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 = 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(NULL); /* Never returns */
723 #ifdef RX_ENABLE_TSFPQ
724 /* no use leaving packets around in this thread's local queue if
725 * it isn't getting donated to the server thread pool.
727 rxi_FlushLocalPacketsTSFPQ();
728 #endif /* RX_ENABLE_TSFPQ */
732 /* Create a new client connection to the specified service, using the
733 * specified security object to implement the security model for this
735 struct rx_connection *
736 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
737 register struct rx_securityClass *securityObject,
738 int serviceSecurityIndex)
742 register struct rx_connection *conn;
747 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", ntohl(shost), ntohs(sport), sservice, securityObject, serviceSecurityIndex));
749 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
750 * the case of kmem_alloc? */
751 conn = rxi_AllocConnection();
752 #ifdef RX_ENABLE_LOCKS
753 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
754 MUTEX_INIT(&conn->conn_data_lock, "conn call lock", MUTEX_DEFAULT, 0);
755 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
758 MUTEX_ENTER(&rx_connHashTable_lock);
759 cid = (rx_nextCid += RX_MAXCALLS);
760 conn->type = RX_CLIENT_CONNECTION;
762 conn->epoch = rx_epoch;
763 conn->peer = rxi_FindPeer(shost, sport, 0, 1);
764 conn->serviceId = sservice;
765 conn->securityObject = securityObject;
766 conn->securityData = (void *) 0;
767 conn->securityIndex = serviceSecurityIndex;
768 rx_SetConnDeadTime(conn, rx_connDeadTime);
769 conn->ackRate = RX_FAST_ACK_RATE;
771 conn->specific = NULL;
772 conn->challengeEvent = NULL;
773 conn->delayedAbortEvent = NULL;
774 conn->abortCount = 0;
777 RXS_NewConnection(securityObject, conn);
779 CONN_HASH(shost, sport, conn->cid, conn->epoch, RX_CLIENT_CONNECTION);
781 conn->refCount++; /* no lock required since only this thread knows... */
782 conn->next = rx_connHashTable[hashindex];
783 rx_connHashTable[hashindex] = conn;
784 rx_MutexIncrement(rx_stats.nClientConns, rx_stats_mutex);
785 MUTEX_EXIT(&rx_connHashTable_lock);
791 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
793 /* The idea is to set the dead time to a value that allows several
794 * keepalives to be dropped without timing out the connection. */
795 conn->secondsUntilDead = MAX(seconds, 6);
796 conn->secondsUntilPing = conn->secondsUntilDead / 6;
799 int rxi_lowPeerRefCount = 0;
800 int rxi_lowConnRefCount = 0;
803 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
804 * NOTE: must not be called with rx_connHashTable_lock held.
807 rxi_CleanupConnection(struct rx_connection *conn)
809 /* Notify the service exporter, if requested, that this connection
810 * is being destroyed */
811 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
812 (*conn->service->destroyConnProc) (conn);
814 /* Notify the security module that this connection is being destroyed */
815 RXS_DestroyConnection(conn->securityObject, conn);
817 /* If this is the last connection using the rx_peer struct, set its
818 * idle time to now. rxi_ReapConnections will reap it if it's still
819 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
821 MUTEX_ENTER(&rx_peerHashTable_lock);
822 if (conn->peer->refCount < 2) {
823 conn->peer->idleWhen = clock_Sec();
824 if (conn->peer->refCount < 1) {
825 conn->peer->refCount = 1;
826 MUTEX_ENTER(&rx_stats_mutex);
827 rxi_lowPeerRefCount++;
828 MUTEX_EXIT(&rx_stats_mutex);
831 conn->peer->refCount--;
832 MUTEX_EXIT(&rx_peerHashTable_lock);
834 if (conn->type == RX_SERVER_CONNECTION)
835 rx_MutexDecrement(rx_stats.nServerConns, rx_stats_mutex);
837 rx_MutexDecrement(rx_stats.nClientConns, rx_stats_mutex);
839 if (conn->specific) {
841 for (i = 0; i < conn->nSpecific; i++) {
842 if (conn->specific[i] && rxi_keyCreate_destructor[i])
843 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
844 conn->specific[i] = NULL;
846 free(conn->specific);
848 conn->specific = NULL;
852 MUTEX_DESTROY(&conn->conn_call_lock);
853 MUTEX_DESTROY(&conn->conn_data_lock);
854 CV_DESTROY(&conn->conn_call_cv);
856 rxi_FreeConnection(conn);
859 /* Destroy the specified connection */
861 rxi_DestroyConnection(register struct rx_connection *conn)
863 MUTEX_ENTER(&rx_connHashTable_lock);
864 rxi_DestroyConnectionNoLock(conn);
865 /* conn should be at the head of the cleanup list */
866 if (conn == rx_connCleanup_list) {
867 rx_connCleanup_list = rx_connCleanup_list->next;
868 MUTEX_EXIT(&rx_connHashTable_lock);
869 rxi_CleanupConnection(conn);
871 #ifdef RX_ENABLE_LOCKS
873 MUTEX_EXIT(&rx_connHashTable_lock);
875 #endif /* RX_ENABLE_LOCKS */
879 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
881 register struct rx_connection **conn_ptr;
882 register int havecalls = 0;
883 struct rx_packet *packet;
890 MUTEX_ENTER(&conn->conn_data_lock);
891 if (conn->refCount > 0)
894 MUTEX_ENTER(&rx_stats_mutex);
895 rxi_lowConnRefCount++;
896 MUTEX_EXIT(&rx_stats_mutex);
899 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
900 /* Busy; wait till the last guy before proceeding */
901 MUTEX_EXIT(&conn->conn_data_lock);
906 /* If the client previously called rx_NewCall, but it is still
907 * waiting, treat this as a running call, and wait to destroy the
908 * connection later when the call completes. */
909 if ((conn->type == RX_CLIENT_CONNECTION)
910 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
911 conn->flags |= RX_CONN_DESTROY_ME;
912 MUTEX_EXIT(&conn->conn_data_lock);
916 MUTEX_EXIT(&conn->conn_data_lock);
918 /* Check for extant references to this connection */
919 for (i = 0; i < RX_MAXCALLS; i++) {
920 register struct rx_call *call = conn->call[i];
923 if (conn->type == RX_CLIENT_CONNECTION) {
924 MUTEX_ENTER(&call->lock);
925 if (call->delayedAckEvent) {
926 /* Push the final acknowledgment out now--there
927 * won't be a subsequent call to acknowledge the
928 * last reply packets */
929 rxevent_Cancel(call->delayedAckEvent, call,
930 RX_CALL_REFCOUNT_DELAY);
931 if (call->state == RX_STATE_PRECALL
932 || call->state == RX_STATE_ACTIVE) {
933 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
935 rxi_AckAll(NULL, call, 0);
938 MUTEX_EXIT(&call->lock);
942 #ifdef RX_ENABLE_LOCKS
944 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
945 MUTEX_EXIT(&conn->conn_data_lock);
947 /* Someone is accessing a packet right now. */
951 #endif /* RX_ENABLE_LOCKS */
954 /* Don't destroy the connection if there are any call
955 * structures still in use */
956 MUTEX_ENTER(&conn->conn_data_lock);
957 conn->flags |= RX_CONN_DESTROY_ME;
958 MUTEX_EXIT(&conn->conn_data_lock);
963 if (conn->delayedAbortEvent) {
964 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
965 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
967 MUTEX_ENTER(&conn->conn_data_lock);
968 rxi_SendConnectionAbort(conn, packet, 0, 1);
969 MUTEX_EXIT(&conn->conn_data_lock);
970 rxi_FreePacket(packet);
974 /* Remove from connection hash table before proceeding */
976 &rx_connHashTable[CONN_HASH
977 (peer->host, peer->port, conn->cid, conn->epoch,
979 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
980 if (*conn_ptr == conn) {
981 *conn_ptr = conn->next;
985 /* if the conn that we are destroying was the last connection, then we
986 * clear rxLastConn as well */
987 if (rxLastConn == conn)
990 /* Make sure the connection is completely reset before deleting it. */
991 /* get rid of pending events that could zap us later */
992 if (conn->challengeEvent)
993 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
994 if (conn->checkReachEvent)
995 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
997 /* Add the connection to the list of destroyed connections that
998 * need to be cleaned up. This is necessary to avoid deadlocks
999 * in the routines we call to inform others that this connection is
1000 * being destroyed. */
1001 conn->next = rx_connCleanup_list;
1002 rx_connCleanup_list = conn;
1005 /* Externally available version */
1007 rx_DestroyConnection(register struct rx_connection *conn)
1012 rxi_DestroyConnection(conn);
1017 rx_GetConnection(register struct rx_connection *conn)
1022 MUTEX_ENTER(&conn->conn_data_lock);
1024 MUTEX_EXIT(&conn->conn_data_lock);
1028 /* Wait for the transmit queue to no longer be busy.
1029 * requires the call->lock to be held */
1030 static void rxi_WaitforTQBusy(struct rx_call *call) {
1031 while (call->flags & RX_CALL_TQ_BUSY) {
1032 call->flags |= RX_CALL_TQ_WAIT;
1034 #ifdef RX_ENABLE_LOCKS
1035 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1036 CV_WAIT(&call->cv_tq, &call->lock);
1037 #else /* RX_ENABLE_LOCKS */
1038 osi_rxSleep(&call->tq);
1039 #endif /* RX_ENABLE_LOCKS */
1041 if (call->tqWaiters == 0) {
1042 call->flags &= ~RX_CALL_TQ_WAIT;
1046 /* Start a new rx remote procedure call, on the specified connection.
1047 * If wait is set to 1, wait for a free call channel; otherwise return
1048 * 0. Maxtime gives the maximum number of seconds this call may take,
1049 * after rx_NewCall returns. After this time interval, a call to any
1050 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1051 * For fine grain locking, we hold the conn_call_lock in order to
1052 * to ensure that we don't get signalle after we found a call in an active
1053 * state and before we go to sleep.
1056 rx_NewCall(register struct rx_connection *conn)
1059 register struct rx_call *call;
1060 struct clock queueTime;
1064 dpf(("rx_NewCall(conn %x)\n", conn));
1067 clock_GetTime(&queueTime);
1068 MUTEX_ENTER(&conn->conn_call_lock);
1071 * Check if there are others waiting for a new call.
1072 * If so, let them go first to avoid starving them.
1073 * This is a fairly simple scheme, and might not be
1074 * a complete solution for large numbers of waiters.
1076 * makeCallWaiters keeps track of the number of
1077 * threads waiting to make calls and the
1078 * RX_CONN_MAKECALL_WAITING flag bit is used to
1079 * indicate that there are indeed calls waiting.
1080 * The flag is set when the waiter is incremented.
1081 * It is only cleared in rx_EndCall when
1082 * makeCallWaiters is 0. This prevents us from
1083 * accidently destroying the connection while it
1084 * is potentially about to be used.
1086 MUTEX_ENTER(&conn->conn_data_lock);
1087 if (conn->makeCallWaiters) {
1088 conn->flags |= RX_CONN_MAKECALL_WAITING;
1089 conn->makeCallWaiters++;
1090 MUTEX_EXIT(&conn->conn_data_lock);
1092 #ifdef RX_ENABLE_LOCKS
1093 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1097 MUTEX_ENTER(&conn->conn_data_lock);
1098 conn->makeCallWaiters--;
1100 MUTEX_EXIT(&conn->conn_data_lock);
1103 for (i = 0; i < RX_MAXCALLS; i++) {
1104 call = conn->call[i];
1106 MUTEX_ENTER(&call->lock);
1107 if (call->state == RX_STATE_DALLY) {
1108 rxi_ResetCall(call, 0);
1109 (*call->callNumber)++;
1112 MUTEX_EXIT(&call->lock);
1114 call = rxi_NewCall(conn, i);
1118 if (i < RX_MAXCALLS) {
1121 MUTEX_ENTER(&conn->conn_data_lock);
1122 conn->flags |= RX_CONN_MAKECALL_WAITING;
1123 conn->makeCallWaiters++;
1124 MUTEX_EXIT(&conn->conn_data_lock);
1126 #ifdef RX_ENABLE_LOCKS
1127 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1131 MUTEX_ENTER(&conn->conn_data_lock);
1132 conn->makeCallWaiters--;
1133 MUTEX_EXIT(&conn->conn_data_lock);
1136 * Wake up anyone else who might be giving us a chance to
1137 * run (see code above that avoids resource starvation).
1139 #ifdef RX_ENABLE_LOCKS
1140 CV_BROADCAST(&conn->conn_call_cv);
1145 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1147 /* Client is initially in send mode */
1148 call->state = RX_STATE_ACTIVE;
1149 call->error = conn->error;
1151 call->mode = RX_MODE_ERROR;
1153 call->mode = RX_MODE_SENDING;
1155 /* remember start time for call in case we have hard dead time limit */
1156 call->queueTime = queueTime;
1157 clock_GetTime(&call->startTime);
1158 hzero(call->bytesSent);
1159 hzero(call->bytesRcvd);
1161 /* Turn on busy protocol. */
1162 rxi_KeepAliveOn(call);
1164 MUTEX_EXIT(&call->lock);
1165 MUTEX_EXIT(&conn->conn_call_lock);
1168 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1169 /* Now, if TQ wasn't cleared earlier, do it now. */
1170 MUTEX_ENTER(&call->lock);
1171 rxi_WaitforTQBusy(call);
1172 if (call->flags & RX_CALL_TQ_CLEARME) {
1173 rxi_ClearTransmitQueue(call, 0);
1174 queue_Init(&call->tq);
1176 MUTEX_EXIT(&call->lock);
1177 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1179 dpf(("rx_NewCall(call %x)\n", call));
1184 rxi_HasActiveCalls(register struct rx_connection *aconn)
1187 register struct rx_call *tcall;
1191 for (i = 0; i < RX_MAXCALLS; i++) {
1192 if ((tcall = aconn->call[i])) {
1193 if ((tcall->state == RX_STATE_ACTIVE)
1194 || (tcall->state == RX_STATE_PRECALL)) {
1205 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1206 register afs_int32 * aint32s)
1209 register struct rx_call *tcall;
1213 for (i = 0; i < RX_MAXCALLS; i++) {
1214 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1215 aint32s[i] = aconn->callNumber[i] + 1;
1217 aint32s[i] = aconn->callNumber[i];
1224 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1225 register afs_int32 * aint32s)
1228 register struct rx_call *tcall;
1232 for (i = 0; i < RX_MAXCALLS; i++) {
1233 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1234 aconn->callNumber[i] = aint32s[i] - 1;
1236 aconn->callNumber[i] = aint32s[i];
1242 /* Advertise a new service. A service is named locally by a UDP port
1243 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1246 char *serviceName; Name for identification purposes (e.g. the
1247 service name might be used for probing for
1250 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1251 char *serviceName, struct rx_securityClass **securityObjects,
1252 int nSecurityObjects,
1253 afs_int32(*serviceProc) (struct rx_call * acall))
1255 osi_socket socket = OSI_NULLSOCKET;
1256 register struct rx_service *tservice;
1262 if (serviceId == 0) {
1264 "rx_NewService: service id for service %s is not non-zero.\n",
1271 "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",
1279 tservice = rxi_AllocService();
1281 for (i = 0; i < RX_MAX_SERVICES; i++) {
1282 register struct rx_service *service = rx_services[i];
1284 if (port == service->servicePort && host == service->serviceHost) {
1285 if (service->serviceId == serviceId) {
1286 /* The identical service has already been
1287 * installed; if the caller was intending to
1288 * change the security classes used by this
1289 * service, he/she loses. */
1291 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1292 serviceName, serviceId, service->serviceName);
1294 rxi_FreeService(tservice);
1297 /* Different service, same port: re-use the socket
1298 * which is bound to the same port */
1299 socket = service->socket;
1302 if (socket == OSI_NULLSOCKET) {
1303 /* If we don't already have a socket (from another
1304 * service on same port) get a new one */
1305 socket = rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
1306 if (socket == OSI_NULLSOCKET) {
1308 rxi_FreeService(tservice);
1313 service->socket = socket;
1314 service->serviceHost = host;
1315 service->servicePort = port;
1316 service->serviceId = serviceId;
1317 service->serviceName = serviceName;
1318 service->nSecurityObjects = nSecurityObjects;
1319 service->securityObjects = securityObjects;
1320 service->minProcs = 0;
1321 service->maxProcs = 1;
1322 service->idleDeadTime = 60;
1323 service->connDeadTime = rx_connDeadTime;
1324 service->executeRequestProc = serviceProc;
1325 service->checkReach = 0;
1326 rx_services[i] = service; /* not visible until now */
1332 rxi_FreeService(tservice);
1333 (osi_Msg "rx_NewService: cannot support > %d services\n",
1338 /* Set configuration options for all of a service's security objects */
1341 rx_SetSecurityConfiguration(struct rx_service *service,
1342 rx_securityConfigVariables type,
1346 for (i = 0; i<service->nSecurityObjects; i++) {
1347 if (service->securityObjects[i]) {
1348 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1356 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1357 struct rx_securityClass **securityObjects, int nSecurityObjects,
1358 afs_int32(*serviceProc) (struct rx_call * acall))
1360 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1363 /* Generic request processing loop. This routine should be called
1364 * by the implementation dependent rx_ServerProc. If socketp is
1365 * non-null, it will be set to the file descriptor that this thread
1366 * is now listening on. If socketp is null, this routine will never
1369 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1371 register struct rx_call *call;
1372 register afs_int32 code;
1373 register struct rx_service *tservice = NULL;
1380 call = rx_GetCall(threadID, tservice, socketp);
1381 if (socketp && *socketp != OSI_NULLSOCKET) {
1382 /* We are now a listener thread */
1387 /* if server is restarting( typically smooth shutdown) then do not
1388 * allow any new calls.
1391 if (rx_tranquil && (call != NULL)) {
1395 MUTEX_ENTER(&call->lock);
1397 rxi_CallError(call, RX_RESTARTING);
1398 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1400 MUTEX_EXIT(&call->lock);
1404 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1405 #ifdef RX_ENABLE_LOCKS
1407 #endif /* RX_ENABLE_LOCKS */
1408 afs_termState = AFSOP_STOP_AFS;
1409 afs_osi_Wakeup(&afs_termState);
1410 #ifdef RX_ENABLE_LOCKS
1412 #endif /* RX_ENABLE_LOCKS */
1417 tservice = call->conn->service;
1419 if (tservice->beforeProc)
1420 (*tservice->beforeProc) (call);
1422 code = call->conn->service->executeRequestProc(call);
1424 if (tservice->afterProc)
1425 (*tservice->afterProc) (call, code);
1427 rx_EndCall(call, code);
1428 MUTEX_ENTER(&rx_stats_mutex);
1430 MUTEX_EXIT(&rx_stats_mutex);
1436 rx_WakeupServerProcs(void)
1438 struct rx_serverQueueEntry *np, *tqp;
1442 MUTEX_ENTER(&rx_serverPool_lock);
1444 #ifdef RX_ENABLE_LOCKS
1445 if (rx_waitForPacket)
1446 CV_BROADCAST(&rx_waitForPacket->cv);
1447 #else /* RX_ENABLE_LOCKS */
1448 if (rx_waitForPacket)
1449 osi_rxWakeup(rx_waitForPacket);
1450 #endif /* RX_ENABLE_LOCKS */
1451 MUTEX_ENTER(&freeSQEList_lock);
1452 for (np = rx_FreeSQEList; np; np = tqp) {
1453 tqp = *(struct rx_serverQueueEntry **)np;
1454 #ifdef RX_ENABLE_LOCKS
1455 CV_BROADCAST(&np->cv);
1456 #else /* RX_ENABLE_LOCKS */
1458 #endif /* RX_ENABLE_LOCKS */
1460 MUTEX_EXIT(&freeSQEList_lock);
1461 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1462 #ifdef RX_ENABLE_LOCKS
1463 CV_BROADCAST(&np->cv);
1464 #else /* RX_ENABLE_LOCKS */
1466 #endif /* RX_ENABLE_LOCKS */
1468 MUTEX_EXIT(&rx_serverPool_lock);
1473 * One thing that seems to happen is that all the server threads get
1474 * tied up on some empty or slow call, and then a whole bunch of calls
1475 * arrive at once, using up the packet pool, so now there are more
1476 * empty calls. The most critical resources here are server threads
1477 * and the free packet pool. The "doreclaim" code seems to help in
1478 * general. I think that eventually we arrive in this state: there
1479 * are lots of pending calls which do have all their packets present,
1480 * so they won't be reclaimed, are multi-packet calls, so they won't
1481 * be scheduled until later, and thus are tying up most of the free
1482 * packet pool for a very long time.
1484 * 1. schedule multi-packet calls if all the packets are present.
1485 * Probably CPU-bound operation, useful to return packets to pool.
1486 * Do what if there is a full window, but the last packet isn't here?
1487 * 3. preserve one thread which *only* runs "best" calls, otherwise
1488 * it sleeps and waits for that type of call.
1489 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1490 * the current dataquota business is badly broken. The quota isn't adjusted
1491 * to reflect how many packets are presently queued for a running call.
1492 * So, when we schedule a queued call with a full window of packets queued
1493 * up for it, that *should* free up a window full of packets for other 2d-class
1494 * calls to be able to use from the packet pool. But it doesn't.
1496 * NB. Most of the time, this code doesn't run -- since idle server threads
1497 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1498 * as a new call arrives.
1500 /* Sleep until a call arrives. Returns a pointer to the call, ready
1501 * for an rx_Read. */
1502 #ifdef RX_ENABLE_LOCKS
1504 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1506 struct rx_serverQueueEntry *sq;
1507 register struct rx_call *call = (struct rx_call *)0;
1508 struct rx_service *service = NULL;
1511 MUTEX_ENTER(&freeSQEList_lock);
1513 if ((sq = rx_FreeSQEList)) {
1514 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1515 MUTEX_EXIT(&freeSQEList_lock);
1516 } else { /* otherwise allocate a new one and return that */
1517 MUTEX_EXIT(&freeSQEList_lock);
1518 sq = (struct rx_serverQueueEntry *)
1519 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1520 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1521 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1524 MUTEX_ENTER(&rx_serverPool_lock);
1525 if (cur_service != NULL) {
1526 ReturnToServerPool(cur_service);
1529 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1530 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1532 /* Scan for eligible incoming calls. A call is not eligible
1533 * if the maximum number of calls for its service type are
1534 * already executing */
1535 /* One thread will process calls FCFS (to prevent starvation),
1536 * while the other threads may run ahead looking for calls which
1537 * have all their input data available immediately. This helps
1538 * keep threads from blocking, waiting for data from the client. */
1539 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1540 service = tcall->conn->service;
1541 if (!QuotaOK(service)) {
1544 if (tno == rxi_fcfs_thread_num
1545 || !tcall->queue_item_header.next) {
1546 /* If we're the fcfs thread , then we'll just use
1547 * this call. If we haven't been able to find an optimal
1548 * choice, and we're at the end of the list, then use a
1549 * 2d choice if one has been identified. Otherwise... */
1550 call = (choice2 ? choice2 : tcall);
1551 service = call->conn->service;
1552 } else if (!queue_IsEmpty(&tcall->rq)) {
1553 struct rx_packet *rp;
1554 rp = queue_First(&tcall->rq, rx_packet);
1555 if (rp->header.seq == 1) {
1557 || (rp->header.flags & RX_LAST_PACKET)) {
1559 } else if (rxi_2dchoice && !choice2
1560 && !(tcall->flags & RX_CALL_CLEARED)
1561 && (tcall->rprev > rxi_HardAckRate)) {
1570 ReturnToServerPool(service);
1577 MUTEX_EXIT(&rx_serverPool_lock);
1578 MUTEX_ENTER(&call->lock);
1580 if (call->flags & RX_CALL_WAIT_PROC) {
1581 call->flags &= ~RX_CALL_WAIT_PROC;
1582 MUTEX_ENTER(&rx_stats_mutex);
1584 MUTEX_EXIT(&rx_stats_mutex);
1587 if (call->state != RX_STATE_PRECALL || call->error) {
1588 MUTEX_EXIT(&call->lock);
1589 MUTEX_ENTER(&rx_serverPool_lock);
1590 ReturnToServerPool(service);
1595 if (queue_IsEmpty(&call->rq)
1596 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1597 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1599 CLEAR_CALL_QUEUE_LOCK(call);
1602 /* If there are no eligible incoming calls, add this process
1603 * to the idle server queue, to wait for one */
1607 *socketp = OSI_NULLSOCKET;
1609 sq->socketp = socketp;
1610 queue_Append(&rx_idleServerQueue, sq);
1611 #ifndef AFS_AIX41_ENV
1612 rx_waitForPacket = sq;
1614 rx_waitingForPacket = sq;
1615 #endif /* AFS_AIX41_ENV */
1617 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1619 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1620 MUTEX_EXIT(&rx_serverPool_lock);
1621 return (struct rx_call *)0;
1624 } while (!(call = sq->newcall)
1625 && !(socketp && *socketp != OSI_NULLSOCKET));
1626 MUTEX_EXIT(&rx_serverPool_lock);
1628 MUTEX_ENTER(&call->lock);
1634 MUTEX_ENTER(&freeSQEList_lock);
1635 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1636 rx_FreeSQEList = sq;
1637 MUTEX_EXIT(&freeSQEList_lock);
1640 clock_GetTime(&call->startTime);
1641 call->state = RX_STATE_ACTIVE;
1642 call->mode = RX_MODE_RECEIVING;
1643 #ifdef RX_KERNEL_TRACE
1644 if (ICL_SETACTIVE(afs_iclSetp)) {
1645 int glockOwner = ISAFS_GLOCK();
1648 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1649 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1656 rxi_calltrace(RX_CALL_START, call);
1657 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1658 call->conn->service->servicePort, call->conn->service->serviceId,
1661 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1662 MUTEX_EXIT(&call->lock);
1664 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1669 #else /* RX_ENABLE_LOCKS */
1671 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1673 struct rx_serverQueueEntry *sq;
1674 register struct rx_call *call = (struct rx_call *)0, *choice2;
1675 struct rx_service *service = NULL;
1679 MUTEX_ENTER(&freeSQEList_lock);
1681 if ((sq = rx_FreeSQEList)) {
1682 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1683 MUTEX_EXIT(&freeSQEList_lock);
1684 } else { /* otherwise allocate a new one and return that */
1685 MUTEX_EXIT(&freeSQEList_lock);
1686 sq = (struct rx_serverQueueEntry *)
1687 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1688 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1689 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1691 MUTEX_ENTER(&sq->lock);
1693 if (cur_service != NULL) {
1694 cur_service->nRequestsRunning--;
1695 if (cur_service->nRequestsRunning < cur_service->minProcs)
1699 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1700 register struct rx_call *tcall, *ncall;
1701 /* Scan for eligible incoming calls. A call is not eligible
1702 * if the maximum number of calls for its service type are
1703 * already executing */
1704 /* One thread will process calls FCFS (to prevent starvation),
1705 * while the other threads may run ahead looking for calls which
1706 * have all their input data available immediately. This helps
1707 * keep threads from blocking, waiting for data from the client. */
1708 choice2 = (struct rx_call *)0;
1709 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1710 service = tcall->conn->service;
1711 if (QuotaOK(service)) {
1712 if (tno == rxi_fcfs_thread_num
1713 || !tcall->queue_item_header.next) {
1714 /* If we're the fcfs thread, then we'll just use
1715 * this call. If we haven't been able to find an optimal
1716 * choice, and we're at the end of the list, then use a
1717 * 2d choice if one has been identified. Otherwise... */
1718 call = (choice2 ? choice2 : tcall);
1719 service = call->conn->service;
1720 } else if (!queue_IsEmpty(&tcall->rq)) {
1721 struct rx_packet *rp;
1722 rp = queue_First(&tcall->rq, rx_packet);
1723 if (rp->header.seq == 1
1725 || (rp->header.flags & RX_LAST_PACKET))) {
1727 } else if (rxi_2dchoice && !choice2
1728 && !(tcall->flags & RX_CALL_CLEARED)
1729 && (tcall->rprev > rxi_HardAckRate)) {
1742 /* we can't schedule a call if there's no data!!! */
1743 /* send an ack if there's no data, if we're missing the
1744 * first packet, or we're missing something between first
1745 * and last -- there's a "hole" in the incoming data. */
1746 if (queue_IsEmpty(&call->rq)
1747 || queue_First(&call->rq, rx_packet)->header.seq != 1
1748 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1749 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1751 call->flags &= (~RX_CALL_WAIT_PROC);
1752 service->nRequestsRunning++;
1753 /* just started call in minProcs pool, need fewer to maintain
1755 if (service->nRequestsRunning <= service->minProcs)
1759 /* MUTEX_EXIT(&call->lock); */
1761 /* If there are no eligible incoming calls, add this process
1762 * to the idle server queue, to wait for one */
1765 *socketp = OSI_NULLSOCKET;
1767 sq->socketp = socketp;
1768 queue_Append(&rx_idleServerQueue, sq);
1772 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1774 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1775 return (struct rx_call *)0;
1778 } while (!(call = sq->newcall)
1779 && !(socketp && *socketp != OSI_NULLSOCKET));
1781 MUTEX_EXIT(&sq->lock);
1783 MUTEX_ENTER(&freeSQEList_lock);
1784 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1785 rx_FreeSQEList = sq;
1786 MUTEX_EXIT(&freeSQEList_lock);
1789 clock_GetTime(&call->startTime);
1790 call->state = RX_STATE_ACTIVE;
1791 call->mode = RX_MODE_RECEIVING;
1792 #ifdef RX_KERNEL_TRACE
1793 if (ICL_SETACTIVE(afs_iclSetp)) {
1794 int glockOwner = ISAFS_GLOCK();
1797 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1798 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1805 rxi_calltrace(RX_CALL_START, call);
1806 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1807 call->conn->service->servicePort, call->conn->service->serviceId,
1810 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1817 #endif /* RX_ENABLE_LOCKS */
1821 /* Establish a procedure to be called when a packet arrives for a
1822 * call. This routine will be called at most once after each call,
1823 * and will also be called if there is an error condition on the or
1824 * the call is complete. Used by multi rx to build a selection
1825 * function which determines which of several calls is likely to be a
1826 * good one to read from.
1827 * NOTE: the way this is currently implemented it is probably only a
1828 * good idea to (1) use it immediately after a newcall (clients only)
1829 * and (2) only use it once. Other uses currently void your warranty
1832 rx_SetArrivalProc(register struct rx_call *call,
1833 register void (*proc) (register struct rx_call * call,
1835 register int index),
1836 register void * handle, register int arg)
1838 call->arrivalProc = proc;
1839 call->arrivalProcHandle = handle;
1840 call->arrivalProcArg = arg;
1843 /* Call is finished (possibly prematurely). Return rc to the peer, if
1844 * appropriate, and return the final error code from the conversation
1848 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1850 register struct rx_connection *conn = call->conn;
1851 register struct rx_service *service;
1857 dpf(("rx_EndCall(call %x rc %d error %d abortCode %d)\n", call, rc, call->error, call->abortCode));
1860 MUTEX_ENTER(&call->lock);
1862 if (rc == 0 && call->error == 0) {
1863 call->abortCode = 0;
1864 call->abortCount = 0;
1867 call->arrivalProc = (void (*)())0;
1868 if (rc && call->error == 0) {
1869 rxi_CallError(call, rc);
1870 /* Send an abort message to the peer if this error code has
1871 * only just been set. If it was set previously, assume the
1872 * peer has already been sent the error code or will request it
1874 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1876 if (conn->type == RX_SERVER_CONNECTION) {
1877 /* Make sure reply or at least dummy reply is sent */
1878 if (call->mode == RX_MODE_RECEIVING) {
1879 rxi_WriteProc(call, 0, 0);
1881 if (call->mode == RX_MODE_SENDING) {
1882 rxi_FlushWrite(call);
1884 service = conn->service;
1885 rxi_calltrace(RX_CALL_END, call);
1886 /* Call goes to hold state until reply packets are acknowledged */
1887 if (call->tfirst + call->nSoftAcked < call->tnext) {
1888 call->state = RX_STATE_HOLD;
1890 call->state = RX_STATE_DALLY;
1891 rxi_ClearTransmitQueue(call, 0);
1892 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1893 rxevent_Cancel(call->keepAliveEvent, call,
1894 RX_CALL_REFCOUNT_ALIVE);
1896 } else { /* Client connection */
1898 /* Make sure server receives input packets, in the case where
1899 * no reply arguments are expected */
1900 if ((call->mode == RX_MODE_SENDING)
1901 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
1902 (void)rxi_ReadProc(call, &dummy, 1);
1905 /* If we had an outstanding delayed ack, be nice to the server
1906 * and force-send it now.
1908 if (call->delayedAckEvent) {
1909 rxevent_Cancel(call->delayedAckEvent, call,
1910 RX_CALL_REFCOUNT_DELAY);
1911 call->delayedAckEvent = NULL;
1912 rxi_SendDelayedAck(NULL, call, NULL);
1915 /* We need to release the call lock since it's lower than the
1916 * conn_call_lock and we don't want to hold the conn_call_lock
1917 * over the rx_ReadProc call. The conn_call_lock needs to be held
1918 * here for the case where rx_NewCall is perusing the calls on
1919 * the connection structure. We don't want to signal until
1920 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
1921 * have checked this call, found it active and by the time it
1922 * goes to sleep, will have missed the signal.
1924 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
1925 * there are threads waiting to use the conn object.
1927 MUTEX_EXIT(&call->lock);
1928 MUTEX_ENTER(&conn->conn_call_lock);
1929 MUTEX_ENTER(&call->lock);
1930 MUTEX_ENTER(&conn->conn_data_lock);
1931 conn->flags |= RX_CONN_BUSY;
1932 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
1933 if (conn->makeCallWaiters == 0)
1934 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
1935 MUTEX_EXIT(&conn->conn_data_lock);
1936 #ifdef RX_ENABLE_LOCKS
1937 CV_BROADCAST(&conn->conn_call_cv);
1942 #ifdef RX_ENABLE_LOCKS
1944 MUTEX_EXIT(&conn->conn_data_lock);
1946 #endif /* RX_ENABLE_LOCKS */
1947 call->state = RX_STATE_DALLY;
1949 error = call->error;
1951 /* currentPacket, nLeft, and NFree must be zeroed here, because
1952 * ResetCall cannot: ResetCall may be called at splnet(), in the
1953 * kernel version, and may interrupt the macros rx_Read or
1954 * rx_Write, which run at normal priority for efficiency. */
1955 if (call->currentPacket) {
1956 queue_Prepend(&call->iovq, call->currentPacket);
1957 call->currentPacket = (struct rx_packet *)0;
1960 call->nLeft = call->nFree = call->curlen = 0;
1962 /* Free any packets from the last call to ReadvProc/WritevProc */
1963 rxi_FreePackets(0, &call->iovq);
1965 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
1966 MUTEX_EXIT(&call->lock);
1967 if (conn->type == RX_CLIENT_CONNECTION) {
1968 MUTEX_EXIT(&conn->conn_call_lock);
1969 conn->flags &= ~RX_CONN_BUSY;
1973 * Map errors to the local host's errno.h format.
1975 error = ntoh_syserr_conv(error);
1979 #if !defined(KERNEL)
1981 /* Call this routine when shutting down a server or client (especially
1982 * clients). This will allow Rx to gracefully garbage collect server
1983 * connections, and reduce the number of retries that a server might
1984 * make to a dead client.
1985 * This is not quite right, since some calls may still be ongoing and
1986 * we can't lock them to destroy them. */
1990 register struct rx_connection **conn_ptr, **conn_end;
1994 if (rxinit_status == 1) {
1996 return; /* Already shutdown. */
1998 rxi_DeleteCachedConnections();
1999 if (rx_connHashTable) {
2000 MUTEX_ENTER(&rx_connHashTable_lock);
2001 for (conn_ptr = &rx_connHashTable[0], conn_end =
2002 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2004 struct rx_connection *conn, *next;
2005 for (conn = *conn_ptr; conn; conn = next) {
2007 if (conn->type == RX_CLIENT_CONNECTION) {
2008 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2010 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2011 #ifdef RX_ENABLE_LOCKS
2012 rxi_DestroyConnectionNoLock(conn);
2013 #else /* RX_ENABLE_LOCKS */
2014 rxi_DestroyConnection(conn);
2015 #endif /* RX_ENABLE_LOCKS */
2019 #ifdef RX_ENABLE_LOCKS
2020 while (rx_connCleanup_list) {
2021 struct rx_connection *conn;
2022 conn = rx_connCleanup_list;
2023 rx_connCleanup_list = rx_connCleanup_list->next;
2024 MUTEX_EXIT(&rx_connHashTable_lock);
2025 rxi_CleanupConnection(conn);
2026 MUTEX_ENTER(&rx_connHashTable_lock);
2028 MUTEX_EXIT(&rx_connHashTable_lock);
2029 #endif /* RX_ENABLE_LOCKS */
2034 afs_winsockCleanup();
2042 /* if we wakeup packet waiter too often, can get in loop with two
2043 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2045 rxi_PacketsUnWait(void)
2047 if (!rx_waitingForPackets) {
2051 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2052 return; /* still over quota */
2055 rx_waitingForPackets = 0;
2056 #ifdef RX_ENABLE_LOCKS
2057 CV_BROADCAST(&rx_waitingForPackets_cv);
2059 osi_rxWakeup(&rx_waitingForPackets);
2065 /* ------------------Internal interfaces------------------------- */
2067 /* Return this process's service structure for the
2068 * specified socket and service */
2070 rxi_FindService(register osi_socket socket, register u_short serviceId)
2072 register struct rx_service **sp;
2073 for (sp = &rx_services[0]; *sp; sp++) {
2074 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2080 /* Allocate a call structure, for the indicated channel of the
2081 * supplied connection. The mode and state of the call must be set by
2082 * the caller. Returns the call with mutex locked. */
2084 rxi_NewCall(register struct rx_connection *conn, register int channel)
2086 register struct rx_call *call;
2087 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2088 register struct rx_call *cp; /* Call pointer temp */
2089 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2090 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2092 dpf(("rxi_NewCall(conn %x, channel %d)\n", conn, channel));
2094 /* Grab an existing call structure, or allocate a new one.
2095 * Existing call structures are assumed to have been left reset by
2097 MUTEX_ENTER(&rx_freeCallQueue_lock);
2099 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2101 * EXCEPT that the TQ might not yet be cleared out.
2102 * Skip over those with in-use TQs.
2105 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2106 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2112 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2113 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2114 call = queue_First(&rx_freeCallQueue, rx_call);
2115 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2117 rx_MutexDecrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2118 MUTEX_EXIT(&rx_freeCallQueue_lock);
2119 MUTEX_ENTER(&call->lock);
2120 CLEAR_CALL_QUEUE_LOCK(call);
2121 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2122 /* Now, if TQ wasn't cleared earlier, do it now. */
2123 if (call->flags & RX_CALL_TQ_CLEARME) {
2124 rxi_ClearTransmitQueue(call, 0);
2125 queue_Init(&call->tq);
2127 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2128 /* Bind the call to its connection structure */
2130 rxi_ResetCall(call, 1);
2132 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2134 MUTEX_EXIT(&rx_freeCallQueue_lock);
2135 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2136 MUTEX_ENTER(&call->lock);
2137 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2138 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2139 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2141 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2142 /* Initialize once-only items */
2143 queue_Init(&call->tq);
2144 queue_Init(&call->rq);
2145 queue_Init(&call->iovq);
2146 /* Bind the call to its connection structure (prereq for reset) */
2148 rxi_ResetCall(call, 1);
2150 call->channel = channel;
2151 call->callNumber = &conn->callNumber[channel];
2152 /* Note that the next expected call number is retained (in
2153 * conn->callNumber[i]), even if we reallocate the call structure
2155 conn->call[channel] = call;
2156 /* if the channel's never been used (== 0), we should start at 1, otherwise
2157 * the call number is valid from the last time this channel was used */
2158 if (*call->callNumber == 0)
2159 *call->callNumber = 1;
2164 /* A call has been inactive long enough that so we can throw away
2165 * state, including the call structure, which is placed on the call
2167 * Call is locked upon entry.
2168 * haveCTLock set if called from rxi_ReapConnections
2170 #ifdef RX_ENABLE_LOCKS
2172 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2173 #else /* RX_ENABLE_LOCKS */
2175 rxi_FreeCall(register struct rx_call *call)
2176 #endif /* RX_ENABLE_LOCKS */
2178 register int channel = call->channel;
2179 register struct rx_connection *conn = call->conn;
2182 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2183 (*call->callNumber)++;
2184 rxi_ResetCall(call, 0);
2185 call->conn->call[channel] = (struct rx_call *)0;
2187 MUTEX_ENTER(&rx_freeCallQueue_lock);
2188 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2189 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2190 /* A call may be free even though its transmit queue is still in use.
2191 * Since we search the call list from head to tail, put busy calls at
2192 * the head of the list, and idle calls at the tail.
2194 if (call->flags & RX_CALL_TQ_BUSY)
2195 queue_Prepend(&rx_freeCallQueue, call);
2197 queue_Append(&rx_freeCallQueue, call);
2198 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2199 queue_Append(&rx_freeCallQueue, call);
2200 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2201 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2202 MUTEX_EXIT(&rx_freeCallQueue_lock);
2204 /* Destroy the connection if it was previously slated for
2205 * destruction, i.e. the Rx client code previously called
2206 * rx_DestroyConnection (client connections), or
2207 * rxi_ReapConnections called the same routine (server
2208 * connections). Only do this, however, if there are no
2209 * outstanding calls. Note that for fine grain locking, there appears
2210 * to be a deadlock in that rxi_FreeCall has a call locked and
2211 * DestroyConnectionNoLock locks each call in the conn. But note a
2212 * few lines up where we have removed this call from the conn.
2213 * If someone else destroys a connection, they either have no
2214 * call lock held or are going through this section of code.
2216 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2217 MUTEX_ENTER(&conn->conn_data_lock);
2219 MUTEX_EXIT(&conn->conn_data_lock);
2220 #ifdef RX_ENABLE_LOCKS
2222 rxi_DestroyConnectionNoLock(conn);
2224 rxi_DestroyConnection(conn);
2225 #else /* RX_ENABLE_LOCKS */
2226 rxi_DestroyConnection(conn);
2227 #endif /* RX_ENABLE_LOCKS */
2231 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2233 rxi_Alloc(register size_t size)
2237 rx_MutexAdd1Increment2(rxi_Allocsize, (afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2238 p = (char *)osi_Alloc(size);
2241 osi_Panic("rxi_Alloc error");
2247 rxi_Free(void *addr, register size_t size)
2249 rx_MutexAdd1Decrement2(rxi_Allocsize, -(afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2250 osi_Free(addr, size);
2253 /* Find the peer process represented by the supplied (host,port)
2254 * combination. If there is no appropriate active peer structure, a
2255 * new one will be allocated and initialized
2256 * The origPeer, if set, is a pointer to a peer structure on which the
2257 * refcount will be be decremented. This is used to replace the peer
2258 * structure hanging off a connection structure */
2260 rxi_FindPeer(register afs_uint32 host, register u_short port,
2261 struct rx_peer *origPeer, int create)
2263 register struct rx_peer *pp;
2265 hashIndex = PEER_HASH(host, port);
2266 MUTEX_ENTER(&rx_peerHashTable_lock);
2267 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2268 if ((pp->host == host) && (pp->port == port))
2273 pp = rxi_AllocPeer(); /* This bzero's *pp */
2274 pp->host = host; /* set here or in InitPeerParams is zero */
2276 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2277 queue_Init(&pp->congestionQueue);
2278 queue_Init(&pp->rpcStats);
2279 pp->next = rx_peerHashTable[hashIndex];
2280 rx_peerHashTable[hashIndex] = pp;
2281 rxi_InitPeerParams(pp);
2282 rx_MutexIncrement(rx_stats.nPeerStructs, rx_stats_mutex);
2289 origPeer->refCount--;
2290 MUTEX_EXIT(&rx_peerHashTable_lock);
2295 /* Find the connection at (host, port) started at epoch, and with the
2296 * given connection id. Creates the server connection if necessary.
2297 * The type specifies whether a client connection or a server
2298 * connection is desired. In both cases, (host, port) specify the
2299 * peer's (host, pair) pair. Client connections are not made
2300 * automatically by this routine. The parameter socket gives the
2301 * socket descriptor on which the packet was received. This is used,
2302 * in the case of server connections, to check that *new* connections
2303 * come via a valid (port, serviceId). Finally, the securityIndex
2304 * parameter must match the existing index for the connection. If a
2305 * server connection is created, it will be created using the supplied
2306 * index, if the index is valid for this service */
2307 struct rx_connection *
2308 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2309 register u_short port, u_short serviceId, afs_uint32 cid,
2310 afs_uint32 epoch, int type, u_int securityIndex)
2312 int hashindex, flag;
2313 register struct rx_connection *conn;
2314 hashindex = CONN_HASH(host, port, cid, epoch, type);
2315 MUTEX_ENTER(&rx_connHashTable_lock);
2316 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2317 rx_connHashTable[hashindex],
2320 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2321 && (epoch == conn->epoch)) {
2322 register struct rx_peer *pp = conn->peer;
2323 if (securityIndex != conn->securityIndex) {
2324 /* this isn't supposed to happen, but someone could forge a packet
2325 * like this, and there seems to be some CM bug that makes this
2326 * happen from time to time -- in which case, the fileserver
2328 MUTEX_EXIT(&rx_connHashTable_lock);
2329 return (struct rx_connection *)0;
2331 if (pp->host == host && pp->port == port)
2333 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2335 /* So what happens when it's a callback connection? */
2336 if ( /*type == RX_CLIENT_CONNECTION && */
2337 (conn->epoch & 0x80000000))
2341 /* the connection rxLastConn that was used the last time is not the
2342 ** one we are looking for now. Hence, start searching in the hash */
2344 conn = rx_connHashTable[hashindex];
2349 struct rx_service *service;
2350 if (type == RX_CLIENT_CONNECTION) {
2351 MUTEX_EXIT(&rx_connHashTable_lock);
2352 return (struct rx_connection *)0;
2354 service = rxi_FindService(socket, serviceId);
2355 if (!service || (securityIndex >= service->nSecurityObjects)
2356 || (service->securityObjects[securityIndex] == 0)) {
2357 MUTEX_EXIT(&rx_connHashTable_lock);
2358 return (struct rx_connection *)0;
2360 conn = rxi_AllocConnection(); /* This bzero's the connection */
2361 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2362 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2363 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2364 conn->next = rx_connHashTable[hashindex];
2365 rx_connHashTable[hashindex] = conn;
2366 conn->peer = rxi_FindPeer(host, port, 0, 1);
2367 conn->type = RX_SERVER_CONNECTION;
2368 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2369 conn->epoch = epoch;
2370 conn->cid = cid & RX_CIDMASK;
2371 /* conn->serial = conn->lastSerial = 0; */
2372 /* conn->timeout = 0; */
2373 conn->ackRate = RX_FAST_ACK_RATE;
2374 conn->service = service;
2375 conn->serviceId = serviceId;
2376 conn->securityIndex = securityIndex;
2377 conn->securityObject = service->securityObjects[securityIndex];
2378 conn->nSpecific = 0;
2379 conn->specific = NULL;
2380 rx_SetConnDeadTime(conn, service->connDeadTime);
2381 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2382 /* Notify security object of the new connection */
2383 RXS_NewConnection(conn->securityObject, conn);
2384 /* XXXX Connection timeout? */
2385 if (service->newConnProc)
2386 (*service->newConnProc) (conn);
2387 rx_MutexIncrement(rx_stats.nServerConns, rx_stats_mutex);
2390 MUTEX_ENTER(&conn->conn_data_lock);
2392 MUTEX_EXIT(&conn->conn_data_lock);
2394 rxLastConn = conn; /* store this connection as the last conn used */
2395 MUTEX_EXIT(&rx_connHashTable_lock);
2399 /* There are two packet tracing routines available for testing and monitoring
2400 * Rx. One is called just after every packet is received and the other is
2401 * called just before every packet is sent. Received packets, have had their
2402 * headers decoded, and packets to be sent have not yet had their headers
2403 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2404 * containing the network address. Both can be modified. The return value, if
2405 * non-zero, indicates that the packet should be dropped. */
2407 int (*rx_justReceived) () = 0;
2408 int (*rx_almostSent) () = 0;
2410 /* A packet has been received off the interface. Np is the packet, socket is
2411 * the socket number it was received from (useful in determining which service
2412 * this packet corresponds to), and (host, port) reflect the host,port of the
2413 * sender. This call returns the packet to the caller if it is finished with
2414 * it, rather than de-allocating it, just as a small performance hack */
2417 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2418 afs_uint32 host, u_short port, int *tnop,
2419 struct rx_call **newcallp)
2421 register struct rx_call *call;
2422 register struct rx_connection *conn;
2424 afs_uint32 currentCallNumber;
2430 struct rx_packet *tnp;
2433 /* We don't print out the packet until now because (1) the time may not be
2434 * accurate enough until now in the lwp implementation (rx_Listener only gets
2435 * the time after the packet is read) and (2) from a protocol point of view,
2436 * this is the first time the packet has been seen */
2437 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2438 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2439 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2440 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2441 np->header.epoch, np->header.cid, np->header.callNumber,
2442 np->header.seq, np->header.flags, np));
2445 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2446 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2449 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2450 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2453 /* If an input tracer function is defined, call it with the packet and
2454 * network address. Note this function may modify its arguments. */
2455 if (rx_justReceived) {
2456 struct sockaddr_in addr;
2458 addr.sin_family = AF_INET;
2459 addr.sin_port = port;
2460 addr.sin_addr.s_addr = host;
2461 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2462 addr.sin_len = sizeof(addr);
2463 #endif /* AFS_OSF_ENV */
2464 drop = (*rx_justReceived) (np, &addr);
2465 /* drop packet if return value is non-zero */
2468 port = addr.sin_port; /* in case fcn changed addr */
2469 host = addr.sin_addr.s_addr;
2473 /* If packet was not sent by the client, then *we* must be the client */
2474 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2475 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2477 /* Find the connection (or fabricate one, if we're the server & if
2478 * necessary) associated with this packet */
2480 rxi_FindConnection(socket, host, port, np->header.serviceId,
2481 np->header.cid, np->header.epoch, type,
2482 np->header.securityIndex);
2485 /* If no connection found or fabricated, just ignore the packet.
2486 * (An argument could be made for sending an abort packet for
2491 MUTEX_ENTER(&conn->conn_data_lock);
2492 if (conn->maxSerial < np->header.serial)
2493 conn->maxSerial = np->header.serial;
2494 MUTEX_EXIT(&conn->conn_data_lock);
2496 /* If the connection is in an error state, send an abort packet and ignore
2497 * the incoming packet */
2499 /* Don't respond to an abort packet--we don't want loops! */
2500 MUTEX_ENTER(&conn->conn_data_lock);
2501 if (np->header.type != RX_PACKET_TYPE_ABORT)
2502 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2504 MUTEX_EXIT(&conn->conn_data_lock);
2508 /* Check for connection-only requests (i.e. not call specific). */
2509 if (np->header.callNumber == 0) {
2510 switch (np->header.type) {
2511 case RX_PACKET_TYPE_ABORT: {
2512 /* What if the supplied error is zero? */
2513 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2514 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2515 rxi_ConnectionError(conn, errcode);
2516 MUTEX_ENTER(&conn->conn_data_lock);
2518 MUTEX_EXIT(&conn->conn_data_lock);
2521 case RX_PACKET_TYPE_CHALLENGE:
2522 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2523 MUTEX_ENTER(&conn->conn_data_lock);
2525 MUTEX_EXIT(&conn->conn_data_lock);
2527 case RX_PACKET_TYPE_RESPONSE:
2528 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2529 MUTEX_ENTER(&conn->conn_data_lock);
2531 MUTEX_EXIT(&conn->conn_data_lock);
2533 case RX_PACKET_TYPE_PARAMS:
2534 case RX_PACKET_TYPE_PARAMS + 1:
2535 case RX_PACKET_TYPE_PARAMS + 2:
2536 /* ignore these packet types for now */
2537 MUTEX_ENTER(&conn->conn_data_lock);
2539 MUTEX_EXIT(&conn->conn_data_lock);
2544 /* Should not reach here, unless the peer is broken: send an
2546 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2547 MUTEX_ENTER(&conn->conn_data_lock);
2548 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2550 MUTEX_EXIT(&conn->conn_data_lock);
2555 channel = np->header.cid & RX_CHANNELMASK;
2556 call = conn->call[channel];
2557 #ifdef RX_ENABLE_LOCKS
2559 MUTEX_ENTER(&call->lock);
2560 /* Test to see if call struct is still attached to conn. */
2561 if (call != conn->call[channel]) {
2563 MUTEX_EXIT(&call->lock);
2564 if (type == RX_SERVER_CONNECTION) {
2565 call = conn->call[channel];
2566 /* If we started with no call attached and there is one now,
2567 * another thread is also running this routine and has gotten
2568 * the connection channel. We should drop this packet in the tests
2569 * below. If there was a call on this connection and it's now
2570 * gone, then we'll be making a new call below.
2571 * If there was previously a call and it's now different then
2572 * the old call was freed and another thread running this routine
2573 * has created a call on this channel. One of these two threads
2574 * has a packet for the old call and the code below handles those
2578 MUTEX_ENTER(&call->lock);
2580 /* This packet can't be for this call. If the new call address is
2581 * 0 then no call is running on this channel. If there is a call
2582 * then, since this is a client connection we're getting data for
2583 * it must be for the previous call.
2585 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2586 MUTEX_ENTER(&conn->conn_data_lock);
2588 MUTEX_EXIT(&conn->conn_data_lock);
2593 currentCallNumber = conn->callNumber[channel];
2595 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2596 if (np->header.callNumber < currentCallNumber) {
2597 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2598 #ifdef RX_ENABLE_LOCKS
2600 MUTEX_EXIT(&call->lock);
2602 MUTEX_ENTER(&conn->conn_data_lock);
2604 MUTEX_EXIT(&conn->conn_data_lock);
2608 MUTEX_ENTER(&conn->conn_call_lock);
2609 call = rxi_NewCall(conn, channel);
2610 MUTEX_EXIT(&conn->conn_call_lock);
2611 *call->callNumber = np->header.callNumber;
2612 if (np->header.callNumber == 0)
2613 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));
2615 call->state = RX_STATE_PRECALL;
2616 clock_GetTime(&call->queueTime);
2617 hzero(call->bytesSent);
2618 hzero(call->bytesRcvd);
2620 * If the number of queued calls exceeds the overload
2621 * threshold then abort this call.
2623 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2624 struct rx_packet *tp;
2626 rxi_CallError(call, rx_BusyError);
2627 tp = rxi_SendCallAbort(call, np, 1, 0);
2628 MUTEX_EXIT(&call->lock);
2629 MUTEX_ENTER(&conn->conn_data_lock);
2631 MUTEX_EXIT(&conn->conn_data_lock);
2632 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2635 rxi_KeepAliveOn(call);
2636 } else if (np->header.callNumber != currentCallNumber) {
2637 /* Wait until the transmit queue is idle before deciding
2638 * whether to reset the current call. Chances are that the
2639 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2642 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2643 while ((call->state == RX_STATE_ACTIVE)
2644 && (call->flags & RX_CALL_TQ_BUSY)) {
2645 call->flags |= RX_CALL_TQ_WAIT;
2647 #ifdef RX_ENABLE_LOCKS
2648 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2649 CV_WAIT(&call->cv_tq, &call->lock);
2650 #else /* RX_ENABLE_LOCKS */
2651 osi_rxSleep(&call->tq);
2652 #endif /* RX_ENABLE_LOCKS */
2654 if (call->tqWaiters == 0)
2655 call->flags &= ~RX_CALL_TQ_WAIT;
2657 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2658 /* If the new call cannot be taken right now send a busy and set
2659 * the error condition in this call, so that it terminates as
2660 * quickly as possible */
2661 if (call->state == RX_STATE_ACTIVE) {
2662 struct rx_packet *tp;
2664 rxi_CallError(call, RX_CALL_DEAD);
2665 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2667 MUTEX_EXIT(&call->lock);
2668 MUTEX_ENTER(&conn->conn_data_lock);
2670 MUTEX_EXIT(&conn->conn_data_lock);
2673 rxi_ResetCall(call, 0);
2674 *call->callNumber = np->header.callNumber;
2675 if (np->header.callNumber == 0)
2676 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));
2678 call->state = RX_STATE_PRECALL;
2679 clock_GetTime(&call->queueTime);
2680 hzero(call->bytesSent);
2681 hzero(call->bytesRcvd);
2683 * If the number of queued calls exceeds the overload
2684 * threshold then abort this call.
2686 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2687 struct rx_packet *tp;
2689 rxi_CallError(call, rx_BusyError);
2690 tp = rxi_SendCallAbort(call, np, 1, 0);
2691 MUTEX_EXIT(&call->lock);
2692 MUTEX_ENTER(&conn->conn_data_lock);
2694 MUTEX_EXIT(&conn->conn_data_lock);
2695 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2698 rxi_KeepAliveOn(call);
2700 /* Continuing call; do nothing here. */
2702 } else { /* we're the client */
2703 /* Ignore all incoming acknowledgements for calls in DALLY state */
2704 if (call && (call->state == RX_STATE_DALLY)
2705 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2706 rx_MutexIncrement(rx_stats.ignorePacketDally, rx_stats_mutex);
2707 #ifdef RX_ENABLE_LOCKS
2709 MUTEX_EXIT(&call->lock);
2712 MUTEX_ENTER(&conn->conn_data_lock);
2714 MUTEX_EXIT(&conn->conn_data_lock);
2718 /* Ignore anything that's not relevant to the current call. If there
2719 * isn't a current call, then no packet is relevant. */
2720 if (!call || (np->header.callNumber != currentCallNumber)) {
2721 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2722 #ifdef RX_ENABLE_LOCKS
2724 MUTEX_EXIT(&call->lock);
2727 MUTEX_ENTER(&conn->conn_data_lock);
2729 MUTEX_EXIT(&conn->conn_data_lock);
2732 /* If the service security object index stamped in the packet does not
2733 * match the connection's security index, ignore the packet */
2734 if (np->header.securityIndex != conn->securityIndex) {
2735 #ifdef RX_ENABLE_LOCKS
2736 MUTEX_EXIT(&call->lock);
2738 MUTEX_ENTER(&conn->conn_data_lock);
2740 MUTEX_EXIT(&conn->conn_data_lock);
2744 /* If we're receiving the response, then all transmit packets are
2745 * implicitly acknowledged. Get rid of them. */
2746 if (np->header.type == RX_PACKET_TYPE_DATA) {
2747 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2748 /* XXX Hack. Because we must release the global rx lock when
2749 * sending packets (osi_NetSend) we drop all acks while we're
2750 * traversing the tq in rxi_Start sending packets out because
2751 * packets may move to the freePacketQueue as result of being here!
2752 * So we drop these packets until we're safely out of the
2753 * traversing. Really ugly!
2754 * For fine grain RX locking, we set the acked field in the
2755 * packets and let rxi_Start remove them from the transmit queue.
2757 if (call->flags & RX_CALL_TQ_BUSY) {
2758 #ifdef RX_ENABLE_LOCKS
2759 rxi_SetAcksInTransmitQueue(call);
2762 return np; /* xmitting; drop packet */
2765 rxi_ClearTransmitQueue(call, 0);
2767 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2768 rxi_ClearTransmitQueue(call, 0);
2769 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2771 if (np->header.type == RX_PACKET_TYPE_ACK) {
2772 /* now check to see if this is an ack packet acknowledging that the
2773 * server actually *lost* some hard-acked data. If this happens we
2774 * ignore this packet, as it may indicate that the server restarted in
2775 * the middle of a call. It is also possible that this is an old ack
2776 * packet. We don't abort the connection in this case, because this
2777 * *might* just be an old ack packet. The right way to detect a server
2778 * restart in the midst of a call is to notice that the server epoch
2780 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2781 * XXX unacknowledged. I think that this is off-by-one, but
2782 * XXX I don't dare change it just yet, since it will
2783 * XXX interact badly with the server-restart detection
2784 * XXX code in receiveackpacket. */
2785 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2786 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2787 MUTEX_EXIT(&call->lock);
2788 MUTEX_ENTER(&conn->conn_data_lock);
2790 MUTEX_EXIT(&conn->conn_data_lock);
2794 } /* else not a data packet */
2797 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2798 /* Set remote user defined status from packet */
2799 call->remoteStatus = np->header.userStatus;
2801 /* Note the gap between the expected next packet and the actual
2802 * packet that arrived, when the new packet has a smaller serial number
2803 * than expected. Rioses frequently reorder packets all by themselves,
2804 * so this will be quite important with very large window sizes.
2805 * Skew is checked against 0 here to avoid any dependence on the type of
2806 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2808 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2809 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2810 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2812 MUTEX_ENTER(&conn->conn_data_lock);
2813 skew = conn->lastSerial - np->header.serial;
2814 conn->lastSerial = np->header.serial;
2815 MUTEX_EXIT(&conn->conn_data_lock);
2817 register struct rx_peer *peer;
2819 if (skew > peer->inPacketSkew) {
2820 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2822 peer->inPacketSkew = skew;
2826 /* Now do packet type-specific processing */
2827 switch (np->header.type) {
2828 case RX_PACKET_TYPE_DATA:
2829 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2832 case RX_PACKET_TYPE_ACK:
2833 /* Respond immediately to ack packets requesting acknowledgement
2835 if (np->header.flags & RX_REQUEST_ACK) {
2837 (void)rxi_SendCallAbort(call, 0, 1, 0);
2839 (void)rxi_SendAck(call, 0, np->header.serial,
2840 RX_ACK_PING_RESPONSE, 1);
2842 np = rxi_ReceiveAckPacket(call, np, 1);
2844 case RX_PACKET_TYPE_ABORT: {
2845 /* An abort packet: reset the call, passing the error up to the user. */
2846 /* What if error is zero? */
2847 /* What if the error is -1? the application will treat it as a timeout. */
2848 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
2849 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
2850 rxi_CallError(call, errdata);
2851 MUTEX_EXIT(&call->lock);
2852 MUTEX_ENTER(&conn->conn_data_lock);
2854 MUTEX_EXIT(&conn->conn_data_lock);
2855 return np; /* xmitting; drop packet */
2857 case RX_PACKET_TYPE_BUSY:
2860 case RX_PACKET_TYPE_ACKALL:
2861 /* All packets acknowledged, so we can drop all packets previously
2862 * readied for sending */
2863 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2864 /* XXX Hack. We because we can't release the global rx lock when
2865 * sending packets (osi_NetSend) we drop all ack pkts while we're
2866 * traversing the tq in rxi_Start sending packets out because
2867 * packets may move to the freePacketQueue as result of being
2868 * here! So we drop these packets until we're safely out of the
2869 * traversing. Really ugly!
2870 * For fine grain RX locking, we set the acked field in the packets
2871 * and let rxi_Start remove the packets from the transmit queue.
2873 if (call->flags & RX_CALL_TQ_BUSY) {
2874 #ifdef RX_ENABLE_LOCKS
2875 rxi_SetAcksInTransmitQueue(call);
2877 #else /* RX_ENABLE_LOCKS */
2878 MUTEX_EXIT(&call->lock);
2879 MUTEX_ENTER(&conn->conn_data_lock);
2881 MUTEX_EXIT(&conn->conn_data_lock);
2882 return np; /* xmitting; drop packet */
2883 #endif /* RX_ENABLE_LOCKS */
2885 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2886 rxi_ClearTransmitQueue(call, 0);
2889 /* Should not reach here, unless the peer is broken: send an abort
2891 rxi_CallError(call, RX_PROTOCOL_ERROR);
2892 np = rxi_SendCallAbort(call, np, 1, 0);
2895 /* Note when this last legitimate packet was received, for keep-alive
2896 * processing. Note, we delay getting the time until now in the hope that
2897 * the packet will be delivered to the user before any get time is required
2898 * (if not, then the time won't actually be re-evaluated here). */
2899 call->lastReceiveTime = clock_Sec();
2900 MUTEX_EXIT(&call->lock);
2901 MUTEX_ENTER(&conn->conn_data_lock);
2903 MUTEX_EXIT(&conn->conn_data_lock);
2907 /* return true if this is an "interesting" connection from the point of view
2908 of someone trying to debug the system */
2910 rxi_IsConnInteresting(struct rx_connection *aconn)
2913 register struct rx_call *tcall;
2915 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
2917 for (i = 0; i < RX_MAXCALLS; i++) {
2918 tcall = aconn->call[i];
2920 if ((tcall->state == RX_STATE_PRECALL)
2921 || (tcall->state == RX_STATE_ACTIVE))
2923 if ((tcall->mode == RX_MODE_SENDING)
2924 || (tcall->mode == RX_MODE_RECEIVING))
2932 /* if this is one of the last few packets AND it wouldn't be used by the
2933 receiving call to immediately satisfy a read request, then drop it on
2934 the floor, since accepting it might prevent a lock-holding thread from
2935 making progress in its reading. If a call has been cleared while in
2936 the precall state then ignore all subsequent packets until the call
2937 is assigned to a thread. */
2940 TooLow(struct rx_packet *ap, struct rx_call *acall)
2943 MUTEX_ENTER(&rx_stats_mutex);
2944 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
2945 && (acall->state == RX_STATE_PRECALL))
2946 || ((rx_nFreePackets < rxi_dataQuota + 2)
2947 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
2948 && (acall->flags & RX_CALL_READER_WAIT)))) {
2951 MUTEX_EXIT(&rx_stats_mutex);
2957 rxi_CheckReachEvent(struct rxevent *event, struct rx_connection *conn,
2958 struct rx_call *acall)
2960 struct rx_call *call = acall;
2964 MUTEX_ENTER(&conn->conn_data_lock);
2965 conn->checkReachEvent = NULL;
2966 waiting = conn->flags & RX_CONN_ATTACHWAIT;
2969 MUTEX_EXIT(&conn->conn_data_lock);
2973 MUTEX_ENTER(&conn->conn_call_lock);
2974 MUTEX_ENTER(&conn->conn_data_lock);
2975 for (i = 0; i < RX_MAXCALLS; i++) {
2976 struct rx_call *tc = conn->call[i];
2977 if (tc && tc->state == RX_STATE_PRECALL) {
2983 /* Indicate that rxi_CheckReachEvent is no longer running by
2984 * clearing the flag. Must be atomic under conn_data_lock to
2985 * avoid a new call slipping by: rxi_CheckConnReach holds
2986 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
2988 conn->flags &= ~RX_CONN_ATTACHWAIT;
2989 MUTEX_EXIT(&conn->conn_data_lock);
2990 MUTEX_EXIT(&conn->conn_call_lock);
2995 MUTEX_ENTER(&call->lock);
2996 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
2998 MUTEX_EXIT(&call->lock);
3000 clock_GetTime(&when);
3001 when.sec += RX_CHECKREACH_TIMEOUT;
3002 MUTEX_ENTER(&conn->conn_data_lock);
3003 if (!conn->checkReachEvent) {
3005 conn->checkReachEvent =
3006 rxevent_Post(&when, rxi_CheckReachEvent, conn, NULL);
3008 MUTEX_EXIT(&conn->conn_data_lock);
3014 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3016 struct rx_service *service = conn->service;
3017 struct rx_peer *peer = conn->peer;
3018 afs_uint32 now, lastReach;
3020 if (service->checkReach == 0)
3024 MUTEX_ENTER(&peer->peer_lock);
3025 lastReach = peer->lastReachTime;
3026 MUTEX_EXIT(&peer->peer_lock);
3027 if (now - lastReach < RX_CHECKREACH_TTL)
3030 MUTEX_ENTER(&conn->conn_data_lock);
3031 if (conn->flags & RX_CONN_ATTACHWAIT) {
3032 MUTEX_EXIT(&conn->conn_data_lock);
3035 conn->flags |= RX_CONN_ATTACHWAIT;
3036 MUTEX_EXIT(&conn->conn_data_lock);
3037 if (!conn->checkReachEvent)
3038 rxi_CheckReachEvent(NULL, conn, call);
3043 /* try to attach call, if authentication is complete */
3045 TryAttach(register struct rx_call *acall, register osi_socket socket,
3046 register int *tnop, register struct rx_call **newcallp,
3049 struct rx_connection *conn = acall->conn;
3051 if (conn->type == RX_SERVER_CONNECTION
3052 && acall->state == RX_STATE_PRECALL) {
3053 /* Don't attach until we have any req'd. authentication. */
3054 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3055 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3056 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3057 /* Note: this does not necessarily succeed; there
3058 * may not any proc available
3061 rxi_ChallengeOn(acall->conn);
3066 /* A data packet has been received off the interface. This packet is
3067 * appropriate to the call (the call is in the right state, etc.). This
3068 * routine can return a packet to the caller, for re-use */
3071 rxi_ReceiveDataPacket(register struct rx_call *call,
3072 register struct rx_packet *np, int istack,
3073 osi_socket socket, afs_uint32 host, u_short port,
3074 int *tnop, struct rx_call **newcallp)
3076 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3080 afs_uint32 seq, serial, flags;
3082 struct rx_packet *tnp;
3084 rx_MutexIncrement(rx_stats.dataPacketsRead, rx_stats_mutex);
3087 /* If there are no packet buffers, drop this new packet, unless we can find
3088 * packet buffers from inactive calls */
3090 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3091 MUTEX_ENTER(&rx_freePktQ_lock);
3092 rxi_NeedMorePackets = TRUE;
3093 MUTEX_EXIT(&rx_freePktQ_lock);
3094 rx_MutexIncrement(rx_stats.noPacketBuffersOnRead, rx_stats_mutex);
3095 call->rprev = np->header.serial;
3096 rxi_calltrace(RX_TRACE_DROP, call);
3097 dpf(("packet %x dropped on receipt - quota problems", np));
3099 rxi_ClearReceiveQueue(call);
3100 clock_GetTime(&when);
3101 clock_Add(&when, &rx_softAckDelay);
3102 if (!call->delayedAckEvent
3103 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3104 rxevent_Cancel(call->delayedAckEvent, call,
3105 RX_CALL_REFCOUNT_DELAY);
3106 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3107 call->delayedAckEvent =
3108 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3110 /* we've damaged this call already, might as well do it in. */
3116 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3117 * packet is one of several packets transmitted as a single
3118 * datagram. Do not send any soft or hard acks until all packets
3119 * in a jumbogram have been processed. Send negative acks right away.
3121 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3122 /* tnp is non-null when there are more packets in the
3123 * current jumbo gram */
3130 seq = np->header.seq;
3131 serial = np->header.serial;
3132 flags = np->header.flags;
3134 /* If the call is in an error state, send an abort message */
3136 return rxi_SendCallAbort(call, np, istack, 0);
3138 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3139 * AFS 3.5 jumbogram. */
3140 if (flags & RX_JUMBO_PACKET) {
3141 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3146 if (np->header.spare != 0) {
3147 MUTEX_ENTER(&call->conn->conn_data_lock);
3148 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3149 MUTEX_EXIT(&call->conn->conn_data_lock);
3152 /* The usual case is that this is the expected next packet */
3153 if (seq == call->rnext) {
3155 /* Check to make sure it is not a duplicate of one already queued */
3156 if (queue_IsNotEmpty(&call->rq)
3157 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3158 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3159 dpf(("packet %x dropped on receipt - duplicate", np));
3160 rxevent_Cancel(call->delayedAckEvent, call,
3161 RX_CALL_REFCOUNT_DELAY);
3162 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3168 /* It's the next packet. Stick it on the receive queue
3169 * for this call. Set newPackets to make sure we wake
3170 * the reader once all packets have been processed */
3171 queue_Prepend(&call->rq, np);
3173 np = NULL; /* We can't use this anymore */
3176 /* If an ack is requested then set a flag to make sure we
3177 * send an acknowledgement for this packet */
3178 if (flags & RX_REQUEST_ACK) {
3179 ackNeeded = RX_ACK_REQUESTED;
3182 /* Keep track of whether we have received the last packet */
3183 if (flags & RX_LAST_PACKET) {
3184 call->flags |= RX_CALL_HAVE_LAST;
3188 /* Check whether we have all of the packets for this call */
3189 if (call->flags & RX_CALL_HAVE_LAST) {
3190 afs_uint32 tseq; /* temporary sequence number */
3191 struct rx_packet *tp; /* Temporary packet pointer */
3192 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3194 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3195 if (tseq != tp->header.seq)
3197 if (tp->header.flags & RX_LAST_PACKET) {
3198 call->flags |= RX_CALL_RECEIVE_DONE;
3205 /* Provide asynchronous notification for those who want it
3206 * (e.g. multi rx) */
3207 if (call->arrivalProc) {
3208 (*call->arrivalProc) (call, call->arrivalProcHandle,
3209 call->arrivalProcArg);
3210 call->arrivalProc = (void (*)())0;
3213 /* Update last packet received */
3216 /* If there is no server process serving this call, grab
3217 * one, if available. We only need to do this once. If a
3218 * server thread is available, this thread becomes a server
3219 * thread and the server thread becomes a listener thread. */
3221 TryAttach(call, socket, tnop, newcallp, 0);
3224 /* This is not the expected next packet. */
3226 /* Determine whether this is a new or old packet, and if it's
3227 * a new one, whether it fits into the current receive window.
3228 * Also figure out whether the packet was delivered in sequence.
3229 * We use the prev variable to determine whether the new packet
3230 * is the successor of its immediate predecessor in the
3231 * receive queue, and the missing flag to determine whether
3232 * any of this packets predecessors are missing. */
3234 afs_uint32 prev; /* "Previous packet" sequence number */
3235 struct rx_packet *tp; /* Temporary packet pointer */
3236 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3237 int missing; /* Are any predecessors missing? */
3239 /* If the new packet's sequence number has been sent to the
3240 * application already, then this is a duplicate */
3241 if (seq < call->rnext) {
3242 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3243 rxevent_Cancel(call->delayedAckEvent, call,
3244 RX_CALL_REFCOUNT_DELAY);
3245 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3251 /* If the sequence number is greater than what can be
3252 * accomodated by the current window, then send a negative
3253 * acknowledge and drop the packet */
3254 if ((call->rnext + call->rwind) <= seq) {
3255 rxevent_Cancel(call->delayedAckEvent, call,
3256 RX_CALL_REFCOUNT_DELAY);
3257 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3264 /* Look for the packet in the queue of old received packets */
3265 for (prev = call->rnext - 1, missing =
3266 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3267 /*Check for duplicate packet */
3268 if (seq == tp->header.seq) {
3269 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3270 rxevent_Cancel(call->delayedAckEvent, call,
3271 RX_CALL_REFCOUNT_DELAY);
3272 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3278 /* If we find a higher sequence packet, break out and
3279 * insert the new packet here. */
3280 if (seq < tp->header.seq)
3282 /* Check for missing packet */
3283 if (tp->header.seq != prev + 1) {
3287 prev = tp->header.seq;
3290 /* Keep track of whether we have received the last packet. */
3291 if (flags & RX_LAST_PACKET) {
3292 call->flags |= RX_CALL_HAVE_LAST;
3295 /* It's within the window: add it to the the receive queue.
3296 * tp is left by the previous loop either pointing at the
3297 * packet before which to insert the new packet, or at the
3298 * queue head if the queue is empty or the packet should be
3300 queue_InsertBefore(tp, np);
3304 /* Check whether we have all of the packets for this call */
3305 if ((call->flags & RX_CALL_HAVE_LAST)
3306 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3307 afs_uint32 tseq; /* temporary sequence number */
3310 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3311 if (tseq != tp->header.seq)
3313 if (tp->header.flags & RX_LAST_PACKET) {
3314 call->flags |= RX_CALL_RECEIVE_DONE;
3321 /* We need to send an ack of the packet is out of sequence,
3322 * or if an ack was requested by the peer. */
3323 if (seq != prev + 1 || missing) {
3324 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3325 } else if (flags & RX_REQUEST_ACK) {
3326 ackNeeded = RX_ACK_REQUESTED;
3329 /* Acknowledge the last packet for each call */
3330 if (flags & RX_LAST_PACKET) {
3341 * If the receiver is waiting for an iovec, fill the iovec
3342 * using the data from the receive queue */
3343 if (call->flags & RX_CALL_IOVEC_WAIT) {
3344 didHardAck = rxi_FillReadVec(call, serial);
3345 /* the call may have been aborted */
3354 /* Wakeup the reader if any */
3355 if ((call->flags & RX_CALL_READER_WAIT)
3356 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3357 || (call->iovNext >= call->iovMax)
3358 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3359 call->flags &= ~RX_CALL_READER_WAIT;
3360 #ifdef RX_ENABLE_LOCKS
3361 CV_BROADCAST(&call->cv_rq);
3363 osi_rxWakeup(&call->rq);
3369 * Send an ack when requested by the peer, or once every
3370 * rxi_SoftAckRate packets until the last packet has been
3371 * received. Always send a soft ack for the last packet in
3372 * the server's reply. */
3374 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3375 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3376 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3377 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3378 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3379 } else if (call->nSoftAcks) {
3380 clock_GetTime(&when);
3381 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3382 clock_Add(&when, &rx_lastAckDelay);
3384 clock_Add(&when, &rx_softAckDelay);
3386 if (!call->delayedAckEvent
3387 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3388 rxevent_Cancel(call->delayedAckEvent, call,
3389 RX_CALL_REFCOUNT_DELAY);
3390 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3391 call->delayedAckEvent =
3392 rxevent_Post(&when, rxi_SendDelayedAck, call, 0);
3394 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3395 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3402 static void rxi_ComputeRate();
3406 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3408 struct rx_peer *peer = conn->peer;
3410 MUTEX_ENTER(&peer->peer_lock);
3411 peer->lastReachTime = clock_Sec();
3412 MUTEX_EXIT(&peer->peer_lock);
3414 MUTEX_ENTER(&conn->conn_data_lock);
3415 if (conn->flags & RX_CONN_ATTACHWAIT) {
3418 conn->flags &= ~RX_CONN_ATTACHWAIT;
3419 MUTEX_EXIT(&conn->conn_data_lock);
3421 for (i = 0; i < RX_MAXCALLS; i++) {
3422 struct rx_call *call = conn->call[i];
3425 MUTEX_ENTER(&call->lock);
3426 /* tnop can be null if newcallp is null */
3427 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3429 MUTEX_EXIT(&call->lock);
3433 MUTEX_EXIT(&conn->conn_data_lock);
3437 rx_ack_reason(int reason)
3440 case RX_ACK_REQUESTED:
3442 case RX_ACK_DUPLICATE:
3444 case RX_ACK_OUT_OF_SEQUENCE:
3446 case RX_ACK_EXCEEDS_WINDOW:
3448 case RX_ACK_NOSPACE:
3452 case RX_ACK_PING_RESPONSE:
3464 /* rxi_ComputePeerNetStats
3466 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3467 * estimates (like RTT and throughput) based on ack packets. Caller
3468 * must ensure that the packet in question is the right one (i.e.
3469 * serial number matches).
3472 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3473 struct rx_ackPacket *ap, struct rx_packet *np)
3475 struct rx_peer *peer = call->conn->peer;
3477 /* Use RTT if not delayed by client. */
3478 if (ap->reason != RX_ACK_DELAY)
3479 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3481 rxi_ComputeRate(peer, call, p, np, ap->reason);
3485 /* The real smarts of the whole thing. */
3487 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3490 struct rx_ackPacket *ap;
3492 register struct rx_packet *tp;
3493 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3494 register struct rx_connection *conn = call->conn;
3495 struct rx_peer *peer = conn->peer;
3498 /* because there are CM's that are bogus, sending weird values for this. */
3499 afs_uint32 skew = 0;
3504 int newAckCount = 0;
3505 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3506 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3508 rx_MutexIncrement(rx_stats.ackPacketsRead, rx_stats_mutex);
3509 ap = (struct rx_ackPacket *)rx_DataOf(np);
3510 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3512 return np; /* truncated ack packet */
3514 /* depends on ack packet struct */
3515 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3516 first = ntohl(ap->firstPacket);
3517 serial = ntohl(ap->serial);
3518 /* temporarily disabled -- needs to degrade over time
3519 * skew = ntohs(ap->maxSkew); */
3521 /* Ignore ack packets received out of order */
3522 if (first < call->tfirst) {
3526 if (np->header.flags & RX_SLOW_START_OK) {
3527 call->flags |= RX_CALL_SLOW_START_OK;
3530 if (ap->reason == RX_ACK_PING_RESPONSE)
3531 rxi_UpdatePeerReach(conn, call);
3535 if (rxdebug_active) {
3539 len = _snprintf(msg, sizeof(msg),
3540 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3541 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3542 ntohl(ap->serial), ntohl(ap->previousPacket),
3543 (unsigned int)np->header.seq, (unsigned int)skew,
3544 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3548 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3549 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3553 OutputDebugString(msg);
3555 #else /* AFS_NT40_ENV */
3558 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3559 ap->reason, ntohl(ap->previousPacket),
3560 (unsigned int)np->header.seq, (unsigned int)serial,
3561 (unsigned int)skew, ntohl(ap->firstPacket));
3564 for (offset = 0; offset < nAcks; offset++)
3565 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3570 #endif /* AFS_NT40_ENV */
3573 /* Update the outgoing packet skew value to the latest value of
3574 * the peer's incoming packet skew value. The ack packet, of
3575 * course, could arrive out of order, but that won't affect things
3577 MUTEX_ENTER(&peer->peer_lock);
3578 peer->outPacketSkew = skew;
3580 /* Check for packets that no longer need to be transmitted, and
3581 * discard them. This only applies to packets positively
3582 * acknowledged as having been sent to the peer's upper level.
3583 * All other packets must be retained. So only packets with
3584 * sequence numbers < ap->firstPacket are candidates. */
3585 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3586 if (tp->header.seq >= first)
3588 call->tfirst = tp->header.seq + 1;
3590 && (tp->header.serial == serial || tp->firstSerial == serial))
3591 rxi_ComputePeerNetStats(call, tp, ap, np);
3592 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3595 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3596 /* XXX Hack. Because we have to release the global rx lock when sending
3597 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3598 * in rxi_Start sending packets out because packets may move to the
3599 * freePacketQueue as result of being here! So we drop these packets until
3600 * we're safely out of the traversing. Really ugly!
3601 * To make it even uglier, if we're using fine grain locking, we can
3602 * set the ack bits in the packets and have rxi_Start remove the packets
3603 * when it's done transmitting.
3605 if (call->flags & RX_CALL_TQ_BUSY) {
3606 #ifdef RX_ENABLE_LOCKS
3607 tp->flags |= RX_PKTFLAG_ACKED;
3608 call->flags |= RX_CALL_TQ_SOME_ACKED;
3609 #else /* RX_ENABLE_LOCKS */
3611 #endif /* RX_ENABLE_LOCKS */
3613 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3616 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3621 /* Give rate detector a chance to respond to ping requests */
3622 if (ap->reason == RX_ACK_PING_RESPONSE) {
3623 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3627 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3629 /* Now go through explicit acks/nacks and record the results in
3630 * the waiting packets. These are packets that can't be released
3631 * yet, even with a positive acknowledge. This positive
3632 * acknowledge only means the packet has been received by the
3633 * peer, not that it will be retained long enough to be sent to
3634 * the peer's upper level. In addition, reset the transmit timers
3635 * of any missing packets (those packets that must be missing
3636 * because this packet was out of sequence) */
3638 call->nSoftAcked = 0;
3639 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3640 /* Update round trip time if the ack was stimulated on receipt
3642 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3643 #ifdef RX_ENABLE_LOCKS
3644 if (tp->header.seq >= first)
3645 #endif /* RX_ENABLE_LOCKS */
3646 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3648 && (tp->header.serial == serial || tp->firstSerial == serial))
3649 rxi_ComputePeerNetStats(call, tp, ap, np);
3651 /* Set the acknowledge flag per packet based on the
3652 * information in the ack packet. An acknowlegded packet can
3653 * be downgraded when the server has discarded a packet it
3654 * soacked previously, or when an ack packet is received
3655 * out of sequence. */
3656 if (tp->header.seq < first) {
3657 /* Implicit ack information */
3658 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3661 tp->flags |= RX_PKTFLAG_ACKED;
3662 } else if (tp->header.seq < first + nAcks) {
3663 /* Explicit ack information: set it in the packet appropriately */
3664 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3665 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3667 tp->flags |= RX_PKTFLAG_ACKED;
3674 } else /* RX_ACK_TYPE_NACK */ {
3675 tp->flags &= ~RX_PKTFLAG_ACKED;
3679 tp->flags &= ~RX_PKTFLAG_ACKED;
3683 /* If packet isn't yet acked, and it has been transmitted at least
3684 * once, reset retransmit time using latest timeout
3685 * ie, this should readjust the retransmit timer for all outstanding
3686 * packets... So we don't just retransmit when we should know better*/
3688 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3689 tp->retryTime = tp->timeSent;
3690 clock_Add(&tp->retryTime, &peer->timeout);
3691 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3692 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3696 /* If the window has been extended by this acknowledge packet,
3697 * then wakeup a sender waiting in alloc for window space, or try
3698 * sending packets now, if he's been sitting on packets due to
3699 * lack of window space */
3700 if (call->tnext < (call->tfirst + call->twind)) {
3701 #ifdef RX_ENABLE_LOCKS
3702 CV_SIGNAL(&call->cv_twind);
3704 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3705 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3706 osi_rxWakeup(&call->twind);
3709 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3710 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3714 /* if the ack packet has a receivelen field hanging off it,
3715 * update our state */
3716 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3719 /* If the ack packet has a "recommended" size that is less than
3720 * what I am using now, reduce my size to match */
3721 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3722 (int)sizeof(afs_int32), &tSize);
3723 tSize = (afs_uint32) ntohl(tSize);
3724 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3726 /* Get the maximum packet size to send to this peer */
3727 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
3729 tSize = (afs_uint32) ntohl(tSize);
3730 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3731 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3733 /* sanity check - peer might have restarted with different params.
3734 * If peer says "send less", dammit, send less... Peer should never
3735 * be unable to accept packets of the size that prior AFS versions would
3736 * send without asking. */
3737 if (peer->maxMTU != tSize) {
3738 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
3740 peer->maxMTU = tSize;
3741 peer->MTU = MIN(tSize, peer->MTU);
3742 call->MTU = MIN(call->MTU, tSize);
3745 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3748 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3749 (int)sizeof(afs_int32), &tSize);
3750 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3751 if (tSize < call->twind) { /* smaller than our send */
3752 call->twind = tSize; /* window, we must send less... */
3753 call->ssthresh = MIN(call->twind, call->ssthresh);
3756 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3757 * network MTU confused with the loopback MTU. Calculate the
3758 * maximum MTU here for use in the slow start code below.
3760 maxMTU = peer->maxMTU;
3761 /* Did peer restart with older RX version? */
3762 if (peer->maxDgramPackets > 1) {
3763 peer->maxDgramPackets = 1;
3765 } else if (np->length >=
3766 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3769 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3770 sizeof(afs_int32), &tSize);
3771 tSize = (afs_uint32) ntohl(tSize);
3773 * As of AFS 3.5 we set the send window to match the receive window.
3775 if (tSize < call->twind) {
3776 call->twind = tSize;
3777 call->ssthresh = MIN(call->twind, call->ssthresh);
3778 } else if (tSize > call->twind) {
3779 call->twind = tSize;
3783 * As of AFS 3.5, a jumbogram is more than one fixed size
3784 * packet transmitted in a single UDP datagram. If the remote
3785 * MTU is smaller than our local MTU then never send a datagram
3786 * larger than the natural MTU.
3789 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3790 sizeof(afs_int32), &tSize);
3791 maxDgramPackets = (afs_uint32) ntohl(tSize);
3792 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3793 maxDgramPackets = MIN(maxDgramPackets, peer->ifDgramPackets);
3794 if (peer->natMTU < peer->ifMTU)
3795 maxDgramPackets = MIN(maxDgramPackets, rxi_AdjustDgramPackets(1, peer->natMTU));
3796 if (maxDgramPackets > 1) {
3797 peer->maxDgramPackets = maxDgramPackets;
3798 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3800 peer->maxDgramPackets = 1;
3801 call->MTU = peer->natMTU;
3803 } else if (peer->maxDgramPackets > 1) {
3804 /* Restarted with lower version of RX */
3805 peer->maxDgramPackets = 1;
3807 } else if (peer->maxDgramPackets > 1
3808 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3809 /* Restarted with lower version of RX */
3810 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3811 peer->natMTU = OLD_MAX_PACKET_SIZE;
3812 peer->MTU = OLD_MAX_PACKET_SIZE;
3813 peer->maxDgramPackets = 1;
3814 peer->nDgramPackets = 1;
3816 call->MTU = OLD_MAX_PACKET_SIZE;
3821 * Calculate how many datagrams were successfully received after
3822 * the first missing packet and adjust the negative ack counter
3827 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3828 if (call->nNacks < nNacked) {
3829 call->nNacks = nNacked;
3838 if (call->flags & RX_CALL_FAST_RECOVER) {
3840 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3842 call->flags &= ~RX_CALL_FAST_RECOVER;
3843 call->cwind = call->nextCwind;
3844 call->nextCwind = 0;
3847 call->nCwindAcks = 0;
3848 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3849 /* Three negative acks in a row trigger congestion recovery */
3850 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3851 MUTEX_EXIT(&peer->peer_lock);
3852 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3853 /* someone else is waiting to start recovery */
3856 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3857 rxi_WaitforTQBusy(call);
3858 MUTEX_ENTER(&peer->peer_lock);
3859 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3860 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3861 call->flags |= RX_CALL_FAST_RECOVER;
3862 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3864 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3865 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3866 call->nextCwind = call->ssthresh;
3869 peer->MTU = call->MTU;
3870 peer->cwind = call->nextCwind;
3871 peer->nDgramPackets = call->nDgramPackets;
3873 call->congestSeq = peer->congestSeq;
3874 /* Reset the resend times on the packets that were nacked
3875 * so we will retransmit as soon as the window permits*/
3876 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
3878 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3879 clock_Zero(&tp->retryTime);
3881 } else if (tp->flags & RX_PKTFLAG_ACKED) {
3886 /* If cwind is smaller than ssthresh, then increase
3887 * the window one packet for each ack we receive (exponential
3889 * If cwind is greater than or equal to ssthresh then increase
3890 * the congestion window by one packet for each cwind acks we
3891 * receive (linear growth). */
3892 if (call->cwind < call->ssthresh) {
3894 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
3895 call->nCwindAcks = 0;
3897 call->nCwindAcks += newAckCount;
3898 if (call->nCwindAcks >= call->cwind) {
3899 call->nCwindAcks = 0;
3900 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3904 * If we have received several acknowledgements in a row then
3905 * it is time to increase the size of our datagrams
3907 if ((int)call->nAcks > rx_nDgramThreshold) {
3908 if (peer->maxDgramPackets > 1) {
3909 if (call->nDgramPackets < peer->maxDgramPackets) {
3910 call->nDgramPackets++;
3912 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
3913 } else if (call->MTU < peer->maxMTU) {
3914 call->MTU += peer->natMTU;
3915 call->MTU = MIN(call->MTU, peer->maxMTU);
3921 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
3923 /* Servers need to hold the call until all response packets have
3924 * been acknowledged. Soft acks are good enough since clients
3925 * are not allowed to clear their receive queues. */
3926 if (call->state == RX_STATE_HOLD
3927 && call->tfirst + call->nSoftAcked >= call->tnext) {
3928 call->state = RX_STATE_DALLY;
3929 rxi_ClearTransmitQueue(call, 0);
3930 } else if (!queue_IsEmpty(&call->tq)) {
3931 rxi_Start(0, call, 0, istack);
3936 /* Received a response to a challenge packet */
3938 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
3939 register struct rx_packet *np, int istack)
3943 /* Ignore the packet if we're the client */
3944 if (conn->type == RX_CLIENT_CONNECTION)
3947 /* If already authenticated, ignore the packet (it's probably a retry) */
3948 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
3951 /* Otherwise, have the security object evaluate the response packet */
3952 error = RXS_CheckResponse(conn->securityObject, conn, np);
3954 /* If the response is invalid, reset the connection, sending
3955 * an abort to the peer */
3959 rxi_ConnectionError(conn, error);
3960 MUTEX_ENTER(&conn->conn_data_lock);
3961 np = rxi_SendConnectionAbort(conn, np, istack, 0);
3962 MUTEX_EXIT(&conn->conn_data_lock);
3965 /* If the response is valid, any calls waiting to attach
3966 * servers can now do so */
3969 for (i = 0; i < RX_MAXCALLS; i++) {
3970 struct rx_call *call = conn->call[i];
3972 MUTEX_ENTER(&call->lock);
3973 if (call->state == RX_STATE_PRECALL)
3974 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
3975 /* tnop can be null if newcallp is null */
3976 MUTEX_EXIT(&call->lock);
3980 /* Update the peer reachability information, just in case
3981 * some calls went into attach-wait while we were waiting
3982 * for authentication..
3984 rxi_UpdatePeerReach(conn, NULL);
3989 /* A client has received an authentication challenge: the security
3990 * object is asked to cough up a respectable response packet to send
3991 * back to the server. The server is responsible for retrying the
3992 * challenge if it fails to get a response. */
3995 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
3996 register struct rx_packet *np, int istack)
4000 /* Ignore the challenge if we're the server */
4001 if (conn->type == RX_SERVER_CONNECTION)
4004 /* Ignore the challenge if the connection is otherwise idle; someone's
4005 * trying to use us as an oracle. */
4006 if (!rxi_HasActiveCalls(conn))
4009 /* Send the security object the challenge packet. It is expected to fill
4010 * in the response. */
4011 error = RXS_GetResponse(conn->securityObject, conn, np);
4013 /* If the security object is unable to return a valid response, reset the
4014 * connection and send an abort to the peer. Otherwise send the response
4015 * packet to the peer connection. */
4017 rxi_ConnectionError(conn, error);
4018 MUTEX_ENTER(&conn->conn_data_lock);
4019 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4020 MUTEX_EXIT(&conn->conn_data_lock);
4022 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4023 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4029 /* Find an available server process to service the current request in
4030 * the given call structure. If one isn't available, queue up this
4031 * call so it eventually gets one */
4033 rxi_AttachServerProc(register struct rx_call *call,
4034 register osi_socket socket, register int *tnop,
4035 register struct rx_call **newcallp)
4037 register struct rx_serverQueueEntry *sq;
4038 register struct rx_service *service = call->conn->service;
4039 register int haveQuota = 0;
4041 /* May already be attached */
4042 if (call->state == RX_STATE_ACTIVE)
4045 MUTEX_ENTER(&rx_serverPool_lock);
4047 haveQuota = QuotaOK(service);
4048 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4049 /* If there are no processes available to service this call,
4050 * put the call on the incoming call queue (unless it's
4051 * already on the queue).
4053 #ifdef RX_ENABLE_LOCKS
4055 ReturnToServerPool(service);
4056 #endif /* RX_ENABLE_LOCKS */
4058 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4059 call->flags |= RX_CALL_WAIT_PROC;
4060 MUTEX_ENTER(&rx_stats_mutex);
4063 MUTEX_EXIT(&rx_stats_mutex);
4064 rxi_calltrace(RX_CALL_ARRIVAL, call);
4065 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4066 queue_Append(&rx_incomingCallQueue, call);
4069 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4071 /* If hot threads are enabled, and both newcallp and sq->socketp
4072 * are non-null, then this thread will process the call, and the
4073 * idle server thread will start listening on this threads socket.
4076 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4079 *sq->socketp = socket;
4080 clock_GetTime(&call->startTime);
4081 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4085 if (call->flags & RX_CALL_WAIT_PROC) {
4086 /* Conservative: I don't think this should happen */
4087 call->flags &= ~RX_CALL_WAIT_PROC;
4088 if (queue_IsOnQueue(call)) {
4090 MUTEX_ENTER(&rx_stats_mutex);
4092 MUTEX_EXIT(&rx_stats_mutex);
4095 call->state = RX_STATE_ACTIVE;
4096 call->mode = RX_MODE_RECEIVING;
4097 #ifdef RX_KERNEL_TRACE
4099 int glockOwner = ISAFS_GLOCK();
4102 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4103 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4109 if (call->flags & RX_CALL_CLEARED) {
4110 /* send an ack now to start the packet flow up again */
4111 call->flags &= ~RX_CALL_CLEARED;
4112 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4114 #ifdef RX_ENABLE_LOCKS
4117 service->nRequestsRunning++;
4118 if (service->nRequestsRunning <= service->minProcs)
4124 MUTEX_EXIT(&rx_serverPool_lock);
4127 /* Delay the sending of an acknowledge event for a short while, while
4128 * a new call is being prepared (in the case of a client) or a reply
4129 * is being prepared (in the case of a server). Rather than sending
4130 * an ack packet, an ACKALL packet is sent. */
4132 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4134 #ifdef RX_ENABLE_LOCKS
4136 MUTEX_ENTER(&call->lock);
4137 call->delayedAckEvent = NULL;
4138 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4140 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4141 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4143 MUTEX_EXIT(&call->lock);
4144 #else /* RX_ENABLE_LOCKS */
4146 call->delayedAckEvent = NULL;
4147 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4148 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4149 #endif /* RX_ENABLE_LOCKS */
4153 rxi_SendDelayedAck(struct rxevent *event, register struct rx_call *call,
4156 #ifdef RX_ENABLE_LOCKS
4158 MUTEX_ENTER(&call->lock);
4159 if (event == call->delayedAckEvent)
4160 call->delayedAckEvent = NULL;
4161 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4163 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4165 MUTEX_EXIT(&call->lock);
4166 #else /* RX_ENABLE_LOCKS */
4168 call->delayedAckEvent = NULL;
4169 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4170 #endif /* RX_ENABLE_LOCKS */
4174 #ifdef RX_ENABLE_LOCKS
4175 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4176 * clearing them out.
4179 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4181 register struct rx_packet *p, *tp;
4184 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4185 p->flags |= RX_PKTFLAG_ACKED;
4189 call->flags |= RX_CALL_TQ_CLEARME;
4190 call->flags |= RX_CALL_TQ_SOME_ACKED;
4193 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4194 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4195 call->tfirst = call->tnext;
4196 call->nSoftAcked = 0;
4198 if (call->flags & RX_CALL_FAST_RECOVER) {
4199 call->flags &= ~RX_CALL_FAST_RECOVER;
4200 call->cwind = call->nextCwind;
4201 call->nextCwind = 0;
4204 CV_SIGNAL(&call->cv_twind);
4206 #endif /* RX_ENABLE_LOCKS */
4208 /* Clear out the transmit queue for the current call (all packets have
4209 * been received by peer) */
4211 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4213 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4214 register struct rx_packet *p, *tp;
4216 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4218 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4219 p->flags |= RX_PKTFLAG_ACKED;
4223 call->flags |= RX_CALL_TQ_CLEARME;
4224 call->flags |= RX_CALL_TQ_SOME_ACKED;
4227 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4228 rxi_FreePackets(0, &call->tq);
4229 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4230 call->flags &= ~RX_CALL_TQ_CLEARME;
4232 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4234 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4235 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4236 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4237 call->nSoftAcked = 0;
4239 if (call->flags & RX_CALL_FAST_RECOVER) {
4240 call->flags &= ~RX_CALL_FAST_RECOVER;
4241 call->cwind = call->nextCwind;
4243 #ifdef RX_ENABLE_LOCKS
4244 CV_SIGNAL(&call->cv_twind);
4246 osi_rxWakeup(&call->twind);
4251 rxi_ClearReceiveQueue(register struct rx_call *call)
4253 if (queue_IsNotEmpty(&call->rq)) {
4254 rx_packetReclaims += rxi_FreePackets(0, &call->rq);
4255 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4257 if (call->state == RX_STATE_PRECALL) {
4258 call->flags |= RX_CALL_CLEARED;
4262 /* Send an abort packet for the specified call */
4264 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4265 int istack, int force)
4273 /* Clients should never delay abort messages */
4274 if (rx_IsClientConn(call->conn))
4277 if (call->abortCode != call->error) {
4278 call->abortCode = call->error;
4279 call->abortCount = 0;
4282 if (force || rxi_callAbortThreshhold == 0
4283 || call->abortCount < rxi_callAbortThreshhold) {
4284 if (call->delayedAbortEvent) {
4285 rxevent_Cancel(call->delayedAbortEvent, call,
4286 RX_CALL_REFCOUNT_ABORT);
4288 error = htonl(call->error);
4291 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4292 (char *)&error, sizeof(error), istack);
4293 } else if (!call->delayedAbortEvent) {
4294 clock_GetTime(&when);
4295 clock_Addmsec(&when, rxi_callAbortDelay);
4296 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4297 call->delayedAbortEvent =
4298 rxevent_Post(&when, rxi_SendDelayedCallAbort, call, 0);
4303 /* Send an abort packet for the specified connection. Packet is an
4304 * optional pointer to a packet that can be used to send the abort.
4305 * Once the number of abort messages reaches the threshhold, an
4306 * event is scheduled to send the abort. Setting the force flag
4307 * overrides sending delayed abort messages.
4309 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4310 * to send the abort packet.
4313 rxi_SendConnectionAbort(register struct rx_connection *conn,
4314 struct rx_packet *packet, int istack, int force)
4322 /* Clients should never delay abort messages */
4323 if (rx_IsClientConn(conn))
4326 if (force || rxi_connAbortThreshhold == 0
4327 || conn->abortCount < rxi_connAbortThreshhold) {
4328 if (conn->delayedAbortEvent) {
4329 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4331 error = htonl(conn->error);
4333 MUTEX_EXIT(&conn->conn_data_lock);
4335 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4336 RX_PACKET_TYPE_ABORT, (char *)&error,
4337 sizeof(error), istack);
4338 MUTEX_ENTER(&conn->conn_data_lock);
4339 } else if (!conn->delayedAbortEvent) {
4340 clock_GetTime(&when);
4341 clock_Addmsec(&when, rxi_connAbortDelay);
4342 conn->delayedAbortEvent =
4343 rxevent_Post(&when, rxi_SendDelayedConnAbort, conn, 0);
4348 /* Associate an error all of the calls owned by a connection. Called
4349 * with error non-zero. This is only for really fatal things, like
4350 * bad authentication responses. The connection itself is set in
4351 * error at this point, so that future packets received will be
4354 rxi_ConnectionError(register struct rx_connection *conn,
4355 register afs_int32 error)
4360 dpf(("rxi_ConnectionError conn %x error %d", conn, error));
4362 MUTEX_ENTER(&conn->conn_data_lock);
4363 if (conn->challengeEvent)
4364 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4365 if (conn->checkReachEvent) {
4366 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4367 conn->checkReachEvent = 0;
4368 conn->flags &= ~RX_CONN_ATTACHWAIT;
4371 MUTEX_EXIT(&conn->conn_data_lock);
4372 for (i = 0; i < RX_MAXCALLS; i++) {
4373 struct rx_call *call = conn->call[i];
4375 MUTEX_ENTER(&call->lock);
4376 rxi_CallError(call, error);
4377 MUTEX_EXIT(&call->lock);
4380 conn->error = error;
4381 rx_MutexIncrement(rx_stats.fatalErrors, rx_stats_mutex);
4386 rxi_CallError(register struct rx_call *call, afs_int32 error)
4388 dpf(("rxi_CallError call %x error %d call->error %d", call, error, call->error));
4390 error = call->error;
4392 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4393 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4394 rxi_ResetCall(call, 0);
4397 rxi_ResetCall(call, 0);
4399 call->error = error;
4400 call->mode = RX_MODE_ERROR;
4403 /* Reset various fields in a call structure, and wakeup waiting
4404 * processes. Some fields aren't changed: state & mode are not
4405 * touched (these must be set by the caller), and bufptr, nLeft, and
4406 * nFree are not reset, since these fields are manipulated by
4407 * unprotected macros, and may only be reset by non-interrupting code.
4410 /* this code requires that call->conn be set properly as a pre-condition. */
4411 #endif /* ADAPT_WINDOW */
4414 rxi_ResetCall(register struct rx_call *call, register int newcall)
4417 register struct rx_peer *peer;
4418 struct rx_packet *packet;
4420 dpf(("rxi_ResetCall(call %x, newcall %d)\n", call, newcall));
4422 /* Notify anyone who is waiting for asynchronous packet arrival */
4423 if (call->arrivalProc) {
4424 (*call->arrivalProc) (call, call->arrivalProcHandle,
4425 call->arrivalProcArg);
4426 call->arrivalProc = (void (*)())0;
4429 if (call->delayedAbortEvent) {
4430 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4431 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4433 rxi_SendCallAbort(call, packet, 0, 1);
4434 rxi_FreePacket(packet);
4439 * Update the peer with the congestion information in this call
4440 * so other calls on this connection can pick up where this call
4441 * left off. If the congestion sequence numbers don't match then
4442 * another call experienced a retransmission.
4444 peer = call->conn->peer;
4445 MUTEX_ENTER(&peer->peer_lock);
4447 if (call->congestSeq == peer->congestSeq) {
4448 peer->cwind = MAX(peer->cwind, call->cwind);
4449 peer->MTU = MAX(peer->MTU, call->MTU);
4450 peer->nDgramPackets =
4451 MAX(peer->nDgramPackets, call->nDgramPackets);
4454 call->abortCode = 0;
4455 call->abortCount = 0;
4457 if (peer->maxDgramPackets > 1) {
4458 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4460 call->MTU = peer->MTU;
4462 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4463 call->ssthresh = rx_maxSendWindow;
4464 call->nDgramPackets = peer->nDgramPackets;
4465 call->congestSeq = peer->congestSeq;
4466 MUTEX_EXIT(&peer->peer_lock);
4468 flags = call->flags;
4469 rxi_ClearReceiveQueue(call);
4470 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4471 if (flags & RX_CALL_TQ_BUSY) {
4472 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4473 call->flags |= (flags & RX_CALL_TQ_WAIT);
4475 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4477 rxi_ClearTransmitQueue(call, 0);
4478 queue_Init(&call->tq);
4479 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4480 dpf(("rcall %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4483 while (call->tqWaiters) {
4484 #ifdef RX_ENABLE_LOCKS
4485 CV_BROADCAST(&call->cv_tq);
4486 #else /* RX_ENABLE_LOCKS */
4487 osi_rxWakeup(&call->tq);
4488 #endif /* RX_ENABLE_LOCKS */
4492 queue_Init(&call->rq);
4494 call->rwind = rx_initReceiveWindow;
4495 call->twind = rx_initSendWindow;
4496 call->nSoftAcked = 0;
4497 call->nextCwind = 0;
4500 call->nCwindAcks = 0;
4501 call->nSoftAcks = 0;
4502 call->nHardAcks = 0;
4504 call->tfirst = call->rnext = call->tnext = 1;
4506 call->lastAcked = 0;
4507 call->localStatus = call->remoteStatus = 0;
4509 if (flags & RX_CALL_READER_WAIT) {
4510 #ifdef RX_ENABLE_LOCKS
4511 CV_BROADCAST(&call->cv_rq);
4513 osi_rxWakeup(&call->rq);
4516 if (flags & RX_CALL_WAIT_PACKETS) {
4517 MUTEX_ENTER(&rx_freePktQ_lock);
4518 rxi_PacketsUnWait(); /* XXX */
4519 MUTEX_EXIT(&rx_freePktQ_lock);
4521 #ifdef RX_ENABLE_LOCKS
4522 CV_SIGNAL(&call->cv_twind);
4524 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4525 osi_rxWakeup(&call->twind);
4528 #ifdef RX_ENABLE_LOCKS
4529 /* The following ensures that we don't mess with any queue while some
4530 * other thread might also be doing so. The call_queue_lock field is
4531 * is only modified under the call lock. If the call is in the process
4532 * of being removed from a queue, the call is not locked until the
4533 * the queue lock is dropped and only then is the call_queue_lock field
4534 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4535 * Note that any other routine which removes a call from a queue has to
4536 * obtain the queue lock before examing the queue and removing the call.
4538 if (call->call_queue_lock) {
4539 MUTEX_ENTER(call->call_queue_lock);
4540 if (queue_IsOnQueue(call)) {
4542 if (flags & RX_CALL_WAIT_PROC) {
4543 MUTEX_ENTER(&rx_stats_mutex);
4545 MUTEX_EXIT(&rx_stats_mutex);
4548 MUTEX_EXIT(call->call_queue_lock);
4549 CLEAR_CALL_QUEUE_LOCK(call);
4551 #else /* RX_ENABLE_LOCKS */
4552 if (queue_IsOnQueue(call)) {
4554 if (flags & RX_CALL_WAIT_PROC)
4557 #endif /* RX_ENABLE_LOCKS */
4559 rxi_KeepAliveOff(call);
4560 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4563 /* Send an acknowledge for the indicated packet (seq,serial) of the
4564 * indicated call, for the indicated reason (reason). This
4565 * acknowledge will specifically acknowledge receiving the packet, and
4566 * will also specify which other packets for this call have been
4567 * received. This routine returns the packet that was used to the
4568 * caller. The caller is responsible for freeing it or re-using it.
4569 * This acknowledgement also returns the highest sequence number
4570 * actually read out by the higher level to the sender; the sender
4571 * promises to keep around packets that have not been read by the
4572 * higher level yet (unless, of course, the sender decides to abort
4573 * the call altogether). Any of p, seq, serial, pflags, or reason may
4574 * be set to zero without ill effect. That is, if they are zero, they
4575 * will not convey any information.
4576 * NOW there is a trailer field, after the ack where it will safely be
4577 * ignored by mundanes, which indicates the maximum size packet this
4578 * host can swallow. */
4580 register struct rx_packet *optionalPacket; use to send ack (or null)
4581 int seq; Sequence number of the packet we are acking
4582 int serial; Serial number of the packet
4583 int pflags; Flags field from packet header
4584 int reason; Reason an acknowledge was prompted
4588 rxi_SendAck(register struct rx_call *call,
4589 register struct rx_packet *optionalPacket, int serial, int reason,
4592 struct rx_ackPacket *ap;
4593 register struct rx_packet *rqp;
4594 register struct rx_packet *nxp; /* For queue_Scan */
4595 register struct rx_packet *p;
4598 #ifdef RX_ENABLE_TSFPQ
4599 struct rx_ts_info_t * rx_ts_info;
4603 * Open the receive window once a thread starts reading packets
4605 if (call->rnext > 1) {
4606 call->rwind = rx_maxReceiveWindow;
4609 call->nHardAcks = 0;
4610 call->nSoftAcks = 0;
4611 if (call->rnext > call->lastAcked)
4612 call->lastAcked = call->rnext;
4616 rx_computelen(p, p->length); /* reset length, you never know */
4617 } /* where that's been... */
4618 #ifdef RX_ENABLE_TSFPQ
4620 RX_TS_INFO_GET(rx_ts_info);
4621 if ((p = rx_ts_info->local_special_packet)) {
4622 rx_computelen(p, p->length);
4623 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4624 rx_ts_info->local_special_packet = p;
4625 } else { /* We won't send the ack, but don't panic. */
4626 return optionalPacket;
4630 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4631 /* We won't send the ack, but don't panic. */
4632 return optionalPacket;
4637 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4640 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4641 #ifndef RX_ENABLE_TSFPQ
4642 if (!optionalPacket)
4645 return optionalPacket;
4647 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4648 if (rx_Contiguous(p) < templ) {
4649 #ifndef RX_ENABLE_TSFPQ
4650 if (!optionalPacket)
4653 return optionalPacket;
4658 /* MTUXXX failing to send an ack is very serious. We should */
4659 /* try as hard as possible to send even a partial ack; it's */
4660 /* better than nothing. */
4661 ap = (struct rx_ackPacket *)rx_DataOf(p);
4662 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4663 ap->reason = reason;
4665 /* The skew computation used to be bogus, I think it's better now. */
4666 /* We should start paying attention to skew. XXX */
4667 ap->serial = htonl(serial);
4668 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4670 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4671 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4673 /* No fear of running out of ack packet here because there can only be at most
4674 * one window full of unacknowledged packets. The window size must be constrained
4675 * to be less than the maximum ack size, of course. Also, an ack should always
4676 * fit into a single packet -- it should not ever be fragmented. */
4677 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4678 if (!rqp || !call->rq.next
4679 || (rqp->header.seq > (call->rnext + call->rwind))) {
4680 #ifndef RX_ENABLE_TSFPQ
4681 if (!optionalPacket)
4684 rxi_CallError(call, RX_CALL_DEAD);
4685 return optionalPacket;
4688 while (rqp->header.seq > call->rnext + offset)
4689 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4690 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4692 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4693 #ifndef RX_ENABLE_TSFPQ
4694 if (!optionalPacket)
4697 rxi_CallError(call, RX_CALL_DEAD);
4698 return optionalPacket;
4703 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4705 /* these are new for AFS 3.3 */
4706 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4707 templ = htonl(templ);
4708 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4709 templ = htonl(call->conn->peer->ifMTU);
4710 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4711 sizeof(afs_int32), &templ);
4713 /* new for AFS 3.4 */
4714 templ = htonl(call->rwind);
4715 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4716 sizeof(afs_int32), &templ);
4718 /* new for AFS 3.5 */
4719 templ = htonl(call->conn->peer->ifDgramPackets);
4720 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4721 sizeof(afs_int32), &templ);
4723 p->header.serviceId = call->conn->serviceId;
4724 p->header.cid = (call->conn->cid | call->channel);
4725 p->header.callNumber = *call->callNumber;
4727 p->header.securityIndex = call->conn->securityIndex;
4728 p->header.epoch = call->conn->epoch;
4729 p->header.type = RX_PACKET_TYPE_ACK;
4730 p->header.flags = RX_SLOW_START_OK;
4731 if (reason == RX_ACK_PING) {
4732 p->header.flags |= RX_REQUEST_ACK;
4734 clock_GetTime(&call->pingRequestTime);
4737 if (call->conn->type == RX_CLIENT_CONNECTION)
4738 p->header.flags |= RX_CLIENT_INITIATED;
4742 if (rxdebug_active) {
4746 len = _snprintf(msg, sizeof(msg),
4747 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4748 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4749 ntohl(ap->serial), ntohl(ap->previousPacket),
4750 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
4751 ap->nAcks, ntohs(ap->bufferSpace) );
4755 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
4756 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4760 OutputDebugString(msg);
4762 #else /* AFS_NT40_ENV */
4764 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
4765 ap->reason, ntohl(ap->previousPacket),
4766 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4768 for (offset = 0; offset < ap->nAcks; offset++)
4769 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4774 #endif /* AFS_NT40_ENV */
4777 register int i, nbytes = p->length;
4779 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4780 if (nbytes <= p->wirevec[i].iov_len) {
4781 register int savelen, saven;
4783 savelen = p->wirevec[i].iov_len;
4785 p->wirevec[i].iov_len = nbytes;
4787 rxi_Send(call, p, istack);
4788 p->wirevec[i].iov_len = savelen;
4792 nbytes -= p->wirevec[i].iov_len;
4795 rx_MutexIncrement(rx_stats.ackPacketsSent, rx_stats_mutex);
4796 #ifndef RX_ENABLE_TSFPQ
4797 if (!optionalPacket)
4800 return optionalPacket; /* Return packet for re-use by caller */
4803 /* Send all of the packets in the list in single datagram */
4805 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4806 int istack, int moreFlag, struct clock *now,
4807 struct clock *retryTime, int resending)
4812 struct rx_connection *conn = call->conn;
4813 struct rx_peer *peer = conn->peer;
4815 MUTEX_ENTER(&peer->peer_lock);
4818 peer->reSends += len;
4819 rx_MutexIncrement(rx_stats.dataPacketsSent, rx_stats_mutex);
4820 MUTEX_EXIT(&peer->peer_lock);
4822 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4826 /* Set the packet flags and schedule the resend events */
4827 /* Only request an ack for the last packet in the list */
4828 for (i = 0; i < len; i++) {
4829 list[i]->retryTime = *retryTime;
4830 if (list[i]->header.serial) {
4831 /* Exponentially backoff retry times */
4832 if (list[i]->backoff < MAXBACKOFF) {
4833 /* so it can't stay == 0 */
4834 list[i]->backoff = (list[i]->backoff << 1) + 1;
4837 clock_Addmsec(&(list[i]->retryTime),
4838 ((afs_uint32) list[i]->backoff) << 8);
4841 /* Wait a little extra for the ack on the last packet */
4842 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4843 clock_Addmsec(&(list[i]->retryTime), 400);
4846 /* Record the time sent */
4847 list[i]->timeSent = *now;
4849 /* Ask for an ack on retransmitted packets, on every other packet
4850 * if the peer doesn't support slow start. Ask for an ack on every
4851 * packet until the congestion window reaches the ack rate. */
4852 if (list[i]->header.serial) {
4854 rx_MutexIncrement(rx_stats.dataPacketsReSent, rx_stats_mutex);
4856 /* improved RTO calculation- not Karn */
4857 list[i]->firstSent = *now;
4858 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
4859 || (!(call->flags & RX_CALL_SLOW_START_OK)
4860 && (list[i]->header.seq & 1)))) {
4865 MUTEX_ENTER(&peer->peer_lock);
4869 rx_MutexIncrement(rx_stats.dataPacketsSent, rx_stats_mutex);
4870 MUTEX_EXIT(&peer->peer_lock);
4872 /* Tag this packet as not being the last in this group,
4873 * for the receiver's benefit */
4874 if (i < len - 1 || moreFlag) {
4875 list[i]->header.flags |= RX_MORE_PACKETS;
4878 /* Install the new retransmit time for the packet, and
4879 * record the time sent */
4880 list[i]->timeSent = *now;
4884 list[len - 1]->header.flags |= RX_REQUEST_ACK;
4887 /* Since we're about to send a data packet to the peer, it's
4888 * safe to nuke any scheduled end-of-packets ack */
4889 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4891 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
4892 MUTEX_EXIT(&call->lock);
4894 rxi_SendPacketList(call, conn, list, len, istack);
4896 rxi_SendPacket(call, conn, list[0], istack);
4898 MUTEX_ENTER(&call->lock);
4899 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
4901 /* Update last send time for this call (for keep-alive
4902 * processing), and for the connection (so that we can discover
4903 * idle connections) */
4904 conn->lastSendTime = call->lastSendTime = clock_Sec();
4907 /* When sending packets we need to follow these rules:
4908 * 1. Never send more than maxDgramPackets in a jumbogram.
4909 * 2. Never send a packet with more than two iovecs in a jumbogram.
4910 * 3. Never send a retransmitted packet in a jumbogram.
4911 * 4. Never send more than cwind/4 packets in a jumbogram
4912 * We always keep the last list we should have sent so we
4913 * can set the RX_MORE_PACKETS flags correctly.
4916 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
4917 int istack, struct clock *now, struct clock *retryTime,
4920 int i, cnt, lastCnt = 0;
4921 struct rx_packet **listP, **lastP = 0;
4922 struct rx_peer *peer = call->conn->peer;
4923 int morePackets = 0;
4925 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
4926 /* Does the current packet force us to flush the current list? */
4928 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
4929 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
4931 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
4933 /* If the call enters an error state stop sending, or if
4934 * we entered congestion recovery mode, stop sending */
4935 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4943 /* Add the current packet to the list if it hasn't been acked.
4944 * Otherwise adjust the list pointer to skip the current packet. */
4945 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
4947 /* Do we need to flush the list? */
4948 if (cnt >= (int)peer->maxDgramPackets
4949 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
4950 || list[i]->header.serial
4951 || list[i]->length != RX_JUMBOBUFFERSIZE) {
4953 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
4954 retryTime, resending);
4955 /* If the call enters an error state stop sending, or if
4956 * we entered congestion recovery mode, stop sending */
4958 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4963 listP = &list[i + 1];
4968 osi_Panic("rxi_SendList error");
4970 listP = &list[i + 1];
4974 /* Send the whole list when the call is in receive mode, when
4975 * the call is in eof mode, when we are in fast recovery mode,
4976 * and when we have the last packet */
4977 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
4978 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
4979 || (call->flags & RX_CALL_FAST_RECOVER)) {
4980 /* Check for the case where the current list contains
4981 * an acked packet. Since we always send retransmissions
4982 * in a separate packet, we only need to check the first
4983 * packet in the list */
4984 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
4988 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
4989 retryTime, resending);
4990 /* If the call enters an error state stop sending, or if
4991 * we entered congestion recovery mode, stop sending */
4992 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
4996 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
4999 } else if (lastCnt > 0) {
5000 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5005 #ifdef RX_ENABLE_LOCKS
5006 /* Call rxi_Start, below, but with the call lock held. */
5008 rxi_StartUnlocked(struct rxevent *event, register struct rx_call *call,
5009 void *arg1, int istack)
5011 MUTEX_ENTER(&call->lock);
5012 rxi_Start(event, call, arg1, istack);
5013 MUTEX_EXIT(&call->lock);
5015 #endif /* RX_ENABLE_LOCKS */
5017 /* This routine is called when new packets are readied for
5018 * transmission and when retransmission may be necessary, or when the
5019 * transmission window or burst count are favourable. This should be
5020 * better optimized for new packets, the usual case, now that we've
5021 * got rid of queues of send packets. XXXXXXXXXXX */
5023 rxi_Start(struct rxevent *event, register struct rx_call *call,
5024 void *arg1, int istack)
5026 struct rx_packet *p;
5027 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
5028 struct rx_peer *peer = call->conn->peer;
5029 struct clock now, retryTime;
5033 struct rx_packet **xmitList;
5036 /* If rxi_Start is being called as a result of a resend event,
5037 * then make sure that the event pointer is removed from the call
5038 * structure, since there is no longer a per-call retransmission
5040 if (event && event == call->resendEvent) {
5041 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5042 call->resendEvent = NULL;
5044 if (queue_IsEmpty(&call->tq)) {
5048 /* Timeouts trigger congestion recovery */
5049 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5050 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5051 /* someone else is waiting to start recovery */
5054 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5055 rxi_WaitforTQBusy(call);
5056 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5057 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5058 call->flags |= RX_CALL_FAST_RECOVER;
5059 if (peer->maxDgramPackets > 1) {
5060 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5062 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5064 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5065 call->nDgramPackets = 1;
5067 call->nextCwind = 1;
5070 MUTEX_ENTER(&peer->peer_lock);
5071 peer->MTU = call->MTU;
5072 peer->cwind = call->cwind;
5073 peer->nDgramPackets = 1;
5075 call->congestSeq = peer->congestSeq;
5076 MUTEX_EXIT(&peer->peer_lock);
5077 /* Clear retry times on packets. Otherwise, it's possible for
5078 * some packets in the queue to force resends at rates faster
5079 * than recovery rates.
5081 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5082 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5083 clock_Zero(&p->retryTime);
5088 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5089 rx_MutexIncrement(rx_tq_debug.rxi_start_in_error, rx_stats_mutex);
5094 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5095 /* Get clock to compute the re-transmit time for any packets
5096 * in this burst. Note, if we back off, it's reasonable to
5097 * back off all of the packets in the same manner, even if
5098 * some of them have been retransmitted more times than more
5099 * recent additions */
5100 clock_GetTime(&now);
5101 retryTime = now; /* initialize before use */
5102 MUTEX_ENTER(&peer->peer_lock);
5103 clock_Add(&retryTime, &peer->timeout);
5104 MUTEX_EXIT(&peer->peer_lock);
5106 /* Send (or resend) any packets that need it, subject to
5107 * window restrictions and congestion burst control
5108 * restrictions. Ask for an ack on the last packet sent in
5109 * this burst. For now, we're relying upon the window being
5110 * considerably bigger than the largest number of packets that
5111 * are typically sent at once by one initial call to
5112 * rxi_Start. This is probably bogus (perhaps we should ask
5113 * for an ack when we're half way through the current
5114 * window?). Also, for non file transfer applications, this
5115 * may end up asking for an ack for every packet. Bogus. XXXX
5118 * But check whether we're here recursively, and let the other guy
5121 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5122 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5123 call->flags |= RX_CALL_TQ_BUSY;
5125 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5127 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5128 call->flags &= ~RX_CALL_NEED_START;
5129 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5131 maxXmitPackets = MIN(call->twind, call->cwind);
5132 xmitList = (struct rx_packet **)
5133 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5134 if (xmitList == NULL)
5135 osi_Panic("rxi_Start, failed to allocate xmit list");
5136 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5137 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5138 /* We shouldn't be sending packets if a thread is waiting
5139 * to initiate congestion recovery */
5143 && (call->flags & RX_CALL_FAST_RECOVER)) {
5144 /* Only send one packet during fast recovery */
5147 if ((p->flags & RX_PKTFLAG_FREE)
5148 || (!queue_IsEnd(&call->tq, nxp)
5149 && (nxp->flags & RX_PKTFLAG_FREE))
5150 || (p == (struct rx_packet *)&rx_freePacketQueue)
5151 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5152 osi_Panic("rxi_Start: xmit queue clobbered");
5154 if (p->flags & RX_PKTFLAG_ACKED) {
5155 rx_MutexIncrement(rx_stats.ignoreAckedPacket, rx_stats_mutex);
5156 continue; /* Ignore this packet if it has been acknowledged */
5159 /* Turn off all flags except these ones, which are the same
5160 * on each transmission */
5161 p->header.flags &= RX_PRESET_FLAGS;
5163 if (p->header.seq >=
5164 call->tfirst + MIN((int)call->twind,
5165 (int)(call->nSoftAcked +
5167 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5168 /* Note: if we're waiting for more window space, we can
5169 * still send retransmits; hence we don't return here, but
5170 * break out to schedule a retransmit event */
5171 dpf(("call %d waiting for window",
5172 *(call->callNumber)));
5176 /* Transmit the packet if it needs to be sent. */
5177 if (!clock_Lt(&now, &p->retryTime)) {
5178 if (nXmitPackets == maxXmitPackets) {
5179 rxi_SendXmitList(call, xmitList, nXmitPackets,
5180 istack, &now, &retryTime,
5182 osi_Free(xmitList, maxXmitPackets *
5183 sizeof(struct rx_packet *));
5186 xmitList[nXmitPackets++] = p;
5190 /* xmitList now hold pointers to all of the packets that are
5191 * ready to send. Now we loop to send the packets */
5192 if (nXmitPackets > 0) {
5193 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5194 &now, &retryTime, resending);
5197 maxXmitPackets * sizeof(struct rx_packet *));
5199 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5201 * TQ references no longer protected by this flag; they must remain
5202 * protected by the global lock.
5204 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5205 call->flags &= ~RX_CALL_TQ_BUSY;
5206 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5207 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5208 #ifdef RX_ENABLE_LOCKS
5209 osirx_AssertMine(&call->lock, "rxi_Start start");
5210 CV_BROADCAST(&call->cv_tq);
5211 #else /* RX_ENABLE_LOCKS */
5212 osi_rxWakeup(&call->tq);
5213 #endif /* RX_ENABLE_LOCKS */
5218 /* We went into the error state while sending packets. Now is
5219 * the time to reset the call. This will also inform the using
5220 * process that the call is in an error state.
5222 rx_MutexIncrement(rx_tq_debug.rxi_start_aborted, rx_stats_mutex);
5223 call->flags &= ~RX_CALL_TQ_BUSY;
5224 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5225 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5226 #ifdef RX_ENABLE_LOCKS
5227 osirx_AssertMine(&call->lock, "rxi_Start middle");
5228 CV_BROADCAST(&call->cv_tq);
5229 #else /* RX_ENABLE_LOCKS */
5230 osi_rxWakeup(&call->tq);
5231 #endif /* RX_ENABLE_LOCKS */
5233 rxi_CallError(call, call->error);
5236 #ifdef RX_ENABLE_LOCKS
5237 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5238 register int missing;
5239 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5240 /* Some packets have received acks. If they all have, we can clear
5241 * the transmit queue.
5244 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5245 if (p->header.seq < call->tfirst
5246 && (p->flags & RX_PKTFLAG_ACKED)) {
5253 call->flags |= RX_CALL_TQ_CLEARME;
5255 #endif /* RX_ENABLE_LOCKS */
5256 /* Don't bother doing retransmits if the TQ is cleared. */
5257 if (call->flags & RX_CALL_TQ_CLEARME) {
5258 rxi_ClearTransmitQueue(call, 1);
5260 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5263 /* Always post a resend event, if there is anything in the
5264 * queue, and resend is possible. There should be at least
5265 * one unacknowledged packet in the queue ... otherwise none
5266 * of these packets should be on the queue in the first place.
5268 if (call->resendEvent) {
5269 /* Cancel the existing event and post a new one */
5270 rxevent_Cancel(call->resendEvent, call,
5271 RX_CALL_REFCOUNT_RESEND);
5274 /* The retry time is the retry time on the first unacknowledged
5275 * packet inside the current window */
5277 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5278 /* Don't set timers for packets outside the window */
5279 if (p->header.seq >= call->tfirst + call->twind) {
5283 if (!(p->flags & RX_PKTFLAG_ACKED)
5284 && !clock_IsZero(&p->retryTime)) {
5286 retryTime = p->retryTime;
5291 /* Post a new event to re-run rxi_Start when retries may be needed */
5292 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5293 #ifdef RX_ENABLE_LOCKS
5294 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5296 rxevent_Post2(&retryTime, rxi_StartUnlocked,
5297 (void *)call, 0, istack);
5298 #else /* RX_ENABLE_LOCKS */
5300 rxevent_Post2(&retryTime, rxi_Start, (void *)call,
5302 #endif /* RX_ENABLE_LOCKS */
5305 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5306 } while (call->flags & RX_CALL_NEED_START);
5308 * TQ references no longer protected by this flag; they must remain
5309 * protected by the global lock.
5311 call->flags &= ~RX_CALL_TQ_BUSY;
5312 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5313 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5314 #ifdef RX_ENABLE_LOCKS
5315 osirx_AssertMine(&call->lock, "rxi_Start end");
5316 CV_BROADCAST(&call->cv_tq);
5317 #else /* RX_ENABLE_LOCKS */
5318 osi_rxWakeup(&call->tq);
5319 #endif /* RX_ENABLE_LOCKS */
5322 call->flags |= RX_CALL_NEED_START;
5324 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5326 if (call->resendEvent) {
5327 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5332 /* Also adjusts the keep alive parameters for the call, to reflect
5333 * that we have just sent a packet (so keep alives aren't sent
5336 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5339 register struct rx_connection *conn = call->conn;
5341 /* Stamp each packet with the user supplied status */
5342 p->header.userStatus = call->localStatus;
5344 /* Allow the security object controlling this call's security to
5345 * make any last-minute changes to the packet */
5346 RXS_SendPacket(conn->securityObject, call, p);
5348 /* Since we're about to send SOME sort of packet to the peer, it's
5349 * safe to nuke any scheduled end-of-packets ack */
5350 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5352 /* Actually send the packet, filling in more connection-specific fields */
5353 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5354 MUTEX_EXIT(&call->lock);
5355 rxi_SendPacket(call, conn, p, istack);
5356 MUTEX_ENTER(&call->lock);
5357 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5359 /* Update last send time for this call (for keep-alive
5360 * processing), and for the connection (so that we can discover
5361 * idle connections) */
5362 conn->lastSendTime = call->lastSendTime = clock_Sec();
5366 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5367 * that things are fine. Also called periodically to guarantee that nothing
5368 * falls through the cracks (e.g. (error + dally) connections have keepalive
5369 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5371 * haveCTLock Set if calling from rxi_ReapConnections
5373 #ifdef RX_ENABLE_LOCKS
5375 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5376 #else /* RX_ENABLE_LOCKS */
5378 rxi_CheckCall(register struct rx_call *call)
5379 #endif /* RX_ENABLE_LOCKS */
5381 register struct rx_connection *conn = call->conn;
5383 afs_uint32 deadTime;
5385 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5386 if (call->flags & RX_CALL_TQ_BUSY) {
5387 /* Call is active and will be reset by rxi_Start if it's
5388 * in an error state.
5393 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5395 (((afs_uint32) conn->secondsUntilDead << 10) +
5396 ((afs_uint32) conn->peer->rtt >> 3) +
5397 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5399 /* These are computed to the second (+- 1 second). But that's
5400 * good enough for these values, which should be a significant
5401 * number of seconds. */
5402 if (now > (call->lastReceiveTime + deadTime)) {
5403 if (call->state == RX_STATE_ACTIVE) {
5404 rxi_CallError(call, RX_CALL_DEAD);
5407 #ifdef RX_ENABLE_LOCKS
5408 /* Cancel pending events */
5409 rxevent_Cancel(call->delayedAckEvent, call,
5410 RX_CALL_REFCOUNT_DELAY);
5411 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5412 rxevent_Cancel(call->keepAliveEvent, call,
5413 RX_CALL_REFCOUNT_ALIVE);
5414 if (call->refCount == 0) {
5415 rxi_FreeCall(call, haveCTLock);
5419 #else /* RX_ENABLE_LOCKS */
5422 #endif /* RX_ENABLE_LOCKS */
5424 /* Non-active calls are destroyed if they are not responding
5425 * to pings; active calls are simply flagged in error, so the
5426 * attached process can die reasonably gracefully. */
5428 /* see if we have a non-activity timeout */
5429 if (call->startWait && conn->idleDeadTime
5430 && ((call->startWait + conn->idleDeadTime) < now)) {
5431 if (call->state == RX_STATE_ACTIVE) {
5432 rxi_CallError(call, RX_CALL_TIMEOUT);
5436 /* see if we have a hard timeout */
5437 if (conn->hardDeadTime
5438 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5439 if (call->state == RX_STATE_ACTIVE)
5440 rxi_CallError(call, RX_CALL_TIMEOUT);
5447 /* When a call is in progress, this routine is called occasionally to
5448 * make sure that some traffic has arrived (or been sent to) the peer.
5449 * If nothing has arrived in a reasonable amount of time, the call is
5450 * declared dead; if nothing has been sent for a while, we send a
5451 * keep-alive packet (if we're actually trying to keep the call alive)
5454 rxi_KeepAliveEvent(struct rxevent *event, register struct rx_call *call,
5457 struct rx_connection *conn;
5460 MUTEX_ENTER(&call->lock);
5461 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5462 if (event == call->keepAliveEvent)
5463 call->keepAliveEvent = NULL;
5466 #ifdef RX_ENABLE_LOCKS
5467 if (rxi_CheckCall(call, 0)) {
5468 MUTEX_EXIT(&call->lock);
5471 #else /* RX_ENABLE_LOCKS */
5472 if (rxi_CheckCall(call))
5474 #endif /* RX_ENABLE_LOCKS */
5476 /* Don't try to keep alive dallying calls */
5477 if (call->state == RX_STATE_DALLY) {
5478 MUTEX_EXIT(&call->lock);
5483 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5484 /* Don't try to send keepalives if there is unacknowledged data */
5485 /* the rexmit code should be good enough, this little hack
5486 * doesn't quite work XXX */
5487 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5489 rxi_ScheduleKeepAliveEvent(call);
5490 MUTEX_EXIT(&call->lock);
5495 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5497 if (!call->keepAliveEvent) {
5499 clock_GetTime(&when);
5500 when.sec += call->conn->secondsUntilPing;
5501 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5502 call->keepAliveEvent =
5503 rxevent_Post(&when, rxi_KeepAliveEvent, call, 0);
5507 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5509 rxi_KeepAliveOn(register struct rx_call *call)
5511 /* Pretend last packet received was received now--i.e. if another
5512 * packet isn't received within the keep alive time, then the call
5513 * will die; Initialize last send time to the current time--even
5514 * if a packet hasn't been sent yet. This will guarantee that a
5515 * keep-alive is sent within the ping time */
5516 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5517 rxi_ScheduleKeepAliveEvent(call);
5520 /* This routine is called to send connection abort messages
5521 * that have been delayed to throttle looping clients. */
5523 rxi_SendDelayedConnAbort(struct rxevent *event,
5524 register struct rx_connection *conn, char *dummy)
5527 struct rx_packet *packet;
5529 MUTEX_ENTER(&conn->conn_data_lock);
5530 conn->delayedAbortEvent = NULL;
5531 error = htonl(conn->error);
5533 MUTEX_EXIT(&conn->conn_data_lock);
5534 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5537 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5538 RX_PACKET_TYPE_ABORT, (char *)&error,
5540 rxi_FreePacket(packet);
5544 /* This routine is called to send call abort messages
5545 * that have been delayed to throttle looping clients. */
5547 rxi_SendDelayedCallAbort(struct rxevent *event, register struct rx_call *call,
5551 struct rx_packet *packet;
5553 MUTEX_ENTER(&call->lock);
5554 call->delayedAbortEvent = NULL;
5555 error = htonl(call->error);
5557 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5560 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5561 (char *)&error, sizeof(error), 0);
5562 rxi_FreePacket(packet);
5564 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5565 MUTEX_EXIT(&call->lock);
5568 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5569 * seconds) to ask the client to authenticate itself. The routine
5570 * issues a challenge to the client, which is obtained from the
5571 * security object associated with the connection */
5573 rxi_ChallengeEvent(struct rxevent *event, register struct rx_connection *conn,
5574 void *arg1, int tries)
5576 conn->challengeEvent = NULL;
5577 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5578 register struct rx_packet *packet;
5582 /* We've failed to authenticate for too long.
5583 * Reset any calls waiting for authentication;
5584 * they are all in RX_STATE_PRECALL.
5588 MUTEX_ENTER(&conn->conn_call_lock);
5589 for (i = 0; i < RX_MAXCALLS; i++) {
5590 struct rx_call *call = conn->call[i];
5592 MUTEX_ENTER(&call->lock);
5593 if (call->state == RX_STATE_PRECALL) {
5594 rxi_CallError(call, RX_CALL_DEAD);
5595 rxi_SendCallAbort(call, NULL, 0, 0);
5597 MUTEX_EXIT(&call->lock);
5600 MUTEX_EXIT(&conn->conn_call_lock);
5604 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5606 /* If there's no packet available, do this later. */
5607 RXS_GetChallenge(conn->securityObject, conn, packet);
5608 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5609 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5610 rxi_FreePacket(packet);
5612 clock_GetTime(&when);
5613 when.sec += RX_CHALLENGE_TIMEOUT;
5614 conn->challengeEvent =
5615 rxevent_Post2(&when, rxi_ChallengeEvent, conn, 0,
5620 /* Call this routine to start requesting the client to authenticate
5621 * itself. This will continue until authentication is established,
5622 * the call times out, or an invalid response is returned. The
5623 * security object associated with the connection is asked to create
5624 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5625 * defined earlier. */
5627 rxi_ChallengeOn(register struct rx_connection *conn)
5629 if (!conn->challengeEvent) {
5630 RXS_CreateChallenge(conn->securityObject, conn);
5631 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5636 /* Compute round trip time of the packet provided, in *rttp.
5639 /* rxi_ComputeRoundTripTime is called with peer locked. */
5640 /* sentp and/or peer may be null */
5642 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5643 register struct clock *sentp,
5644 register struct rx_peer *peer)
5646 struct clock thisRtt, *rttp = &thisRtt;
5648 register int rtt_timeout;
5650 clock_GetTime(rttp);
5652 if (clock_Lt(rttp, sentp)) {
5654 return; /* somebody set the clock back, don't count this time. */
5656 clock_Sub(rttp, sentp);
5657 MUTEX_ENTER(&rx_stats_mutex);
5658 if (clock_Lt(rttp, &rx_stats.minRtt))
5659 rx_stats.minRtt = *rttp;
5660 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5661 if (rttp->sec > 60) {
5662 MUTEX_EXIT(&rx_stats_mutex);
5663 return; /* somebody set the clock ahead */
5665 rx_stats.maxRtt = *rttp;
5667 clock_Add(&rx_stats.totalRtt, rttp);
5668 rx_stats.nRttSamples++;
5669 MUTEX_EXIT(&rx_stats_mutex);
5671 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5673 /* Apply VanJacobson round-trip estimations */
5678 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5679 * srtt is stored as fixed point with 3 bits after the binary
5680 * point (i.e., scaled by 8). The following magic is
5681 * equivalent to the smoothing algorithm in rfc793 with an
5682 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5683 * srtt*8 = srtt*8 + rtt - srtt
5684 * srtt = srtt + rtt/8 - srtt/8
5687 delta = MSEC(rttp) - (peer->rtt >> 3);
5691 * We accumulate a smoothed rtt variance (actually, a smoothed
5692 * mean difference), then set the retransmit timer to smoothed
5693 * rtt + 4 times the smoothed variance (was 2x in van's original
5694 * paper, but 4x works better for me, and apparently for him as
5696 * rttvar is stored as
5697 * fixed point with 2 bits after the binary point (scaled by
5698 * 4). The following is equivalent to rfc793 smoothing with
5699 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5700 * replaces rfc793's wired-in beta.
5701 * dev*4 = dev*4 + (|actual - expected| - dev)
5707 delta -= (peer->rtt_dev >> 2);
5708 peer->rtt_dev += delta;
5710 /* I don't have a stored RTT so I start with this value. Since I'm
5711 * probably just starting a call, and will be pushing more data down
5712 * this, I expect congestion to increase rapidly. So I fudge a
5713 * little, and I set deviance to half the rtt. In practice,
5714 * deviance tends to approach something a little less than
5715 * half the smoothed rtt. */
5716 peer->rtt = (MSEC(rttp) << 3) + 8;
5717 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5719 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5720 * the other of these connections is usually in a user process, and can
5721 * be switched and/or swapped out. So on fast, reliable networks, the
5722 * timeout would otherwise be too short.
5724 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5725 clock_Zero(&(peer->timeout));
5726 clock_Addmsec(&(peer->timeout), rtt_timeout);
5728 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)));
5732 /* Find all server connections that have not been active for a long time, and
5735 rxi_ReapConnections(void)
5738 clock_GetTime(&now);
5740 /* Find server connection structures that haven't been used for
5741 * greater than rx_idleConnectionTime */
5743 struct rx_connection **conn_ptr, **conn_end;
5744 int i, havecalls = 0;
5745 MUTEX_ENTER(&rx_connHashTable_lock);
5746 for (conn_ptr = &rx_connHashTable[0], conn_end =
5747 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5749 struct rx_connection *conn, *next;
5750 struct rx_call *call;
5754 for (conn = *conn_ptr; conn; conn = next) {
5755 /* XXX -- Shouldn't the connection be locked? */
5758 for (i = 0; i < RX_MAXCALLS; i++) {
5759 call = conn->call[i];
5762 MUTEX_ENTER(&call->lock);
5763 #ifdef RX_ENABLE_LOCKS
5764 result = rxi_CheckCall(call, 1);
5765 #else /* RX_ENABLE_LOCKS */
5766 result = rxi_CheckCall(call);
5767 #endif /* RX_ENABLE_LOCKS */
5768 MUTEX_EXIT(&call->lock);
5770 /* If CheckCall freed the call, it might
5771 * have destroyed the connection as well,
5772 * which screws up the linked lists.
5778 if (conn->type == RX_SERVER_CONNECTION) {
5779 /* This only actually destroys the connection if
5780 * there are no outstanding calls */
5781 MUTEX_ENTER(&conn->conn_data_lock);
5782 if (!havecalls && !conn->refCount
5783 && ((conn->lastSendTime + rx_idleConnectionTime) <
5785 conn->refCount++; /* it will be decr in rx_DestroyConn */
5786 MUTEX_EXIT(&conn->conn_data_lock);
5787 #ifdef RX_ENABLE_LOCKS
5788 rxi_DestroyConnectionNoLock(conn);
5789 #else /* RX_ENABLE_LOCKS */
5790 rxi_DestroyConnection(conn);
5791 #endif /* RX_ENABLE_LOCKS */
5793 #ifdef RX_ENABLE_LOCKS
5795 MUTEX_EXIT(&conn->conn_data_lock);
5797 #endif /* RX_ENABLE_LOCKS */
5801 #ifdef RX_ENABLE_LOCKS
5802 while (rx_connCleanup_list) {
5803 struct rx_connection *conn;
5804 conn = rx_connCleanup_list;
5805 rx_connCleanup_list = rx_connCleanup_list->next;
5806 MUTEX_EXIT(&rx_connHashTable_lock);
5807 rxi_CleanupConnection(conn);
5808 MUTEX_ENTER(&rx_connHashTable_lock);
5810 MUTEX_EXIT(&rx_connHashTable_lock);
5811 #endif /* RX_ENABLE_LOCKS */
5814 /* Find any peer structures that haven't been used (haven't had an
5815 * associated connection) for greater than rx_idlePeerTime */
5817 struct rx_peer **peer_ptr, **peer_end;
5819 MUTEX_ENTER(&rx_rpc_stats);
5820 MUTEX_ENTER(&rx_peerHashTable_lock);
5821 for (peer_ptr = &rx_peerHashTable[0], peer_end =
5822 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
5824 struct rx_peer *peer, *next, *prev;
5825 for (prev = peer = *peer_ptr; peer; peer = next) {
5827 code = MUTEX_TRYENTER(&peer->peer_lock);
5828 if ((code) && (peer->refCount == 0)
5829 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
5830 rx_interface_stat_p rpc_stat, nrpc_stat;
5832 MUTEX_EXIT(&peer->peer_lock);
5833 MUTEX_DESTROY(&peer->peer_lock);
5835 (&peer->rpcStats, rpc_stat, nrpc_stat,
5836 rx_interface_stat)) {
5837 unsigned int num_funcs;
5840 queue_Remove(&rpc_stat->queue_header);
5841 queue_Remove(&rpc_stat->all_peers);
5842 num_funcs = rpc_stat->stats[0].func_total;
5844 sizeof(rx_interface_stat_t) +
5845 rpc_stat->stats[0].func_total *
5846 sizeof(rx_function_entry_v1_t);
5848 rxi_Free(rpc_stat, space);
5849 rxi_rpc_peer_stat_cnt -= num_funcs;
5852 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
5853 if (peer == *peer_ptr) {
5860 MUTEX_EXIT(&peer->peer_lock);
5866 MUTEX_EXIT(&rx_peerHashTable_lock);
5867 MUTEX_EXIT(&rx_rpc_stats);
5870 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
5871 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
5872 * GC, just below. Really, we shouldn't have to keep moving packets from
5873 * one place to another, but instead ought to always know if we can
5874 * afford to hold onto a packet in its particular use. */
5875 MUTEX_ENTER(&rx_freePktQ_lock);
5876 if (rx_waitingForPackets) {
5877 rx_waitingForPackets = 0;
5878 #ifdef RX_ENABLE_LOCKS
5879 CV_BROADCAST(&rx_waitingForPackets_cv);
5881 osi_rxWakeup(&rx_waitingForPackets);
5884 MUTEX_EXIT(&rx_freePktQ_lock);
5886 now.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
5887 rxevent_Post(&now, rxi_ReapConnections, 0, 0);
5891 /* rxs_Release - This isn't strictly necessary but, since the macro name from
5892 * rx.h is sort of strange this is better. This is called with a security
5893 * object before it is discarded. Each connection using a security object has
5894 * its own refcount to the object so it won't actually be freed until the last
5895 * connection is destroyed.
5897 * This is the only rxs module call. A hold could also be written but no one
5901 rxs_Release(struct rx_securityClass *aobj)
5903 return RXS_Close(aobj);
5907 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
5908 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
5909 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
5910 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
5912 /* Adjust our estimate of the transmission rate to this peer, given
5913 * that the packet p was just acked. We can adjust peer->timeout and
5914 * call->twind. Pragmatically, this is called
5915 * only with packets of maximal length.
5916 * Called with peer and call locked.
5920 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
5921 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
5923 afs_int32 xferSize, xferMs;
5924 register afs_int32 minTime;
5927 /* Count down packets */
5928 if (peer->rateFlag > 0)
5930 /* Do nothing until we're enabled */
5931 if (peer->rateFlag != 0)
5936 /* Count only when the ack seems legitimate */
5937 switch (ackReason) {
5938 case RX_ACK_REQUESTED:
5940 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
5944 case RX_ACK_PING_RESPONSE:
5945 if (p) /* want the response to ping-request, not data send */
5947 clock_GetTime(&newTO);
5948 if (clock_Gt(&newTO, &call->pingRequestTime)) {
5949 clock_Sub(&newTO, &call->pingRequestTime);
5950 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
5954 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
5961 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));
5963 /* Track only packets that are big enough. */
5964 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
5968 /* absorb RTT data (in milliseconds) for these big packets */
5969 if (peer->smRtt == 0) {
5970 peer->smRtt = xferMs;
5972 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
5977 if (peer->countDown) {
5981 peer->countDown = 10; /* recalculate only every so often */
5983 /* In practice, we can measure only the RTT for full packets,
5984 * because of the way Rx acks the data that it receives. (If it's
5985 * smaller than a full packet, it often gets implicitly acked
5986 * either by the call response (from a server) or by the next call
5987 * (from a client), and either case confuses transmission times
5988 * with processing times.) Therefore, replace the above
5989 * more-sophisticated processing with a simpler version, where the
5990 * smoothed RTT is kept for full-size packets, and the time to
5991 * transmit a windowful of full-size packets is simply RTT *
5992 * windowSize. Again, we take two steps:
5993 - ensure the timeout is large enough for a single packet's RTT;
5994 - ensure that the window is small enough to fit in the desired timeout.*/
5996 /* First, the timeout check. */
5997 minTime = peer->smRtt;
5998 /* Get a reasonable estimate for a timeout period */
6000 newTO.sec = minTime / 1000;
6001 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6003 /* Increase the timeout period so that we can always do at least
6004 * one packet exchange */
6005 if (clock_Gt(&newTO, &peer->timeout)) {
6007 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));
6009 peer->timeout = newTO;
6012 /* Now, get an estimate for the transmit window size. */
6013 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6014 /* Now, convert to the number of full packets that could fit in a
6015 * reasonable fraction of that interval */
6016 minTime /= (peer->smRtt << 1);
6017 xferSize = minTime; /* (make a copy) */
6019 /* Now clamp the size to reasonable bounds. */
6022 else if (minTime > rx_Window)
6023 minTime = rx_Window;
6024 /* if (minTime != peer->maxWindow) {
6025 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6026 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6027 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6029 peer->maxWindow = minTime;
6030 elide... call->twind = minTime;
6034 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6035 * Discern this by calculating the timeout necessary for rx_Window
6037 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6038 /* calculate estimate for transmission interval in milliseconds */
6039 minTime = rx_Window * peer->smRtt;
6040 if (minTime < 1000) {
6041 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
6042 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6043 peer->timeout.usec, peer->smRtt, peer->packetSize));
6045 newTO.sec = 0; /* cut back on timeout by half a second */
6046 newTO.usec = 500000;
6047 clock_Sub(&peer->timeout, &newTO);
6052 } /* end of rxi_ComputeRate */
6053 #endif /* ADAPT_WINDOW */
6061 #define TRACE_OPTION_DEBUGLOG 4
6069 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6070 0, KEY_QUERY_VALUE, &parmKey);
6071 if (code != ERROR_SUCCESS)
6074 dummyLen = sizeof(TraceOption);
6075 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6076 (BYTE *) &TraceOption, &dummyLen);
6077 if (code == ERROR_SUCCESS) {
6078 rxdebug_active = (TraceOption & TRACE_OPTION_DEBUGLOG) ? 1 : 0;
6080 RegCloseKey (parmKey);
6081 #endif /* AFS_NT40_ENV */
6086 rx_DebugOnOff(int on)
6088 rxdebug_active = on;
6090 #endif /* AFS_NT40_ENV */
6093 /* Don't call this debugging routine directly; use dpf */
6095 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
6096 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
6104 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6107 len = _snprintf(msg, sizeof(msg)-2,
6108 tformat, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10,
6109 a11, a12, a13, a14, a15);
6111 if (msg[len-1] != '\n') {
6115 OutputDebugString(msg);
6120 clock_GetTime(&now);
6121 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
6122 (unsigned int)now.usec / 1000);
6123 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
6130 * This function is used to process the rx_stats structure that is local
6131 * to a process as well as an rx_stats structure received from a remote
6132 * process (via rxdebug). Therefore, it needs to do minimal version
6136 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
6137 afs_int32 freePackets, char version)
6141 if (size != sizeof(struct rx_stats)) {
6143 "Unexpected size of stats structure: was %d, expected %d\n",
6144 size, sizeof(struct rx_stats));
6147 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6150 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6151 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6152 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6153 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6154 s->specialPktAllocFailures);
6156 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6157 s->receivePktAllocFailures, s->sendPktAllocFailures,
6158 s->specialPktAllocFailures);
6162 " greedy %d, " "bogusReads %d (last from host %x), "
6163 "noPackets %d, " "noBuffers %d, " "selects %d, "
6164 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6165 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6166 s->selects, s->sendSelects);
6168 fprintf(file, " packets read: ");
6169 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6170 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6172 fprintf(file, "\n");
6175 " other read counters: data %d, " "ack %d, " "dup %d "
6176 "spurious %d " "dally %d\n", s->dataPacketsRead,
6177 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6178 s->ignorePacketDally);
6180 fprintf(file, " packets sent: ");
6181 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6182 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6184 fprintf(file, "\n");
6187 " other send counters: ack %d, " "data %d (not resends), "
6188 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6189 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6190 s->dataPacketsPushed, s->ignoreAckedPacket);
6193 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6194 s->netSendFailures, (int)s->fatalErrors);
6196 if (s->nRttSamples) {
6197 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6198 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6200 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6201 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6205 " %d server connections, " "%d client connections, "
6206 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6207 s->nServerConns, s->nClientConns, s->nPeerStructs,
6208 s->nCallStructs, s->nFreeCallStructs);
6210 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6211 fprintf(file, " %d clock updates\n", clock_nUpdates);
6216 /* for backward compatibility */
6218 rx_PrintStats(FILE * file)
6220 MUTEX_ENTER(&rx_stats_mutex);
6221 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6223 MUTEX_EXIT(&rx_stats_mutex);
6227 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6229 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6230 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6231 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6234 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6235 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6236 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6239 " Packet size %d, " "max in packet skew %d, "
6240 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6241 (int)peer->outPacketSkew);
6244 #ifdef AFS_PTHREAD_ENV
6246 * This mutex protects the following static variables:
6250 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6251 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6253 #define LOCK_RX_DEBUG
6254 #define UNLOCK_RX_DEBUG
6255 #endif /* AFS_PTHREAD_ENV */
6258 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6259 u_char type, void *inputData, size_t inputLength,
6260 void *outputData, size_t outputLength)
6262 static afs_int32 counter = 100;
6263 time_t waitTime, waitCount, startTime;
6264 struct rx_header theader;
6266 register afs_int32 code;
6267 struct timeval tv_now, tv_wake, tv_delta;
6268 struct sockaddr_in taddr, faddr;
6273 startTime = time(0);
6279 tp = &tbuffer[sizeof(struct rx_header)];
6280 taddr.sin_family = AF_INET;
6281 taddr.sin_port = remotePort;
6282 taddr.sin_addr.s_addr = remoteAddr;
6283 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6284 taddr.sin_len = sizeof(struct sockaddr_in);
6287 memset(&theader, 0, sizeof(theader));
6288 theader.epoch = htonl(999);
6290 theader.callNumber = htonl(counter);
6293 theader.type = type;
6294 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6295 theader.serviceId = 0;
6297 memcpy(tbuffer, &theader, sizeof(theader));
6298 memcpy(tp, inputData, inputLength);
6300 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6301 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6303 /* see if there's a packet available */
6304 gettimeofday(&tv_wake,0);
6305 tv_wake.tv_sec += waitTime;
6308 FD_SET(socket, &imask);
6309 tv_delta.tv_sec = tv_wake.tv_sec;
6310 tv_delta.tv_usec = tv_wake.tv_usec;
6311 gettimeofday(&tv_now, 0);
6313 if (tv_delta.tv_usec < tv_now.tv_usec) {
6315 tv_delta.tv_usec += 1000000;
6318 tv_delta.tv_usec -= tv_now.tv_usec;
6320 if (tv_delta.tv_sec < tv_now.tv_sec) {
6324 tv_delta.tv_sec -= tv_now.tv_sec;
6326 code = select(socket + 1, &imask, 0, 0, &tv_delta);
6327 if (code == 1 && FD_ISSET(socket, &imask)) {
6328 /* now receive a packet */
6329 faddrLen = sizeof(struct sockaddr_in);
6331 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6332 (struct sockaddr *)&faddr, &faddrLen);
6335 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6336 if (counter == ntohl(theader.callNumber))
6344 /* see if we've timed out */
6352 code -= sizeof(struct rx_header);
6353 if (code > outputLength)
6354 code = outputLength;
6355 memcpy(outputData, tp, code);
6360 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6361 afs_uint16 remotePort, struct rx_debugStats * stat,
6362 afs_uint32 * supportedValues)
6364 struct rx_debugIn in;
6367 *supportedValues = 0;
6368 in.type = htonl(RX_DEBUGI_GETSTATS);
6371 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6372 &in, sizeof(in), stat, sizeof(*stat));
6375 * If the call was successful, fixup the version and indicate
6376 * what contents of the stat structure are valid.
6377 * Also do net to host conversion of fields here.
6381 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6382 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6384 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6385 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6387 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6388 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6390 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6391 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6393 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6394 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6396 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6397 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6399 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6400 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6402 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6403 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6406 stat->nFreePackets = ntohl(stat->nFreePackets);
6407 stat->packetReclaims = ntohl(stat->packetReclaims);
6408 stat->callsExecuted = ntohl(stat->callsExecuted);
6409 stat->nWaiting = ntohl(stat->nWaiting);
6410 stat->idleThreads = ntohl(stat->idleThreads);
6417 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6418 afs_uint16 remotePort, struct rx_stats * stat,
6419 afs_uint32 * supportedValues)
6421 struct rx_debugIn in;
6422 afs_int32 *lp = (afs_int32 *) stat;
6427 * supportedValues is currently unused, but added to allow future
6428 * versioning of this function.
6431 *supportedValues = 0;
6432 in.type = htonl(RX_DEBUGI_RXSTATS);
6434 memset(stat, 0, sizeof(*stat));
6436 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6437 &in, sizeof(in), stat, sizeof(*stat));
6442 * Do net to host conversion here
6445 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6454 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6455 afs_uint16 remotePort, size_t version_length,
6459 return MakeDebugCall(socket, remoteAddr, remotePort,
6460 RX_PACKET_TYPE_VERSION, a, 1, version,
6465 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6466 afs_uint16 remotePort, afs_int32 * nextConnection,
6467 int allConnections, afs_uint32 debugSupportedValues,
6468 struct rx_debugConn * conn,
6469 afs_uint32 * supportedValues)
6471 struct rx_debugIn in;
6476 * supportedValues is currently unused, but added to allow future
6477 * versioning of this function.
6480 *supportedValues = 0;
6481 if (allConnections) {
6482 in.type = htonl(RX_DEBUGI_GETALLCONN);
6484 in.type = htonl(RX_DEBUGI_GETCONN);
6486 in.index = htonl(*nextConnection);
6487 memset(conn, 0, sizeof(*conn));
6489 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6490 &in, sizeof(in), conn, sizeof(*conn));
6493 *nextConnection += 1;
6496 * Convert old connection format to new structure.
6499 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6500 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6501 #define MOVEvL(a) (conn->a = vL->a)
6503 /* any old or unrecognized version... */
6504 for (i = 0; i < RX_MAXCALLS; i++) {
6505 MOVEvL(callState[i]);
6506 MOVEvL(callMode[i]);
6507 MOVEvL(callFlags[i]);
6508 MOVEvL(callOther[i]);
6510 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6511 MOVEvL(secStats.type);
6512 MOVEvL(secStats.level);
6513 MOVEvL(secStats.flags);
6514 MOVEvL(secStats.expires);
6515 MOVEvL(secStats.packetsReceived);
6516 MOVEvL(secStats.packetsSent);
6517 MOVEvL(secStats.bytesReceived);
6518 MOVEvL(secStats.bytesSent);
6523 * Do net to host conversion here
6525 * I don't convert host or port since we are most likely
6526 * going to want these in NBO.
6528 conn->cid = ntohl(conn->cid);
6529 conn->serial = ntohl(conn->serial);
6530 for (i = 0; i < RX_MAXCALLS; i++) {
6531 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6533 conn->error = ntohl(conn->error);
6534 conn->secStats.flags = ntohl(conn->secStats.flags);
6535 conn->secStats.expires = ntohl(conn->secStats.expires);
6536 conn->secStats.packetsReceived =
6537 ntohl(conn->secStats.packetsReceived);
6538 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6539 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6540 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6541 conn->epoch = ntohl(conn->epoch);
6542 conn->natMTU = ntohl(conn->natMTU);
6549 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6550 afs_uint16 remotePort, afs_int32 * nextPeer,
6551 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6552 afs_uint32 * supportedValues)
6554 struct rx_debugIn in;
6558 * supportedValues is currently unused, but added to allow future
6559 * versioning of this function.
6562 *supportedValues = 0;
6563 in.type = htonl(RX_DEBUGI_GETPEER);
6564 in.index = htonl(*nextPeer);
6565 memset(peer, 0, sizeof(*peer));
6567 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6568 &in, sizeof(in), peer, sizeof(*peer));
6574 * Do net to host conversion here
6576 * I don't convert host or port since we are most likely
6577 * going to want these in NBO.
6579 peer->ifMTU = ntohs(peer->ifMTU);
6580 peer->idleWhen = ntohl(peer->idleWhen);
6581 peer->refCount = ntohs(peer->refCount);
6582 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6583 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6584 peer->rtt = ntohl(peer->rtt);
6585 peer->rtt_dev = ntohl(peer->rtt_dev);
6586 peer->timeout.sec = ntohl(peer->timeout.sec);
6587 peer->timeout.usec = ntohl(peer->timeout.usec);
6588 peer->nSent = ntohl(peer->nSent);
6589 peer->reSends = ntohl(peer->reSends);
6590 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6591 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6592 peer->rateFlag = ntohl(peer->rateFlag);
6593 peer->natMTU = ntohs(peer->natMTU);
6594 peer->maxMTU = ntohs(peer->maxMTU);
6595 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6596 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6597 peer->MTU = ntohs(peer->MTU);
6598 peer->cwind = ntohs(peer->cwind);
6599 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6600 peer->congestSeq = ntohs(peer->congestSeq);
6601 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6602 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6603 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6604 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6609 #endif /* RXDEBUG */
6614 struct rx_serverQueueEntry *np;
6617 register struct rx_call *call;
6618 register struct rx_serverQueueEntry *sq;
6622 if (rxinit_status == 1) {
6624 return; /* Already shutdown. */
6628 #ifndef AFS_PTHREAD_ENV
6629 FD_ZERO(&rx_selectMask);
6630 #endif /* AFS_PTHREAD_ENV */
6631 rxi_dataQuota = RX_MAX_QUOTA;
6632 #ifndef AFS_PTHREAD_ENV
6634 #endif /* AFS_PTHREAD_ENV */
6637 #ifndef AFS_PTHREAD_ENV
6638 #ifndef AFS_USE_GETTIMEOFDAY
6640 #endif /* AFS_USE_GETTIMEOFDAY */
6641 #endif /* AFS_PTHREAD_ENV */
6643 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6644 call = queue_First(&rx_freeCallQueue, rx_call);
6646 rxi_Free(call, sizeof(struct rx_call));
6649 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6650 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6656 struct rx_peer **peer_ptr, **peer_end;
6657 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6658 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6660 struct rx_peer *peer, *next;
6661 for (peer = *peer_ptr; peer; peer = next) {
6662 rx_interface_stat_p rpc_stat, nrpc_stat;
6665 (&peer->rpcStats, rpc_stat, nrpc_stat,
6666 rx_interface_stat)) {
6667 unsigned int num_funcs;
6670 queue_Remove(&rpc_stat->queue_header);
6671 queue_Remove(&rpc_stat->all_peers);
6672 num_funcs = rpc_stat->stats[0].func_total;
6674 sizeof(rx_interface_stat_t) +
6675 rpc_stat->stats[0].func_total *
6676 sizeof(rx_function_entry_v1_t);
6678 rxi_Free(rpc_stat, space);
6679 MUTEX_ENTER(&rx_rpc_stats);
6680 rxi_rpc_peer_stat_cnt -= num_funcs;
6681 MUTEX_EXIT(&rx_rpc_stats);
6685 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6689 for (i = 0; i < RX_MAX_SERVICES; i++) {
6691 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6693 for (i = 0; i < rx_hashTableSize; i++) {
6694 register struct rx_connection *tc, *ntc;
6695 MUTEX_ENTER(&rx_connHashTable_lock);
6696 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6698 for (j = 0; j < RX_MAXCALLS; j++) {
6700 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6703 rxi_Free(tc, sizeof(*tc));
6705 MUTEX_EXIT(&rx_connHashTable_lock);
6708 MUTEX_ENTER(&freeSQEList_lock);
6710 while ((np = rx_FreeSQEList)) {
6711 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6712 MUTEX_DESTROY(&np->lock);
6713 rxi_Free(np, sizeof(*np));
6716 MUTEX_EXIT(&freeSQEList_lock);
6717 MUTEX_DESTROY(&freeSQEList_lock);
6718 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6719 MUTEX_DESTROY(&rx_connHashTable_lock);
6720 MUTEX_DESTROY(&rx_peerHashTable_lock);
6721 MUTEX_DESTROY(&rx_serverPool_lock);
6723 osi_Free(rx_connHashTable,
6724 rx_hashTableSize * sizeof(struct rx_connection *));
6725 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6727 UNPIN(rx_connHashTable,
6728 rx_hashTableSize * sizeof(struct rx_connection *));
6729 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6731 rxi_FreeAllPackets();
6733 MUTEX_ENTER(&rx_stats_mutex);
6734 rxi_dataQuota = RX_MAX_QUOTA;
6735 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6736 MUTEX_EXIT(&rx_stats_mutex);
6742 #ifdef RX_ENABLE_LOCKS
6744 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6746 if (!MUTEX_ISMINE(lockaddr))
6747 osi_Panic("Lock not held: %s", msg);
6749 #endif /* RX_ENABLE_LOCKS */
6754 * Routines to implement connection specific data.
6758 rx_KeyCreate(rx_destructor_t rtn)
6761 MUTEX_ENTER(&rxi_keyCreate_lock);
6762 key = rxi_keyCreate_counter++;
6763 rxi_keyCreate_destructor = (rx_destructor_t *)
6764 realloc((void *)rxi_keyCreate_destructor,
6765 (key + 1) * sizeof(rx_destructor_t));
6766 rxi_keyCreate_destructor[key] = rtn;
6767 MUTEX_EXIT(&rxi_keyCreate_lock);
6772 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6775 MUTEX_ENTER(&conn->conn_data_lock);
6776 if (!conn->specific) {
6777 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
6778 for (i = 0; i < key; i++)
6779 conn->specific[i] = NULL;
6780 conn->nSpecific = key + 1;
6781 conn->specific[key] = ptr;
6782 } else if (key >= conn->nSpecific) {
6783 conn->specific = (void **)
6784 realloc(conn->specific, (key + 1) * sizeof(void *));
6785 for (i = conn->nSpecific; i < key; i++)
6786 conn->specific[i] = NULL;
6787 conn->nSpecific = key + 1;
6788 conn->specific[key] = ptr;
6790 if (conn->specific[key] && rxi_keyCreate_destructor[key])
6791 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
6792 conn->specific[key] = ptr;
6794 MUTEX_EXIT(&conn->conn_data_lock);
6798 rx_GetSpecific(struct rx_connection *conn, int key)
6801 MUTEX_ENTER(&conn->conn_data_lock);
6802 if (key >= conn->nSpecific)
6805 ptr = conn->specific[key];
6806 MUTEX_EXIT(&conn->conn_data_lock);
6810 #endif /* !KERNEL */
6813 * processStats is a queue used to store the statistics for the local
6814 * process. Its contents are similar to the contents of the rpcStats
6815 * queue on a rx_peer structure, but the actual data stored within
6816 * this queue contains totals across the lifetime of the process (assuming
6817 * the stats have not been reset) - unlike the per peer structures
6818 * which can come and go based upon the peer lifetime.
6821 static struct rx_queue processStats = { &processStats, &processStats };
6824 * peerStats is a queue used to store the statistics for all peer structs.
6825 * Its contents are the union of all the peer rpcStats queues.
6828 static struct rx_queue peerStats = { &peerStats, &peerStats };
6831 * rxi_monitor_processStats is used to turn process wide stat collection
6835 static int rxi_monitor_processStats = 0;
6838 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
6841 static int rxi_monitor_peerStats = 0;
6844 * rxi_AddRpcStat - given all of the information for a particular rpc
6845 * call, create (if needed) and update the stat totals for the rpc.
6849 * IN stats - the queue of stats that will be updated with the new value
6851 * IN rxInterface - a unique number that identifies the rpc interface
6853 * IN currentFunc - the index of the function being invoked
6855 * IN totalFunc - the total number of functions in this interface
6857 * IN queueTime - the amount of time this function waited for a thread
6859 * IN execTime - the amount of time this function invocation took to execute
6861 * IN bytesSent - the number bytes sent by this invocation
6863 * IN bytesRcvd - the number bytes received by this invocation
6865 * IN isServer - if true, this invocation was made to a server
6867 * IN remoteHost - the ip address of the remote host
6869 * IN remotePort - the port of the remote host
6871 * IN addToPeerList - if != 0, add newly created stat to the global peer list
6873 * INOUT counter - if a new stats structure is allocated, the counter will
6874 * be updated with the new number of allocated stat structures
6882 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
6883 afs_uint32 currentFunc, afs_uint32 totalFunc,
6884 struct clock *queueTime, struct clock *execTime,
6885 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
6886 afs_uint32 remoteHost, afs_uint32 remotePort,
6887 int addToPeerList, unsigned int *counter)
6890 rx_interface_stat_p rpc_stat, nrpc_stat;
6893 * See if there's already a structure for this interface
6896 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
6897 if ((rpc_stat->stats[0].interfaceId == rxInterface)
6898 && (rpc_stat->stats[0].remote_is_server == isServer))
6903 * Didn't find a match so allocate a new structure and add it to the
6907 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
6908 || (rpc_stat->stats[0].interfaceId != rxInterface)
6909 || (rpc_stat->stats[0].remote_is_server != isServer)) {
6914 sizeof(rx_interface_stat_t) +
6915 totalFunc * sizeof(rx_function_entry_v1_t);
6917 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
6918 if (rpc_stat == NULL) {
6922 *counter += totalFunc;
6923 for (i = 0; i < totalFunc; i++) {
6924 rpc_stat->stats[i].remote_peer = remoteHost;
6925 rpc_stat->stats[i].remote_port = remotePort;
6926 rpc_stat->stats[i].remote_is_server = isServer;
6927 rpc_stat->stats[i].interfaceId = rxInterface;
6928 rpc_stat->stats[i].func_total = totalFunc;
6929 rpc_stat->stats[i].func_index = i;
6930 hzero(rpc_stat->stats[i].invocations);
6931 hzero(rpc_stat->stats[i].bytes_sent);
6932 hzero(rpc_stat->stats[i].bytes_rcvd);
6933 rpc_stat->stats[i].queue_time_sum.sec = 0;
6934 rpc_stat->stats[i].queue_time_sum.usec = 0;
6935 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
6936 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
6937 rpc_stat->stats[i].queue_time_min.sec = 9999999;
6938 rpc_stat->stats[i].queue_time_min.usec = 9999999;
6939 rpc_stat->stats[i].queue_time_max.sec = 0;
6940 rpc_stat->stats[i].queue_time_max.usec = 0;
6941 rpc_stat->stats[i].execution_time_sum.sec = 0;
6942 rpc_stat->stats[i].execution_time_sum.usec = 0;
6943 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
6944 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
6945 rpc_stat->stats[i].execution_time_min.sec = 9999999;
6946 rpc_stat->stats[i].execution_time_min.usec = 9999999;
6947 rpc_stat->stats[i].execution_time_max.sec = 0;
6948 rpc_stat->stats[i].execution_time_max.usec = 0;
6950 queue_Prepend(stats, rpc_stat);
6951 if (addToPeerList) {
6952 queue_Prepend(&peerStats, &rpc_stat->all_peers);
6957 * Increment the stats for this function
6960 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
6961 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
6962 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
6963 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
6964 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
6965 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
6966 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
6968 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
6969 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
6971 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
6972 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
6974 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
6975 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
6977 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
6978 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
6986 * rx_IncrementTimeAndCount - increment the times and count for a particular
6991 * IN peer - the peer who invoked the rpc
6993 * IN rxInterface - a unique number that identifies the rpc interface
6995 * IN currentFunc - the index of the function being invoked
6997 * IN totalFunc - the total number of functions in this interface
6999 * IN queueTime - the amount of time this function waited for a thread
7001 * IN execTime - the amount of time this function invocation took to execute
7003 * IN bytesSent - the number bytes sent by this invocation
7005 * IN bytesRcvd - the number bytes received by this invocation
7007 * IN isServer - if true, this invocation was made to a server
7015 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7016 afs_uint32 currentFunc, afs_uint32 totalFunc,
7017 struct clock *queueTime, struct clock *execTime,
7018 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7022 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
7025 MUTEX_ENTER(&rx_rpc_stats);
7026 MUTEX_ENTER(&peer->peer_lock);
7028 if (rxi_monitor_peerStats) {
7029 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7030 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7031 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7034 if (rxi_monitor_processStats) {
7035 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7036 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7037 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7040 MUTEX_EXIT(&peer->peer_lock);
7041 MUTEX_EXIT(&rx_rpc_stats);
7046 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7050 * IN callerVersion - the rpc stat version of the caller.
7052 * IN count - the number of entries to marshall.
7054 * IN stats - pointer to stats to be marshalled.
7056 * OUT ptr - Where to store the marshalled data.
7063 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7064 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7070 * We only support the first version
7072 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7073 *(ptr++) = stats->remote_peer;
7074 *(ptr++) = stats->remote_port;
7075 *(ptr++) = stats->remote_is_server;
7076 *(ptr++) = stats->interfaceId;
7077 *(ptr++) = stats->func_total;
7078 *(ptr++) = stats->func_index;
7079 *(ptr++) = hgethi(stats->invocations);
7080 *(ptr++) = hgetlo(stats->invocations);
7081 *(ptr++) = hgethi(stats->bytes_sent);
7082 *(ptr++) = hgetlo(stats->bytes_sent);
7083 *(ptr++) = hgethi(stats->bytes_rcvd);
7084 *(ptr++) = hgetlo(stats->bytes_rcvd);
7085 *(ptr++) = stats->queue_time_sum.sec;
7086 *(ptr++) = stats->queue_time_sum.usec;
7087 *(ptr++) = stats->queue_time_sum_sqr.sec;
7088 *(ptr++) = stats->queue_time_sum_sqr.usec;
7089 *(ptr++) = stats->queue_time_min.sec;
7090 *(ptr++) = stats->queue_time_min.usec;
7091 *(ptr++) = stats->queue_time_max.sec;
7092 *(ptr++) = stats->queue_time_max.usec;
7093 *(ptr++) = stats->execution_time_sum.sec;
7094 *(ptr++) = stats->execution_time_sum.usec;
7095 *(ptr++) = stats->execution_time_sum_sqr.sec;
7096 *(ptr++) = stats->execution_time_sum_sqr.usec;
7097 *(ptr++) = stats->execution_time_min.sec;
7098 *(ptr++) = stats->execution_time_min.usec;
7099 *(ptr++) = stats->execution_time_max.sec;
7100 *(ptr++) = stats->execution_time_max.usec;
7106 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7111 * IN callerVersion - the rpc stat version of the caller
7113 * OUT myVersion - the rpc stat version of this function
7115 * OUT clock_sec - local time seconds
7117 * OUT clock_usec - local time microseconds
7119 * OUT allocSize - the number of bytes allocated to contain stats
7121 * OUT statCount - the number stats retrieved from this process.
7123 * OUT stats - the actual stats retrieved from this process.
7127 * Returns void. If successful, stats will != NULL.
7131 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7132 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7133 size_t * allocSize, afs_uint32 * statCount,
7134 afs_uint32 ** stats)
7144 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7147 * Check to see if stats are enabled
7150 MUTEX_ENTER(&rx_rpc_stats);
7151 if (!rxi_monitor_processStats) {
7152 MUTEX_EXIT(&rx_rpc_stats);
7156 clock_GetTime(&now);
7157 *clock_sec = now.sec;
7158 *clock_usec = now.usec;
7161 * Allocate the space based upon the caller version
7163 * If the client is at an older version than we are,
7164 * we return the statistic data in the older data format, but
7165 * we still return our version number so the client knows we
7166 * are maintaining more data than it can retrieve.
7169 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7170 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7171 *statCount = rxi_rpc_process_stat_cnt;
7174 * This can't happen yet, but in the future version changes
7175 * can be handled by adding additional code here
7179 if (space > (size_t) 0) {
7181 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7184 rx_interface_stat_p rpc_stat, nrpc_stat;
7188 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7190 * Copy the data based upon the caller version
7192 rx_MarshallProcessRPCStats(callerVersion,
7193 rpc_stat->stats[0].func_total,
7194 rpc_stat->stats, &ptr);
7200 MUTEX_EXIT(&rx_rpc_stats);
7205 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7209 * IN callerVersion - the rpc stat version of the caller
7211 * OUT myVersion - the rpc stat version of this function
7213 * OUT clock_sec - local time seconds
7215 * OUT clock_usec - local time microseconds
7217 * OUT allocSize - the number of bytes allocated to contain stats
7219 * OUT statCount - the number of stats retrieved from the individual
7222 * OUT stats - the actual stats retrieved from the individual peer structures.
7226 * Returns void. If successful, stats will != NULL.
7230 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7231 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7232 size_t * allocSize, afs_uint32 * statCount,
7233 afs_uint32 ** stats)
7243 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7246 * Check to see if stats are enabled
7249 MUTEX_ENTER(&rx_rpc_stats);
7250 if (!rxi_monitor_peerStats) {
7251 MUTEX_EXIT(&rx_rpc_stats);
7255 clock_GetTime(&now);
7256 *clock_sec = now.sec;
7257 *clock_usec = now.usec;
7260 * Allocate the space based upon the caller version
7262 * If the client is at an older version than we are,
7263 * we return the statistic data in the older data format, but
7264 * we still return our version number so the client knows we
7265 * are maintaining more data than it can retrieve.
7268 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7269 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7270 *statCount = rxi_rpc_peer_stat_cnt;
7273 * This can't happen yet, but in the future version changes
7274 * can be handled by adding additional code here
7278 if (space > (size_t) 0) {
7280 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7283 rx_interface_stat_p rpc_stat, nrpc_stat;
7287 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7289 * We have to fix the offset of rpc_stat since we are
7290 * keeping this structure on two rx_queues. The rx_queue
7291 * package assumes that the rx_queue member is the first
7292 * member of the structure. That is, rx_queue assumes that
7293 * any one item is only on one queue at a time. We are
7294 * breaking that assumption and so we have to do a little
7295 * math to fix our pointers.
7298 fix_offset = (char *)rpc_stat;
7299 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7300 rpc_stat = (rx_interface_stat_p) fix_offset;
7303 * Copy the data based upon the caller version
7305 rx_MarshallProcessRPCStats(callerVersion,
7306 rpc_stat->stats[0].func_total,
7307 rpc_stat->stats, &ptr);
7313 MUTEX_EXIT(&rx_rpc_stats);
7318 * rx_FreeRPCStats - free memory allocated by
7319 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7323 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7324 * rx_RetrievePeerRPCStats
7326 * IN allocSize - the number of bytes in stats.
7334 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7336 rxi_Free(stats, allocSize);
7340 * rx_queryProcessRPCStats - see if process rpc stat collection is
7341 * currently enabled.
7347 * Returns 0 if stats are not enabled != 0 otherwise
7351 rx_queryProcessRPCStats(void)
7354 MUTEX_ENTER(&rx_rpc_stats);
7355 rc = rxi_monitor_processStats;
7356 MUTEX_EXIT(&rx_rpc_stats);
7361 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7367 * Returns 0 if stats are not enabled != 0 otherwise
7371 rx_queryPeerRPCStats(void)
7374 MUTEX_ENTER(&rx_rpc_stats);
7375 rc = rxi_monitor_peerStats;
7376 MUTEX_EXIT(&rx_rpc_stats);
7381 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7391 rx_enableProcessRPCStats(void)
7393 MUTEX_ENTER(&rx_rpc_stats);
7394 rx_enable_stats = 1;
7395 rxi_monitor_processStats = 1;
7396 MUTEX_EXIT(&rx_rpc_stats);
7400 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7410 rx_enablePeerRPCStats(void)
7412 MUTEX_ENTER(&rx_rpc_stats);
7413 rx_enable_stats = 1;
7414 rxi_monitor_peerStats = 1;
7415 MUTEX_EXIT(&rx_rpc_stats);
7419 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7429 rx_disableProcessRPCStats(void)
7431 rx_interface_stat_p rpc_stat, nrpc_stat;
7434 MUTEX_ENTER(&rx_rpc_stats);
7437 * Turn off process statistics and if peer stats is also off, turn
7441 rxi_monitor_processStats = 0;
7442 if (rxi_monitor_peerStats == 0) {
7443 rx_enable_stats = 0;
7446 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7447 unsigned int num_funcs = 0;
7450 queue_Remove(rpc_stat);
7451 num_funcs = rpc_stat->stats[0].func_total;
7453 sizeof(rx_interface_stat_t) +
7454 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7456 rxi_Free(rpc_stat, space);
7457 rxi_rpc_process_stat_cnt -= num_funcs;
7459 MUTEX_EXIT(&rx_rpc_stats);
7463 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7473 rx_disablePeerRPCStats(void)
7475 struct rx_peer **peer_ptr, **peer_end;
7478 MUTEX_ENTER(&rx_rpc_stats);
7481 * Turn off peer statistics and if process stats is also off, turn
7485 rxi_monitor_peerStats = 0;
7486 if (rxi_monitor_processStats == 0) {
7487 rx_enable_stats = 0;
7490 MUTEX_ENTER(&rx_peerHashTable_lock);
7491 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7492 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7494 struct rx_peer *peer, *next, *prev;
7495 for (prev = peer = *peer_ptr; peer; peer = next) {
7497 code = MUTEX_TRYENTER(&peer->peer_lock);
7499 rx_interface_stat_p rpc_stat, nrpc_stat;
7502 (&peer->rpcStats, rpc_stat, nrpc_stat,
7503 rx_interface_stat)) {
7504 unsigned int num_funcs = 0;
7507 queue_Remove(&rpc_stat->queue_header);
7508 queue_Remove(&rpc_stat->all_peers);
7509 num_funcs = rpc_stat->stats[0].func_total;
7511 sizeof(rx_interface_stat_t) +
7512 rpc_stat->stats[0].func_total *
7513 sizeof(rx_function_entry_v1_t);
7515 rxi_Free(rpc_stat, space);
7516 rxi_rpc_peer_stat_cnt -= num_funcs;
7518 MUTEX_EXIT(&peer->peer_lock);
7519 if (prev == *peer_ptr) {
7529 MUTEX_EXIT(&rx_peerHashTable_lock);
7530 MUTEX_EXIT(&rx_rpc_stats);
7534 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7539 * IN clearFlag - flag indicating which stats to clear
7547 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7549 rx_interface_stat_p rpc_stat, nrpc_stat;
7551 MUTEX_ENTER(&rx_rpc_stats);
7553 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7554 unsigned int num_funcs = 0, i;
7555 num_funcs = rpc_stat->stats[0].func_total;
7556 for (i = 0; i < num_funcs; i++) {
7557 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7558 hzero(rpc_stat->stats[i].invocations);
7560 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7561 hzero(rpc_stat->stats[i].bytes_sent);
7563 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7564 hzero(rpc_stat->stats[i].bytes_rcvd);
7566 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7567 rpc_stat->stats[i].queue_time_sum.sec = 0;
7568 rpc_stat->stats[i].queue_time_sum.usec = 0;
7570 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7571 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7572 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7574 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7575 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7576 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7578 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7579 rpc_stat->stats[i].queue_time_max.sec = 0;
7580 rpc_stat->stats[i].queue_time_max.usec = 0;
7582 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7583 rpc_stat->stats[i].execution_time_sum.sec = 0;
7584 rpc_stat->stats[i].execution_time_sum.usec = 0;
7586 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7587 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7588 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7590 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7591 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7592 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7594 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7595 rpc_stat->stats[i].execution_time_max.sec = 0;
7596 rpc_stat->stats[i].execution_time_max.usec = 0;
7601 MUTEX_EXIT(&rx_rpc_stats);
7605 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7610 * IN clearFlag - flag indicating which stats to clear
7618 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7620 rx_interface_stat_p rpc_stat, nrpc_stat;
7622 MUTEX_ENTER(&rx_rpc_stats);
7624 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7625 unsigned int num_funcs = 0, i;
7628 * We have to fix the offset of rpc_stat since we are
7629 * keeping this structure on two rx_queues. The rx_queue
7630 * package assumes that the rx_queue member is the first
7631 * member of the structure. That is, rx_queue assumes that
7632 * any one item is only on one queue at a time. We are
7633 * breaking that assumption and so we have to do a little
7634 * math to fix our pointers.
7637 fix_offset = (char *)rpc_stat;
7638 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7639 rpc_stat = (rx_interface_stat_p) fix_offset;
7641 num_funcs = rpc_stat->stats[0].func_total;
7642 for (i = 0; i < num_funcs; i++) {
7643 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7644 hzero(rpc_stat->stats[i].invocations);
7646 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7647 hzero(rpc_stat->stats[i].bytes_sent);
7649 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7650 hzero(rpc_stat->stats[i].bytes_rcvd);
7652 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7653 rpc_stat->stats[i].queue_time_sum.sec = 0;
7654 rpc_stat->stats[i].queue_time_sum.usec = 0;
7656 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7657 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7658 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7660 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7661 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7662 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7664 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7665 rpc_stat->stats[i].queue_time_max.sec = 0;
7666 rpc_stat->stats[i].queue_time_max.usec = 0;
7668 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7669 rpc_stat->stats[i].execution_time_sum.sec = 0;
7670 rpc_stat->stats[i].execution_time_sum.usec = 0;
7672 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7673 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7674 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7676 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7677 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7678 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7680 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7681 rpc_stat->stats[i].execution_time_max.sec = 0;
7682 rpc_stat->stats[i].execution_time_max.usec = 0;
7687 MUTEX_EXIT(&rx_rpc_stats);
7691 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7692 * is authorized to enable/disable/clear RX statistics.
7694 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7697 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7699 rxi_rxstat_userok = proc;
7703 rx_RxStatUserOk(struct rx_call *call)
7705 if (!rxi_rxstat_userok)
7707 return rxi_rxstat_userok(call);
7712 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
7713 * function in the MSVC runtime DLL (msvcrt.dll).
7715 * Note: the system serializes calls to this function.
7718 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
7719 DWORD reason, /* reason function is being called */
7720 LPVOID reserved) /* reserved for future use */
7723 case DLL_PROCESS_ATTACH:
7724 /* library is being attached to a process */
7728 case DLL_PROCESS_DETACH: