5609 rxi_FreePacket(packet);
5613 /* This routine is called to send call abort messages
5614 * that have been delayed to throttle looping clients. */
5616 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5620 struct rx_packet *packet;
5622 MUTEX_ENTER(&call->lock);
5623 call->delayedAbortEvent = NULL;
5624 error = htonl(call->error);
5626 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5629 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5630 (char *)&error, sizeof(error), 0);
5631 rxi_FreePacket(packet);
5633 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5634 MUTEX_EXIT(&call->lock);
5637 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5638 * seconds) to ask the client to authenticate itself. The routine
5639 * issues a challenge to the client, which is obtained from the
5640 * security object associated with the connection */
5642 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5643 void *arg1, int tries)
5645 conn->challengeEvent = NULL;
5646 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5647 register struct rx_packet *packet;
5651 /* We've failed to authenticate for too long.
5652 * Reset any calls waiting for authentication;
5653 * they are all in RX_STATE_PRECALL.
5657 MUTEX_ENTER(&conn->conn_call_lock);
5658 for (i = 0; i < RX_MAXCALLS; i++) {
5659 struct rx_call *call = conn->call[i];
5661 MUTEX_ENTER(&call->lock);
5662 if (call->state == RX_STATE_PRECALL) {
5663 rxi_CallError(call, RX_CALL_DEAD);
5664 rxi_SendCallAbort(call, NULL, 0, 0);
5666 MUTEX_EXIT(&call->lock);
5669 MUTEX_EXIT(&conn->conn_call_lock);
5673 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5675 /* If there's no packet available, do this later. */
5676 RXS_GetChallenge(conn->securityObject, conn, packet);
5677 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5678 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5679 rxi_FreePacket(packet);
5681 clock_GetTime(&when);
5682 when.sec += RX_CHALLENGE_TIMEOUT;
5683 conn->challengeEvent =
5684 rxevent_Post2(&when, rxi_ChallengeEvent, conn, 0,
5689 /* Call this routine to start requesting the client to authenticate
5690 * itself. This will continue until authentication is established,
5691 * the call times out, or an invalid response is returned. The
5692 * security object associated with the connection is asked to create
5693 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5694 * defined earlier. */
5696 rxi_ChallengeOn(register struct rx_connection *conn)
5698 if (!conn->challengeEvent) {
5699 RXS_CreateChallenge(conn->securityObject, conn);
5700 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5705 /* Compute round trip time of the packet provided, in *rttp.
5708 /* rxi_ComputeRoundTripTime is called with peer locked. */
5709 /* sentp and/or peer may be null */
5711 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5712 register struct clock *sentp,
5713 register struct rx_peer *peer)
5715 struct clock thisRtt, *rttp = &thisRtt;
5717 register int rtt_timeout;
5719 clock_GetTime(rttp);
5721 if (clock_Lt(rttp, sentp)) {
5723 return; /* somebody set the clock back, don't count this time. */
5725 clock_Sub(rttp, sentp);
5726 MUTEX_ENTER(&rx_stats_mutex);
5727 if (clock_Lt(rttp, &rx_stats.minRtt))
5728 rx_stats.minRtt = *rttp;
5729 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5730 if (rttp->sec > 60) {
5731 MUTEX_EXIT(&rx_stats_mutex);
5732 return; /* somebody set the clock ahead */
5734 rx_stats.maxRtt = *rttp;
5736 clock_Add(&rx_stats.totalRtt, rttp);
5737 rx_stats.nRttSamples++;
5738 MUTEX_EXIT(&rx_stats_mutex);
5740 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5742 /* Apply VanJacobson round-trip estimations */
5747 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5748 * srtt is stored as fixed point with 3 bits after the binary
5749 * point (i.e., scaled by 8). The following magic is
5750 * equivalent to the smoothing algorithm in rfc793 with an
5751 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5752 * srtt*8 = srtt*8 + rtt - srtt
5753 * srtt = srtt + rtt/8 - srtt/8
5756 delta = MSEC(rttp) - (peer->rtt >> 3);
5760 * We accumulate a smoothed rtt variance (actually, a smoothed
5761 * mean difference), then set the retransmit timer to smoothed
5762 * rtt + 4 times the smoothed variance (was 2x in van's original
5763 * paper, but 4x works better for me, and apparently for him as
5765 * rttvar is stored as
5766 * fixed point with 2 bits after the binary point (scaled by
5767 * 4). The following is equivalent to rfc793 smoothing with
5768 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5769 * replaces rfc793's wired-in beta.
5770 * dev*4 = dev*4 + (|actual - expected| - dev)
5776 delta -= (peer->rtt_dev >> 2);
5777 peer->rtt_dev += delta;
5779 /* I don't have a stored RTT so I start with this value. Since I'm
5780 * probably just starting a call, and will be pushing more data down
5781 * this, I expect congestion to increase rapidly. So I fudge a
5782 * little, and I set deviance to half the rtt. In practice,
5783 * deviance tends to approach something a little less than
5784 * half the smoothed rtt. */
5785 peer->rtt = (MSEC(rttp) << 3) + 8;
5786 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5788 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5789 * the other of these connections is usually in a user process, and can
5790 * be switched and/or swapped out. So on fast, reliable networks, the
5791 * timeout would otherwise be too short.
5793 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5794 clock_Zero(&(peer->timeout));
5795 clock_Addmsec(&(peer->timeout), rtt_timeout);
5797 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)));
5801 /* Find all server connections that have not been active for a long time, and
5804 rxi_ReapConnections(void)
5807 clock_GetTime(&now);
5809 /* Find server connection structures that haven't been used for
5810 * greater than rx_idleConnectionTime */
5812 struct rx_connection **conn_ptr, **conn_end;
5813 int i, havecalls = 0;
5814 MUTEX_ENTER(&rx_connHashTable_lock);
5815 for (conn_ptr = &rx_connHashTable[0], conn_end =
5816 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5818 struct rx_connection *conn, *next;
5819 struct rx_call *call;
5823 for (conn = *conn_ptr; conn; conn = next) {
5824 /* XXX -- Shouldn't the connection be locked? */
5827 for (i = 0; i < RX_MAXCALLS; i++) {
5828 call = conn->call[i];
5831 MUTEX_ENTER(&call->lock);
5832 #ifdef RX_ENABLE_LOCKS
5833 result = rxi_CheckCall(call, 1);
5834 #else /* RX_ENABLE_LOCKS */
5835 result = rxi_CheckCall(call);
5836 #endif /* RX_ENABLE_LOCKS */
5837 MUTEX_EXIT(&call->lock);
5839 /* If CheckCall freed the call, it might
5840 * have destroyed the connection as well,
5841 * which screws up the linked lists.
5847 if (conn->type == RX_SERVER_CONNECTION) {
5848 /* This only actually destroys the connection if
5849 * there are no outstanding calls */
5850 MUTEX_ENTER(&conn->conn_data_lock);
5851 if (!havecalls && !conn->refCount
5852 && ((conn->lastSendTime + rx_idleConnectionTime) <
5854 conn->refCount++; /* it will be decr in rx_DestroyConn */
5855 MUTEX_EXIT(&conn->conn_data_lock);
5856 #ifdef RX_ENABLE_LOCKS
5857 rxi_DestroyConnectionNoLock(conn);
5858 #else /* RX_ENABLE_LOCKS */
5859 rxi_DestroyConnection(conn);
5860 #endif /* RX_ENABLE_LOCKS */
5862 #ifdef RX_ENABLE_LOCKS
5864 MUTEX_EXIT(&conn->conn_data_lock);
5866 #endif /* RX_ENABLE_LOCKS */
5870 #ifdef RX_ENABLE_LOCKS
5871 while (rx_connCleanup_list) {
5872 struct rx_connection *conn;
5873 conn = rx_connCleanup_list;
5874 rx_connCleanup_list = rx_connCleanup_list->next;
5875 MUTEX_EXIT(&rx_connHashTable_lock);
5876 rxi_CleanupConnection(conn);
5877 MUTEX_ENTER(&rx_connHashTable_lock);
5879 MUTEX_EXIT(&rx_connHashTable_lock);
5880 #endif /* RX_ENABLE_LOCKS */
5883 /* Find any peer structures that haven't been used (haven't had an
5884 * associated connection) for greater than rx_idlePeerTime */
5886 struct rx_peer **peer_ptr, **peer_end;
5888 MUTEX_ENTER(&rx_rpc_stats);
5889 MUTEX_ENTER(&rx_peerHashTable_lock);
5890 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5891 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5893 struct rx_peer *peer, *next, *prev;
5894 for (prev = peer = *peer_ptr; peer; peer = next) {
5896 code = MUTEX_TRYENTER(&peer->peer_lock);
5897 if ((code) && (peer->refCount == 0)
5898 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5899 rx_interface_stat_p rpc_stat, nrpc_stat;
5901 MUTEX_EXIT(&peer->peer_lock);
5902 MUTEX_DESTROY(&peer->peer_lock);
5904 (&peer->rpcStats, rpc_stat, nrpc_stat,
5905 rx_interface_stat)) {
5906 unsigned int num_funcs;
5909 queue_Remove(&rpc_stat->queue_header);
5910 queue_Remove(&rpc_stat->all_peers);
5911 num_funcs = rpc_stat->stats[0].func_total;
5913 sizeof(rx_interface_stat_t) +
5914 rpc_stat->stats[0].func_total *
5915 sizeof(rx_function_entry_v1_t);
5917 rxi_Free(rpc_stat, space);
5918 rxi_rpc_peer_stat_cnt -= num_funcs;
5921 MUTEX_ENTER(&rx_stats_mutex);
5922 rx_stats.nPeerStructs--;
5923 MUTEX_EXIT(&rx_stats_mutex);
5924 if (peer == *peer_ptr) {
5931 MUTEX_EXIT(&peer->peer_lock);
5937 MUTEX_EXIT(&rx_peerHashTable_lock);
5938 MUTEX_EXIT(&rx_rpc_stats);
5941 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5942 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5943 * GC, just below. Really, we shouldn't have to keep moving packets from
5944 * one place to another, but instead ought to always know if we can
5945 * afford to hold onto a packet in its particular use. */
5946 MUTEX_ENTER(&rx_freePktQ_lock);
5947 if (rx_waitingForPackets) {
5948 rx_waitingForPackets = 0;
5949 #ifdef RX_ENABLE_LOCKS
5950 CV_BROADCAST(&rx_waitingForPackets_cv);
5952 osi_rxWakeup(&rx_waitingForPackets);
5955 MUTEX_EXIT(&rx_freePktQ_lock);
5957 now.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5958 rxevent_Post(&now, rxi_ReapConnections, 0, 0);
5962 /* rxs_Release - This isn't strictly necessary but, since the macro name from
5963 * rx.h is sort of strange this is better. This is called with a security
5964 * object before it is discarded. Each connection using a security object has
5965 * its own refcount to the object so it won't actually be freed until the last
5966 * connection is destroyed.
5968 * This is the only rxs module call. A hold could also be written but no one
5972 rxs_Release(struct rx_securityClass *aobj)
5974 return RXS_Close(aobj);
5978 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
5979 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
5980 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
5981 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
5983 /* Adjust our estimate of the transmission rate to this peer, given
5984 * that the packet p was just acked. We can adjust peer->timeout and
5985 * call->twind. Pragmatically, this is called
5986 * only with packets of maximal length.
5987 * Called with peer and call locked.
5991 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
5992 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
5994 afs_int32 xferSize, xferMs;
5995 register afs_int32 minTime;
5998 /* Count down packets */
5999 if (peer->rateFlag > 0)
6001 /* Do nothing until we're enabled */
6002 if (peer->rateFlag != 0)
6007 /* Count only when the ack seems legitimate */
6008 switch (ackReason) {
6009 case RX_ACK_REQUESTED:
6011 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6015 case RX_ACK_PING_RESPONSE:
6016 if (p) /* want the response to ping-request, not data send */
6018 clock_GetTime(&newTO);
6019 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6020 clock_Sub(&newTO, &call->pingRequestTime);
6021 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6025 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6032 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));
6034 /* Track only packets that are big enough. */
6035 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6039 /* absorb RTT data (in milliseconds) for these big packets */
6040 if (peer->smRtt == 0) {
6041 peer->smRtt = xferMs;
6043 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6048 if (peer->countDown) {
6052 peer->countDown = 10; /* recalculate only every so often */
6054 /* In practice, we can measure only the RTT for full packets,
6055 * because of the way Rx acks the data that it receives. (If it's
6056 * smaller than a full packet, it often gets implicitly acked
6057 * either by the call response (from a server) or by the next call
6058 * (from a client), and either case confuses transmission times
6059 * with processing times.) Therefore, replace the above
6060 * more-sophisticated processing with a simpler version, where the
6061 * smoothed RTT is kept for full-size packets, and the time to
6062 * transmit a windowful of full-size packets is simply RTT *
6063 * windowSize. Again, we take two steps:
6064 - ensure the timeout is large enough for a single packet's RTT;
6065 - ensure that the window is small enough to fit in the desired timeout.*/
6067 /* First, the timeout check. */
6068 minTime = peer->smRtt;
6069 /* Get a reasonable estimate for a timeout period */
6071 newTO.sec = minTime / 1000;
6072 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6074 /* Increase the timeout period so that we can always do at least
6075 * one packet exchange */
6076 if (clock_Gt(&newTO, &peer->timeout)) {
6078 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));
6080 peer->timeout = newTO;
6083 /* Now, get an estimate for the transmit window size. */
6084 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6085 /* Now, convert to the number of full packets that could fit in a
6086 * reasonable fraction of that interval */
6087 minTime /= (peer->smRtt << 1);
6088 xferSize = minTime; /* (make a copy) */
6090 /* Now clamp the size to reasonable bounds. */
6093 else if (minTime > rx_Window)
6094 minTime = rx_Window;
6095 /* if (minTime != peer->maxWindow) {
6096 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6097 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6098 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6100 peer->maxWindow = minTime;
6101 elide... call->twind = minTime;
6105 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6106 * Discern this by calculating the timeout necessary for rx_Window
6108 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6109 /* calculate estimate for transmission interval in milliseconds */
6110 minTime = rx_Window * peer->smRtt;
6111 if (minTime < 1000) {
6112 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
6113 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6114 peer->timeout.usec, peer->smRtt, peer->packetSize));
6116 newTO.sec = 0; /* cut back on timeout by half a second */
6117 newTO.usec = 500000;
6118 clock_Sub(&peer->timeout, &newTO);
6123 } /* end of rxi_ComputeRate */
6124 #endif /* ADAPT_WINDOW */
6132 #define TRACE_OPTION_DEBUGLOG 4
6140 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6141 0, KEY_QUERY_VALUE, &parmKey);
6142 if (code != ERROR_SUCCESS)
6145 dummyLen = sizeof(TraceOption);
6146 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6147 (BYTE *) &TraceOption, &dummyLen);
6148 if (code == ERROR_SUCCESS) {
6149 rxdebug_active = (TraceOption & TRACE_OPTION_DEBUGLOG) ? 1 : 0;
6151 RegCloseKey (parmKey);
6152 #endif /* AFS_NT40_ENV */
6157 rx_DebugOnOff(int on)
6159 rxdebug_active = on;
6161 #endif /* AFS_NT40_ENV */
6164 /* Don't call this debugging routine directly; use dpf */
6166 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
6167 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
6175 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6178 len = _snprintf(msg, sizeof(msg)-2,
6179 tformat, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10,
6180 a11, a12, a13, a14, a15);
6182 if (msg[len-1] != '\n') {
6186 OutputDebugString(msg);
6191 clock_GetTime(&now);
6192 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
6193 (unsigned int)now.usec / 1000);
6194 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
6201 * This function is used to process the rx_stats structure that is local
6202 * to a process as well as an rx_stats structure received from a remote
6203 * process (via rxdebug). Therefore, it needs to do minimal version
6207 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
6208 afs_int32 freePackets, char version)
6212 if (size != sizeof(struct rx_stats)) {
6214 "Unexpected size of stats structure: was %d, expected %d\n",
6215 size, sizeof(struct rx_stats));
6218 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6221 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6222 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6223 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6224 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6225 s->specialPktAllocFailures);
6227 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6228 s->receivePktAllocFailures, s->sendPktAllocFailures,
6229 s->specialPktAllocFailures);
6233 " greedy %d, " "bogusReads %d (last from host %x), "
6234 "noPackets %d, " "noBuffers %d, " "selects %d, "
6235 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6236 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6237 s->selects, s->sendSelects);
6239 fprintf(file, " packets read: ");
6240 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6241 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6243 fprintf(file, "\n");
6246 " other read counters: data %d, " "ack %d, " "dup %d "
6247 "spurious %d " "dally %d\n", s->dataPacketsRead,
6248 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6249 s->ignorePacketDally);
6251 fprintf(file, " packets sent: ");
6252 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6253 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6255 fprintf(file, "\n");
6258 " other send counters: ack %d, " "data %d (not resends), "
6259 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6260 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6261 s->dataPacketsPushed, s->ignoreAckedPacket);
6264 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6265 s->netSendFailures, (int)s->fatalErrors);
6267 if (s->nRttSamples) {
6268 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6269 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6271 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6272 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6276 " %d server connections, " "%d client connections, "
6277 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6278 s->nServerConns, s->nClientConns, s->nPeerStructs,
6279 s->nCallStructs, s->nFreeCallStructs);
6281 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6282 fprintf(file, " %d clock updates\n", clock_nUpdates);
6287 /* for backward compatibility */
6289 rx_PrintStats(FILE * file)
6291 MUTEX_ENTER(&rx_stats_mutex);
6292 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6294 MUTEX_EXIT(&rx_stats_mutex);
6298 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6300 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6301 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6302 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6305 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6306 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6307 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6310 " Packet size %d, " "max in packet skew %d, "
6311 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6312 (int)peer->outPacketSkew);
6315 #ifdef AFS_PTHREAD_ENV
6317 * This mutex protects the following static variables:
6321 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6322 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6324 #define LOCK_RX_DEBUG
6325 #define UNLOCK_RX_DEBUG
6326 #endif /* AFS_PTHREAD_ENV */
6329 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6330 u_char type, void *inputData, size_t inputLength,
6331 void *outputData, size_t outputLength)
6333 static afs_int32 counter = 100;
6334 time_t waitTime, waitCount, startTime, endTime;
6335 struct rx_header theader;
6337 register afs_int32 code;
6338 struct timeval tv_now, tv_wake, tv_delta;
6339 struct sockaddr_in taddr, faddr;
6344 startTime = time(0);
6350 tp = &tbuffer[sizeof(struct rx_header)];
6351 taddr.sin_family = AF_INET;
6352 taddr.sin_port = remotePort;
6353 taddr.sin_addr.s_addr = remoteAddr;
6354 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6355 taddr.sin_len = sizeof(struct sockaddr_in);
6358 memset(&theader, 0, sizeof(theader));
6359 theader.epoch = htonl(999);
6361 theader.callNumber = htonl(counter);
6364 theader.type = type;
6365 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6366 theader.serviceId = 0;
6368 memcpy(tbuffer, &theader, sizeof(theader));
6369 memcpy(tp, inputData, inputLength);
6371 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6372 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6374 /* see if there's a packet available */
6375 gettimeofday(&tv_wake,0);
6376 tv_wake.tv_sec += waitTime;
6379 FD_SET(socket, &imask);
6380 tv_delta.tv_sec = tv_wake.tv_sec;
6381 tv_delta.tv_usec = tv_wake.tv_usec;
6382 gettimeofday(&tv_now, 0);
6384 if (tv_delta.tv_usec < tv_now.tv_usec) {
6386 tv_delta.tv_usec += 1000000;
6389 tv_delta.tv_usec -= tv_now.tv_usec;
6391 if (tv_delta.tv_sec < tv_now.tv_sec) {
6395 tv_delta.tv_sec -= tv_now.tv_sec;
6397 code = select(socket + 1, &imask, 0, 0, &tv_delta);
6398 if (code == 1 && FD_ISSET(socket, &imask)) {
6399 /* now receive a packet */
6400 faddrLen = sizeof(struct sockaddr_in);
6402 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6403 (struct sockaddr *)&faddr, &faddrLen);
6406 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6407 if (counter == ntohl(theader.callNumber))
6415 /* see if we've timed out */
6423 code -= sizeof(struct rx_header);
6424 if (code > outputLength)
6425 code = outputLength;
6426 memcpy(outputData, tp, code);
6431 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6432 afs_uint16 remotePort, struct rx_debugStats * stat,
6433 afs_uint32 * supportedValues)
6435 struct rx_debugIn in;
6438 *supportedValues = 0;
6439 in.type = htonl(RX_DEBUGI_GETSTATS);
6442 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6443 &in, sizeof(in), stat, sizeof(*stat));
6446 * If the call was successful, fixup the version and indicate
6447 * what contents of the stat structure are valid.
6448 * Also do net to host conversion of fields here.
6452 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6453 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6455 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6456 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6458 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6459 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6461 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6462 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6464 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6465 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6467 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6468 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6470 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6471 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6473 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6474 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6477 stat->nFreePackets = ntohl(stat->nFreePackets);
6478 stat->packetReclaims = ntohl(stat->packetReclaims);
6479 stat->callsExecuted = ntohl(stat->callsExecuted);
6480 stat->nWaiting = ntohl(stat->nWaiting);
6481 stat->idleThreads = ntohl(stat->idleThreads);
6488 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6489 afs_uint16 remotePort, struct rx_stats * stat,
6490 afs_uint32 * supportedValues)
6492 struct rx_debugIn in;
6493 afs_int32 *lp = (afs_int32 *) stat;
6498 * supportedValues is currently unused, but added to allow future
6499 * versioning of this function.
6502 *supportedValues = 0;
6503 in.type = htonl(RX_DEBUGI_RXSTATS);
6505 memset(stat, 0, sizeof(*stat));
6507 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6508 &in, sizeof(in), stat, sizeof(*stat));
6513 * Do net to host conversion here
6516 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6525 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6526 afs_uint16 remotePort, size_t version_length,
6530 return MakeDebugCall(socket, remoteAddr, remotePort,
6531 RX_PACKET_TYPE_VERSION, a, 1, version,
6536 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6537 afs_uint16 remotePort, afs_int32 * nextConnection,
6538 int allConnections, afs_uint32 debugSupportedValues,
6539 struct rx_debugConn * conn,
6540 afs_uint32 * supportedValues)
6542 struct rx_debugIn in;
6547 * supportedValues is currently unused, but added to allow future
6548 * versioning of this function.
6551 *supportedValues = 0;
6552 if (allConnections) {
6553 in.type = htonl(RX_DEBUGI_GETALLCONN);
6555 in.type = htonl(RX_DEBUGI_GETCONN);
6557 in.index = htonl(*nextConnection);
6558 memset(conn, 0, sizeof(*conn));
6560 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6561 &in, sizeof(in), conn, sizeof(*conn));
6564 *nextConnection += 1;
6567 * Convert old connection format to new structure.
6570 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6571 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6572 #define MOVEvL(a) (conn->a = vL->a)
6574 /* any old or unrecognized version... */
6575 for (i = 0; i < RX_MAXCALLS; i++) {
6576 MOVEvL(callState[i]);
6577 MOVEvL(callMode[i]);
6578 MOVEvL(callFlags[i]);
6579 MOVEvL(callOther[i]);
6581 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6582 MOVEvL(secStats.type);
6583 MOVEvL(secStats.level);
6584 MOVEvL(secStats.flags);
6585 MOVEvL(secStats.expires);
6586 MOVEvL(secStats.packetsReceived);
6587 MOVEvL(secStats.packetsSent);
6588 MOVEvL(secStats.bytesReceived);
6589 MOVEvL(secStats.bytesSent);
6594 * Do net to host conversion here
6596 * I don't convert host or port since we are most likely
6597 * going to want these in NBO.
6599 conn->cid = ntohl(conn->cid);
6600 conn->serial = ntohl(conn->serial);
6601 for (i = 0; i < RX_MAXCALLS; i++) {
6602 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6604 conn->error = ntohl(conn->error);
6605 conn->secStats.flags = ntohl(conn->secStats.flags);
6606 conn->secStats.expires = ntohl(conn->secStats.expires);
6607 conn->secStats.packetsReceived =
6608 ntohl(conn->secStats.packetsReceived);
6609 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6610 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6611 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6612 conn->epoch = ntohl(conn->epoch);
6613 conn->natMTU = ntohl(conn->natMTU);
6620 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6621 afs_uint16 remotePort, afs_int32 * nextPeer,
6622 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6623 afs_uint32 * supportedValues)
6625 struct rx_debugIn in;
6629 * supportedValues is currently unused, but added to allow future
6630 * versioning of this function.
6633 *supportedValues = 0;
6634 in.type = htonl(RX_DEBUGI_GETPEER);
6635 in.index = htonl(*nextPeer);
6636 memset(peer, 0, sizeof(*peer));
6638 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6639 &in, sizeof(in), peer, sizeof(*peer));
6645 * Do net to host conversion here
6647 * I don't convert host or port since we are most likely
6648 * going to want these in NBO.
6650 peer->ifMTU = ntohs(peer->ifMTU);
6651 peer->idleWhen = ntohl(peer->idleWhen);
6652 peer->refCount = ntohs(peer->refCount);
6653 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6654 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6655 peer->rtt = ntohl(peer->rtt);
6656 peer->rtt_dev = ntohl(peer->rtt_dev);
6657 peer->timeout.sec = ntohl(peer->timeout.sec);
6658 peer->timeout.usec = ntohl(peer->timeout.usec);
6659 peer->nSent = ntohl(peer->nSent);
6660 peer->reSends = ntohl(peer->reSends);
6661 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6662 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6663 peer->rateFlag = ntohl(peer->rateFlag);
6664 peer->natMTU = ntohs(peer->natMTU);
6665 peer->maxMTU = ntohs(peer->maxMTU);
6666 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6667 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6668 peer->MTU = ntohs(peer->MTU);
6669 peer->cwind = ntohs(peer->cwind);
6670 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6671 peer->congestSeq = ntohs(peer->congestSeq);
6672 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6673 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6674 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6675 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6680 #endif /* RXDEBUG */
6685 struct rx_serverQueueEntry *np;
6688 register struct rx_call *call;
6689 register struct rx_serverQueueEntry *sq;
6693 if (rxinit_status == 1) {
6695 return; /* Already shutdown. */
6699 #ifndef AFS_PTHREAD_ENV
6700 FD_ZERO(&rx_selectMask);
6701 #endif /* AFS_PTHREAD_ENV */
6702 rxi_dataQuota = RX_MAX_QUOTA;
6703 #ifndef AFS_PTHREAD_ENV
6705 #endif /* AFS_PTHREAD_ENV */
6708 #ifndef AFS_PTHREAD_ENV
6709 #ifndef AFS_USE_GETTIMEOFDAY
6711 #endif /* AFS_USE_GETTIMEOFDAY */
6712 #endif /* AFS_PTHREAD_ENV */
6714 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6715 call = queue_First(&rx_freeCallQueue, rx_call);
6717 rxi_Free(call, sizeof(struct rx_call));
6720 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6721 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6727 struct rx_peer **peer_ptr, **peer_end;
6728 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6729 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6731 struct rx_peer *peer, *next;
6732 for (peer = *peer_ptr; peer; peer = next) {
6733 rx_interface_stat_p rpc_stat, nrpc_stat;
6736 (&peer->rpcStats, rpc_stat, nrpc_stat,
6737 rx_interface_stat)) {
6738 unsigned int num_funcs;
6741 queue_Remove(&rpc_stat->queue_header);
6742 queue_Remove(&rpc_stat->all_peers);
6743 num_funcs = rpc_stat->stats[0].func_total;
6745 sizeof(rx_interface_stat_t) +
6746 rpc_stat->stats[0].func_total *
6747 sizeof(rx_function_entry_v1_t);
6749 rxi_Free(rpc_stat, space);
6750 MUTEX_ENTER(&rx_rpc_stats);
6751 rxi_rpc_peer_stat_cnt -= num_funcs;
6752 MUTEX_EXIT(&rx_rpc_stats);
6756 MUTEX_ENTER(&rx_stats_mutex);
6757 rx_stats.nPeerStructs--;
6758 MUTEX_EXIT(&rx_stats_mutex);
6762 for (i = 0; i < RX_MAX_SERVICES; i++) {
6764 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6766 for (i = 0; i < rx_hashTableSize; i++) {
6767 register struct rx_connection *tc, *ntc;
6768 MUTEX_ENTER(&rx_connHashTable_lock);
6769 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6771 for (j = 0; j < RX_MAXCALLS; j++) {
6773 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6776 rxi_Free(tc, sizeof(*tc));
6778 MUTEX_EXIT(&rx_connHashTable_lock);
6781 MUTEX_ENTER(&freeSQEList_lock);
6783 while ((np = rx_FreeSQEList)) {
6784 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6785 MUTEX_DESTROY(&np->lock);
6786 rxi_Free(np, sizeof(*np));
6789 MUTEX_EXIT(&freeSQEList_lock);
6790 MUTEX_DESTROY(&freeSQEList_lock);
6791 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6792 MUTEX_DESTROY(&rx_connHashTable_lock);
6793 MUTEX_DESTROY(&rx_peerHashTable_lock);
6794 MUTEX_DESTROY(&rx_serverPool_lock);
6796 osi_Free(rx_connHashTable,
6797 rx_hashTableSize * sizeof(struct rx_connection *));
6798 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6800 UNPIN(rx_connHashTable,
6801 rx_hashTableSize * sizeof(struct rx_connection *));
6802 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6804 rxi_FreeAllPackets();
6806 MUTEX_ENTER(&rx_stats_mutex);
6807 rxi_dataQuota = RX_MAX_QUOTA;
6808 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6809 MUTEX_EXIT(&rx_stats_mutex);
6815 #ifdef RX_ENABLE_LOCKS
6817 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6819 if (!MUTEX_ISMINE(lockaddr))
6820 osi_Panic("Lock not held: %s", msg);
6822 #endif /* RX_ENABLE_LOCKS */
6827 * Routines to implement connection specific data.
6831 rx_KeyCreate(rx_destructor_t rtn)
6834 MUTEX_ENTER(&rxi_keyCreate_lock);
6835 key = rxi_keyCreate_counter++;
6836 rxi_keyCreate_destructor = (rx_destructor_t *)
6837 realloc((void *)rxi_keyCreate_destructor,
6838 (key + 1) * sizeof(rx_destructor_t));
6839 rxi_keyCreate_destructor[key] = rtn;
6840 MUTEX_EXIT(&rxi_keyCreate_lock);
6845 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6848 MUTEX_ENTER(&conn->conn_data_lock);
6849 if (!conn->specific) {
6850 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6851 for (i = 0; i < key; i++)
6852 conn->specific[i] = NULL;
6853 conn->nSpecific = key + 1;
6854 conn->specific[key] = ptr;
6855 } else if (key >= conn->nSpecific) {
6856 conn->specific = (void **)
6857 realloc(conn->specific, (key + 1) * sizeof(void *));
6858 for (i = conn->nSpecific; i < key; i++)
6859 conn->specific[i] = NULL;
6860 conn->nSpecific = key + 1;
6861 conn->specific[key] = ptr;
6863 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6864 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6865 conn->specific[key] = ptr;
6867 MUTEX_EXIT(&conn->conn_data_lock);
6871 rx_GetSpecific(struct rx_connection *conn, int key)
6874 MUTEX_ENTER(&conn->conn_data_lock);
6875 if (key >= conn->nSpecific)
6878 ptr = conn->specific[key];
6879 MUTEX_EXIT(&conn->conn_data_lock);
6883 #endif /* !KERNEL */
6886 * processStats is a queue used to store the statistics for the local
6887 * process. Its contents are similar to the contents of the rpcStats
6888 * queue on a rx_peer structure, but the actual data stored within
6889 * this queue contains totals across the lifetime of the process (assuming
6890 * the stats have not been reset) - unlike the per peer structures
6891 * which can come and go based upon the peer lifetime.
6894 static struct rx_queue processStats = { &processStats, &processStats };
6897 * peerStats is a queue used to store the statistics for all peer structs.
6898 * Its contents are the union of all the peer rpcStats queues.
6901 static struct rx_queue peerStats = { &peerStats, &peerStats };
6904 * rxi_monitor_processStats is used to turn process wide stat collection
6908 static int rxi_monitor_processStats = 0;
6911 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6914 static int rxi_monitor_peerStats = 0;
6917 * rxi_AddRpcStat - given all of the information for a particular rpc
6918 * call, create (if needed) and update the stat totals for the rpc.
6922 * IN stats - the queue of stats that will be updated with the new value
6924 * IN rxInterface - a unique number that identifies the rpc interface
6926 * IN currentFunc - the index of the function being invoked
6928 * IN totalFunc - the total number of functions in this interface
6930 * IN queueTime - the amount of time this function waited for a thread
6932 * IN execTime - the amount of time this function invocation took to execute
6934 * IN bytesSent - the number bytes sent by this invocation
6936 * IN bytesRcvd - the number bytes received by this invocation
6938 * IN isServer - if true, this invocation was made to a server
6940 * IN remoteHost - the ip address of the remote host
6942 * IN remotePort - the port of the remote host
6944 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6946 * INOUT counter - if a new stats structure is allocated, the counter will
6947 * be updated with the new number of allocated stat structures
6955 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6956 afs_uint32 currentFunc, afs_uint32 totalFunc,
6957 struct clock *queueTime, struct clock *execTime,
6958 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6959 afs_uint32 remoteHost, afs_uint32 remotePort,
6960 int addToPeerList, unsigned int *counter)
6963 rx_interface_stat_p rpc_stat, nrpc_stat;
6966 * See if there's already a structure for this interface
6969 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6970 if ((rpc_stat->stats[0].interfaceId == rxInterface)
6971 && (rpc_stat->stats[0].remote_is_server == isServer))
6976 * Didn't find a match so allocate a new structure and add it to the
6980 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
6981 || (rpc_stat->stats[0].interfaceId != rxInterface)
6982 || (rpc_stat->stats[0].remote_is_server != isServer)) {
6987 sizeof(rx_interface_stat_t) +
6988 totalFunc * sizeof(rx_function_entry_v1_t);
6990 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
6991 if (rpc_stat == NULL) {
6995 *counter += totalFunc;
6996 for (i = 0; i < totalFunc; i++) {
6997 rpc_stat->stats[i].remote_peer = remoteHost;
6998 rpc_stat->stats[i].remote_port = remotePort;
6999 rpc_stat->stats[i].remote_is_server = isServer;
7000 rpc_stat->stats[i].interfaceId = rxInterface;
7001 rpc_stat->stats[i].func_total = totalFunc;
7002 rpc_stat->stats[i].func_index = i;
7003 hzero(rpc_stat->stats[i].invocations);
7004 hzero(rpc_stat->stats[i].bytes_sent);
7005 hzero(rpc_stat->stats[i].bytes_rcvd);
7006 rpc_stat->stats[i].queue_time_sum.sec = 0;
7007 rpc_stat->stats[i].queue_time_sum.usec = 0;
7008 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7009 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7010 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7011 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7012 rpc_stat->stats[i].queue_time_max.sec = 0;
7013 rpc_stat->stats[i].queue_time_max.usec = 0;
7014 rpc_stat->stats[i].execution_time_sum.sec = 0;
7015 rpc_stat->stats[i].execution_time_sum.usec = 0;
7016 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7017 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7018 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7019 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7020 rpc_stat->stats[i].execution_time_max.sec = 0;
7021 rpc_stat->stats[i].execution_time_max.usec = 0;
7023 queue_Prepend(stats, rpc_stat);
7024 if (addToPeerList) {
7025 queue_Prepend(&peerStats, &rpc_stat->all_peers);
7030 * Increment the stats for this function
7033 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
7034 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
7035 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
7036 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
7037 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7038 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7039 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7041 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7042 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7044 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7045 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7047 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7048 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7050 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7051 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7059 * rx_IncrementTimeAndCount - increment the times and count for a particular
7064 * IN peer - the peer who invoked the rpc
7066 * IN rxInterface - a unique number that identifies the rpc interface
7068 * IN currentFunc - the index of the function being invoked
7070 * IN totalFunc - the total number of functions in this interface
7072 * IN queueTime - the amount of time this function waited for a thread
7074 * IN execTime - the amount of time this function invocation took to execute
7076 * IN bytesSent - the number bytes sent by this invocation
7078 * IN bytesRcvd - the number bytes received by this invocation
7080 * IN isServer - if true, this invocation was made to a server
7088 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7089 afs_uint32 currentFunc, afs_uint32 totalFunc,
7090 struct clock *queueTime, struct clock *execTime,
7091 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7095 MUTEX_ENTER(&rx_rpc_stats);
7096 MUTEX_ENTER(&peer->peer_lock);
7098 if (rxi_monitor_peerStats) {
7099 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7100 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7101 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7104 if (rxi_monitor_processStats) {
7105 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7106 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7107 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7110 MUTEX_EXIT(&peer->peer_lock);
7111 MUTEX_EXIT(&rx_rpc_stats);
7116 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7120 * IN callerVersion - the rpc stat version of the caller.
7122 * IN count - the number of entries to marshall.
7124 * IN stats - pointer to stats to be marshalled.
7126 * OUT ptr - Where to store the marshalled data.
7133 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7134 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7140 * We only support the first version
7142 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7143 *(ptr++) = stats->remote_peer;
7144 *(ptr++) = stats->remote_port;
7145 *(ptr++) = stats->remote_is_server;
7146 *(ptr++) = stats->interfaceId;
7147 *(ptr++) = stats->func_total;
7148 *(ptr++) = stats->func_index;
7149 *(ptr++) = hgethi(stats->invocations);
7150 *(ptr++) = hgetlo(stats->invocations);
7151 *(ptr++) = hgethi(stats->bytes_sent);
7152 *(ptr++) = hgetlo(stats->bytes_sent);
7153 *(ptr++) = hgethi(stats->bytes_rcvd);
7154 *(ptr++) = hgetlo(stats->bytes_rcvd);
7155 *(ptr++) = stats->queue_time_sum.sec;
7156 *(ptr++) = stats->queue_time_sum.usec;
7157 *(ptr++) = stats->queue_time_sum_sqr.sec;
7158 *(ptr++) = stats->queue_time_sum_sqr.usec;
7159 *(ptr++) = stats->queue_time_min.sec;
7160 *(ptr++) = stats->queue_time_min.usec;
7161 *(ptr++) = stats->queue_time_max.sec;
7162 *(ptr++) = stats->queue_time_max.usec;
7163 *(ptr++) = stats->execution_time_sum.sec;
7164 *(ptr++) = stats->execution_time_sum.usec;
7165 *(ptr++) = stats->execution_time_sum_sqr.sec;
7166 *(ptr++) = stats->execution_time_sum_sqr.usec;
7167 *(ptr++) = stats->execution_time_min.sec;
7168 *(ptr++) = stats->execution_time_min.usec;
7169 *(ptr++) = stats->execution_time_max.sec;
7170 *(ptr++) = stats->execution_time_max.usec;
7176 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7181 * IN callerVersion - the rpc stat version of the caller
7183 * OUT myVersion - the rpc stat version of this function
7185 * OUT clock_sec - local time seconds
7187 * OUT clock_usec - local time microseconds
7189 * OUT allocSize - the number of bytes allocated to contain stats
7191 * OUT statCount - the number stats retrieved from this process.
7193 * OUT stats - the actual stats retrieved from this process.
7197 * Returns void. If successful, stats will != NULL.
7201 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7202 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7203 size_t * allocSize, afs_uint32 * statCount,
7204 afs_uint32 ** stats)
7214 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7217 * Check to see if stats are enabled
7220 MUTEX_ENTER(&rx_rpc_stats);
7221 if (!rxi_monitor_processStats) {
7222 MUTEX_EXIT(&rx_rpc_stats);
7226 clock_GetTime(&now);
7227 *clock_sec = now.sec;
7228 *clock_usec = now.usec;
7231 * Allocate the space based upon the caller version
7233 * If the client is at an older version than we are,
7234 * we return the statistic data in the older data format, but
7235 * we still return our version number so the client knows we
7236 * are maintaining more data than it can retrieve.
7239 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7240 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7241 *statCount = rxi_rpc_process_stat_cnt;
7244 * This can't happen yet, but in the future version changes
7245 * can be handled by adding additional code here
7249 if (space > (size_t) 0) {
7251 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7254 rx_interface_stat_p rpc_stat, nrpc_stat;
7258 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7260 * Copy the data based upon the caller version
7262 rx_MarshallProcessRPCStats(callerVersion,
7263 rpc_stat->stats[0].func_total,
7264 rpc_stat->stats, &ptr);
7270 MUTEX_EXIT(&rx_rpc_stats);
7275 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7279 * IN callerVersion - the rpc stat version of the caller
7281 * OUT myVersion - the rpc stat version of this function
7283 * OUT clock_sec - local time seconds
7285 * OUT clock_usec - local time microseconds
7287 * OUT allocSize - the number of bytes allocated to contain stats
7289 * OUT statCount - the number of stats retrieved from the individual
7292 * OUT stats - the actual stats retrieved from the individual peer structures.
7296 * Returns void. If successful, stats will != NULL.
7300 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7301 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7302 size_t * allocSize, afs_uint32 * statCount,
7303 afs_uint32 ** stats)
7313 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7316 * Check to see if stats are enabled
7319 MUTEX_ENTER(&rx_rpc_stats);
7320 if (!rxi_monitor_peerStats) {
7321 MUTEX_EXIT(&rx_rpc_stats);
7325 clock_GetTime(&now);
7326 *clock_sec = now.sec;
7327 *clock_usec = now.usec;
7330 * Allocate the space based upon the caller version
7332 * If the client is at an older version than we are,
7333 * we return the statistic data in the older data format, but
7334 * we still return our version number so the client knows we
7335 * are maintaining more data than it can retrieve.
7338 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7339 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7340 *statCount = rxi_rpc_peer_stat_cnt;
7343 * This can't happen yet, but in the future version changes
7344 * can be handled by adding additional code here
7348 if (space > (size_t) 0) {
7350 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7353 rx_interface_stat_p rpc_stat, nrpc_stat;
7357 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7359 * We have to fix the offset of rpc_stat since we are
7360 * keeping this structure on two rx_queues. The rx_queue
7361 * package assumes that the rx_queue member is the first
7362 * member of the structure. That is, rx_queue assumes that
7363 * any one item is only on one queue at a time. We are
7364 * breaking that assumption and so we have to do a little
7365 * math to fix our pointers.
7368 fix_offset = (char *)rpc_stat;
7369 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7370 rpc_stat = (rx_interface_stat_p) fix_offset;
7373 * Copy the data based upon the caller version
7375 rx_MarshallProcessRPCStats(callerVersion,
7376 rpc_stat->stats[0].func_total,
7377 rpc_stat->stats, &ptr);
7383 MUTEX_EXIT(&rx_rpc_stats);
7388 * rx_FreeRPCStats - free memory allocated by
7389 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7393 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7394 * rx_RetrievePeerRPCStats
7396 * IN allocSize - the number of bytes in stats.
7404 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7406 rxi_Free(stats, allocSize);
7410 * rx_queryProcessRPCStats - see if process rpc stat collection is
7411 * currently enabled.
7417 * Returns 0 if stats are not enabled != 0 otherwise
7421 rx_queryProcessRPCStats(void)
7424 MUTEX_ENTER(&rx_rpc_stats);
7425 rc = rxi_monitor_processStats;
7426 MUTEX_EXIT(&rx_rpc_stats);
7431 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7437 * Returns 0 if stats are not enabled != 0 otherwise
7441 rx_queryPeerRPCStats(void)
7444 MUTEX_ENTER(&rx_rpc_stats);
7445 rc = rxi_monitor_peerStats;
7446 MUTEX_EXIT(&rx_rpc_stats);
7451 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7461 rx_enableProcessRPCStats(void)
7463 MUTEX_ENTER(&rx_rpc_stats);
7464 rx_enable_stats = 1;
7465 rxi_monitor_processStats = 1;
7466 MUTEX_EXIT(&rx_rpc_stats);
7470 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7480 rx_enablePeerRPCStats(void)
7482 MUTEX_ENTER(&rx_rpc_stats);
7483 rx_enable_stats = 1;
7484 rxi_monitor_peerStats = 1;
7485 MUTEX_EXIT(&rx_rpc_stats);
7489 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7499 rx_disableProcessRPCStats(void)
7501 rx_interface_stat_p rpc_stat, nrpc_stat;
7504 MUTEX_ENTER(&rx_rpc_stats);
7507 * Turn off process statistics and if peer stats is also off, turn
7511 rxi_monitor_processStats = 0;
7512 if (rxi_monitor_peerStats == 0) {
7513 rx_enable_stats = 0;
7516 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7517 unsigned int num_funcs = 0;
7520 queue_Remove(rpc_stat);
7521 num_funcs = rpc_stat->stats[0].func_total;
7523 sizeof(rx_interface_stat_t) +
7524 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7526 rxi_Free(rpc_stat, space);
7527 rxi_rpc_process_stat_cnt -= num_funcs;
7529 MUTEX_EXIT(&rx_rpc_stats);
7533 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7543 rx_disablePeerRPCStats(void)
7545 struct rx_peer **peer_ptr, **peer_end;
7548 MUTEX_ENTER(&rx_rpc_stats);
7551 * Turn off peer statistics and if process stats is also off, turn
7555 rxi_monitor_peerStats = 0;
7556 if (rxi_monitor_processStats == 0) {
7557 rx_enable_stats = 0;
7560 MUTEX_ENTER(&rx_peerHashTable_lock);
7561 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7562 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7564 struct rx_peer *peer, *next, *prev;
7565 for (prev = peer = *peer_ptr; peer; peer = next) {
7567 code = MUTEX_TRYENTER(&peer->peer_lock);
7569 rx_interface_stat_p rpc_stat, nrpc_stat;
7572 (&peer->rpcStats, rpc_stat, nrpc_stat,
7573 rx_interface_stat)) {
7574 unsigned int num_funcs = 0;
7577 queue_Remove(&rpc_stat->queue_header);
7578 queue_Remove(&rpc_stat->all_peers);
7579 num_funcs = rpc_stat->stats[0].func_total;
7581 sizeof(rx_interface_stat_t) +
7582 rpc_stat->stats[0].func_total *
7583 sizeof(rx_function_entry_v1_t);
7585 rxi_Free(rpc_stat, space);
7586 rxi_rpc_peer_stat_cnt -= num_funcs;
7588 MUTEX_EXIT(&peer->peer_lock);
7589 if (prev == *peer_ptr) {
7599 MUTEX_EXIT(&rx_peerHashTable_lock);
7600 MUTEX_EXIT(&rx_rpc_stats);
7604 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7609 * IN clearFlag - flag indicating which stats to clear
7617 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7619 rx_interface_stat_p rpc_stat, nrpc_stat;
7621 MUTEX_ENTER(&rx_rpc_stats);
7623 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7624 unsigned int num_funcs = 0, i;
7625 num_funcs = rpc_stat->stats[0].func_total;
7626 for (i = 0; i < num_funcs; i++) {
7627 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7628 hzero(rpc_stat->stats[i].invocations);
7630 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7631 hzero(rpc_stat->stats[i].bytes_sent);
7633 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7634 hzero(rpc_stat->stats[i].bytes_rcvd);
7636 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7637 rpc_stat->stats[i].queue_time_sum.sec = 0;
7638 rpc_stat->stats[i].queue_time_sum.usec = 0;
7640 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7641 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7642 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7644 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7645 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7646 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7648 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7649 rpc_stat->stats[i].queue_time_max.sec = 0;
7650 rpc_stat->stats[i].queue_time_max.usec = 0;
7652 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7653 rpc_stat->stats[i].execution_time_sum.sec = 0;
7654 rpc_stat->stats[i].execution_time_sum.usec = 0;
7656 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7657 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7658 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7660 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7661 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7662 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7664 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7665 rpc_stat->stats[i].execution_time_max.sec = 0;
7666 rpc_stat->stats[i].execution_time_max.usec = 0;
7671 MUTEX_EXIT(&rx_rpc_stats);
7675 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7680 * IN clearFlag - flag indicating which stats to clear
7688 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7690 rx_interface_stat_p rpc_stat, nrpc_stat;
7692 MUTEX_ENTER(&rx_rpc_stats);
7694 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7695 unsigned int num_funcs = 0, i;
7698 * We have to fix the offset of rpc_stat since we are
7699 * keeping this structure on two rx_queues. The rx_queue
7700 * package assumes that the rx_queue member is the first
7701 * member of the structure. That is, rx_queue assumes that
7702 * any one item is only on one queue at a time. We are
7703 * breaking that assumption and so we have to do a little
7704 * math to fix our pointers.
7707 fix_offset = (char *)rpc_stat;
7708 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7709 rpc_stat = (rx_interface_stat_p) fix_offset;
7711 num_funcs = rpc_stat->stats[0].func_total;
7712 for (i = 0; i < num_funcs; i++) {
7713 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7714 hzero(rpc_stat->stats[i].invocations);
7716 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7717 hzero(rpc_stat->stats[i].bytes_sent);
7719 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7720 hzero(rpc_stat->stats[i].bytes_rcvd);
7722 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7723 rpc_stat->stats[i].queue_time_sum.sec = 0;
7724 rpc_stat->stats[i].queue_time_sum.usec = 0;
7726 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7727 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7728 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7730 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7731 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7732 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7734 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7735 rpc_stat->stats[i].queue_time_max.sec = 0;
7736 rpc_stat->stats[i].queue_time_max.usec = 0;
7738 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7739 rpc_stat->stats[i].execution_time_sum.sec = 0;
7740 rpc_stat->stats[i].execution_time_sum.usec = 0;
7742 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7743 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7744 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7746 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7747 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7748 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7750 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7751 rpc_stat->stats[i].execution_time_max.sec = 0;
7752 rpc_stat->stats[i].execution_time_max.usec = 0;
7757 MUTEX_EXIT(&rx_rpc_stats);
7761 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7762 * is authorized to enable/disable/clear RX statistics.
7764 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7767 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7769 rxi_rxstat_userok = proc;
7773 rx_RxStatUserOk(struct rx_call *call)
7775 if (!rxi_rxstat_userok)
7777 return rxi_rxstat_userok(call);
7782 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
7783 * function in the MSVC runtime DLL (msvcrt.dll).
7785 * Note: the system serializes calls to this function.
7788 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
7789 DWORD reason, /* reason function is being called */
7790 LPVOID reserved) /* reserved for future use */
7793 case DLL_PROCESS_ATTACH:
7794 /* library is being attached to a process */
7798 case DLL_PROCESS_DETACH: