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"
37 #include "inet/common.h"
39 #include "inet/ip_ire.h"
41 #include "afs/afs_args.h"
42 #include "afs/afs_osi.h"
43 #ifdef RX_KERNEL_TRACE
44 #include "rx_kcommon.h"
46 #if (defined(AFS_AUX_ENV) || defined(AFS_AIX_ENV))
50 #undef RXDEBUG /* turn off debugging */
52 #if defined(AFS_SGI_ENV)
53 #include "sys/debug.h"
62 #endif /* AFS_OSF_ENV */
64 #include "afs/sysincludes.h"
65 #include "afsincludes.h"
68 #include "rx_kmutex.h"
69 #include "rx_kernel.h"
73 #include "rx_globals.h"
75 #define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
76 #define AFSOP_STOP_AFS 211 /* Stop AFS process */
77 #define AFSOP_STOP_BKG 212 /* Stop BKG process */
79 extern afs_int32 afs_termState;
81 #include "sys/lockl.h"
82 #include "sys/lock_def.h"
83 #endif /* AFS_AIX41_ENV */
84 # include "rxgen_consts.h"
86 # include <sys/types.h>
93 # include <afs/afsutil.h>
94 # include <WINNT\afsreg.h>
96 # include <sys/socket.h>
97 # include <sys/file.h>
99 # include <sys/stat.h>
100 # include <netinet/in.h>
101 # include <sys/time.h>
104 # include "rx_user.h"
105 # include "rx_clock.h"
106 # include "rx_queue.h"
107 # include "rx_globals.h"
108 # include "rx_trace.h"
109 # include <afs/rxgen_consts.h>
113 #ifdef AFS_PTHREAD_ENV
114 int (*registerProgram) (pid_t, char *) = 0;
115 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
117 int (*registerProgram) (PROCESS, char *) = 0;
118 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
122 /* Local static routines */
123 static void rxi_DestroyConnectionNoLock(register struct rx_connection *conn);
124 #ifdef RX_ENABLE_LOCKS
125 static void rxi_SetAcksInTransmitQueue(register struct rx_call *call);
128 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
130 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
131 afs_int32 rxi_start_in_error;
133 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
136 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
137 * currently allocated within rx. This number is used to allocate the
138 * memory required to return the statistics when queried.
141 static unsigned int rxi_rpc_peer_stat_cnt;
144 * rxi_rpc_process_stat_cnt counts the total number of local process stat
145 * structures currently allocated within rx. The number is used to allocate
146 * the memory required to return the statistics when queried.
149 static unsigned int rxi_rpc_process_stat_cnt;
151 #if !defined(offsetof)
152 #include <stddef.h> /* for definition of offsetof() */
155 #ifdef AFS_PTHREAD_ENV
159 * Use procedural initialization of mutexes/condition variables
163 extern pthread_mutex_t rx_stats_mutex;
164 extern pthread_mutex_t des_init_mutex;
165 extern pthread_mutex_t des_random_mutex;
166 extern pthread_mutex_t rx_clock_mutex;
167 extern pthread_mutex_t rxi_connCacheMutex;
168 extern pthread_mutex_t rx_event_mutex;
169 extern pthread_mutex_t osi_malloc_mutex;
170 extern pthread_mutex_t event_handler_mutex;
171 extern pthread_mutex_t listener_mutex;
172 extern pthread_mutex_t rx_if_init_mutex;
173 extern pthread_mutex_t rx_if_mutex;
174 extern pthread_mutex_t rxkad_client_uid_mutex;
175 extern pthread_mutex_t rxkad_random_mutex;
177 extern pthread_cond_t rx_event_handler_cond;
178 extern pthread_cond_t rx_listener_cond;
180 static pthread_mutex_t epoch_mutex;
181 static pthread_mutex_t rx_init_mutex;
182 static pthread_mutex_t rx_debug_mutex;
183 static pthread_mutex_t rx_rpc_stats;
186 rxi_InitPthread(void)
188 assert(pthread_mutex_init(&rx_clock_mutex, (const pthread_mutexattr_t *)0)
190 assert(pthread_mutex_init(&rx_stats_mutex, (const pthread_mutexattr_t *)0)
192 assert(pthread_mutex_init
193 (&rxi_connCacheMutex, (const pthread_mutexattr_t *)0) == 0);
194 assert(pthread_mutex_init(&rx_init_mutex, (const pthread_mutexattr_t *)0)
196 assert(pthread_mutex_init(&epoch_mutex, (const pthread_mutexattr_t *)0) ==
198 assert(pthread_mutex_init(&rx_event_mutex, (const pthread_mutexattr_t *)0)
200 assert(pthread_mutex_init(&des_init_mutex, (const pthread_mutexattr_t *)0)
202 assert(pthread_mutex_init
203 (&des_random_mutex, (const pthread_mutexattr_t *)0) == 0);
204 assert(pthread_mutex_init
205 (&osi_malloc_mutex, (const pthread_mutexattr_t *)0) == 0);
206 assert(pthread_mutex_init
207 (&event_handler_mutex, (const pthread_mutexattr_t *)0) == 0);
208 assert(pthread_mutex_init(&listener_mutex, (const pthread_mutexattr_t *)0)
210 assert(pthread_mutex_init
211 (&rx_if_init_mutex, (const pthread_mutexattr_t *)0) == 0);
212 assert(pthread_mutex_init(&rx_if_mutex, (const pthread_mutexattr_t *)0) ==
214 assert(pthread_mutex_init
215 (&rxkad_client_uid_mutex, (const pthread_mutexattr_t *)0) == 0);
216 assert(pthread_mutex_init
217 (&rxkad_random_mutex, (const pthread_mutexattr_t *)0) == 0);
218 assert(pthread_mutex_init(&rx_debug_mutex, (const pthread_mutexattr_t *)0)
221 assert(pthread_cond_init
222 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
223 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
225 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
226 assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
228 rxkad_global_stats_init();
230 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
231 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
232 #ifdef RX_ENABLE_LOCKS
235 #endif /* RX_LOCKS_DB */
236 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
237 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
239 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
241 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
243 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
245 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
246 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
247 #endif /* RX_ENABLE_LOCKS */
250 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
251 #define INIT_PTHREAD_LOCKS \
252 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
254 * The rx_stats_mutex mutex protects the following global variables:
259 * rxi_lowConnRefCount
260 * rxi_lowPeerRefCount
269 #define INIT_PTHREAD_LOCKS
273 /* Variables for handling the minProcs implementation. availProcs gives the
274 * number of threads available in the pool at this moment (not counting dudes
275 * executing right now). totalMin gives the total number of procs required
276 * for handling all minProcs requests. minDeficit is a dynamic variable
277 * tracking the # of procs required to satisfy all of the remaining minProcs
279 * For fine grain locking to work, the quota check and the reservation of
280 * a server thread has to come while rxi_availProcs and rxi_minDeficit
281 * are locked. To this end, the code has been modified under #ifdef
282 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
283 * same time. A new function, ReturnToServerPool() returns the allocation.
285 * A call can be on several queue's (but only one at a time). When
286 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
287 * that no one else is touching the queue. To this end, we store the address
288 * of the queue lock in the call structure (under the call lock) when we
289 * put the call on a queue, and we clear the call_queue_lock when the
290 * call is removed from a queue (once the call lock has been obtained).
291 * This allows rxi_ResetCall to safely synchronize with others wishing
292 * to manipulate the queue.
295 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
296 static afs_kmutex_t rx_rpc_stats;
297 void rxi_StartUnlocked(struct rxevent *event, void *call,
298 void *arg1, int istack);
301 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
302 ** pretty good that the next packet coming in is from the same connection
303 ** as the last packet, since we're send multiple packets in a transmit window.
305 struct rx_connection *rxLastConn = 0;
307 #ifdef RX_ENABLE_LOCKS
308 /* The locking hierarchy for rx fine grain locking is composed of these
311 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
312 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
313 * call->lock - locks call data fields.
314 * These are independent of each other:
315 * rx_freeCallQueue_lock
320 * serverQueueEntry->lock
322 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
323 * peer->lock - locks peer data fields.
324 * conn_data_lock - that more than one thread is not updating a conn data
325 * field at the same time.
333 * Do we need a lock to protect the peer field in the conn structure?
334 * conn->peer was previously a constant for all intents and so has no
335 * lock protecting this field. The multihomed client delta introduced
336 * a RX code change : change the peer field in the connection structure
337 * to that remote inetrface from which the last packet for this
338 * connection was sent out. This may become an issue if further changes
341 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
342 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
344 /* rxdb_fileID is used to identify the lock location, along with line#. */
345 static int rxdb_fileID = RXDB_FILE_RX;
346 #endif /* RX_LOCKS_DB */
347 #else /* RX_ENABLE_LOCKS */
348 #define SET_CALL_QUEUE_LOCK(C, L)
349 #define CLEAR_CALL_QUEUE_LOCK(C)
350 #endif /* RX_ENABLE_LOCKS */
351 struct rx_serverQueueEntry *rx_waitForPacket = 0;
352 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
354 /* ------------Exported Interfaces------------- */
356 /* This function allows rxkad to set the epoch to a suitably random number
357 * which rx_NewConnection will use in the future. The principle purpose is to
358 * get rxnull connections to use the same epoch as the rxkad connections do, at
359 * least once the first rxkad connection is established. This is important now
360 * that the host/port addresses aren't used in FindConnection: the uniqueness
361 * of epoch/cid matters and the start time won't do. */
363 #ifdef AFS_PTHREAD_ENV
365 * This mutex protects the following global variables:
369 #define LOCK_EPOCH assert(pthread_mutex_lock(&epoch_mutex)==0)
370 #define UNLOCK_EPOCH assert(pthread_mutex_unlock(&epoch_mutex)==0)
374 #endif /* AFS_PTHREAD_ENV */
377 rx_SetEpoch(afs_uint32 epoch)
384 /* Initialize rx. A port number may be mentioned, in which case this
385 * becomes the default port number for any service installed later.
386 * If 0 is provided for the port number, a random port will be chosen
387 * by the kernel. Whether this will ever overlap anything in
388 * /etc/services is anybody's guess... Returns 0 on success, -1 on
393 int rxinit_status = 1;
394 #ifdef AFS_PTHREAD_ENV
396 * This mutex protects the following global variables:
400 #define LOCK_RX_INIT assert(pthread_mutex_lock(&rx_init_mutex)==0)
401 #define UNLOCK_RX_INIT assert(pthread_mutex_unlock(&rx_init_mutex)==0)
404 #define UNLOCK_RX_INIT
408 rx_InitHost(u_int host, u_int port)
415 char *htable, *ptable;
422 if (rxinit_status == 0) {
423 tmp_status = rxinit_status;
425 return tmp_status; /* Already started; return previous error code. */
431 if (afs_winsockInit() < 0)
437 * Initialize anything necessary to provide a non-premptive threading
440 rxi_InitializeThreadSupport();
443 /* Allocate and initialize a socket for client and perhaps server
446 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
447 if (rx_socket == OSI_NULLSOCKET) {
451 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
454 #endif /* RX_LOCKS_DB */
455 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
456 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
457 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
458 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
459 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
461 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
463 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
465 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
467 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
468 #if defined(AFS_HPUX110_ENV)
470 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
471 #endif /* AFS_HPUX110_ENV */
472 #endif /* RX_ENABLE_LOCKS && KERNEL */
475 rx_connDeadTime = 12;
476 rx_tranquil = 0; /* reset flag */
477 memset((char *)&rx_stats, 0, sizeof(struct rx_stats));
479 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
480 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
481 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
482 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
483 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
484 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
486 /* Malloc up a bunch of packets & buffers */
488 queue_Init(&rx_freePacketQueue);
489 rxi_NeedMorePackets = FALSE;
490 #ifdef RX_ENABLE_TSFPQ
491 rx_nPackets = 0; /* in TSFPQ version, rx_nPackets is managed by rxi_MorePackets* */
492 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
493 #else /* RX_ENABLE_TSFPQ */
494 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
495 rxi_MorePackets(rx_nPackets);
496 #endif /* RX_ENABLE_TSFPQ */
503 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
504 tv.tv_sec = clock_now.sec;
505 tv.tv_usec = clock_now.usec;
506 srand((unsigned int)tv.tv_usec);
513 #if defined(KERNEL) && !defined(UKERNEL)
514 /* Really, this should never happen in a real kernel */
517 struct sockaddr_in addr;
518 socklen_t addrlen = sizeof(addr);
519 if (getsockname((int)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
523 rx_port = addr.sin_port;
526 rx_stats.minRtt.sec = 9999999;
528 rx_SetEpoch(tv.tv_sec | 0x80000000);
530 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
531 * will provide a randomer value. */
533 MUTEX_ENTER(&rx_stats_mutex);
534 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
535 MUTEX_EXIT(&rx_stats_mutex);
536 /* *Slightly* random start time for the cid. This is just to help
537 * out with the hashing function at the peer */
538 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
539 rx_connHashTable = (struct rx_connection **)htable;
540 rx_peerHashTable = (struct rx_peer **)ptable;
542 rx_lastAckDelay.sec = 0;
543 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
544 rx_hardAckDelay.sec = 0;
545 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
546 rx_softAckDelay.sec = 0;
547 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
549 rxevent_Init(20, rxi_ReScheduleEvents);
551 /* Initialize various global queues */
552 queue_Init(&rx_idleServerQueue);
553 queue_Init(&rx_incomingCallQueue);
554 queue_Init(&rx_freeCallQueue);
556 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
557 /* Initialize our list of usable IP addresses. */
561 /* Start listener process (exact function is dependent on the
562 * implementation environment--kernel or user space) */
566 tmp_status = rxinit_status = 0;
574 return rx_InitHost(htonl(INADDR_ANY), port);
577 /* called with unincremented nRequestsRunning to see if it is OK to start
578 * a new thread in this service. Could be "no" for two reasons: over the
579 * max quota, or would prevent others from reaching their min quota.
581 #ifdef RX_ENABLE_LOCKS
582 /* This verion of QuotaOK reserves quota if it's ok while the
583 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
586 QuotaOK(register struct rx_service *aservice)
588 /* check if over max quota */
589 if (aservice->nRequestsRunning >= aservice->maxProcs) {
593 /* under min quota, we're OK */
594 /* otherwise, can use only if there are enough to allow everyone
595 * to go to their min quota after this guy starts.
597 MUTEX_ENTER(&rx_stats_mutex);
598 if ((aservice->nRequestsRunning < aservice->minProcs)
599 || (rxi_availProcs > rxi_minDeficit)) {
600 aservice->nRequestsRunning++;
601 /* just started call in minProcs pool, need fewer to maintain
603 if (aservice->nRequestsRunning <= aservice->minProcs)
606 MUTEX_EXIT(&rx_stats_mutex);
609 MUTEX_EXIT(&rx_stats_mutex);
615 ReturnToServerPool(register struct rx_service *aservice)
617 aservice->nRequestsRunning--;
618 MUTEX_ENTER(&rx_stats_mutex);
619 if (aservice->nRequestsRunning < aservice->minProcs)
622 MUTEX_EXIT(&rx_stats_mutex);
625 #else /* RX_ENABLE_LOCKS */
627 QuotaOK(register struct rx_service *aservice)
630 /* under min quota, we're OK */
631 if (aservice->nRequestsRunning < aservice->minProcs)
634 /* check if over max quota */
635 if (aservice->nRequestsRunning >= aservice->maxProcs)
638 /* otherwise, can use only if there are enough to allow everyone
639 * to go to their min quota after this guy starts.
641 if (rxi_availProcs > rxi_minDeficit)
645 #endif /* RX_ENABLE_LOCKS */
648 /* Called by rx_StartServer to start up lwp's to service calls.
649 NExistingProcs gives the number of procs already existing, and which
650 therefore needn't be created. */
652 rxi_StartServerProcs(int nExistingProcs)
654 register struct rx_service *service;
659 /* For each service, reserve N processes, where N is the "minimum"
660 * number of processes that MUST be able to execute a request in parallel,
661 * at any time, for that process. Also compute the maximum difference
662 * between any service's maximum number of processes that can run
663 * (i.e. the maximum number that ever will be run, and a guarantee
664 * that this number will run if other services aren't running), and its
665 * minimum number. The result is the extra number of processes that
666 * we need in order to provide the latter guarantee */
667 for (i = 0; i < RX_MAX_SERVICES; i++) {
669 service = rx_services[i];
670 if (service == (struct rx_service *)0)
672 nProcs += service->minProcs;
673 diff = service->maxProcs - service->minProcs;
677 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
678 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
679 for (i = 0; i < nProcs; i++) {
680 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
686 /* This routine is only required on Windows */
688 rx_StartClientThread(void)
690 #ifdef AFS_PTHREAD_ENV
692 pid = pthread_self();
693 #endif /* AFS_PTHREAD_ENV */
695 #endif /* AFS_NT40_ENV */
697 /* This routine must be called if any services are exported. If the
698 * donateMe flag is set, the calling process is donated to the server
701 rx_StartServer(int donateMe)
703 register struct rx_service *service;
709 /* Start server processes, if necessary (exact function is dependent
710 * on the implementation environment--kernel or user space). DonateMe
711 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
712 * case, one less new proc will be created rx_StartServerProcs.
714 rxi_StartServerProcs(donateMe);
716 /* count up the # of threads in minProcs, and add set the min deficit to
717 * be that value, too.
719 for (i = 0; i < RX_MAX_SERVICES; i++) {
720 service = rx_services[i];
721 if (service == (struct rx_service *)0)
723 MUTEX_ENTER(&rx_stats_mutex);
724 rxi_totalMin += service->minProcs;
725 /* below works even if a thread is running, since minDeficit would
726 * still have been decremented and later re-incremented.
728 rxi_minDeficit += service->minProcs;
729 MUTEX_EXIT(&rx_stats_mutex);
732 /* Turn on reaping of idle server connections */
733 rxi_ReapConnections(NULL, NULL, NULL);
742 #ifdef AFS_PTHREAD_ENV
744 pid = (pid_t) pthread_self();
745 #else /* AFS_PTHREAD_ENV */
747 LWP_CurrentProcess(&pid);
748 #endif /* AFS_PTHREAD_ENV */
750 sprintf(name, "srv_%d", ++nProcs);
752 (*registerProgram) (pid, name);
754 #endif /* AFS_NT40_ENV */
755 rx_ServerProc(NULL); /* Never returns */
757 #ifdef RX_ENABLE_TSFPQ
758 /* no use leaving packets around in this thread's local queue if
759 * it isn't getting donated to the server thread pool.
761 rxi_FlushLocalPacketsTSFPQ();
762 #endif /* RX_ENABLE_TSFPQ */
766 /* Create a new client connection to the specified service, using the
767 * specified security object to implement the security model for this
769 struct rx_connection *
770 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
771 register struct rx_securityClass *securityObject,
772 int serviceSecurityIndex)
775 afs_int32 cid, cix, nclones;
776 register struct rx_connection *conn, *tconn, *ptconn;
781 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", ntohl(shost), ntohs(sport), sservice, securityObject, serviceSecurityIndex));
784 nclones = rx_max_clones_per_connection;
786 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
787 * the case of kmem_alloc? */
790 MUTEX_ENTER(&rx_connHashTable_lock);
792 /* send in the clones */
793 for(cix = 0; cix <= nclones; ++cix) {
796 tconn = rxi_AllocConnection();
797 tconn->type = RX_CLIENT_CONNECTION;
798 tconn->epoch = rx_epoch;
799 tconn->peer = rxi_FindPeer(shost, sport, 0, 1);
800 tconn->serviceId = sservice;
801 tconn->securityObject = securityObject;
802 tconn->securityData = (void *) 0;
803 tconn->securityIndex = serviceSecurityIndex;
804 tconn->ackRate = RX_FAST_ACK_RATE;
805 tconn->nSpecific = 0;
806 tconn->specific = NULL;
807 tconn->challengeEvent = NULL;
808 tconn->delayedAbortEvent = NULL;
809 tconn->abortCount = 0;
811 for (i = 0; i < RX_MAXCALLS; i++) {
812 tconn->twind[i] = rx_initSendWindow;
813 tconn->rwind[i] = rx_initReceiveWindow;
816 tconn->next_clone = 0;
817 tconn->nclones = nclones;
818 rx_SetConnDeadTime(tconn, rx_connDeadTime);
823 tconn->flags |= RX_CLONED_CONNECTION;
824 tconn->parent = conn;
825 ptconn->next_clone = tconn;
828 /* generic connection setup */
829 #ifdef RX_ENABLE_LOCKS
830 MUTEX_INIT(&tconn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
831 MUTEX_INIT(&tconn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
832 CV_INIT(&tconn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
834 cid = (rx_nextCid += RX_MAXCALLS);
836 RXS_NewConnection(securityObject, tconn);
838 CONN_HASH(shost, sport, tconn->cid, tconn->epoch,
839 RX_CLIENT_CONNECTION);
840 tconn->refCount++; /* no lock required since only this thread knows */
841 tconn->next = rx_connHashTable[hashindex];
842 rx_connHashTable[hashindex] = tconn;
843 rx_MutexIncrement(rx_stats.nClientConns, rx_stats_mutex);
846 MUTEX_EXIT(&rx_connHashTable_lock);
852 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
854 /* The idea is to set the dead time to a value that allows several
855 * keepalives to be dropped without timing out the connection. */
856 struct rx_connection *tconn;
859 tconn->secondsUntilDead = MAX(seconds, 6);
860 tconn->secondsUntilPing = tconn->secondsUntilDead / 6;
861 } while(tconn->next_clone && (tconn = tconn->next_clone));
864 int rxi_lowPeerRefCount = 0;
865 int rxi_lowConnRefCount = 0;
868 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
869 * NOTE: must not be called with rx_connHashTable_lock held.
872 rxi_CleanupConnection(struct rx_connection *conn)
874 /* Notify the service exporter, if requested, that this connection
875 * is being destroyed */
876 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
877 (*conn->service->destroyConnProc) (conn);
879 /* Notify the security module that this connection is being destroyed */
880 RXS_DestroyConnection(conn->securityObject, conn);
882 /* If this is the last connection using the rx_peer struct, set its
883 * idle time to now. rxi_ReapConnections will reap it if it's still
884 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
886 MUTEX_ENTER(&rx_peerHashTable_lock);
887 if (conn->peer->refCount < 2) {
888 conn->peer->idleWhen = clock_Sec();
889 if (conn->peer->refCount < 1) {
890 conn->peer->refCount = 1;
891 MUTEX_ENTER(&rx_stats_mutex);
892 rxi_lowPeerRefCount++;
893 MUTEX_EXIT(&rx_stats_mutex);
896 conn->peer->refCount--;
897 MUTEX_EXIT(&rx_peerHashTable_lock);
899 if (conn->type == RX_SERVER_CONNECTION)
900 rx_MutexDecrement(rx_stats.nServerConns, rx_stats_mutex);
902 rx_MutexDecrement(rx_stats.nClientConns, rx_stats_mutex);
904 if (conn->specific) {
906 for (i = 0; i < conn->nSpecific; i++) {
907 if (conn->specific[i] && rxi_keyCreate_destructor[i])
908 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
909 conn->specific[i] = NULL;
911 free(conn->specific);
913 conn->specific = NULL;
917 MUTEX_DESTROY(&conn->conn_call_lock);
918 MUTEX_DESTROY(&conn->conn_data_lock);
919 CV_DESTROY(&conn->conn_call_cv);
921 rxi_FreeConnection(conn);
924 /* Destroy the specified connection */
926 rxi_DestroyConnection(register struct rx_connection *conn)
928 register struct rx_connection *tconn, *dtconn;
930 MUTEX_ENTER(&rx_connHashTable_lock);
932 if(!(conn->flags & RX_CLONED_CONNECTION)) {
933 tconn = conn->next_clone;
934 conn->next_clone = 0; /* once */
938 tconn = tconn->next_clone;
939 rxi_DestroyConnectionNoLock(dtconn);
941 if (dtconn == rx_connCleanup_list) {
942 rx_connCleanup_list = rx_connCleanup_list->next;
943 MUTEX_EXIT(&rx_connHashTable_lock);
944 /* rxi_CleanupConnection will free tconn */
945 rxi_CleanupConnection(dtconn);
946 MUTEX_ENTER(&rx_connHashTable_lock);
953 rxi_DestroyConnectionNoLock(conn);
954 /* conn should be at the head of the cleanup list */
955 if (conn == rx_connCleanup_list) {
956 rx_connCleanup_list = rx_connCleanup_list->next;
957 MUTEX_EXIT(&rx_connHashTable_lock);
958 rxi_CleanupConnection(conn);
960 #ifdef RX_ENABLE_LOCKS
962 MUTEX_EXIT(&rx_connHashTable_lock);
964 #endif /* RX_ENABLE_LOCKS */
968 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
970 register struct rx_connection **conn_ptr;
971 register int havecalls = 0;
972 struct rx_packet *packet;
979 MUTEX_ENTER(&conn->conn_data_lock);
980 if (conn->refCount > 0)
983 MUTEX_ENTER(&rx_stats_mutex);
984 rxi_lowConnRefCount++;
985 MUTEX_EXIT(&rx_stats_mutex);
988 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
989 /* Busy; wait till the last guy before proceeding */
990 MUTEX_EXIT(&conn->conn_data_lock);
995 /* If the client previously called rx_NewCall, but it is still
996 * waiting, treat this as a running call, and wait to destroy the
997 * connection later when the call completes. */
998 if ((conn->type == RX_CLIENT_CONNECTION)
999 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
1000 conn->flags |= RX_CONN_DESTROY_ME;
1001 MUTEX_EXIT(&conn->conn_data_lock);
1005 MUTEX_EXIT(&conn->conn_data_lock);
1007 /* Check for extant references to this connection */
1008 for (i = 0; i < RX_MAXCALLS; i++) {
1009 register struct rx_call *call = conn->call[i];
1012 if (conn->type == RX_CLIENT_CONNECTION) {
1013 MUTEX_ENTER(&call->lock);
1014 if (call->delayedAckEvent) {
1015 /* Push the final acknowledgment out now--there
1016 * won't be a subsequent call to acknowledge the
1017 * last reply packets */
1018 rxevent_Cancel(call->delayedAckEvent, call,
1019 RX_CALL_REFCOUNT_DELAY);
1020 if (call->state == RX_STATE_PRECALL
1021 || call->state == RX_STATE_ACTIVE) {
1022 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1024 rxi_AckAll(NULL, call, 0);
1027 MUTEX_EXIT(&call->lock);
1031 #ifdef RX_ENABLE_LOCKS
1033 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1034 MUTEX_EXIT(&conn->conn_data_lock);
1036 /* Someone is accessing a packet right now. */
1040 #endif /* RX_ENABLE_LOCKS */
1043 /* Don't destroy the connection if there are any call
1044 * structures still in use */
1045 MUTEX_ENTER(&conn->conn_data_lock);
1046 conn->flags |= RX_CONN_DESTROY_ME;
1047 MUTEX_EXIT(&conn->conn_data_lock);
1052 if (conn->delayedAbortEvent) {
1053 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
1054 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1056 MUTEX_ENTER(&conn->conn_data_lock);
1057 rxi_SendConnectionAbort(conn, packet, 0, 1);
1058 MUTEX_EXIT(&conn->conn_data_lock);
1059 rxi_FreePacket(packet);
1063 /* Remove from connection hash table before proceeding */
1065 &rx_connHashTable[CONN_HASH
1066 (peer->host, peer->port, conn->cid, conn->epoch,
1068 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1069 if (*conn_ptr == conn) {
1070 *conn_ptr = conn->next;
1074 /* if the conn that we are destroying was the last connection, then we
1075 * clear rxLastConn as well */
1076 if (rxLastConn == conn)
1079 /* Make sure the connection is completely reset before deleting it. */
1080 /* get rid of pending events that could zap us later */
1081 if (conn->challengeEvent)
1082 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1083 if (conn->checkReachEvent)
1084 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1086 /* Add the connection to the list of destroyed connections that
1087 * need to be cleaned up. This is necessary to avoid deadlocks
1088 * in the routines we call to inform others that this connection is
1089 * being destroyed. */
1090 conn->next = rx_connCleanup_list;
1091 rx_connCleanup_list = conn;
1094 /* Externally available version */
1096 rx_DestroyConnection(register struct rx_connection *conn)
1101 rxi_DestroyConnection(conn);
1106 rx_GetConnection(register struct rx_connection *conn)
1111 MUTEX_ENTER(&conn->conn_data_lock);
1113 MUTEX_EXIT(&conn->conn_data_lock);
1117 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1118 /* Wait for the transmit queue to no longer be busy.
1119 * requires the call->lock to be held */
1120 static void rxi_WaitforTQBusy(struct rx_call *call) {
1121 while (call->flags & RX_CALL_TQ_BUSY) {
1122 call->flags |= RX_CALL_TQ_WAIT;
1124 #ifdef RX_ENABLE_LOCKS
1125 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1126 CV_WAIT(&call->cv_tq, &call->lock);
1127 #else /* RX_ENABLE_LOCKS */
1128 osi_rxSleep(&call->tq);
1129 #endif /* RX_ENABLE_LOCKS */
1131 if (call->tqWaiters == 0) {
1132 call->flags &= ~RX_CALL_TQ_WAIT;
1138 /* Start a new rx remote procedure call, on the specified connection.
1139 * If wait is set to 1, wait for a free call channel; otherwise return
1140 * 0. Maxtime gives the maximum number of seconds this call may take,
1141 * after rx_NewCall returns. After this time interval, a call to any
1142 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1143 * For fine grain locking, we hold the conn_call_lock in order to
1144 * to ensure that we don't get signalle after we found a call in an active
1145 * state and before we go to sleep.
1148 rx_NewCall(register struct rx_connection *conn)
1151 register struct rx_call *call;
1152 register struct rx_connection *tconn;
1153 struct clock queueTime;
1157 dpf(("rx_NewCall(conn %x)\n", conn));
1160 clock_GetTime(&queueTime);
1161 MUTEX_ENTER(&conn->conn_call_lock);
1164 * Check if there are others waiting for a new call.
1165 * If so, let them go first to avoid starving them.
1166 * This is a fairly simple scheme, and might not be
1167 * a complete solution for large numbers of waiters.
1169 * makeCallWaiters keeps track of the number of
1170 * threads waiting to make calls and the
1171 * RX_CONN_MAKECALL_WAITING flag bit is used to
1172 * indicate that there are indeed calls waiting.
1173 * The flag is set when the waiter is incremented.
1174 * It is only cleared in rx_EndCall when
1175 * makeCallWaiters is 0. This prevents us from
1176 * accidently destroying the connection while it
1177 * is potentially about to be used.
1179 MUTEX_ENTER(&conn->conn_data_lock);
1180 if (conn->makeCallWaiters) {
1181 conn->flags |= RX_CONN_MAKECALL_WAITING;
1182 conn->makeCallWaiters++;
1183 MUTEX_EXIT(&conn->conn_data_lock);
1185 #ifdef RX_ENABLE_LOCKS
1186 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1190 MUTEX_ENTER(&conn->conn_data_lock);
1191 conn->makeCallWaiters--;
1193 MUTEX_EXIT(&conn->conn_data_lock);
1195 /* search for next free call on this connection or
1196 * its clones, if any */
1200 for (i = 0; i < RX_MAXCALLS; i++) {
1201 call = tconn->call[i];
1203 MUTEX_ENTER(&call->lock);
1204 if (call->state == RX_STATE_DALLY) {
1205 rxi_ResetCall(call, 0);
1206 (*call->callNumber)++;
1209 MUTEX_EXIT(&call->lock);
1211 call = rxi_NewCall(tconn, i);
1214 } /* for i < RX_MAXCALLS */
1215 } while (tconn->next_clone && (tconn = tconn->next_clone));
1219 if (i < RX_MAXCALLS) {
1223 /* to be here, all available calls for this connection (and all
1224 * its clones) must be in use */
1226 MUTEX_ENTER(&conn->conn_data_lock);
1227 conn->flags |= RX_CONN_MAKECALL_WAITING;
1228 conn->makeCallWaiters++;
1229 MUTEX_EXIT(&conn->conn_data_lock);
1231 #ifdef RX_ENABLE_LOCKS
1232 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1236 MUTEX_ENTER(&conn->conn_data_lock);
1237 conn->makeCallWaiters--;
1238 MUTEX_EXIT(&conn->conn_data_lock);
1241 * Wake up anyone else who might be giving us a chance to
1242 * run (see code above that avoids resource starvation).
1244 #ifdef RX_ENABLE_LOCKS
1245 CV_BROADCAST(&conn->conn_call_cv);
1250 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1252 /* Client is initially in send mode */
1253 call->state = RX_STATE_ACTIVE;
1254 call->error = conn->error;
1256 call->mode = RX_MODE_ERROR;
1258 call->mode = RX_MODE_SENDING;
1260 /* remember start time for call in case we have hard dead time limit */
1261 call->queueTime = queueTime;
1262 clock_GetTime(&call->startTime);
1263 hzero(call->bytesSent);
1264 hzero(call->bytesRcvd);
1266 /* Turn on busy protocol. */
1267 rxi_KeepAliveOn(call);
1269 MUTEX_EXIT(&call->lock);
1270 MUTEX_EXIT(&conn->conn_call_lock);
1273 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1274 /* Now, if TQ wasn't cleared earlier, do it now. */
1275 MUTEX_ENTER(&call->lock);
1276 rxi_WaitforTQBusy(call);
1277 if (call->flags & RX_CALL_TQ_CLEARME) {
1278 rxi_ClearTransmitQueue(call, 0);
1279 queue_Init(&call->tq);
1281 MUTEX_EXIT(&call->lock);
1282 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1284 dpf(("rx_NewCall(call %x)\n", call));
1289 rxi_HasActiveCalls(register struct rx_connection *aconn)
1292 register struct rx_call *tcall;
1296 for (i = 0; i < RX_MAXCALLS; i++) {
1297 if ((tcall = aconn->call[i])) {
1298 if ((tcall->state == RX_STATE_ACTIVE)
1299 || (tcall->state == RX_STATE_PRECALL)) {
1310 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1311 register afs_int32 * aint32s)
1314 register struct rx_call *tcall;
1318 for (i = 0; i < RX_MAXCALLS; i++) {
1319 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1320 aint32s[i] = aconn->callNumber[i] + 1;
1322 aint32s[i] = aconn->callNumber[i];
1329 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1330 register afs_int32 * aint32s)
1333 register struct rx_call *tcall;
1337 for (i = 0; i < RX_MAXCALLS; i++) {
1338 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1339 aconn->callNumber[i] = aint32s[i] - 1;
1341 aconn->callNumber[i] = aint32s[i];
1347 /* Advertise a new service. A service is named locally by a UDP port
1348 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1351 char *serviceName; Name for identification purposes (e.g. the
1352 service name might be used for probing for
1355 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1356 char *serviceName, struct rx_securityClass **securityObjects,
1357 int nSecurityObjects,
1358 afs_int32(*serviceProc) (struct rx_call * acall))
1360 osi_socket socket = OSI_NULLSOCKET;
1361 register struct rx_service *tservice;
1367 if (serviceId == 0) {
1369 "rx_NewService: service id for service %s is not non-zero.\n",
1376 "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",
1384 tservice = rxi_AllocService();
1386 for (i = 0; i < RX_MAX_SERVICES; i++) {
1387 register struct rx_service *service = rx_services[i];
1389 if (port == service->servicePort && host == service->serviceHost) {
1390 if (service->serviceId == serviceId) {
1391 /* The identical service has already been
1392 * installed; if the caller was intending to
1393 * change the security classes used by this
1394 * service, he/she loses. */
1396 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1397 serviceName, serviceId, service->serviceName);
1399 rxi_FreeService(tservice);
1402 /* Different service, same port: re-use the socket
1403 * which is bound to the same port */
1404 socket = service->socket;
1407 if (socket == OSI_NULLSOCKET) {
1408 /* If we don't already have a socket (from another
1409 * service on same port) get a new one */
1410 socket = rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
1411 if (socket == OSI_NULLSOCKET) {
1413 rxi_FreeService(tservice);
1418 service->socket = socket;
1419 service->serviceHost = host;
1420 service->servicePort = port;
1421 service->serviceId = serviceId;
1422 service->serviceName = serviceName;
1423 service->nSecurityObjects = nSecurityObjects;
1424 service->securityObjects = securityObjects;
1425 service->minProcs = 0;
1426 service->maxProcs = 1;
1427 service->idleDeadTime = 60;
1428 service->idleDeadErr = 0;
1429 service->connDeadTime = rx_connDeadTime;
1430 service->executeRequestProc = serviceProc;
1431 service->checkReach = 0;
1432 rx_services[i] = service; /* not visible until now */
1438 rxi_FreeService(tservice);
1439 (osi_Msg "rx_NewService: cannot support > %d services\n",
1444 /* Set configuration options for all of a service's security objects */
1447 rx_SetSecurityConfiguration(struct rx_service *service,
1448 rx_securityConfigVariables type,
1452 for (i = 0; i<service->nSecurityObjects; i++) {
1453 if (service->securityObjects[i]) {
1454 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1462 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1463 struct rx_securityClass **securityObjects, int nSecurityObjects,
1464 afs_int32(*serviceProc) (struct rx_call * acall))
1466 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1469 /* Generic request processing loop. This routine should be called
1470 * by the implementation dependent rx_ServerProc. If socketp is
1471 * non-null, it will be set to the file descriptor that this thread
1472 * is now listening on. If socketp is null, this routine will never
1475 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1477 register struct rx_call *call;
1478 register afs_int32 code;
1479 register struct rx_service *tservice = NULL;
1486 call = rx_GetCall(threadID, tservice, socketp);
1487 if (socketp && *socketp != OSI_NULLSOCKET) {
1488 /* We are now a listener thread */
1493 /* if server is restarting( typically smooth shutdown) then do not
1494 * allow any new calls.
1497 if (rx_tranquil && (call != NULL)) {
1501 MUTEX_ENTER(&call->lock);
1503 rxi_CallError(call, RX_RESTARTING);
1504 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1506 MUTEX_EXIT(&call->lock);
1510 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1511 #ifdef RX_ENABLE_LOCKS
1513 #endif /* RX_ENABLE_LOCKS */
1514 afs_termState = AFSOP_STOP_AFS;
1515 afs_osi_Wakeup(&afs_termState);
1516 #ifdef RX_ENABLE_LOCKS
1518 #endif /* RX_ENABLE_LOCKS */
1523 tservice = call->conn->service;
1525 if (tservice->beforeProc)
1526 (*tservice->beforeProc) (call);
1528 code = call->conn->service->executeRequestProc(call);
1530 if (tservice->afterProc)
1531 (*tservice->afterProc) (call, code);
1533 rx_EndCall(call, code);
1534 MUTEX_ENTER(&rx_stats_mutex);
1536 MUTEX_EXIT(&rx_stats_mutex);
1542 rx_WakeupServerProcs(void)
1544 struct rx_serverQueueEntry *np, *tqp;
1548 MUTEX_ENTER(&rx_serverPool_lock);
1550 #ifdef RX_ENABLE_LOCKS
1551 if (rx_waitForPacket)
1552 CV_BROADCAST(&rx_waitForPacket->cv);
1553 #else /* RX_ENABLE_LOCKS */
1554 if (rx_waitForPacket)
1555 osi_rxWakeup(rx_waitForPacket);
1556 #endif /* RX_ENABLE_LOCKS */
1557 MUTEX_ENTER(&freeSQEList_lock);
1558 for (np = rx_FreeSQEList; np; np = tqp) {
1559 tqp = *(struct rx_serverQueueEntry **)np;
1560 #ifdef RX_ENABLE_LOCKS
1561 CV_BROADCAST(&np->cv);
1562 #else /* RX_ENABLE_LOCKS */
1564 #endif /* RX_ENABLE_LOCKS */
1566 MUTEX_EXIT(&freeSQEList_lock);
1567 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1568 #ifdef RX_ENABLE_LOCKS
1569 CV_BROADCAST(&np->cv);
1570 #else /* RX_ENABLE_LOCKS */
1572 #endif /* RX_ENABLE_LOCKS */
1574 MUTEX_EXIT(&rx_serverPool_lock);
1579 * One thing that seems to happen is that all the server threads get
1580 * tied up on some empty or slow call, and then a whole bunch of calls
1581 * arrive at once, using up the packet pool, so now there are more
1582 * empty calls. The most critical resources here are server threads
1583 * and the free packet pool. The "doreclaim" code seems to help in
1584 * general. I think that eventually we arrive in this state: there
1585 * are lots of pending calls which do have all their packets present,
1586 * so they won't be reclaimed, are multi-packet calls, so they won't
1587 * be scheduled until later, and thus are tying up most of the free
1588 * packet pool for a very long time.
1590 * 1. schedule multi-packet calls if all the packets are present.
1591 * Probably CPU-bound operation, useful to return packets to pool.
1592 * Do what if there is a full window, but the last packet isn't here?
1593 * 3. preserve one thread which *only* runs "best" calls, otherwise
1594 * it sleeps and waits for that type of call.
1595 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1596 * the current dataquota business is badly broken. The quota isn't adjusted
1597 * to reflect how many packets are presently queued for a running call.
1598 * So, when we schedule a queued call with a full window of packets queued
1599 * up for it, that *should* free up a window full of packets for other 2d-class
1600 * calls to be able to use from the packet pool. But it doesn't.
1602 * NB. Most of the time, this code doesn't run -- since idle server threads
1603 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1604 * as a new call arrives.
1606 /* Sleep until a call arrives. Returns a pointer to the call, ready
1607 * for an rx_Read. */
1608 #ifdef RX_ENABLE_LOCKS
1610 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1612 struct rx_serverQueueEntry *sq;
1613 register struct rx_call *call = (struct rx_call *)0;
1614 struct rx_service *service = NULL;
1617 MUTEX_ENTER(&freeSQEList_lock);
1619 if ((sq = rx_FreeSQEList)) {
1620 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1621 MUTEX_EXIT(&freeSQEList_lock);
1622 } else { /* otherwise allocate a new one and return that */
1623 MUTEX_EXIT(&freeSQEList_lock);
1624 sq = (struct rx_serverQueueEntry *)
1625 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1626 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1627 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1630 MUTEX_ENTER(&rx_serverPool_lock);
1631 if (cur_service != NULL) {
1632 ReturnToServerPool(cur_service);
1635 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1636 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1638 /* Scan for eligible incoming calls. A call is not eligible
1639 * if the maximum number of calls for its service type are
1640 * already executing */
1641 /* One thread will process calls FCFS (to prevent starvation),
1642 * while the other threads may run ahead looking for calls which
1643 * have all their input data available immediately. This helps
1644 * keep threads from blocking, waiting for data from the client. */
1645 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1646 service = tcall->conn->service;
1647 if (!QuotaOK(service)) {
1650 if (tno == rxi_fcfs_thread_num
1651 || !tcall->queue_item_header.next) {
1652 /* If we're the fcfs thread , then we'll just use
1653 * this call. If we haven't been able to find an optimal
1654 * choice, and we're at the end of the list, then use a
1655 * 2d choice if one has been identified. Otherwise... */
1656 call = (choice2 ? choice2 : tcall);
1657 service = call->conn->service;
1658 } else if (!queue_IsEmpty(&tcall->rq)) {
1659 struct rx_packet *rp;
1660 rp = queue_First(&tcall->rq, rx_packet);
1661 if (rp->header.seq == 1) {
1663 || (rp->header.flags & RX_LAST_PACKET)) {
1665 } else if (rxi_2dchoice && !choice2
1666 && !(tcall->flags & RX_CALL_CLEARED)
1667 && (tcall->rprev > rxi_HardAckRate)) {
1676 ReturnToServerPool(service);
1683 MUTEX_EXIT(&rx_serverPool_lock);
1684 MUTEX_ENTER(&call->lock);
1686 if (call->flags & RX_CALL_WAIT_PROC) {
1687 call->flags &= ~RX_CALL_WAIT_PROC;
1688 MUTEX_ENTER(&rx_stats_mutex);
1690 MUTEX_EXIT(&rx_stats_mutex);
1693 if (call->state != RX_STATE_PRECALL || call->error) {
1694 MUTEX_EXIT(&call->lock);
1695 MUTEX_ENTER(&rx_serverPool_lock);
1696 ReturnToServerPool(service);
1701 if (queue_IsEmpty(&call->rq)
1702 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1703 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1705 CLEAR_CALL_QUEUE_LOCK(call);
1708 /* If there are no eligible incoming calls, add this process
1709 * to the idle server queue, to wait for one */
1713 *socketp = OSI_NULLSOCKET;
1715 sq->socketp = socketp;
1716 queue_Append(&rx_idleServerQueue, sq);
1717 #ifndef AFS_AIX41_ENV
1718 rx_waitForPacket = sq;
1720 rx_waitingForPacket = sq;
1721 #endif /* AFS_AIX41_ENV */
1723 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1725 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1726 MUTEX_EXIT(&rx_serverPool_lock);
1727 return (struct rx_call *)0;
1730 } while (!(call = sq->newcall)
1731 && !(socketp && *socketp != OSI_NULLSOCKET));
1732 MUTEX_EXIT(&rx_serverPool_lock);
1734 MUTEX_ENTER(&call->lock);
1740 MUTEX_ENTER(&freeSQEList_lock);
1741 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1742 rx_FreeSQEList = sq;
1743 MUTEX_EXIT(&freeSQEList_lock);
1746 clock_GetTime(&call->startTime);
1747 call->state = RX_STATE_ACTIVE;
1748 call->mode = RX_MODE_RECEIVING;
1749 #ifdef RX_KERNEL_TRACE
1750 if (ICL_SETACTIVE(afs_iclSetp)) {
1751 int glockOwner = ISAFS_GLOCK();
1754 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1755 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1762 rxi_calltrace(RX_CALL_START, call);
1763 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1764 call->conn->service->servicePort, call->conn->service->serviceId,
1767 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1768 MUTEX_EXIT(&call->lock);
1770 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1775 #else /* RX_ENABLE_LOCKS */
1777 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1779 struct rx_serverQueueEntry *sq;
1780 register struct rx_call *call = (struct rx_call *)0, *choice2;
1781 struct rx_service *service = NULL;
1785 MUTEX_ENTER(&freeSQEList_lock);
1787 if ((sq = rx_FreeSQEList)) {
1788 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1789 MUTEX_EXIT(&freeSQEList_lock);
1790 } else { /* otherwise allocate a new one and return that */
1791 MUTEX_EXIT(&freeSQEList_lock);
1792 sq = (struct rx_serverQueueEntry *)
1793 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1794 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1795 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1797 MUTEX_ENTER(&sq->lock);
1799 if (cur_service != NULL) {
1800 cur_service->nRequestsRunning--;
1801 if (cur_service->nRequestsRunning < cur_service->minProcs)
1805 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1806 register struct rx_call *tcall, *ncall;
1807 /* Scan for eligible incoming calls. A call is not eligible
1808 * if the maximum number of calls for its service type are
1809 * already executing */
1810 /* One thread will process calls FCFS (to prevent starvation),
1811 * while the other threads may run ahead looking for calls which
1812 * have all their input data available immediately. This helps
1813 * keep threads from blocking, waiting for data from the client. */
1814 choice2 = (struct rx_call *)0;
1815 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1816 service = tcall->conn->service;
1817 if (QuotaOK(service)) {
1818 if (tno == rxi_fcfs_thread_num
1819 || !tcall->queue_item_header.next) {
1820 /* If we're the fcfs thread, then we'll just use
1821 * this call. If we haven't been able to find an optimal
1822 * choice, and we're at the end of the list, then use a
1823 * 2d choice if one has been identified. Otherwise... */
1824 call = (choice2 ? choice2 : tcall);
1825 service = call->conn->service;
1826 } else if (!queue_IsEmpty(&tcall->rq)) {
1827 struct rx_packet *rp;
1828 rp = queue_First(&tcall->rq, rx_packet);
1829 if (rp->header.seq == 1
1831 || (rp->header.flags & RX_LAST_PACKET))) {
1833 } else if (rxi_2dchoice && !choice2
1834 && !(tcall->flags & RX_CALL_CLEARED)
1835 && (tcall->rprev > rxi_HardAckRate)) {
1848 /* we can't schedule a call if there's no data!!! */
1849 /* send an ack if there's no data, if we're missing the
1850 * first packet, or we're missing something between first
1851 * and last -- there's a "hole" in the incoming data. */
1852 if (queue_IsEmpty(&call->rq)
1853 || queue_First(&call->rq, rx_packet)->header.seq != 1
1854 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1855 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1857 call->flags &= (~RX_CALL_WAIT_PROC);
1858 service->nRequestsRunning++;
1859 /* just started call in minProcs pool, need fewer to maintain
1861 if (service->nRequestsRunning <= service->minProcs)
1865 /* MUTEX_EXIT(&call->lock); */
1867 /* If there are no eligible incoming calls, add this process
1868 * to the idle server queue, to wait for one */
1871 *socketp = OSI_NULLSOCKET;
1873 sq->socketp = socketp;
1874 queue_Append(&rx_idleServerQueue, sq);
1878 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1880 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1881 return (struct rx_call *)0;
1884 } while (!(call = sq->newcall)
1885 && !(socketp && *socketp != OSI_NULLSOCKET));
1887 MUTEX_EXIT(&sq->lock);
1889 MUTEX_ENTER(&freeSQEList_lock);
1890 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1891 rx_FreeSQEList = sq;
1892 MUTEX_EXIT(&freeSQEList_lock);
1895 clock_GetTime(&call->startTime);
1896 call->state = RX_STATE_ACTIVE;
1897 call->mode = RX_MODE_RECEIVING;
1898 #ifdef RX_KERNEL_TRACE
1899 if (ICL_SETACTIVE(afs_iclSetp)) {
1900 int glockOwner = ISAFS_GLOCK();
1903 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1904 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1911 rxi_calltrace(RX_CALL_START, call);
1912 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1913 call->conn->service->servicePort, call->conn->service->serviceId,
1916 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1923 #endif /* RX_ENABLE_LOCKS */
1927 /* Establish a procedure to be called when a packet arrives for a
1928 * call. This routine will be called at most once after each call,
1929 * and will also be called if there is an error condition on the or
1930 * the call is complete. Used by multi rx to build a selection
1931 * function which determines which of several calls is likely to be a
1932 * good one to read from.
1933 * NOTE: the way this is currently implemented it is probably only a
1934 * good idea to (1) use it immediately after a newcall (clients only)
1935 * and (2) only use it once. Other uses currently void your warranty
1938 rx_SetArrivalProc(register struct rx_call *call,
1939 register void (*proc) (register struct rx_call * call,
1941 register int index),
1942 register void * handle, register int arg)
1944 call->arrivalProc = proc;
1945 call->arrivalProcHandle = handle;
1946 call->arrivalProcArg = arg;
1949 /* Call is finished (possibly prematurely). Return rc to the peer, if
1950 * appropriate, and return the final error code from the conversation
1954 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1956 register struct rx_connection *conn = call->conn;
1957 register struct rx_service *service;
1963 dpf(("rx_EndCall(call %x rc %d error %d abortCode %d)\n", call, rc, call->error, call->abortCode));
1966 MUTEX_ENTER(&call->lock);
1968 if (rc == 0 && call->error == 0) {
1969 call->abortCode = 0;
1970 call->abortCount = 0;
1973 call->arrivalProc = (void (*)())0;
1974 if (rc && call->error == 0) {
1975 rxi_CallError(call, rc);
1976 /* Send an abort message to the peer if this error code has
1977 * only just been set. If it was set previously, assume the
1978 * peer has already been sent the error code or will request it
1980 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1982 if (conn->type == RX_SERVER_CONNECTION) {
1983 /* Make sure reply or at least dummy reply is sent */
1984 if (call->mode == RX_MODE_RECEIVING) {
1985 rxi_WriteProc(call, 0, 0);
1987 if (call->mode == RX_MODE_SENDING) {
1988 rxi_FlushWrite(call);
1990 service = conn->service;
1991 rxi_calltrace(RX_CALL_END, call);
1992 /* Call goes to hold state until reply packets are acknowledged */
1993 if (call->tfirst + call->nSoftAcked < call->tnext) {
1994 call->state = RX_STATE_HOLD;
1996 call->state = RX_STATE_DALLY;
1997 rxi_ClearTransmitQueue(call, 0);
1998 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1999 rxevent_Cancel(call->keepAliveEvent, call,
2000 RX_CALL_REFCOUNT_ALIVE);
2002 } else { /* Client connection */
2004 /* Make sure server receives input packets, in the case where
2005 * no reply arguments are expected */
2006 if ((call->mode == RX_MODE_SENDING)
2007 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2008 (void)rxi_ReadProc(call, &dummy, 1);
2011 /* If we had an outstanding delayed ack, be nice to the server
2012 * and force-send it now.
2014 if (call->delayedAckEvent) {
2015 rxevent_Cancel(call->delayedAckEvent, call,
2016 RX_CALL_REFCOUNT_DELAY);
2017 call->delayedAckEvent = NULL;
2018 rxi_SendDelayedAck(NULL, call, NULL);
2021 /* We need to release the call lock since it's lower than the
2022 * conn_call_lock and we don't want to hold the conn_call_lock
2023 * over the rx_ReadProc call. The conn_call_lock needs to be held
2024 * here for the case where rx_NewCall is perusing the calls on
2025 * the connection structure. We don't want to signal until
2026 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2027 * have checked this call, found it active and by the time it
2028 * goes to sleep, will have missed the signal.
2030 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
2031 * there are threads waiting to use the conn object.
2033 MUTEX_EXIT(&call->lock);
2034 MUTEX_ENTER(&conn->conn_call_lock);
2035 MUTEX_ENTER(&call->lock);
2036 MUTEX_ENTER(&conn->conn_data_lock);
2037 conn->flags |= RX_CONN_BUSY;
2038 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2039 if (conn->makeCallWaiters == 0)
2040 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
2041 MUTEX_EXIT(&conn->conn_data_lock);
2042 #ifdef RX_ENABLE_LOCKS
2043 CV_BROADCAST(&conn->conn_call_cv);
2048 #ifdef RX_ENABLE_LOCKS
2050 MUTEX_EXIT(&conn->conn_data_lock);
2052 #endif /* RX_ENABLE_LOCKS */
2053 call->state = RX_STATE_DALLY;
2055 error = call->error;
2057 /* currentPacket, nLeft, and NFree must be zeroed here, because
2058 * ResetCall cannot: ResetCall may be called at splnet(), in the
2059 * kernel version, and may interrupt the macros rx_Read or
2060 * rx_Write, which run at normal priority for efficiency. */
2061 if (call->currentPacket) {
2062 queue_Prepend(&call->iovq, call->currentPacket);
2063 call->currentPacket = (struct rx_packet *)0;
2066 call->nLeft = call->nFree = call->curlen = 0;
2068 /* Free any packets from the last call to ReadvProc/WritevProc */
2069 rxi_FreePackets(0, &call->iovq);
2071 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2072 MUTEX_EXIT(&call->lock);
2073 if (conn->type == RX_CLIENT_CONNECTION) {
2074 MUTEX_EXIT(&conn->conn_call_lock);
2075 conn->flags &= ~RX_CONN_BUSY;
2079 * Map errors to the local host's errno.h format.
2081 error = ntoh_syserr_conv(error);
2085 #if !defined(KERNEL)
2087 /* Call this routine when shutting down a server or client (especially
2088 * clients). This will allow Rx to gracefully garbage collect server
2089 * connections, and reduce the number of retries that a server might
2090 * make to a dead client.
2091 * This is not quite right, since some calls may still be ongoing and
2092 * we can't lock them to destroy them. */
2096 register struct rx_connection **conn_ptr, **conn_end;
2100 if (rxinit_status == 1) {
2102 return; /* Already shutdown. */
2104 rxi_DeleteCachedConnections();
2105 if (rx_connHashTable) {
2106 MUTEX_ENTER(&rx_connHashTable_lock);
2107 for (conn_ptr = &rx_connHashTable[0], conn_end =
2108 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2110 struct rx_connection *conn, *next;
2111 for (conn = *conn_ptr; conn; conn = next) {
2113 if (conn->type == RX_CLIENT_CONNECTION) {
2114 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2116 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2117 #ifdef RX_ENABLE_LOCKS
2118 rxi_DestroyConnectionNoLock(conn);
2119 #else /* RX_ENABLE_LOCKS */
2120 rxi_DestroyConnection(conn);
2121 #endif /* RX_ENABLE_LOCKS */
2125 #ifdef RX_ENABLE_LOCKS
2126 while (rx_connCleanup_list) {
2127 struct rx_connection *conn;
2128 conn = rx_connCleanup_list;
2129 rx_connCleanup_list = rx_connCleanup_list->next;
2130 MUTEX_EXIT(&rx_connHashTable_lock);
2131 rxi_CleanupConnection(conn);
2132 MUTEX_ENTER(&rx_connHashTable_lock);
2134 MUTEX_EXIT(&rx_connHashTable_lock);
2135 #endif /* RX_ENABLE_LOCKS */
2140 afs_winsockCleanup();
2148 /* if we wakeup packet waiter too often, can get in loop with two
2149 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2151 rxi_PacketsUnWait(void)
2153 if (!rx_waitingForPackets) {
2157 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2158 return; /* still over quota */
2161 rx_waitingForPackets = 0;
2162 #ifdef RX_ENABLE_LOCKS
2163 CV_BROADCAST(&rx_waitingForPackets_cv);
2165 osi_rxWakeup(&rx_waitingForPackets);
2171 /* ------------------Internal interfaces------------------------- */
2173 /* Return this process's service structure for the
2174 * specified socket and service */
2176 rxi_FindService(register osi_socket socket, register u_short serviceId)
2178 register struct rx_service **sp;
2179 for (sp = &rx_services[0]; *sp; sp++) {
2180 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2186 /* Allocate a call structure, for the indicated channel of the
2187 * supplied connection. The mode and state of the call must be set by
2188 * the caller. Returns the call with mutex locked. */
2190 rxi_NewCall(register struct rx_connection *conn, register int channel)
2192 register struct rx_call *call;
2193 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2194 register struct rx_call *cp; /* Call pointer temp */
2195 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2196 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2198 dpf(("rxi_NewCall(conn %x, channel %d)\n", conn, channel));
2200 /* Grab an existing call structure, or allocate a new one.
2201 * Existing call structures are assumed to have been left reset by
2203 MUTEX_ENTER(&rx_freeCallQueue_lock);
2205 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2207 * EXCEPT that the TQ might not yet be cleared out.
2208 * Skip over those with in-use TQs.
2211 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2212 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2218 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2219 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2220 call = queue_First(&rx_freeCallQueue, rx_call);
2221 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2223 rx_MutexDecrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2224 MUTEX_EXIT(&rx_freeCallQueue_lock);
2225 MUTEX_ENTER(&call->lock);
2226 CLEAR_CALL_QUEUE_LOCK(call);
2227 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2228 /* Now, if TQ wasn't cleared earlier, do it now. */
2229 if (call->flags & RX_CALL_TQ_CLEARME) {
2230 rxi_ClearTransmitQueue(call, 0);
2231 queue_Init(&call->tq);
2233 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2234 /* Bind the call to its connection structure */
2236 rxi_ResetCall(call, 1);
2239 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2241 MUTEX_EXIT(&rx_freeCallQueue_lock);
2242 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2243 MUTEX_ENTER(&call->lock);
2244 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2245 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2246 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2248 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2249 /* Initialize once-only items */
2250 queue_Init(&call->tq);
2251 queue_Init(&call->rq);
2252 queue_Init(&call->iovq);
2253 /* Bind the call to its connection structure (prereq for reset) */
2255 rxi_ResetCall(call, 1);
2257 call->channel = channel;
2258 call->callNumber = &conn->callNumber[channel];
2259 call->rwind = conn->rwind[channel];
2260 call->twind = conn->twind[channel];
2261 /* Note that the next expected call number is retained (in
2262 * conn->callNumber[i]), even if we reallocate the call structure
2264 conn->call[channel] = call;
2265 /* if the channel's never been used (== 0), we should start at 1, otherwise
2266 * the call number is valid from the last time this channel was used */
2267 if (*call->callNumber == 0)
2268 *call->callNumber = 1;
2273 /* A call has been inactive long enough that so we can throw away
2274 * state, including the call structure, which is placed on the call
2276 * Call is locked upon entry.
2277 * haveCTLock set if called from rxi_ReapConnections
2279 #ifdef RX_ENABLE_LOCKS
2281 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2282 #else /* RX_ENABLE_LOCKS */
2284 rxi_FreeCall(register struct rx_call *call)
2285 #endif /* RX_ENABLE_LOCKS */
2287 register int channel = call->channel;
2288 register struct rx_connection *conn = call->conn;
2291 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2292 (*call->callNumber)++;
2293 rxi_ResetCall(call, 0);
2294 call->conn->call[channel] = (struct rx_call *)0;
2296 MUTEX_ENTER(&rx_freeCallQueue_lock);
2297 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2298 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2299 /* A call may be free even though its transmit queue is still in use.
2300 * Since we search the call list from head to tail, put busy calls at
2301 * the head of the list, and idle calls at the tail.
2303 if (call->flags & RX_CALL_TQ_BUSY)
2304 queue_Prepend(&rx_freeCallQueue, call);
2306 queue_Append(&rx_freeCallQueue, call);
2307 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2308 queue_Append(&rx_freeCallQueue, call);
2309 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2310 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2311 MUTEX_EXIT(&rx_freeCallQueue_lock);
2313 /* Destroy the connection if it was previously slated for
2314 * destruction, i.e. the Rx client code previously called
2315 * rx_DestroyConnection (client connections), or
2316 * rxi_ReapConnections called the same routine (server
2317 * connections). Only do this, however, if there are no
2318 * outstanding calls. Note that for fine grain locking, there appears
2319 * to be a deadlock in that rxi_FreeCall has a call locked and
2320 * DestroyConnectionNoLock locks each call in the conn. But note a
2321 * few lines up where we have removed this call from the conn.
2322 * If someone else destroys a connection, they either have no
2323 * call lock held or are going through this section of code.
2325 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2326 MUTEX_ENTER(&conn->conn_data_lock);
2328 MUTEX_EXIT(&conn->conn_data_lock);
2329 #ifdef RX_ENABLE_LOCKS
2331 rxi_DestroyConnectionNoLock(conn);
2333 rxi_DestroyConnection(conn);
2334 #else /* RX_ENABLE_LOCKS */
2335 rxi_DestroyConnection(conn);
2336 #endif /* RX_ENABLE_LOCKS */
2340 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2342 rxi_Alloc(register size_t size)
2346 rx_MutexAdd1Increment2(rxi_Allocsize, (afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2349 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2350 afs_osi_Alloc_NoSleep(size);
2355 osi_Panic("rxi_Alloc error");
2361 rxi_Free(void *addr, register size_t size)
2363 rx_MutexAdd1Decrement2(rxi_Allocsize, -(afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2364 osi_Free(addr, size);
2368 rxi_SetPeerMtu(register afs_uint32 host, register afs_uint32 port, int mtu)
2370 struct rx_peer **peer_ptr, **peer_end;
2373 MUTEX_ENTER(&rx_peerHashTable_lock);
2375 for (peer_ptr = &rx_peerHashTable[0], peer_end =
2376 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
2378 struct rx_peer *peer, *next;
2379 for (peer = *peer_ptr; peer; peer = next) {
2381 if (host == peer->host) {
2382 MUTEX_ENTER(&peer->peer_lock);
2383 peer->ifMTU=MIN(mtu, peer->ifMTU);
2384 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2385 MUTEX_EXIT(&peer->peer_lock);
2390 struct rx_peer *peer, *next;
2391 hashIndex = PEER_HASH(host, port);
2392 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2393 if ((peer->host == host) && (peer->port == port)) {
2394 MUTEX_ENTER(&peer->peer_lock);
2395 peer->ifMTU=MIN(mtu, peer->ifMTU);
2396 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2397 MUTEX_EXIT(&peer->peer_lock);
2401 MUTEX_EXIT(&rx_peerHashTable_lock);
2404 /* Find the peer process represented by the supplied (host,port)
2405 * combination. If there is no appropriate active peer structure, a
2406 * new one will be allocated and initialized
2407 * The origPeer, if set, is a pointer to a peer structure on which the
2408 * refcount will be be decremented. This is used to replace the peer
2409 * structure hanging off a connection structure */
2411 rxi_FindPeer(register afs_uint32 host, register u_short port,
2412 struct rx_peer *origPeer, int create)
2414 register struct rx_peer *pp;
2416 hashIndex = PEER_HASH(host, port);
2417 MUTEX_ENTER(&rx_peerHashTable_lock);
2418 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2419 if ((pp->host == host) && (pp->port == port))
2424 pp = rxi_AllocPeer(); /* This bzero's *pp */
2425 pp->host = host; /* set here or in InitPeerParams is zero */
2427 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2428 queue_Init(&pp->congestionQueue);
2429 queue_Init(&pp->rpcStats);
2430 pp->next = rx_peerHashTable[hashIndex];
2431 rx_peerHashTable[hashIndex] = pp;
2432 rxi_InitPeerParams(pp);
2433 rx_MutexIncrement(rx_stats.nPeerStructs, rx_stats_mutex);
2440 origPeer->refCount--;
2441 MUTEX_EXIT(&rx_peerHashTable_lock);
2446 /* Find the connection at (host, port) started at epoch, and with the
2447 * given connection id. Creates the server connection if necessary.
2448 * The type specifies whether a client connection or a server
2449 * connection is desired. In both cases, (host, port) specify the
2450 * peer's (host, pair) pair. Client connections are not made
2451 * automatically by this routine. The parameter socket gives the
2452 * socket descriptor on which the packet was received. This is used,
2453 * in the case of server connections, to check that *new* connections
2454 * come via a valid (port, serviceId). Finally, the securityIndex
2455 * parameter must match the existing index for the connection. If a
2456 * server connection is created, it will be created using the supplied
2457 * index, if the index is valid for this service */
2458 struct rx_connection *
2459 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2460 register u_short port, u_short serviceId, afs_uint32 cid,
2461 afs_uint32 epoch, int type, u_int securityIndex)
2463 int hashindex, flag, i;
2464 register struct rx_connection *conn;
2465 hashindex = CONN_HASH(host, port, cid, epoch, type);
2466 MUTEX_ENTER(&rx_connHashTable_lock);
2467 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2468 rx_connHashTable[hashindex],
2471 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2472 && (epoch == conn->epoch)) {
2473 register struct rx_peer *pp = conn->peer;
2474 if (securityIndex != conn->securityIndex) {
2475 /* this isn't supposed to happen, but someone could forge a packet
2476 * like this, and there seems to be some CM bug that makes this
2477 * happen from time to time -- in which case, the fileserver
2479 MUTEX_EXIT(&rx_connHashTable_lock);
2480 return (struct rx_connection *)0;
2482 if (pp->host == host && pp->port == port)
2484 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2486 /* So what happens when it's a callback connection? */
2487 if ( /*type == RX_CLIENT_CONNECTION && */
2488 (conn->epoch & 0x80000000))
2492 /* the connection rxLastConn that was used the last time is not the
2493 ** one we are looking for now. Hence, start searching in the hash */
2495 conn = rx_connHashTable[hashindex];
2500 struct rx_service *service;
2501 if (type == RX_CLIENT_CONNECTION) {
2502 MUTEX_EXIT(&rx_connHashTable_lock);
2503 return (struct rx_connection *)0;
2505 service = rxi_FindService(socket, serviceId);
2506 if (!service || (securityIndex >= service->nSecurityObjects)
2507 || (service->securityObjects[securityIndex] == 0)) {
2508 MUTEX_EXIT(&rx_connHashTable_lock);
2509 return (struct rx_connection *)0;
2511 conn = rxi_AllocConnection(); /* This bzero's the connection */
2512 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2513 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2514 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2515 conn->next = rx_connHashTable[hashindex];
2516 rx_connHashTable[hashindex] = conn;
2517 conn->peer = rxi_FindPeer(host, port, 0, 1);
2518 conn->type = RX_SERVER_CONNECTION;
2519 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2520 conn->epoch = epoch;
2521 conn->cid = cid & RX_CIDMASK;
2522 /* conn->serial = conn->lastSerial = 0; */
2523 /* conn->timeout = 0; */
2524 conn->ackRate = RX_FAST_ACK_RATE;
2525 conn->service = service;
2526 conn->serviceId = serviceId;
2527 conn->securityIndex = securityIndex;
2528 conn->securityObject = service->securityObjects[securityIndex];
2529 conn->nSpecific = 0;
2530 conn->specific = NULL;
2531 rx_SetConnDeadTime(conn, service->connDeadTime);
2532 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2533 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2534 for (i = 0; i < RX_MAXCALLS; i++) {
2535 conn->twind[i] = rx_initSendWindow;
2536 conn->rwind[i] = rx_initReceiveWindow;
2538 /* Notify security object of the new connection */
2539 RXS_NewConnection(conn->securityObject, conn);
2540 /* XXXX Connection timeout? */
2541 if (service->newConnProc)
2542 (*service->newConnProc) (conn);
2543 rx_MutexIncrement(rx_stats.nServerConns, rx_stats_mutex);
2546 MUTEX_ENTER(&conn->conn_data_lock);
2548 MUTEX_EXIT(&conn->conn_data_lock);
2550 rxLastConn = conn; /* store this connection as the last conn used */
2551 MUTEX_EXIT(&rx_connHashTable_lock);
2555 /* There are two packet tracing routines available for testing and monitoring
2556 * Rx. One is called just after every packet is received and the other is
2557 * called just before every packet is sent. Received packets, have had their
2558 * headers decoded, and packets to be sent have not yet had their headers
2559 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2560 * containing the network address. Both can be modified. The return value, if
2561 * non-zero, indicates that the packet should be dropped. */
2563 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
2564 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
2566 /* A packet has been received off the interface. Np is the packet, socket is
2567 * the socket number it was received from (useful in determining which service
2568 * this packet corresponds to), and (host, port) reflect the host,port of the
2569 * sender. This call returns the packet to the caller if it is finished with
2570 * it, rather than de-allocating it, just as a small performance hack */
2573 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2574 afs_uint32 host, u_short port, int *tnop,
2575 struct rx_call **newcallp)
2577 register struct rx_call *call;
2578 register struct rx_connection *conn;
2580 afs_uint32 currentCallNumber;
2586 struct rx_packet *tnp;
2589 /* We don't print out the packet until now because (1) the time may not be
2590 * accurate enough until now in the lwp implementation (rx_Listener only gets
2591 * the time after the packet is read) and (2) from a protocol point of view,
2592 * this is the first time the packet has been seen */
2593 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2594 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2595 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2596 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2597 np->header.epoch, np->header.cid, np->header.callNumber,
2598 np->header.seq, np->header.flags, np));
2601 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2602 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2605 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2606 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2609 /* If an input tracer function is defined, call it with the packet and
2610 * network address. Note this function may modify its arguments. */
2611 if (rx_justReceived) {
2612 struct sockaddr_in addr;
2614 addr.sin_family = AF_INET;
2615 addr.sin_port = port;
2616 addr.sin_addr.s_addr = host;
2617 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2618 addr.sin_len = sizeof(addr);
2619 #endif /* AFS_OSF_ENV */
2620 drop = (*rx_justReceived) (np, &addr);
2621 /* drop packet if return value is non-zero */
2624 port = addr.sin_port; /* in case fcn changed addr */
2625 host = addr.sin_addr.s_addr;
2629 /* If packet was not sent by the client, then *we* must be the client */
2630 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2631 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2633 /* Find the connection (or fabricate one, if we're the server & if
2634 * necessary) associated with this packet */
2636 rxi_FindConnection(socket, host, port, np->header.serviceId,
2637 np->header.cid, np->header.epoch, type,
2638 np->header.securityIndex);
2641 /* If no connection found or fabricated, just ignore the packet.
2642 * (An argument could be made for sending an abort packet for
2647 MUTEX_ENTER(&conn->conn_data_lock);
2648 if (conn->maxSerial < np->header.serial)
2649 conn->maxSerial = np->header.serial;
2650 MUTEX_EXIT(&conn->conn_data_lock);
2652 /* If the connection is in an error state, send an abort packet and ignore
2653 * the incoming packet */
2655 /* Don't respond to an abort packet--we don't want loops! */
2656 MUTEX_ENTER(&conn->conn_data_lock);
2657 if (np->header.type != RX_PACKET_TYPE_ABORT)
2658 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2660 MUTEX_EXIT(&conn->conn_data_lock);
2664 /* Check for connection-only requests (i.e. not call specific). */
2665 if (np->header.callNumber == 0) {
2666 switch (np->header.type) {
2667 case RX_PACKET_TYPE_ABORT: {
2668 /* What if the supplied error is zero? */
2669 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2670 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2671 rxi_ConnectionError(conn, errcode);
2672 MUTEX_ENTER(&conn->conn_data_lock);
2674 MUTEX_EXIT(&conn->conn_data_lock);
2677 case RX_PACKET_TYPE_CHALLENGE:
2678 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2679 MUTEX_ENTER(&conn->conn_data_lock);
2681 MUTEX_EXIT(&conn->conn_data_lock);
2683 case RX_PACKET_TYPE_RESPONSE:
2684 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2685 MUTEX_ENTER(&conn->conn_data_lock);
2687 MUTEX_EXIT(&conn->conn_data_lock);
2689 case RX_PACKET_TYPE_PARAMS:
2690 case RX_PACKET_TYPE_PARAMS + 1:
2691 case RX_PACKET_TYPE_PARAMS + 2:
2692 /* ignore these packet types for now */
2693 MUTEX_ENTER(&conn->conn_data_lock);
2695 MUTEX_EXIT(&conn->conn_data_lock);
2700 /* Should not reach here, unless the peer is broken: send an
2702 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2703 MUTEX_ENTER(&conn->conn_data_lock);
2704 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2706 MUTEX_EXIT(&conn->conn_data_lock);
2711 channel = np->header.cid & RX_CHANNELMASK;
2712 call = conn->call[channel];
2713 #ifdef RX_ENABLE_LOCKS
2715 MUTEX_ENTER(&call->lock);
2716 /* Test to see if call struct is still attached to conn. */
2717 if (call != conn->call[channel]) {
2719 MUTEX_EXIT(&call->lock);
2720 if (type == RX_SERVER_CONNECTION) {
2721 call = conn->call[channel];
2722 /* If we started with no call attached and there is one now,
2723 * another thread is also running this routine and has gotten
2724 * the connection channel. We should drop this packet in the tests
2725 * below. If there was a call on this connection and it's now
2726 * gone, then we'll be making a new call below.
2727 * If there was previously a call and it's now different then
2728 * the old call was freed and another thread running this routine
2729 * has created a call on this channel. One of these two threads
2730 * has a packet for the old call and the code below handles those
2734 MUTEX_ENTER(&call->lock);
2736 /* This packet can't be for this call. If the new call address is
2737 * 0 then no call is running on this channel. If there is a call
2738 * then, since this is a client connection we're getting data for
2739 * it must be for the previous call.
2741 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2742 MUTEX_ENTER(&conn->conn_data_lock);
2744 MUTEX_EXIT(&conn->conn_data_lock);
2749 currentCallNumber = conn->callNumber[channel];
2751 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2752 if (np->header.callNumber < currentCallNumber) {
2753 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2754 #ifdef RX_ENABLE_LOCKS
2756 MUTEX_EXIT(&call->lock);
2758 MUTEX_ENTER(&conn->conn_data_lock);
2760 MUTEX_EXIT(&conn->conn_data_lock);
2764 MUTEX_ENTER(&conn->conn_call_lock);
2765 call = rxi_NewCall(conn, channel);
2766 MUTEX_EXIT(&conn->conn_call_lock);
2767 *call->callNumber = np->header.callNumber;
2768 if (np->header.callNumber == 0)
2769 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));
2771 call->state = RX_STATE_PRECALL;
2772 clock_GetTime(&call->queueTime);
2773 hzero(call->bytesSent);
2774 hzero(call->bytesRcvd);
2776 * If the number of queued calls exceeds the overload
2777 * threshold then abort this call.
2779 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2780 struct rx_packet *tp;
2782 rxi_CallError(call, rx_BusyError);
2783 tp = rxi_SendCallAbort(call, np, 1, 0);
2784 MUTEX_EXIT(&call->lock);
2785 MUTEX_ENTER(&conn->conn_data_lock);
2787 MUTEX_EXIT(&conn->conn_data_lock);
2788 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2791 rxi_KeepAliveOn(call);
2792 } else if (np->header.callNumber != currentCallNumber) {
2793 /* Wait until the transmit queue is idle before deciding
2794 * whether to reset the current call. Chances are that the
2795 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2798 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2799 while ((call->state == RX_STATE_ACTIVE)
2800 && (call->flags & RX_CALL_TQ_BUSY)) {
2801 call->flags |= RX_CALL_TQ_WAIT;
2803 #ifdef RX_ENABLE_LOCKS
2804 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2805 CV_WAIT(&call->cv_tq, &call->lock);
2806 #else /* RX_ENABLE_LOCKS */
2807 osi_rxSleep(&call->tq);
2808 #endif /* RX_ENABLE_LOCKS */
2810 if (call->tqWaiters == 0)
2811 call->flags &= ~RX_CALL_TQ_WAIT;
2813 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2814 /* If the new call cannot be taken right now send a busy and set
2815 * the error condition in this call, so that it terminates as
2816 * quickly as possible */
2817 if (call->state == RX_STATE_ACTIVE) {
2818 struct rx_packet *tp;
2820 rxi_CallError(call, RX_CALL_DEAD);
2821 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2823 MUTEX_EXIT(&call->lock);
2824 MUTEX_ENTER(&conn->conn_data_lock);
2826 MUTEX_EXIT(&conn->conn_data_lock);
2829 rxi_ResetCall(call, 0);
2830 *call->callNumber = np->header.callNumber;
2831 if (np->header.callNumber == 0)
2832 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));
2834 call->state = RX_STATE_PRECALL;
2835 clock_GetTime(&call->queueTime);
2836 hzero(call->bytesSent);
2837 hzero(call->bytesRcvd);
2839 * If the number of queued calls exceeds the overload
2840 * threshold then abort this call.
2842 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2843 struct rx_packet *tp;
2845 rxi_CallError(call, rx_BusyError);
2846 tp = rxi_SendCallAbort(call, np, 1, 0);
2847 MUTEX_EXIT(&call->lock);
2848 MUTEX_ENTER(&conn->conn_data_lock);
2850 MUTEX_EXIT(&conn->conn_data_lock);
2851 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2854 rxi_KeepAliveOn(call);
2856 /* Continuing call; do nothing here. */
2858 } else { /* we're the client */
2859 /* Ignore all incoming acknowledgements for calls in DALLY state */
2860 if (call && (call->state == RX_STATE_DALLY)
2861 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2862 rx_MutexIncrement(rx_stats.ignorePacketDally, rx_stats_mutex);
2863 #ifdef RX_ENABLE_LOCKS
2865 MUTEX_EXIT(&call->lock);
2868 MUTEX_ENTER(&conn->conn_data_lock);
2870 MUTEX_EXIT(&conn->conn_data_lock);
2874 /* Ignore anything that's not relevant to the current call. If there
2875 * isn't a current call, then no packet is relevant. */
2876 if (!call || (np->header.callNumber != currentCallNumber)) {
2877 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2878 #ifdef RX_ENABLE_LOCKS
2880 MUTEX_EXIT(&call->lock);
2883 MUTEX_ENTER(&conn->conn_data_lock);
2885 MUTEX_EXIT(&conn->conn_data_lock);
2888 /* If the service security object index stamped in the packet does not
2889 * match the connection's security index, ignore the packet */
2890 if (np->header.securityIndex != conn->securityIndex) {
2891 #ifdef RX_ENABLE_LOCKS
2892 MUTEX_EXIT(&call->lock);
2894 MUTEX_ENTER(&conn->conn_data_lock);
2896 MUTEX_EXIT(&conn->conn_data_lock);
2900 /* If we're receiving the response, then all transmit packets are
2901 * implicitly acknowledged. Get rid of them. */
2902 if (np->header.type == RX_PACKET_TYPE_DATA) {
2903 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2904 /* XXX Hack. Because we must release the global rx lock when
2905 * sending packets (osi_NetSend) we drop all acks while we're
2906 * traversing the tq in rxi_Start sending packets out because
2907 * packets may move to the freePacketQueue as result of being here!
2908 * So we drop these packets until we're safely out of the
2909 * traversing. Really ugly!
2910 * For fine grain RX locking, we set the acked field in the
2911 * packets and let rxi_Start remove them from the transmit queue.
2913 if (call->flags & RX_CALL_TQ_BUSY) {
2914 #ifdef RX_ENABLE_LOCKS
2915 rxi_SetAcksInTransmitQueue(call);
2918 return np; /* xmitting; drop packet */
2921 rxi_ClearTransmitQueue(call, 0);
2923 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2924 rxi_ClearTransmitQueue(call, 0);
2925 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2927 if (np->header.type == RX_PACKET_TYPE_ACK) {
2928 /* now check to see if this is an ack packet acknowledging that the
2929 * server actually *lost* some hard-acked data. If this happens we
2930 * ignore this packet, as it may indicate that the server restarted in
2931 * the middle of a call. It is also possible that this is an old ack
2932 * packet. We don't abort the connection in this case, because this
2933 * *might* just be an old ack packet. The right way to detect a server
2934 * restart in the midst of a call is to notice that the server epoch
2936 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2937 * XXX unacknowledged. I think that this is off-by-one, but
2938 * XXX I don't dare change it just yet, since it will
2939 * XXX interact badly with the server-restart detection
2940 * XXX code in receiveackpacket. */
2941 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2942 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2943 MUTEX_EXIT(&call->lock);
2944 MUTEX_ENTER(&conn->conn_data_lock);
2946 MUTEX_EXIT(&conn->conn_data_lock);
2950 } /* else not a data packet */
2953 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2954 /* Set remote user defined status from packet */
2955 call->remoteStatus = np->header.userStatus;
2957 /* Note the gap between the expected next packet and the actual
2958 * packet that arrived, when the new packet has a smaller serial number
2959 * than expected. Rioses frequently reorder packets all by themselves,
2960 * so this will be quite important with very large window sizes.
2961 * Skew is checked against 0 here to avoid any dependence on the type of
2962 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2964 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2965 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2966 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2968 MUTEX_ENTER(&conn->conn_data_lock);
2969 skew = conn->lastSerial - np->header.serial;
2970 conn->lastSerial = np->header.serial;
2971 MUTEX_EXIT(&conn->conn_data_lock);
2973 register struct rx_peer *peer;
2975 if (skew > peer->inPacketSkew) {
2976 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2978 peer->inPacketSkew = skew;
2982 /* Now do packet type-specific processing */
2983 switch (np->header.type) {
2984 case RX_PACKET_TYPE_DATA:
2985 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2988 case RX_PACKET_TYPE_ACK:
2989 /* Respond immediately to ack packets requesting acknowledgement
2991 if (np->header.flags & RX_REQUEST_ACK) {
2993 (void)rxi_SendCallAbort(call, 0, 1, 0);
2995 (void)rxi_SendAck(call, 0, np->header.serial,
2996 RX_ACK_PING_RESPONSE, 1);
2998 np = rxi_ReceiveAckPacket(call, np, 1);
3000 case RX_PACKET_TYPE_ABORT: {
3001 /* An abort packet: reset the call, passing the error up to the user. */
3002 /* What if error is zero? */
3003 /* What if the error is -1? the application will treat it as a timeout. */
3004 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3005 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
3006 rxi_CallError(call, errdata);
3007 MUTEX_EXIT(&call->lock);
3008 MUTEX_ENTER(&conn->conn_data_lock);
3010 MUTEX_EXIT(&conn->conn_data_lock);
3011 return np; /* xmitting; drop packet */
3013 case RX_PACKET_TYPE_BUSY:
3016 case RX_PACKET_TYPE_ACKALL:
3017 /* All packets acknowledged, so we can drop all packets previously
3018 * readied for sending */
3019 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3020 /* XXX Hack. We because we can't release the global rx lock when
3021 * sending packets (osi_NetSend) we drop all ack pkts while we're
3022 * traversing the tq in rxi_Start sending packets out because
3023 * packets may move to the freePacketQueue as result of being
3024 * here! So we drop these packets until we're safely out of the
3025 * traversing. Really ugly!
3026 * For fine grain RX locking, we set the acked field in the packets
3027 * and let rxi_Start remove the packets from the transmit queue.
3029 if (call->flags & RX_CALL_TQ_BUSY) {
3030 #ifdef RX_ENABLE_LOCKS
3031 rxi_SetAcksInTransmitQueue(call);
3033 #else /* RX_ENABLE_LOCKS */
3034 MUTEX_EXIT(&call->lock);
3035 MUTEX_ENTER(&conn->conn_data_lock);
3037 MUTEX_EXIT(&conn->conn_data_lock);
3038 return np; /* xmitting; drop packet */
3039 #endif /* RX_ENABLE_LOCKS */
3041 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3042 rxi_ClearTransmitQueue(call, 0);
3043 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
3046 /* Should not reach here, unless the peer is broken: send an abort
3048 rxi_CallError(call, RX_PROTOCOL_ERROR);
3049 np = rxi_SendCallAbort(call, np, 1, 0);
3052 /* Note when this last legitimate packet was received, for keep-alive
3053 * processing. Note, we delay getting the time until now in the hope that
3054 * the packet will be delivered to the user before any get time is required
3055 * (if not, then the time won't actually be re-evaluated here). */
3056 call->lastReceiveTime = clock_Sec();
3057 MUTEX_EXIT(&call->lock);
3058 MUTEX_ENTER(&conn->conn_data_lock);
3060 MUTEX_EXIT(&conn->conn_data_lock);
3064 /* return true if this is an "interesting" connection from the point of view
3065 of someone trying to debug the system */
3067 rxi_IsConnInteresting(struct rx_connection *aconn)
3070 register struct rx_call *tcall;
3072 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3074 for (i = 0; i < RX_MAXCALLS; i++) {
3075 tcall = aconn->call[i];
3077 if ((tcall->state == RX_STATE_PRECALL)
3078 || (tcall->state == RX_STATE_ACTIVE))
3080 if ((tcall->mode == RX_MODE_SENDING)
3081 || (tcall->mode == RX_MODE_RECEIVING))
3089 /* if this is one of the last few packets AND it wouldn't be used by the
3090 receiving call to immediately satisfy a read request, then drop it on
3091 the floor, since accepting it might prevent a lock-holding thread from
3092 making progress in its reading. If a call has been cleared while in
3093 the precall state then ignore all subsequent packets until the call
3094 is assigned to a thread. */
3097 TooLow(struct rx_packet *ap, struct rx_call *acall)
3100 MUTEX_ENTER(&rx_stats_mutex);
3101 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3102 && (acall->state == RX_STATE_PRECALL))
3103 || ((rx_nFreePackets < rxi_dataQuota + 2)
3104 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3105 && (acall->flags & RX_CALL_READER_WAIT)))) {
3108 MUTEX_EXIT(&rx_stats_mutex);
3114 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
3116 struct rx_connection *conn = arg1;
3117 struct rx_call *acall = arg2;
3118 struct rx_call *call = acall;
3119 struct clock when, now;
3122 MUTEX_ENTER(&conn->conn_data_lock);
3123 conn->checkReachEvent = NULL;
3124 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3127 MUTEX_EXIT(&conn->conn_data_lock);
3131 MUTEX_ENTER(&conn->conn_call_lock);
3132 MUTEX_ENTER(&conn->conn_data_lock);
3133 for (i = 0; i < RX_MAXCALLS; i++) {
3134 struct rx_call *tc = conn->call[i];
3135 if (tc && tc->state == RX_STATE_PRECALL) {
3141 /* Indicate that rxi_CheckReachEvent is no longer running by
3142 * clearing the flag. Must be atomic under conn_data_lock to
3143 * avoid a new call slipping by: rxi_CheckConnReach holds
3144 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3146 conn->flags &= ~RX_CONN_ATTACHWAIT;
3147 MUTEX_EXIT(&conn->conn_data_lock);
3148 MUTEX_EXIT(&conn->conn_call_lock);
3153 MUTEX_ENTER(&call->lock);
3154 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3156 MUTEX_EXIT(&call->lock);
3158 clock_GetTime(&now);
3160 when.sec += RX_CHECKREACH_TIMEOUT;
3161 MUTEX_ENTER(&conn->conn_data_lock);
3162 if (!conn->checkReachEvent) {
3164 conn->checkReachEvent =
3165 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3168 MUTEX_EXIT(&conn->conn_data_lock);
3174 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3176 struct rx_service *service = conn->service;
3177 struct rx_peer *peer = conn->peer;
3178 afs_uint32 now, lastReach;
3180 if (service->checkReach == 0)
3184 MUTEX_ENTER(&peer->peer_lock);
3185 lastReach = peer->lastReachTime;
3186 MUTEX_EXIT(&peer->peer_lock);
3187 if (now - lastReach < RX_CHECKREACH_TTL)
3190 MUTEX_ENTER(&conn->conn_data_lock);
3191 if (conn->flags & RX_CONN_ATTACHWAIT) {
3192 MUTEX_EXIT(&conn->conn_data_lock);
3195 conn->flags |= RX_CONN_ATTACHWAIT;
3196 MUTEX_EXIT(&conn->conn_data_lock);
3197 if (!conn->checkReachEvent)
3198 rxi_CheckReachEvent(NULL, conn, call);
3203 /* try to attach call, if authentication is complete */
3205 TryAttach(register struct rx_call *acall, register osi_socket socket,
3206 register int *tnop, register struct rx_call **newcallp,
3209 struct rx_connection *conn = acall->conn;
3211 if (conn->type == RX_SERVER_CONNECTION
3212 && acall->state == RX_STATE_PRECALL) {
3213 /* Don't attach until we have any req'd. authentication. */
3214 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3215 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3216 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3217 /* Note: this does not necessarily succeed; there
3218 * may not any proc available
3221 rxi_ChallengeOn(acall->conn);
3226 /* A data packet has been received off the interface. This packet is
3227 * appropriate to the call (the call is in the right state, etc.). This
3228 * routine can return a packet to the caller, for re-use */
3231 rxi_ReceiveDataPacket(register struct rx_call *call,
3232 register struct rx_packet *np, int istack,
3233 osi_socket socket, afs_uint32 host, u_short port,
3234 int *tnop, struct rx_call **newcallp)
3236 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3240 afs_uint32 seq, serial, flags;
3242 struct rx_packet *tnp;
3243 struct clock when, now;
3244 rx_MutexIncrement(rx_stats.dataPacketsRead, rx_stats_mutex);
3247 /* If there are no packet buffers, drop this new packet, unless we can find
3248 * packet buffers from inactive calls */
3250 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3251 MUTEX_ENTER(&rx_freePktQ_lock);
3252 rxi_NeedMorePackets = TRUE;
3253 MUTEX_EXIT(&rx_freePktQ_lock);
3254 rx_MutexIncrement(rx_stats.noPacketBuffersOnRead, rx_stats_mutex);
3255 call->rprev = np->header.serial;
3256 rxi_calltrace(RX_TRACE_DROP, call);
3257 dpf(("packet %x dropped on receipt - quota problems", np));
3259 rxi_ClearReceiveQueue(call);
3260 clock_GetTime(&now);
3262 clock_Add(&when, &rx_softAckDelay);
3263 if (!call->delayedAckEvent
3264 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3265 rxevent_Cancel(call->delayedAckEvent, call,
3266 RX_CALL_REFCOUNT_DELAY);
3267 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3268 call->delayedAckEvent =
3269 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3271 /* we've damaged this call already, might as well do it in. */
3277 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3278 * packet is one of several packets transmitted as a single
3279 * datagram. Do not send any soft or hard acks until all packets
3280 * in a jumbogram have been processed. Send negative acks right away.
3282 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3283 /* tnp is non-null when there are more packets in the
3284 * current jumbo gram */
3291 seq = np->header.seq;
3292 serial = np->header.serial;
3293 flags = np->header.flags;
3295 /* If the call is in an error state, send an abort message */
3297 return rxi_SendCallAbort(call, np, istack, 0);
3299 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3300 * AFS 3.5 jumbogram. */
3301 if (flags & RX_JUMBO_PACKET) {
3302 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3307 if (np->header.spare != 0) {
3308 MUTEX_ENTER(&call->conn->conn_data_lock);
3309 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3310 MUTEX_EXIT(&call->conn->conn_data_lock);
3313 /* The usual case is that this is the expected next packet */
3314 if (seq == call->rnext) {
3316 /* Check to make sure it is not a duplicate of one already queued */
3317 if (queue_IsNotEmpty(&call->rq)
3318 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3319 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3320 dpf(("packet %x dropped on receipt - duplicate", np));
3321 rxevent_Cancel(call->delayedAckEvent, call,
3322 RX_CALL_REFCOUNT_DELAY);
3323 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3329 /* It's the next packet. Stick it on the receive queue
3330 * for this call. Set newPackets to make sure we wake
3331 * the reader once all packets have been processed */
3332 queue_Prepend(&call->rq, np);
3334 np = NULL; /* We can't use this anymore */
3337 /* If an ack is requested then set a flag to make sure we
3338 * send an acknowledgement for this packet */
3339 if (flags & RX_REQUEST_ACK) {
3340 ackNeeded = RX_ACK_REQUESTED;
3343 /* Keep track of whether we have received the last packet */
3344 if (flags & RX_LAST_PACKET) {
3345 call->flags |= RX_CALL_HAVE_LAST;
3349 /* Check whether we have all of the packets for this call */
3350 if (call->flags & RX_CALL_HAVE_LAST) {
3351 afs_uint32 tseq; /* temporary sequence number */
3352 struct rx_packet *tp; /* Temporary packet pointer */
3353 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3355 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3356 if (tseq != tp->header.seq)
3358 if (tp->header.flags & RX_LAST_PACKET) {
3359 call->flags |= RX_CALL_RECEIVE_DONE;
3366 /* Provide asynchronous notification for those who want it
3367 * (e.g. multi rx) */
3368 if (call->arrivalProc) {
3369 (*call->arrivalProc) (call, call->arrivalProcHandle,
3370 call->arrivalProcArg);
3371 call->arrivalProc = (void (*)())0;
3374 /* Update last packet received */
3377 /* If there is no server process serving this call, grab
3378 * one, if available. We only need to do this once. If a
3379 * server thread is available, this thread becomes a server
3380 * thread and the server thread becomes a listener thread. */
3382 TryAttach(call, socket, tnop, newcallp, 0);
3385 /* This is not the expected next packet. */
3387 /* Determine whether this is a new or old packet, and if it's
3388 * a new one, whether it fits into the current receive window.
3389 * Also figure out whether the packet was delivered in sequence.
3390 * We use the prev variable to determine whether the new packet
3391 * is the successor of its immediate predecessor in the
3392 * receive queue, and the missing flag to determine whether
3393 * any of this packets predecessors are missing. */
3395 afs_uint32 prev; /* "Previous packet" sequence number */
3396 struct rx_packet *tp; /* Temporary packet pointer */
3397 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3398 int missing; /* Are any predecessors missing? */
3400 /* If the new packet's sequence number has been sent to the
3401 * application already, then this is a duplicate */
3402 if (seq < call->rnext) {
3403 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3404 rxevent_Cancel(call->delayedAckEvent, call,
3405 RX_CALL_REFCOUNT_DELAY);
3406 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3412 /* If the sequence number is greater than what can be
3413 * accomodated by the current window, then send a negative
3414 * acknowledge and drop the packet */
3415 if ((call->rnext + call->rwind) <= seq) {
3416 rxevent_Cancel(call->delayedAckEvent, call,
3417 RX_CALL_REFCOUNT_DELAY);
3418 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3425 /* Look for the packet in the queue of old received packets */
3426 for (prev = call->rnext - 1, missing =
3427 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3428 /*Check for duplicate packet */
3429 if (seq == tp->header.seq) {
3430 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3431 rxevent_Cancel(call->delayedAckEvent, call,
3432 RX_CALL_REFCOUNT_DELAY);
3433 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3439 /* If we find a higher sequence packet, break out and
3440 * insert the new packet here. */
3441 if (seq < tp->header.seq)
3443 /* Check for missing packet */
3444 if (tp->header.seq != prev + 1) {
3448 prev = tp->header.seq;
3451 /* Keep track of whether we have received the last packet. */
3452 if (flags & RX_LAST_PACKET) {
3453 call->flags |= RX_CALL_HAVE_LAST;
3456 /* It's within the window: add it to the the receive queue.
3457 * tp is left by the previous loop either pointing at the
3458 * packet before which to insert the new packet, or at the
3459 * queue head if the queue is empty or the packet should be
3461 queue_InsertBefore(tp, np);
3465 /* Check whether we have all of the packets for this call */
3466 if ((call->flags & RX_CALL_HAVE_LAST)
3467 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3468 afs_uint32 tseq; /* temporary sequence number */
3471 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3472 if (tseq != tp->header.seq)
3474 if (tp->header.flags & RX_LAST_PACKET) {
3475 call->flags |= RX_CALL_RECEIVE_DONE;
3482 /* We need to send an ack of the packet is out of sequence,
3483 * or if an ack was requested by the peer. */
3484 if (seq != prev + 1 || missing) {
3485 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3486 } else if (flags & RX_REQUEST_ACK) {
3487 ackNeeded = RX_ACK_REQUESTED;
3490 /* Acknowledge the last packet for each call */
3491 if (flags & RX_LAST_PACKET) {
3502 * If the receiver is waiting for an iovec, fill the iovec
3503 * using the data from the receive queue */
3504 if (call->flags & RX_CALL_IOVEC_WAIT) {
3505 didHardAck = rxi_FillReadVec(call, serial);
3506 /* the call may have been aborted */
3515 /* Wakeup the reader if any */
3516 if ((call->flags & RX_CALL_READER_WAIT)
3517 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3518 || (call->iovNext >= call->iovMax)
3519 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3520 call->flags &= ~RX_CALL_READER_WAIT;
3521 #ifdef RX_ENABLE_LOCKS
3522 CV_BROADCAST(&call->cv_rq);
3524 osi_rxWakeup(&call->rq);
3530 * Send an ack when requested by the peer, or once every
3531 * rxi_SoftAckRate packets until the last packet has been
3532 * received. Always send a soft ack for the last packet in
3533 * the server's reply. */
3535 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3536 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3537 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3538 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3539 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3540 } else if (call->nSoftAcks) {
3541 clock_GetTime(&now);
3543 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3544 clock_Add(&when, &rx_lastAckDelay);
3546 clock_Add(&when, &rx_softAckDelay);
3548 if (!call->delayedAckEvent
3549 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3550 rxevent_Cancel(call->delayedAckEvent, call,
3551 RX_CALL_REFCOUNT_DELAY);
3552 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3553 call->delayedAckEvent =
3554 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3556 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3557 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3564 static void rxi_ComputeRate();
3568 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3570 struct rx_peer *peer = conn->peer;
3572 MUTEX_ENTER(&peer->peer_lock);
3573 peer->lastReachTime = clock_Sec();
3574 MUTEX_EXIT(&peer->peer_lock);
3576 MUTEX_ENTER(&conn->conn_data_lock);
3577 if (conn->flags & RX_CONN_ATTACHWAIT) {
3580 conn->flags &= ~RX_CONN_ATTACHWAIT;
3581 MUTEX_EXIT(&conn->conn_data_lock);
3583 for (i = 0; i < RX_MAXCALLS; i++) {
3584 struct rx_call *call = conn->call[i];
3587 MUTEX_ENTER(&call->lock);
3588 /* tnop can be null if newcallp is null */
3589 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3591 MUTEX_EXIT(&call->lock);
3595 MUTEX_EXIT(&conn->conn_data_lock);
3599 rx_ack_reason(int reason)
3602 case RX_ACK_REQUESTED:
3604 case RX_ACK_DUPLICATE:
3606 case RX_ACK_OUT_OF_SEQUENCE:
3608 case RX_ACK_EXCEEDS_WINDOW:
3610 case RX_ACK_NOSPACE:
3614 case RX_ACK_PING_RESPONSE:
3626 /* rxi_ComputePeerNetStats
3628 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3629 * estimates (like RTT and throughput) based on ack packets. Caller
3630 * must ensure that the packet in question is the right one (i.e.
3631 * serial number matches).
3634 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3635 struct rx_ackPacket *ap, struct rx_packet *np)
3637 struct rx_peer *peer = call->conn->peer;
3639 /* Use RTT if not delayed by client. */
3640 if (ap->reason != RX_ACK_DELAY)
3641 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3643 rxi_ComputeRate(peer, call, p, np, ap->reason);
3647 /* The real smarts of the whole thing. */
3649 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3652 struct rx_ackPacket *ap;
3654 register struct rx_packet *tp;
3655 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3656 register struct rx_connection *conn = call->conn;
3657 struct rx_peer *peer = conn->peer;
3660 /* because there are CM's that are bogus, sending weird values for this. */
3661 afs_uint32 skew = 0;
3666 int newAckCount = 0;
3667 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3668 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3670 rx_MutexIncrement(rx_stats.ackPacketsRead, rx_stats_mutex);
3671 ap = (struct rx_ackPacket *)rx_DataOf(np);
3672 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3674 return np; /* truncated ack packet */
3676 /* depends on ack packet struct */
3677 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3678 first = ntohl(ap->firstPacket);
3679 serial = ntohl(ap->serial);
3680 /* temporarily disabled -- needs to degrade over time
3681 * skew = ntohs(ap->maxSkew); */
3683 /* Ignore ack packets received out of order */
3684 if (first < call->tfirst) {
3688 if (np->header.flags & RX_SLOW_START_OK) {
3689 call->flags |= RX_CALL_SLOW_START_OK;
3692 if (ap->reason == RX_ACK_PING_RESPONSE)
3693 rxi_UpdatePeerReach(conn, call);
3697 if (rxdebug_active) {
3701 len = _snprintf(msg, sizeof(msg),
3702 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3703 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3704 ntohl(ap->serial), ntohl(ap->previousPacket),
3705 (unsigned int)np->header.seq, (unsigned int)skew,
3706 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3710 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3711 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3715 OutputDebugString(msg);
3717 #else /* AFS_NT40_ENV */
3720 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3721 ap->reason, ntohl(ap->previousPacket),
3722 (unsigned int)np->header.seq, (unsigned int)serial,
3723 (unsigned int)skew, ntohl(ap->firstPacket));
3726 for (offset = 0; offset < nAcks; offset++)
3727 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3732 #endif /* AFS_NT40_ENV */
3735 /* Update the outgoing packet skew value to the latest value of
3736 * the peer's incoming packet skew value. The ack packet, of
3737 * course, could arrive out of order, but that won't affect things
3739 MUTEX_ENTER(&peer->peer_lock);
3740 peer->outPacketSkew = skew;
3742 /* Check for packets that no longer need to be transmitted, and
3743 * discard them. This only applies to packets positively
3744 * acknowledged as having been sent to the peer's upper level.
3745 * All other packets must be retained. So only packets with
3746 * sequence numbers < ap->firstPacket are candidates. */
3747 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3748 if (tp->header.seq >= first)
3750 call->tfirst = tp->header.seq + 1;
3752 && (tp->header.serial == serial || tp->firstSerial == serial))
3753 rxi_ComputePeerNetStats(call, tp, ap, np);
3754 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3757 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3758 /* XXX Hack. Because we have to release the global rx lock when sending
3759 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3760 * in rxi_Start sending packets out because packets may move to the
3761 * freePacketQueue as result of being here! So we drop these packets until
3762 * we're safely out of the traversing. Really ugly!
3763 * To make it even uglier, if we're using fine grain locking, we can
3764 * set the ack bits in the packets and have rxi_Start remove the packets
3765 * when it's done transmitting.
3767 if (call->flags & RX_CALL_TQ_BUSY) {
3768 #ifdef RX_ENABLE_LOCKS
3769 tp->flags |= RX_PKTFLAG_ACKED;
3770 call->flags |= RX_CALL_TQ_SOME_ACKED;
3771 #else /* RX_ENABLE_LOCKS */
3773 #endif /* RX_ENABLE_LOCKS */
3775 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3778 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3783 /* Give rate detector a chance to respond to ping requests */
3784 if (ap->reason == RX_ACK_PING_RESPONSE) {
3785 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3789 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3791 /* Now go through explicit acks/nacks and record the results in
3792 * the waiting packets. These are packets that can't be released
3793 * yet, even with a positive acknowledge. This positive
3794 * acknowledge only means the packet has been received by the
3795 * peer, not that it will be retained long enough to be sent to
3796 * the peer's upper level. In addition, reset the transmit timers
3797 * of any missing packets (those packets that must be missing
3798 * because this packet was out of sequence) */
3800 call->nSoftAcked = 0;
3801 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3802 /* Update round trip time if the ack was stimulated on receipt
3804 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3805 #ifdef RX_ENABLE_LOCKS
3806 if (tp->header.seq >= first)
3807 #endif /* RX_ENABLE_LOCKS */
3808 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3810 && (tp->header.serial == serial || tp->firstSerial == serial))
3811 rxi_ComputePeerNetStats(call, tp, ap, np);
3813 /* Set the acknowledge flag per packet based on the
3814 * information in the ack packet. An acknowlegded packet can
3815 * be downgraded when the server has discarded a packet it
3816 * soacked previously, or when an ack packet is received
3817 * out of sequence. */
3818 if (tp->header.seq < first) {
3819 /* Implicit ack information */
3820 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3823 tp->flags |= RX_PKTFLAG_ACKED;
3824 } else if (tp->header.seq < first + nAcks) {
3825 /* Explicit ack information: set it in the packet appropriately */
3826 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3827 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3829 tp->flags |= RX_PKTFLAG_ACKED;
3836 } else /* RX_ACK_TYPE_NACK */ {
3837 tp->flags &= ~RX_PKTFLAG_ACKED;
3841 tp->flags &= ~RX_PKTFLAG_ACKED;
3845 /* If packet isn't yet acked, and it has been transmitted at least
3846 * once, reset retransmit time using latest timeout
3847 * ie, this should readjust the retransmit timer for all outstanding
3848 * packets... So we don't just retransmit when we should know better*/
3850 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3851 tp->retryTime = tp->timeSent;
3852 clock_Add(&tp->retryTime, &peer->timeout);
3853 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3854 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3858 /* If the window has been extended by this acknowledge packet,
3859 * then wakeup a sender waiting in alloc for window space, or try
3860 * sending packets now, if he's been sitting on packets due to
3861 * lack of window space */
3862 if (call->tnext < (call->tfirst + call->twind)) {
3863 #ifdef RX_ENABLE_LOCKS
3864 CV_SIGNAL(&call->cv_twind);
3866 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3867 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3868 osi_rxWakeup(&call->twind);
3871 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3872 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3876 /* if the ack packet has a receivelen field hanging off it,
3877 * update our state */
3878 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3881 /* If the ack packet has a "recommended" size that is less than
3882 * what I am using now, reduce my size to match */
3883 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3884 (int)sizeof(afs_int32), &tSize);
3885 tSize = (afs_uint32) ntohl(tSize);
3886 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3888 /* Get the maximum packet size to send to this peer */
3889 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
3891 tSize = (afs_uint32) ntohl(tSize);
3892 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3893 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3895 /* sanity check - peer might have restarted with different params.
3896 * If peer says "send less", dammit, send less... Peer should never
3897 * be unable to accept packets of the size that prior AFS versions would
3898 * send without asking. */
3899 if (peer->maxMTU != tSize) {
3900 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
3902 peer->maxMTU = tSize;
3903 peer->MTU = MIN(tSize, peer->MTU);
3904 call->MTU = MIN(call->MTU, tSize);
3907 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3910 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3911 (int)sizeof(afs_int32), &tSize);
3912 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3913 if (tSize < call->twind) { /* smaller than our send */
3914 call->twind = tSize; /* window, we must send less... */
3915 call->ssthresh = MIN(call->twind, call->ssthresh);
3916 call->conn->twind[call->channel] = call->twind;
3919 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3920 * network MTU confused with the loopback MTU. Calculate the
3921 * maximum MTU here for use in the slow start code below.
3923 maxMTU = peer->maxMTU;
3924 /* Did peer restart with older RX version? */
3925 if (peer->maxDgramPackets > 1) {
3926 peer->maxDgramPackets = 1;
3928 } else if (np->length >=
3929 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3932 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3933 sizeof(afs_int32), &tSize);
3934 tSize = (afs_uint32) ntohl(tSize);
3936 * As of AFS 3.5 we set the send window to match the receive window.
3938 if (tSize < call->twind) {
3939 call->twind = tSize;
3940 call->conn->twind[call->channel] = call->twind;
3941 call->ssthresh = MIN(call->twind, call->ssthresh);
3942 } else if (tSize > call->twind) {
3943 call->twind = tSize;
3944 call->conn->twind[call->channel] = call->twind;
3948 * As of AFS 3.5, a jumbogram is more than one fixed size
3949 * packet transmitted in a single UDP datagram. If the remote
3950 * MTU is smaller than our local MTU then never send a datagram
3951 * larger than the natural MTU.
3954 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3955 sizeof(afs_int32), &tSize);
3956 maxDgramPackets = (afs_uint32) ntohl(tSize);
3957 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3958 maxDgramPackets = MIN(maxDgramPackets, peer->ifDgramPackets);
3959 if (peer->natMTU < peer->ifMTU)
3960 maxDgramPackets = MIN(maxDgramPackets, rxi_AdjustDgramPackets(1, peer->natMTU));
3961 if (maxDgramPackets > 1) {
3962 peer->maxDgramPackets = maxDgramPackets;
3963 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3965 peer->maxDgramPackets = 1;
3966 call->MTU = peer->natMTU;
3968 } else if (peer->maxDgramPackets > 1) {
3969 /* Restarted with lower version of RX */
3970 peer->maxDgramPackets = 1;
3972 } else if (peer->maxDgramPackets > 1
3973 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3974 /* Restarted with lower version of RX */
3975 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3976 peer->natMTU = OLD_MAX_PACKET_SIZE;
3977 peer->MTU = OLD_MAX_PACKET_SIZE;
3978 peer->maxDgramPackets = 1;
3979 peer->nDgramPackets = 1;
3981 call->MTU = OLD_MAX_PACKET_SIZE;
3986 * Calculate how many datagrams were successfully received after
3987 * the first missing packet and adjust the negative ack counter
3992 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3993 if (call->nNacks < nNacked) {
3994 call->nNacks = nNacked;
3997 call->nAcks += newAckCount;
4001 if (call->flags & RX_CALL_FAST_RECOVER) {
4003 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4005 call->flags &= ~RX_CALL_FAST_RECOVER;
4006 call->cwind = call->nextCwind;
4007 call->nextCwind = 0;
4010 call->nCwindAcks = 0;
4011 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4012 /* Three negative acks in a row trigger congestion recovery */
4013 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4014 MUTEX_EXIT(&peer->peer_lock);
4015 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4016 /* someone else is waiting to start recovery */
4019 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4020 rxi_WaitforTQBusy(call);
4021 MUTEX_ENTER(&peer->peer_lock);
4022 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4023 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4024 call->flags |= RX_CALL_FAST_RECOVER;
4025 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4027 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4028 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4029 call->nextCwind = call->ssthresh;
4032 peer->MTU = call->MTU;
4033 peer->cwind = call->nextCwind;
4034 peer->nDgramPackets = call->nDgramPackets;
4036 call->congestSeq = peer->congestSeq;
4037 /* Reset the resend times on the packets that were nacked
4038 * so we will retransmit as soon as the window permits*/
4039 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4041 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4042 clock_Zero(&tp->retryTime);
4044 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4049 /* If cwind is smaller than ssthresh, then increase
4050 * the window one packet for each ack we receive (exponential
4052 * If cwind is greater than or equal to ssthresh then increase
4053 * the congestion window by one packet for each cwind acks we
4054 * receive (linear growth). */
4055 if (call->cwind < call->ssthresh) {
4057 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4058 call->nCwindAcks = 0;
4060 call->nCwindAcks += newAckCount;
4061 if (call->nCwindAcks >= call->cwind) {
4062 call->nCwindAcks = 0;
4063 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4067 * If we have received several acknowledgements in a row then
4068 * it is time to increase the size of our datagrams
4070 if ((int)call->nAcks > rx_nDgramThreshold) {
4071 if (peer->maxDgramPackets > 1) {
4072 if (call->nDgramPackets < peer->maxDgramPackets) {
4073 call->nDgramPackets++;
4075 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4076 } else if (call->MTU < peer->maxMTU) {
4077 call->MTU += peer->natMTU;
4078 call->MTU = MIN(call->MTU, peer->maxMTU);
4084 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4086 /* Servers need to hold the call until all response packets have
4087 * been acknowledged. Soft acks are good enough since clients
4088 * are not allowed to clear their receive queues. */
4089 if (call->state == RX_STATE_HOLD
4090 && call->tfirst + call->nSoftAcked >= call->tnext) {
4091 call->state = RX_STATE_DALLY;
4092 rxi_ClearTransmitQueue(call, 0);
4093 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4094 } else if (!queue_IsEmpty(&call->tq)) {
4095 rxi_Start(0, call, 0, istack);
4100 /* Received a response to a challenge packet */
4102 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
4103 register struct rx_packet *np, int istack)
4107 /* Ignore the packet if we're the client */
4108 if (conn->type == RX_CLIENT_CONNECTION)
4111 /* If already authenticated, ignore the packet (it's probably a retry) */
4112 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4115 /* Otherwise, have the security object evaluate the response packet */
4116 error = RXS_CheckResponse(conn->securityObject, conn, np);
4118 /* If the response is invalid, reset the connection, sending
4119 * an abort to the peer */
4123 rxi_ConnectionError(conn, error);
4124 MUTEX_ENTER(&conn->conn_data_lock);
4125 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4126 MUTEX_EXIT(&conn->conn_data_lock);
4129 /* If the response is valid, any calls waiting to attach
4130 * servers can now do so */
4133 for (i = 0; i < RX_MAXCALLS; i++) {
4134 struct rx_call *call = conn->call[i];
4136 MUTEX_ENTER(&call->lock);
4137 if (call->state == RX_STATE_PRECALL)
4138 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4139 /* tnop can be null if newcallp is null */
4140 MUTEX_EXIT(&call->lock);
4144 /* Update the peer reachability information, just in case
4145 * some calls went into attach-wait while we were waiting
4146 * for authentication..
4148 rxi_UpdatePeerReach(conn, NULL);
4153 /* A client has received an authentication challenge: the security
4154 * object is asked to cough up a respectable response packet to send
4155 * back to the server. The server is responsible for retrying the
4156 * challenge if it fails to get a response. */
4159 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
4160 register struct rx_packet *np, int istack)
4164 /* Ignore the challenge if we're the server */
4165 if (conn->type == RX_SERVER_CONNECTION)
4168 /* Ignore the challenge if the connection is otherwise idle; someone's
4169 * trying to use us as an oracle. */
4170 if (!rxi_HasActiveCalls(conn))
4173 /* Send the security object the challenge packet. It is expected to fill
4174 * in the response. */
4175 error = RXS_GetResponse(conn->securityObject, conn, np);
4177 /* If the security object is unable to return a valid response, reset the
4178 * connection and send an abort to the peer. Otherwise send the response
4179 * packet to the peer connection. */
4181 rxi_ConnectionError(conn, error);
4182 MUTEX_ENTER(&conn->conn_data_lock);
4183 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4184 MUTEX_EXIT(&conn->conn_data_lock);
4186 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4187 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4193 /* Find an available server process to service the current request in
4194 * the given call structure. If one isn't available, queue up this
4195 * call so it eventually gets one */
4197 rxi_AttachServerProc(register struct rx_call *call,
4198 register osi_socket socket, register int *tnop,
4199 register struct rx_call **newcallp)
4201 register struct rx_serverQueueEntry *sq;
4202 register struct rx_service *service = call->conn->service;
4203 register int haveQuota = 0;
4205 /* May already be attached */
4206 if (call->state == RX_STATE_ACTIVE)
4209 MUTEX_ENTER(&rx_serverPool_lock);
4211 haveQuota = QuotaOK(service);
4212 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4213 /* If there are no processes available to service this call,
4214 * put the call on the incoming call queue (unless it's
4215 * already on the queue).
4217 #ifdef RX_ENABLE_LOCKS
4219 ReturnToServerPool(service);
4220 #endif /* RX_ENABLE_LOCKS */
4222 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4223 call->flags |= RX_CALL_WAIT_PROC;
4224 MUTEX_ENTER(&rx_stats_mutex);
4227 MUTEX_EXIT(&rx_stats_mutex);
4228 rxi_calltrace(RX_CALL_ARRIVAL, call);
4229 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4230 queue_Append(&rx_incomingCallQueue, call);
4233 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4235 /* If hot threads are enabled, and both newcallp and sq->socketp
4236 * are non-null, then this thread will process the call, and the
4237 * idle server thread will start listening on this threads socket.
4240 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4243 *sq->socketp = socket;
4244 clock_GetTime(&call->startTime);
4245 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4249 if (call->flags & RX_CALL_WAIT_PROC) {
4250 /* Conservative: I don't think this should happen */
4251 call->flags &= ~RX_CALL_WAIT_PROC;
4252 if (queue_IsOnQueue(call)) {
4254 MUTEX_ENTER(&rx_stats_mutex);
4256 MUTEX_EXIT(&rx_stats_mutex);
4259 call->state = RX_STATE_ACTIVE;
4260 call->mode = RX_MODE_RECEIVING;
4261 #ifdef RX_KERNEL_TRACE
4263 int glockOwner = ISAFS_GLOCK();
4266 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4267 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4273 if (call->flags & RX_CALL_CLEARED) {
4274 /* send an ack now to start the packet flow up again */
4275 call->flags &= ~RX_CALL_CLEARED;
4276 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4278 #ifdef RX_ENABLE_LOCKS
4281 service->nRequestsRunning++;
4282 if (service->nRequestsRunning <= service->minProcs)
4288 MUTEX_EXIT(&rx_serverPool_lock);
4291 /* Delay the sending of an acknowledge event for a short while, while
4292 * a new call is being prepared (in the case of a client) or a reply
4293 * is being prepared (in the case of a server). Rather than sending
4294 * an ack packet, an ACKALL packet is sent. */
4296 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4298 #ifdef RX_ENABLE_LOCKS
4300 MUTEX_ENTER(&call->lock);
4301 call->delayedAckEvent = NULL;
4302 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4304 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4305 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4307 MUTEX_EXIT(&call->lock);
4308 #else /* RX_ENABLE_LOCKS */
4310 call->delayedAckEvent = NULL;
4311 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4312 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4313 #endif /* RX_ENABLE_LOCKS */
4317 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused)
4319 struct rx_call *call = arg1;
4320 #ifdef RX_ENABLE_LOCKS
4322 MUTEX_ENTER(&call->lock);
4323 if (event == call->delayedAckEvent)
4324 call->delayedAckEvent = NULL;
4325 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4327 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4329 MUTEX_EXIT(&call->lock);
4330 #else /* RX_ENABLE_LOCKS */
4332 call->delayedAckEvent = NULL;
4333 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4334 #endif /* RX_ENABLE_LOCKS */
4338 #ifdef RX_ENABLE_LOCKS
4339 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4340 * clearing them out.
4343 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4345 register struct rx_packet *p, *tp;
4348 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4349 p->flags |= RX_PKTFLAG_ACKED;
4353 call->flags |= RX_CALL_TQ_CLEARME;
4354 call->flags |= RX_CALL_TQ_SOME_ACKED;
4357 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4358 call->tfirst = call->tnext;
4359 call->nSoftAcked = 0;
4361 if (call->flags & RX_CALL_FAST_RECOVER) {
4362 call->flags &= ~RX_CALL_FAST_RECOVER;
4363 call->cwind = call->nextCwind;
4364 call->nextCwind = 0;
4367 CV_SIGNAL(&call->cv_twind);
4369 #endif /* RX_ENABLE_LOCKS */
4371 /* Clear out the transmit queue for the current call (all packets have
4372 * been received by peer) */
4374 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4376 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4377 register struct rx_packet *p, *tp;
4379 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4381 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4382 p->flags |= RX_PKTFLAG_ACKED;
4386 call->flags |= RX_CALL_TQ_CLEARME;
4387 call->flags |= RX_CALL_TQ_SOME_ACKED;
4390 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4391 rxi_FreePackets(0, &call->tq);
4392 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4393 call->flags &= ~RX_CALL_TQ_CLEARME;
4395 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4397 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4398 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4399 call->nSoftAcked = 0;
4401 if (call->flags & RX_CALL_FAST_RECOVER) {
4402 call->flags &= ~RX_CALL_FAST_RECOVER;
4403 call->cwind = call->nextCwind;
4405 #ifdef RX_ENABLE_LOCKS
4406 CV_SIGNAL(&call->cv_twind);
4408 osi_rxWakeup(&call->twind);
4413 rxi_ClearReceiveQueue(register struct rx_call *call)
4415 if (queue_IsNotEmpty(&call->rq)) {
4416 rx_packetReclaims += rxi_FreePackets(0, &call->rq);
4417 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4419 if (call->state == RX_STATE_PRECALL) {
4420 call->flags |= RX_CALL_CLEARED;
4424 /* Send an abort packet for the specified call */
4426 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4427 int istack, int force)
4430 struct clock when, now;
4435 /* Clients should never delay abort messages */
4436 if (rx_IsClientConn(call->conn))
4439 if (call->abortCode != call->error) {
4440 call->abortCode = call->error;
4441 call->abortCount = 0;
4444 if (force || rxi_callAbortThreshhold == 0
4445 || call->abortCount < rxi_callAbortThreshhold) {
4446 if (call->delayedAbortEvent) {
4447 rxevent_Cancel(call->delayedAbortEvent, call,
4448 RX_CALL_REFCOUNT_ABORT);
4450 error = htonl(call->error);
4453 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4454 (char *)&error, sizeof(error), istack);
4455 } else if (!call->delayedAbortEvent) {
4456 clock_GetTime(&now);
4458 clock_Addmsec(&when, rxi_callAbortDelay);
4459 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4460 call->delayedAbortEvent =
4461 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4466 /* Send an abort packet for the specified connection. Packet is an
4467 * optional pointer to a packet that can be used to send the abort.
4468 * Once the number of abort messages reaches the threshhold, an
4469 * event is scheduled to send the abort. Setting the force flag
4470 * overrides sending delayed abort messages.
4472 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4473 * to send the abort packet.
4476 rxi_SendConnectionAbort(register struct rx_connection *conn,
4477 struct rx_packet *packet, int istack, int force)
4480 struct clock when, now;
4485 /* Clients should never delay abort messages */
4486 if (rx_IsClientConn(conn))
4489 if (force || rxi_connAbortThreshhold == 0
4490 || conn->abortCount < rxi_connAbortThreshhold) {
4491 if (conn->delayedAbortEvent) {
4492 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4494 error = htonl(conn->error);
4496 MUTEX_EXIT(&conn->conn_data_lock);
4498 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4499 RX_PACKET_TYPE_ABORT, (char *)&error,
4500 sizeof(error), istack);
4501 MUTEX_ENTER(&conn->conn_data_lock);
4502 } else if (!conn->delayedAbortEvent) {
4503 clock_GetTime(&now);
4505 clock_Addmsec(&when, rxi_connAbortDelay);
4506 conn->delayedAbortEvent =
4507 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4512 /* Associate an error all of the calls owned by a connection. Called
4513 * with error non-zero. This is only for really fatal things, like
4514 * bad authentication responses. The connection itself is set in
4515 * error at this point, so that future packets received will be
4518 rxi_ConnectionError(register struct rx_connection *conn,
4519 register afs_int32 error)
4524 dpf(("rxi_ConnectionError conn %x error %d", conn, error));
4526 MUTEX_ENTER(&conn->conn_data_lock);
4527 if (conn->challengeEvent)
4528 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4529 if (conn->checkReachEvent) {
4530 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4531 conn->checkReachEvent = 0;
4532 conn->flags &= ~RX_CONN_ATTACHWAIT;
4535 MUTEX_EXIT(&conn->conn_data_lock);
4536 for (i = 0; i < RX_MAXCALLS; i++) {
4537 struct rx_call *call = conn->call[i];
4539 MUTEX_ENTER(&call->lock);
4540 rxi_CallError(call, error);
4541 MUTEX_EXIT(&call->lock);
4544 conn->error = error;
4545 rx_MutexIncrement(rx_stats.fatalErrors, rx_stats_mutex);
4550 rxi_CallError(register struct rx_call *call, afs_int32 error)
4552 dpf(("rxi_CallError call %x error %d call->error %d", call, error, call->error));
4554 error = call->error;
4556 #ifdef RX_GLOBAL_RXLOCK_KERNEL
4557 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4558 rxi_ResetCall(call, 0);
4561 rxi_ResetCall(call, 0);
4563 call->error = error;
4564 call->mode = RX_MODE_ERROR;
4567 /* Reset various fields in a call structure, and wakeup waiting
4568 * processes. Some fields aren't changed: state & mode are not
4569 * touched (these must be set by the caller), and bufptr, nLeft, and
4570 * nFree are not reset, since these fields are manipulated by
4571 * unprotected macros, and may only be reset by non-interrupting code.
4574 /* this code requires that call->conn be set properly as a pre-condition. */
4575 #endif /* ADAPT_WINDOW */
4578 rxi_ResetCall(register struct rx_call *call, register int newcall)
4581 register struct rx_peer *peer;
4582 struct rx_packet *packet;
4584 dpf(("rxi_ResetCall(call %x, newcall %d)\n", call, newcall));
4586 /* Notify anyone who is waiting for asynchronous packet arrival */
4587 if (call->arrivalProc) {
4588 (*call->arrivalProc) (call, call->arrivalProcHandle,
4589 call->arrivalProcArg);
4590 call->arrivalProc = (void (*)())0;
4593 if (call->delayedAbortEvent) {
4594 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4595 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4597 rxi_SendCallAbort(call, packet, 0, 1);
4598 rxi_FreePacket(packet);
4603 * Update the peer with the congestion information in this call
4604 * so other calls on this connection can pick up where this call
4605 * left off. If the congestion sequence numbers don't match then
4606 * another call experienced a retransmission.
4608 peer = call->conn->peer;
4609 MUTEX_ENTER(&peer->peer_lock);
4611 if (call->congestSeq == peer->congestSeq) {
4612 peer->cwind = MAX(peer->cwind, call->cwind);
4613 peer->MTU = MAX(peer->MTU, call->MTU);
4614 peer->nDgramPackets =
4615 MAX(peer->nDgramPackets, call->nDgramPackets);
4618 call->abortCode = 0;
4619 call->abortCount = 0;
4621 if (peer->maxDgramPackets > 1) {
4622 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4624 call->MTU = peer->MTU;
4626 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4627 call->ssthresh = rx_maxSendWindow;
4628 call->nDgramPackets = peer->nDgramPackets;
4629 call->congestSeq = peer->congestSeq;
4630 MUTEX_EXIT(&peer->peer_lock);
4632 flags = call->flags;
4633 rxi_ClearReceiveQueue(call);
4634 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4635 if (flags & RX_CALL_TQ_BUSY) {
4636 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4637 call->flags |= (flags & RX_CALL_TQ_WAIT);
4639 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4641 rxi_ClearTransmitQueue(call, 0);
4642 queue_Init(&call->tq);
4643 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4644 dpf(("rcall %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4647 while (call->tqWaiters) {
4648 #ifdef RX_ENABLE_LOCKS
4649 CV_BROADCAST(&call->cv_tq);
4650 #else /* RX_ENABLE_LOCKS */
4651 osi_rxWakeup(&call->tq);
4652 #endif /* RX_ENABLE_LOCKS */
4656 queue_Init(&call->rq);
4658 call->twind = call->conn->twind[call->channel];
4659 call->rwind = call->conn->rwind[call->channel];
4660 call->nSoftAcked = 0;
4661 call->nextCwind = 0;
4664 call->nCwindAcks = 0;
4665 call->nSoftAcks = 0;
4666 call->nHardAcks = 0;
4668 call->tfirst = call->rnext = call->tnext = 1;
4670 call->lastAcked = 0;
4671 call->localStatus = call->remoteStatus = 0;
4673 if (flags & RX_CALL_READER_WAIT) {
4674 #ifdef RX_ENABLE_LOCKS
4675 CV_BROADCAST(&call->cv_rq);
4677 osi_rxWakeup(&call->rq);
4680 if (flags & RX_CALL_WAIT_PACKETS) {
4681 MUTEX_ENTER(&rx_freePktQ_lock);
4682 rxi_PacketsUnWait(); /* XXX */
4683 MUTEX_EXIT(&rx_freePktQ_lock);
4685 #ifdef RX_ENABLE_LOCKS
4686 CV_SIGNAL(&call->cv_twind);
4688 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4689 osi_rxWakeup(&call->twind);
4692 #ifdef RX_ENABLE_LOCKS
4693 /* The following ensures that we don't mess with any queue while some
4694 * other thread might also be doing so. The call_queue_lock field is
4695 * is only modified under the call lock. If the call is in the process
4696 * of being removed from a queue, the call is not locked until the
4697 * the queue lock is dropped and only then is the call_queue_lock field
4698 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4699 * Note that any other routine which removes a call from a queue has to
4700 * obtain the queue lock before examing the queue and removing the call.
4702 if (call->call_queue_lock) {
4703 MUTEX_ENTER(call->call_queue_lock);
4704 if (queue_IsOnQueue(call)) {
4706 if (flags & RX_CALL_WAIT_PROC) {
4707 MUTEX_ENTER(&rx_stats_mutex);
4709 MUTEX_EXIT(&rx_stats_mutex);
4712 MUTEX_EXIT(call->call_queue_lock);
4713 CLEAR_CALL_QUEUE_LOCK(call);
4715 #else /* RX_ENABLE_LOCKS */
4716 if (queue_IsOnQueue(call)) {
4718 if (flags & RX_CALL_WAIT_PROC)
4721 #endif /* RX_ENABLE_LOCKS */
4723 rxi_KeepAliveOff(call);
4724 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4727 /* Send an acknowledge for the indicated packet (seq,serial) of the
4728 * indicated call, for the indicated reason (reason). This
4729 * acknowledge will specifically acknowledge receiving the packet, and
4730 * will also specify which other packets for this call have been
4731 * received. This routine returns the packet that was used to the
4732 * caller. The caller is responsible for freeing it or re-using it.
4733 * This acknowledgement also returns the highest sequence number
4734 * actually read out by the higher level to the sender; the sender
4735 * promises to keep around packets that have not been read by the
4736 * higher level yet (unless, of course, the sender decides to abort
4737 * the call altogether). Any of p, seq, serial, pflags, or reason may
4738 * be set to zero without ill effect. That is, if they are zero, they
4739 * will not convey any information.
4740 * NOW there is a trailer field, after the ack where it will safely be
4741 * ignored by mundanes, which indicates the maximum size packet this
4742 * host can swallow. */
4744 register struct rx_packet *optionalPacket; use to send ack (or null)
4745 int seq; Sequence number of the packet we are acking
4746 int serial; Serial number of the packet
4747 int pflags; Flags field from packet header
4748 int reason; Reason an acknowledge was prompted
4752 rxi_SendAck(register struct rx_call *call,
4753 register struct rx_packet *optionalPacket, int serial, int reason,
4756 struct rx_ackPacket *ap;
4757 register struct rx_packet *rqp;
4758 register struct rx_packet *nxp; /* For queue_Scan */
4759 register struct rx_packet *p;
4762 #ifdef RX_ENABLE_TSFPQ
4763 struct rx_ts_info_t * rx_ts_info;
4767 * Open the receive window once a thread starts reading packets
4769 if (call->rnext > 1) {
4770 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
4773 call->nHardAcks = 0;
4774 call->nSoftAcks = 0;
4775 if (call->rnext > call->lastAcked)
4776 call->lastAcked = call->rnext;
4780 rx_computelen(p, p->length); /* reset length, you never know */
4781 } /* where that's been... */
4782 #ifdef RX_ENABLE_TSFPQ
4784 RX_TS_INFO_GET(rx_ts_info);
4785 if ((p = rx_ts_info->local_special_packet)) {
4786 rx_computelen(p, p->length);
4787 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4788 rx_ts_info->local_special_packet = p;
4789 } else { /* We won't send the ack, but don't panic. */
4790 return optionalPacket;
4794 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4795 /* We won't send the ack, but don't panic. */
4796 return optionalPacket;
4801 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4804 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4805 #ifndef RX_ENABLE_TSFPQ
4806 if (!optionalPacket)
4809 return optionalPacket;
4811 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4812 if (rx_Contiguous(p) < templ) {
4813 #ifndef RX_ENABLE_TSFPQ
4814 if (!optionalPacket)
4817 return optionalPacket;
4822 /* MTUXXX failing to send an ack is very serious. We should */
4823 /* try as hard as possible to send even a partial ack; it's */
4824 /* better than nothing. */
4825 ap = (struct rx_ackPacket *)rx_DataOf(p);
4826 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4827 ap->reason = reason;
4829 /* The skew computation used to be bogus, I think it's better now. */
4830 /* We should start paying attention to skew. XXX */
4831 ap->serial = htonl(serial);
4832 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4834 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4835 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4837 /* No fear of running out of ack packet here because there can only be at most
4838 * one window full of unacknowledged packets. The window size must be constrained
4839 * to be less than the maximum ack size, of course. Also, an ack should always
4840 * fit into a single packet -- it should not ever be fragmented. */
4841 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4842 if (!rqp || !call->rq.next
4843 || (rqp->header.seq > (call->rnext + call->rwind))) {
4844 #ifndef RX_ENABLE_TSFPQ
4845 if (!optionalPacket)
4848 rxi_CallError(call, RX_CALL_DEAD);
4849 return optionalPacket;
4852 while (rqp->header.seq > call->rnext + offset)
4853 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4854 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4856 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4857 #ifndef RX_ENABLE_TSFPQ
4858 if (!optionalPacket)
4861 rxi_CallError(call, RX_CALL_DEAD);
4862 return optionalPacket;
4867 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4869 /* these are new for AFS 3.3 */
4870 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4871 templ = htonl(templ);
4872 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4873 templ = htonl(call->conn->peer->ifMTU);
4874 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4875 sizeof(afs_int32), &templ);
4877 /* new for AFS 3.4 */
4878 templ = htonl(call->rwind);
4879 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4880 sizeof(afs_int32), &templ);
4882 /* new for AFS 3.5 */
4883 templ = htonl(call->conn->peer->ifDgramPackets);
4884 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4885 sizeof(afs_int32), &templ);
4887 p->header.serviceId = call->conn->serviceId;
4888 p->header.cid = (call->conn->cid | call->channel);
4889 p->header.callNumber = *call->callNumber;
4891 p->header.securityIndex = call->conn->securityIndex;
4892 p->header.epoch = call->conn->epoch;
4893 p->header.type = RX_PACKET_TYPE_ACK;
4894 p->header.flags = RX_SLOW_START_OK;
4895 if (reason == RX_ACK_PING) {
4896 p->header.flags |= RX_REQUEST_ACK;
4898 clock_GetTime(&call->pingRequestTime);
4901 if (call->conn->type == RX_CLIENT_CONNECTION)
4902 p->header.flags |= RX_CLIENT_INITIATED;
4906 if (rxdebug_active) {
4910 len = _snprintf(msg, sizeof(msg),
4911 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4912 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4913 ntohl(ap->serial), ntohl(ap->previousPacket),
4914 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
4915 ap->nAcks, ntohs(ap->bufferSpace) );
4919 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
4920 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4924 OutputDebugString(msg);
4926 #else /* AFS_NT40_ENV */
4928 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
4929 ap->reason, ntohl(ap->previousPacket),
4930 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4932 for (offset = 0; offset < ap->nAcks; offset++)
4933 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4938 #endif /* AFS_NT40_ENV */
4941 register int i, nbytes = p->length;
4943 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4944 if (nbytes <= p->wirevec[i].iov_len) {
4945 register int savelen, saven;
4947 savelen = p->wirevec[i].iov_len;
4949 p->wirevec[i].iov_len = nbytes;
4951 rxi_Send(call, p, istack);
4952 p->wirevec[i].iov_len = savelen;
4956 nbytes -= p->wirevec[i].iov_len;
4959 rx_MutexIncrement(rx_stats.ackPacketsSent, rx_stats_mutex);
4960 #ifndef RX_ENABLE_TSFPQ
4961 if (!optionalPacket)
4964 return optionalPacket; /* Return packet for re-use by caller */
4967 /* Send all of the packets in the list in single datagram */
4969 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4970 int istack, int moreFlag, struct clock *now,
4971 struct clock *retryTime, int resending)
4976 struct rx_connection *conn = call->conn;
4977 struct rx_peer *peer = conn->peer;
4979 MUTEX_ENTER(&peer->peer_lock);
4982 peer->reSends += len;
4983 rx_MutexIncrement(rx_stats.dataPacketsSent, rx_stats_mutex);
4984 MUTEX_EXIT(&peer->peer_lock);
4986 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4990 /* Set the packet flags and schedule the resend events */
4991 /* Only request an ack for the last packet in the list */
4992 for (i = 0; i < len; i++) {
4993 list[i]->retryTime = *retryTime;
4994 if (list[i]->header.serial) {
4995 /* Exponentially backoff retry times */
4996 if (list[i]->backoff < MAXBACKOFF) {
4997 /* so it can't stay == 0 */
4998 list[i]->backoff = (list[i]->backoff << 1) + 1;
5001 clock_Addmsec(&(list[i]->retryTime),
5002 ((afs_uint32) list[i]->backoff) << 8);
5005 /* Wait a little extra for the ack on the last packet */
5006 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
5007 clock_Addmsec(&(list[i]->retryTime), 400);
5010 /* Record the time sent */
5011 list[i]->timeSent = *now;
5013 /* Ask for an ack on retransmitted packets, on every other packet
5014 * if the peer doesn't support slow start. Ask for an ack on every
5015 * packet until the congestion window reaches the ack rate. */
5016 if (list[i]->header.serial) {
5018 rx_MutexIncrement(rx_stats.dataPacketsReSent, rx_stats_mutex);
5020 /* improved RTO calculation- not Karn */
5021 list[i]->firstSent = *now;
5022 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5023 || (!(call->flags & RX_CALL_SLOW_START_OK)
5024 && (list[i]->header.seq & 1)))) {
5029 MUTEX_ENTER(&peer->peer_lock);
5033 rx_MutexIncrement(rx_stats.dataPacketsSent, rx_stats_mutex);
5034 MUTEX_EXIT(&peer->peer_lock);
5036 /* Tag this packet as not being the last in this group,
5037 * for the receiver's benefit */
5038 if (i < len - 1 || moreFlag) {
5039 list[i]->header.flags |= RX_MORE_PACKETS;
5042 /* Install the new retransmit time for the packet, and
5043 * record the time sent */
5044 list[i]->timeSent = *now;
5048 list[len - 1]->header.flags |= RX_REQUEST_ACK;
5051 /* Since we're about to send a data packet to the peer, it's
5052 * safe to nuke any scheduled end-of-packets ack */
5053 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5055 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5056 MUTEX_EXIT(&call->lock);
5058 rxi_SendPacketList(call, conn, list, len, istack);
5060 rxi_SendPacket(call, conn, list[0], istack);
5062 MUTEX_ENTER(&call->lock);
5063 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5065 /* Update last send time for this call (for keep-alive
5066 * processing), and for the connection (so that we can discover
5067 * idle connections) */
5068 call->lastSendData = conn->lastSendTime = call->lastSendTime = clock_Sec();
5071 /* When sending packets we need to follow these rules:
5072 * 1. Never send more than maxDgramPackets in a jumbogram.
5073 * 2. Never send a packet with more than two iovecs in a jumbogram.
5074 * 3. Never send a retransmitted packet in a jumbogram.
5075 * 4. Never send more than cwind/4 packets in a jumbogram
5076 * We always keep the last list we should have sent so we
5077 * can set the RX_MORE_PACKETS flags correctly.
5080 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5081 int istack, struct clock *now, struct clock *retryTime,
5084 int i, cnt, lastCnt = 0;
5085 struct rx_packet **listP, **lastP = 0;
5086 struct rx_peer *peer = call->conn->peer;
5087 int morePackets = 0;
5089 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
5090 /* Does the current packet force us to flush the current list? */
5092 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5093 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5095 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
5097 /* If the call enters an error state stop sending, or if
5098 * we entered congestion recovery mode, stop sending */
5099 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5107 /* Add the current packet to the list if it hasn't been acked.
5108 * Otherwise adjust the list pointer to skip the current packet. */
5109 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5111 /* Do we need to flush the list? */
5112 if (cnt >= (int)peer->maxDgramPackets
5113 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5114 || list[i]->header.serial
5115 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5117 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5118 retryTime, resending);
5119 /* If the call enters an error state stop sending, or if
5120 * we entered congestion recovery mode, stop sending */
5122 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5127 listP = &list[i + 1];
5132 osi_Panic("rxi_SendList error");
5134 listP = &list[i + 1];
5138 /* Send the whole list when the call is in receive mode, when
5139 * the call is in eof mode, when we are in fast recovery mode,
5140 * and when we have the last packet */
5141 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5142 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5143 || (call->flags & RX_CALL_FAST_RECOVER)) {
5144 /* Check for the case where the current list contains
5145 * an acked packet. Since we always send retransmissions
5146 * in a separate packet, we only need to check the first
5147 * packet in the list */
5148 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5152 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5153 retryTime, resending);
5154 /* If the call enters an error state stop sending, or if
5155 * we entered congestion recovery mode, stop sending */
5156 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5160 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5163 } else if (lastCnt > 0) {
5164 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5169 #ifdef RX_ENABLE_LOCKS
5170 /* Call rxi_Start, below, but with the call lock held. */
5172 rxi_StartUnlocked(struct rxevent *event,
5173 void *arg0, void *arg1, int istack)
5175 struct rx_call *call = arg0;
5177 MUTEX_ENTER(&call->lock);
5178 rxi_Start(event, call, arg1, istack);
5179 MUTEX_EXIT(&call->lock);
5181 #endif /* RX_ENABLE_LOCKS */
5183 /* This routine is called when new packets are readied for
5184 * transmission and when retransmission may be necessary, or when the
5185 * transmission window or burst count are favourable. This should be
5186 * better optimized for new packets, the usual case, now that we've
5187 * got rid of queues of send packets. XXXXXXXXXXX */
5189 rxi_Start(struct rxevent *event,
5190 void *arg0, void *arg1, int istack)
5192 struct rx_call *call = arg0;
5194 struct rx_packet *p;
5195 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
5196 struct rx_peer *peer = call->conn->peer;
5197 struct clock now, usenow, retryTime;
5201 struct rx_packet **xmitList;
5204 /* If rxi_Start is being called as a result of a resend event,
5205 * then make sure that the event pointer is removed from the call
5206 * structure, since there is no longer a per-call retransmission
5208 if (event && event == call->resendEvent) {
5209 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5210 call->resendEvent = NULL;
5212 if (queue_IsEmpty(&call->tq)) {
5216 /* Timeouts trigger congestion recovery */
5217 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5218 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5219 /* someone else is waiting to start recovery */
5222 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5223 rxi_WaitforTQBusy(call);
5224 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5225 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5226 call->flags |= RX_CALL_FAST_RECOVER;
5227 if (peer->maxDgramPackets > 1) {
5228 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5230 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5232 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5233 call->nDgramPackets = 1;
5235 call->nextCwind = 1;
5238 MUTEX_ENTER(&peer->peer_lock);
5239 peer->MTU = call->MTU;
5240 peer->cwind = call->cwind;
5241 peer->nDgramPackets = 1;
5243 call->congestSeq = peer->congestSeq;
5244 MUTEX_EXIT(&peer->peer_lock);
5245 /* Clear retry times on packets. Otherwise, it's possible for
5246 * some packets in the queue to force resends at rates faster
5247 * than recovery rates.
5249 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5250 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5251 clock_Zero(&p->retryTime);
5256 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5257 rx_MutexIncrement(rx_tq_debug.rxi_start_in_error, rx_stats_mutex);
5262 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5263 /* Get clock to compute the re-transmit time for any packets
5264 * in this burst. Note, if we back off, it's reasonable to
5265 * back off all of the packets in the same manner, even if
5266 * some of them have been retransmitted more times than more
5268 * Do a dance to avoid blocking after setting now. */
5269 clock_Zero(&retryTime);
5270 MUTEX_ENTER(&peer->peer_lock);
5271 clock_Add(&retryTime, &peer->timeout);
5272 MUTEX_EXIT(&peer->peer_lock);
5273 clock_GetTime(&now);
5274 clock_Add(&retryTime, &now);
5276 /* Send (or resend) any packets that need it, subject to
5277 * window restrictions and congestion burst control
5278 * restrictions. Ask for an ack on the last packet sent in
5279 * this burst. For now, we're relying upon the window being
5280 * considerably bigger than the largest number of packets that
5281 * are typically sent at once by one initial call to
5282 * rxi_Start. This is probably bogus (perhaps we should ask
5283 * for an ack when we're half way through the current
5284 * window?). Also, for non file transfer applications, this
5285 * may end up asking for an ack for every packet. Bogus. XXXX
5288 * But check whether we're here recursively, and let the other guy
5291 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5292 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5293 call->flags |= RX_CALL_TQ_BUSY;
5295 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5297 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5298 call->flags &= ~RX_CALL_NEED_START;
5299 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5301 maxXmitPackets = MIN(call->twind, call->cwind);
5302 xmitList = (struct rx_packet **)
5303 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
5304 /* XXXX else we must drop any mtx we hold */
5305 afs_osi_Alloc_NoSleep(maxXmitPackets * sizeof(struct rx_packet *));
5307 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5309 if (xmitList == NULL)
5310 osi_Panic("rxi_Start, failed to allocate xmit list");
5311 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5312 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5313 /* We shouldn't be sending packets if a thread is waiting
5314 * to initiate congestion recovery */
5318 && (call->flags & RX_CALL_FAST_RECOVER)) {
5319 /* Only send one packet during fast recovery */
5322 if ((p->flags & RX_PKTFLAG_FREE)
5323 || (!queue_IsEnd(&call->tq, nxp)
5324 && (nxp->flags & RX_PKTFLAG_FREE))
5325 || (p == (struct rx_packet *)&rx_freePacketQueue)
5326 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5327 osi_Panic("rxi_Start: xmit queue clobbered");
5329 if (p->flags & RX_PKTFLAG_ACKED) {
5330 /* Since we may block, don't trust this */
5331 usenow.sec = usenow.usec = 0;
5332 rx_MutexIncrement(rx_stats.ignoreAckedPacket, rx_stats_mutex);
5333 continue; /* Ignore this packet if it has been acknowledged */
5336 /* Turn off all flags except these ones, which are the same
5337 * on each transmission */
5338 p->header.flags &= RX_PRESET_FLAGS;
5340 if (p->header.seq >=
5341 call->tfirst + MIN((int)call->twind,
5342 (int)(call->nSoftAcked +
5344 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5345 /* Note: if we're waiting for more window space, we can
5346 * still send retransmits; hence we don't return here, but
5347 * break out to schedule a retransmit event */
5348 dpf(("call %d waiting for window",
5349 *(call->callNumber)));
5353 /* Transmit the packet if it needs to be sent. */
5354 if (!clock_Lt(&now, &p->retryTime)) {
5355 if (nXmitPackets == maxXmitPackets) {
5356 rxi_SendXmitList(call, xmitList, nXmitPackets,
5357 istack, &now, &retryTime,
5359 osi_Free(xmitList, maxXmitPackets *
5360 sizeof(struct rx_packet *));
5363 xmitList[nXmitPackets++] = p;
5367 /* xmitList now hold pointers to all of the packets that are
5368 * ready to send. Now we loop to send the packets */
5369 if (nXmitPackets > 0) {
5370 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5371 &now, &retryTime, resending);
5374 maxXmitPackets * sizeof(struct rx_packet *));
5376 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5378 * TQ references no longer protected by this flag; they must remain
5379 * protected by the global lock.
5381 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5382 call->flags &= ~RX_CALL_TQ_BUSY;
5383 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5384 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5385 #ifdef RX_ENABLE_LOCKS
5386 osirx_AssertMine(&call->lock, "rxi_Start start");
5387 CV_BROADCAST(&call->cv_tq);
5388 #else /* RX_ENABLE_LOCKS */
5389 osi_rxWakeup(&call->tq);
5390 #endif /* RX_ENABLE_LOCKS */
5395 /* We went into the error state while sending packets. Now is
5396 * the time to reset the call. This will also inform the using
5397 * process that the call is in an error state.
5399 rx_MutexIncrement(rx_tq_debug.rxi_start_aborted, rx_stats_mutex);
5400 call->flags &= ~RX_CALL_TQ_BUSY;
5401 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5402 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5403 #ifdef RX_ENABLE_LOCKS
5404 osirx_AssertMine(&call->lock, "rxi_Start middle");
5405 CV_BROADCAST(&call->cv_tq);
5406 #else /* RX_ENABLE_LOCKS */
5407 osi_rxWakeup(&call->tq);
5408 #endif /* RX_ENABLE_LOCKS */
5410 rxi_CallError(call, call->error);
5413 #ifdef RX_ENABLE_LOCKS
5414 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5415 register int missing;
5416 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5417 /* Some packets have received acks. If they all have, we can clear
5418 * the transmit queue.
5421 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5422 if (p->header.seq < call->tfirst
5423 && (p->flags & RX_PKTFLAG_ACKED)) {
5430 call->flags |= RX_CALL_TQ_CLEARME;
5432 #endif /* RX_ENABLE_LOCKS */
5433 /* Don't bother doing retransmits if the TQ is cleared. */
5434 if (call->flags & RX_CALL_TQ_CLEARME) {
5435 rxi_ClearTransmitQueue(call, 1);
5437 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5440 /* Always post a resend event, if there is anything in the
5441 * queue, and resend is possible. There should be at least
5442 * one unacknowledged packet in the queue ... otherwise none
5443 * of these packets should be on the queue in the first place.
5445 if (call->resendEvent) {
5446 /* Cancel the existing event and post a new one */
5447 rxevent_Cancel(call->resendEvent, call,
5448 RX_CALL_REFCOUNT_RESEND);
5451 /* The retry time is the retry time on the first unacknowledged
5452 * packet inside the current window */
5454 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5455 /* Don't set timers for packets outside the window */
5456 if (p->header.seq >= call->tfirst + call->twind) {
5460 if (!(p->flags & RX_PKTFLAG_ACKED)
5461 && !clock_IsZero(&p->retryTime)) {
5463 retryTime = p->retryTime;
5468 /* Post a new event to re-run rxi_Start when retries may be needed */
5469 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5470 #ifdef RX_ENABLE_LOCKS
5471 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5473 rxevent_PostNow2(&retryTime, &usenow,
5475 (void *)call, 0, istack);
5476 #else /* RX_ENABLE_LOCKS */
5478 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5479 (void *)call, 0, istack);
5480 #endif /* RX_ENABLE_LOCKS */
5483 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5484 } while (call->flags & RX_CALL_NEED_START);
5486 * TQ references no longer protected by this flag; they must remain
5487 * protected by the global lock.
5489 call->flags &= ~RX_CALL_TQ_BUSY;
5490 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5491 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5492 #ifdef RX_ENABLE_LOCKS
5493 osirx_AssertMine(&call->lock, "rxi_Start end");
5494 CV_BROADCAST(&call->cv_tq);
5495 #else /* RX_ENABLE_LOCKS */
5496 osi_rxWakeup(&call->tq);
5497 #endif /* RX_ENABLE_LOCKS */
5500 call->flags |= RX_CALL_NEED_START;
5502 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5504 if (call->resendEvent) {
5505 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5510 /* Also adjusts the keep alive parameters for the call, to reflect
5511 * that we have just sent a packet (so keep alives aren't sent
5514 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5517 register struct rx_connection *conn = call->conn;
5519 /* Stamp each packet with the user supplied status */
5520 p->header.userStatus = call->localStatus;
5522 /* Allow the security object controlling this call's security to
5523 * make any last-minute changes to the packet */
5524 RXS_SendPacket(conn->securityObject, call, p);
5526 /* Since we're about to send SOME sort of packet to the peer, it's
5527 * safe to nuke any scheduled end-of-packets ack */
5528 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5530 /* Actually send the packet, filling in more connection-specific fields */
5531 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5532 MUTEX_EXIT(&call->lock);
5533 rxi_SendPacket(call, conn, p, istack);
5534 MUTEX_ENTER(&call->lock);
5535 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5537 /* Update last send time for this call (for keep-alive
5538 * processing), and for the connection (so that we can discover
5539 * idle connections) */
5540 conn->lastSendTime = call->lastSendTime = clock_Sec();
5541 /* Don't count keepalives here, so idleness can be tracked. */
5542 if (p->header.type != RX_PACKET_TYPE_ACK)
5543 call->lastSendData = call->lastSendTime;
5547 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5548 * that things are fine. Also called periodically to guarantee that nothing
5549 * falls through the cracks (e.g. (error + dally) connections have keepalive
5550 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5552 * haveCTLock Set if calling from rxi_ReapConnections
5554 #ifdef RX_ENABLE_LOCKS
5556 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5557 #else /* RX_ENABLE_LOCKS */
5559 rxi_CheckCall(register struct rx_call *call)
5560 #endif /* RX_ENABLE_LOCKS */
5562 register struct rx_connection *conn = call->conn;
5564 afs_uint32 deadTime;
5566 #ifdef RX_GLOBAL_RXLOCK_KERNEL
5567 if (call->flags & RX_CALL_TQ_BUSY) {
5568 /* Call is active and will be reset by rxi_Start if it's
5569 * in an error state.
5574 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5576 (((afs_uint32) conn->secondsUntilDead << 10) +
5577 ((afs_uint32) conn->peer->rtt >> 3) +
5578 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5580 /* These are computed to the second (+- 1 second). But that's
5581 * good enough for these values, which should be a significant
5582 * number of seconds. */
5583 if (now > (call->lastReceiveTime + deadTime)) {
5584 if (call->state == RX_STATE_ACTIVE) {
5586 #if defined(KERNEL) && defined(AFS_SUN57_ENV)
5588 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5589 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
5590 ip_stack_t *ipst = ns->netstack_ip;
5592 ire = ire_cache_lookup(call->conn->peer->host
5593 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
5595 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
5597 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5604 if (ire && ire->ire_max_frag > 0)
5605 rxi_SetPeerMtu(call->conn->peer->host, 0, ire->ire_max_frag);
5606 #if defined(GLOBAL_NETSTACKID)
5610 #endif /* ADAPT_PMTU */
5611 rxi_CallError(call, RX_CALL_DEAD);
5614 #ifdef RX_ENABLE_LOCKS
5615 /* Cancel pending events */
5616 rxevent_Cancel(call->delayedAckEvent, call,
5617 RX_CALL_REFCOUNT_DELAY);
5618 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5619 rxevent_Cancel(call->keepAliveEvent, call,
5620 RX_CALL_REFCOUNT_ALIVE);
5621 if (call->refCount == 0) {
5622 rxi_FreeCall(call, haveCTLock);
5626 #else /* RX_ENABLE_LOCKS */
5629 #endif /* RX_ENABLE_LOCKS */
5631 /* Non-active calls are destroyed if they are not responding
5632 * to pings; active calls are simply flagged in error, so the
5633 * attached process can die reasonably gracefully. */
5635 /* see if we have a non-activity timeout */
5636 if (call->startWait && conn->idleDeadTime
5637 && ((call->startWait + conn->idleDeadTime) < now)) {
5638 if (call->state == RX_STATE_ACTIVE) {
5639 rxi_CallError(call, RX_CALL_TIMEOUT);
5643 if (call->lastSendData && conn->idleDeadTime && (conn->idleDeadErr != 0)
5644 && ((call->lastSendData + conn->idleDeadTime) < now)) {
5645 if (call->state == RX_STATE_ACTIVE) {
5646 rxi_CallError(call, conn->idleDeadErr);
5650 /* see if we have a hard timeout */
5651 if (conn->hardDeadTime
5652 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5653 if (call->state == RX_STATE_ACTIVE)
5654 rxi_CallError(call, RX_CALL_TIMEOUT);
5661 /* When a call is in progress, this routine is called occasionally to
5662 * make sure that some traffic has arrived (or been sent to) the peer.
5663 * If nothing has arrived in a reasonable amount of time, the call is
5664 * declared dead; if nothing has been sent for a while, we send a
5665 * keep-alive packet (if we're actually trying to keep the call alive)
5668 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
5670 struct rx_call *call = arg1;
5671 struct rx_connection *conn;
5674 MUTEX_ENTER(&call->lock);
5675 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5676 if (event == call->keepAliveEvent)
5677 call->keepAliveEvent = NULL;
5680 #ifdef RX_ENABLE_LOCKS
5681 if (rxi_CheckCall(call, 0)) {
5682 MUTEX_EXIT(&call->lock);
5685 #else /* RX_ENABLE_LOCKS */
5686 if (rxi_CheckCall(call))
5688 #endif /* RX_ENABLE_LOCKS */
5690 /* Don't try to keep alive dallying calls */
5691 if (call->state == RX_STATE_DALLY) {
5692 MUTEX_EXIT(&call->lock);
5697 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5698 /* Don't try to send keepalives if there is unacknowledged data */
5699 /* the rexmit code should be good enough, this little hack
5700 * doesn't quite work XXX */
5701 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5703 rxi_ScheduleKeepAliveEvent(call);
5704 MUTEX_EXIT(&call->lock);
5709 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5711 if (!call->keepAliveEvent) {
5712 struct clock when, now;
5713 clock_GetTime(&now);
5715 when.sec += call->conn->secondsUntilPing;
5716 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5717 call->keepAliveEvent =
5718 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
5722 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5724 rxi_KeepAliveOn(register struct rx_call *call)
5726 /* Pretend last packet received was received now--i.e. if another
5727 * packet isn't received within the keep alive time, then the call
5728 * will die; Initialize last send time to the current time--even
5729 * if a packet hasn't been sent yet. This will guarantee that a
5730 * keep-alive is sent within the ping time */
5731 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5732 rxi_ScheduleKeepAliveEvent(call);
5735 /* This routine is called to send connection abort messages
5736 * that have been delayed to throttle looping clients. */
5738 rxi_SendDelayedConnAbort(struct rxevent *event,
5739 void *arg1, void *unused)
5741 struct rx_connection *conn = arg1;
5744 struct rx_packet *packet;
5746 MUTEX_ENTER(&conn->conn_data_lock);
5747 conn->delayedAbortEvent = NULL;
5748 error = htonl(conn->error);
5750 MUTEX_EXIT(&conn->conn_data_lock);
5751 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5754 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5755 RX_PACKET_TYPE_ABORT, (char *)&error,
5757 rxi_FreePacket(packet);
5761 /* This routine is called to send call abort messages
5762 * that have been delayed to throttle looping clients. */
5764 rxi_SendDelayedCallAbort(struct rxevent *event,
5765 void *arg1, void *dummy)
5767 struct rx_call *call = arg1;
5770 struct rx_packet *packet;
5772 MUTEX_ENTER(&call->lock);
5773 call->delayedAbortEvent = NULL;
5774 error = htonl(call->error);
5776 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5779 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5780 (char *)&error, sizeof(error), 0);
5781 rxi_FreePacket(packet);
5783 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5784 MUTEX_EXIT(&call->lock);
5787 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5788 * seconds) to ask the client to authenticate itself. The routine
5789 * issues a challenge to the client, which is obtained from the
5790 * security object associated with the connection */
5792 rxi_ChallengeEvent(struct rxevent *event,
5793 void *arg0, void *arg1, int tries)
5795 struct rx_connection *conn = arg0;
5797 conn->challengeEvent = NULL;
5798 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5799 register struct rx_packet *packet;
5800 struct clock when, now;
5803 /* We've failed to authenticate for too long.
5804 * Reset any calls waiting for authentication;
5805 * they are all in RX_STATE_PRECALL.
5809 MUTEX_ENTER(&conn->conn_call_lock);
5810 for (i = 0; i < RX_MAXCALLS; i++) {
5811 struct rx_call *call = conn->call[i];
5813 MUTEX_ENTER(&call->lock);
5814 if (call->state == RX_STATE_PRECALL) {
5815 rxi_CallError(call, RX_CALL_DEAD);
5816 rxi_SendCallAbort(call, NULL, 0, 0);
5818 MUTEX_EXIT(&call->lock);
5821 MUTEX_EXIT(&conn->conn_call_lock);
5825 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5827 /* If there's no packet available, do this later. */
5828 RXS_GetChallenge(conn->securityObject, conn, packet);
5829 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5830 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5831 rxi_FreePacket(packet);
5833 clock_GetTime(&now);
5835 when.sec += RX_CHALLENGE_TIMEOUT;
5836 conn->challengeEvent =
5837 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
5842 /* Call this routine to start requesting the client to authenticate
5843 * itself. This will continue until authentication is established,
5844 * the call times out, or an invalid response is returned. The
5845 * security object associated with the connection is asked to create
5846 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5847 * defined earlier. */
5849 rxi_ChallengeOn(register struct rx_connection *conn)
5851 if (!conn->challengeEvent) {
5852 RXS_CreateChallenge(conn->securityObject, conn);
5853 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5858 /* Compute round trip time of the packet provided, in *rttp.
5861 /* rxi_ComputeRoundTripTime is called with peer locked. */
5862 /* sentp and/or peer may be null */
5864 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5865 register struct clock *sentp,
5866 register struct rx_peer *peer)
5868 struct clock thisRtt, *rttp = &thisRtt;
5870 register int rtt_timeout;
5872 clock_GetTime(rttp);
5874 if (clock_Lt(rttp, sentp)) {
5876 return; /* somebody set the clock back, don't count this time. */
5878 clock_Sub(rttp, sentp);
5879 MUTEX_ENTER(&rx_stats_mutex);
5880 if (clock_Lt(rttp, &rx_stats.minRtt))
5881 rx_stats.minRtt = *rttp;
5882 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5883 if (rttp->sec > 60) {
5884 MUTEX_EXIT(&rx_stats_mutex);
5885 return; /* somebody set the clock ahead */
5887 rx_stats.maxRtt = *rttp;
5889 clock_Add(&rx_stats.totalRtt, rttp);
5890 rx_stats.nRttSamples++;
5891 MUTEX_EXIT(&rx_stats_mutex);
5893 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5895 /* Apply VanJacobson round-trip estimations */
5900 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5901 * srtt is stored as fixed point with 3 bits after the binary
5902 * point (i.e., scaled by 8). The following magic is
5903 * equivalent to the smoothing algorithm in rfc793 with an
5904 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5905 * srtt*8 = srtt*8 + rtt - srtt
5906 * srtt = srtt + rtt/8 - srtt/8
5909 delta = MSEC(rttp) - (peer->rtt >> 3);
5913 * We accumulate a smoothed rtt variance (actually, a smoothed
5914 * mean difference), then set the retransmit timer to smoothed
5915 * rtt + 4 times the smoothed variance (was 2x in van's original
5916 * paper, but 4x works better for me, and apparently for him as
5918 * rttvar is stored as
5919 * fixed point with 2 bits after the binary point (scaled by
5920 * 4). The following is equivalent to rfc793 smoothing with
5921 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5922 * replaces rfc793's wired-in beta.
5923 * dev*4 = dev*4 + (|actual - expected| - dev)
5929 delta -= (peer->rtt_dev >> 2);
5930 peer->rtt_dev += delta;
5932 /* I don't have a stored RTT so I start with this value. Since I'm
5933 * probably just starting a call, and will be pushing more data down
5934 * this, I expect congestion to increase rapidly. So I fudge a
5935 * little, and I set deviance to half the rtt. In practice,
5936 * deviance tends to approach something a little less than
5937 * half the smoothed rtt. */
5938 peer->rtt = (MSEC(rttp) << 3) + 8;
5939 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5941 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5942 * the other of these connections is usually in a user process, and can
5943 * be switched and/or swapped out. So on fast, reliable networks, the
5944 * timeout would otherwise be too short.
5946 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5947 clock_Zero(&(peer->timeout));
5948 clock_Addmsec(&(peer->timeout), rtt_timeout);
5950 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)));
5954 /* Find all server connections that have not been active for a long time, and
5957 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2)
5959 struct clock now, when;
5960 clock_GetTime(&now);
5962 /* Find server connection structures that haven't been used for
5963 * greater than rx_idleConnectionTime */
5965 struct rx_connection **conn_ptr, **conn_end;
5966 int i, havecalls = 0;
5967 MUTEX_ENTER(&rx_connHashTable_lock);
5968 for (conn_ptr = &rx_connHashTable[0], conn_end =
5969 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5971 struct rx_connection *conn, *next;
5972 struct rx_call *call;
5976 for (conn = *conn_ptr; conn; conn = next) {
5977 /* XXX -- Shouldn't the connection be locked? */
5980 for (i = 0; i < RX_MAXCALLS; i++) {
5981 call = conn->call[i];
5984 MUTEX_ENTER(&call->lock);
5985 #ifdef RX_ENABLE_LOCKS
5986 result = rxi_CheckCall(call, 1);
5987 #else /* RX_ENABLE_LOCKS */
5988 result = rxi_CheckCall(call);
5989 #endif /* RX_ENABLE_LOCKS */
5990 MUTEX_EXIT(&call->lock);
5992 /* If CheckCall freed the call, it might
5993 * have destroyed the connection as well,
5994 * which screws up the linked lists.
6000 if (conn->type == RX_SERVER_CONNECTION) {
6001 /* This only actually destroys the connection if
6002 * there are no outstanding calls */
6003 MUTEX_ENTER(&conn->conn_data_lock);
6004 if (!havecalls && !conn->refCount
6005 && ((conn->lastSendTime + rx_idleConnectionTime) <
6007 conn->refCount++; /* it will be decr in rx_DestroyConn */
6008 MUTEX_EXIT(&conn->conn_data_lock);
6009 #ifdef RX_ENABLE_LOCKS
6010 rxi_DestroyConnectionNoLock(conn);
6011 #else /* RX_ENABLE_LOCKS */
6012 rxi_DestroyConnection(conn);
6013 #endif /* RX_ENABLE_LOCKS */
6015 #ifdef RX_ENABLE_LOCKS
6017 MUTEX_EXIT(&conn->conn_data_lock);
6019 #endif /* RX_ENABLE_LOCKS */
6023 #ifdef RX_ENABLE_LOCKS
6024 while (rx_connCleanup_list) {
6025 struct rx_connection *conn;
6026 conn = rx_connCleanup_list;
6027 rx_connCleanup_list = rx_connCleanup_list->next;
6028 MUTEX_EXIT(&rx_connHashTable_lock);
6029 rxi_CleanupConnection(conn);
6030 MUTEX_ENTER(&rx_connHashTable_lock);
6032 MUTEX_EXIT(&rx_connHashTable_lock);
6033 #endif /* RX_ENABLE_LOCKS */
6036 /* Find any peer structures that haven't been used (haven't had an
6037 * associated connection) for greater than rx_idlePeerTime */
6039 struct rx_peer **peer_ptr, **peer_end;
6041 MUTEX_ENTER(&rx_rpc_stats);
6042 MUTEX_ENTER(&rx_peerHashTable_lock);
6043 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6044 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6046 struct rx_peer *peer, *next, *prev;
6047 for (prev = peer = *peer_ptr; peer; peer = next) {
6049 code = MUTEX_TRYENTER(&peer->peer_lock);
6050 if ((code) && (peer->refCount == 0)
6051 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6052 rx_interface_stat_p rpc_stat, nrpc_stat;
6054 MUTEX_EXIT(&peer->peer_lock);
6055 MUTEX_DESTROY(&peer->peer_lock);
6057 (&peer->rpcStats, rpc_stat, nrpc_stat,
6058 rx_interface_stat)) {
6059 unsigned int num_funcs;
6062 queue_Remove(&rpc_stat->queue_header);
6063 queue_Remove(&rpc_stat->all_peers);
6064 num_funcs = rpc_stat->stats[0].func_total;
6066 sizeof(rx_interface_stat_t) +
6067 rpc_stat->stats[0].func_total *
6068 sizeof(rx_function_entry_v1_t);
6070 rxi_Free(rpc_stat, space);
6071 rxi_rpc_peer_stat_cnt -= num_funcs;
6074 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6075 if (peer == *peer_ptr) {
6082 MUTEX_EXIT(&peer->peer_lock);
6088 MUTEX_EXIT(&rx_peerHashTable_lock);
6089 MUTEX_EXIT(&rx_rpc_stats);
6092 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
6093 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
6094 * GC, just below. Really, we shouldn't have to keep moving packets from
6095 * one place to another, but instead ought to always know if we can
6096 * afford to hold onto a packet in its particular use. */
6097 MUTEX_ENTER(&rx_freePktQ_lock);
6098 if (rx_waitingForPackets) {
6099 rx_waitingForPackets = 0;
6100 #ifdef RX_ENABLE_LOCKS
6101 CV_BROADCAST(&rx_waitingForPackets_cv);
6103 osi_rxWakeup(&rx_waitingForPackets);
6106 MUTEX_EXIT(&rx_freePktQ_lock);
6109 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
6110 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
6114 /* rxs_Release - This isn't strictly necessary but, since the macro name from
6115 * rx.h is sort of strange this is better. This is called with a security
6116 * object before it is discarded. Each connection using a security object has
6117 * its own refcount to the object so it won't actually be freed until the last
6118 * connection is destroyed.
6120 * This is the only rxs module call. A hold could also be written but no one
6124 rxs_Release(struct rx_securityClass *aobj)
6126 return RXS_Close(aobj);
6130 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
6131 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
6132 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
6133 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
6135 /* Adjust our estimate of the transmission rate to this peer, given
6136 * that the packet p was just acked. We can adjust peer->timeout and
6137 * call->twind. Pragmatically, this is called
6138 * only with packets of maximal length.
6139 * Called with peer and call locked.
6143 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
6144 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
6146 afs_int32 xferSize, xferMs;
6147 register afs_int32 minTime;
6150 /* Count down packets */
6151 if (peer->rateFlag > 0)
6153 /* Do nothing until we're enabled */
6154 if (peer->rateFlag != 0)
6159 /* Count only when the ack seems legitimate */
6160 switch (ackReason) {
6161 case RX_ACK_REQUESTED:
6163 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6167 case RX_ACK_PING_RESPONSE:
6168 if (p) /* want the response to ping-request, not data send */
6170 clock_GetTime(&newTO);
6171 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6172 clock_Sub(&newTO, &call->pingRequestTime);
6173 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6177 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6184 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));
6186 /* Track only packets that are big enough. */
6187 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6191 /* absorb RTT data (in milliseconds) for these big packets */
6192 if (peer->smRtt == 0) {
6193 peer->smRtt = xferMs;
6195 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6200 if (peer->countDown) {
6204 peer->countDown = 10; /* recalculate only every so often */
6206 /* In practice, we can measure only the RTT for full packets,
6207 * because of the way Rx acks the data that it receives. (If it's
6208 * smaller than a full packet, it often gets implicitly acked
6209 * either by the call response (from a server) or by the next call
6210 * (from a client), and either case confuses transmission times
6211 * with processing times.) Therefore, replace the above
6212 * more-sophisticated processing with a simpler version, where the
6213 * smoothed RTT is kept for full-size packets, and the time to
6214 * transmit a windowful of full-size packets is simply RTT *
6215 * windowSize. Again, we take two steps:
6216 - ensure the timeout is large enough for a single packet's RTT;
6217 - ensure that the window is small enough to fit in the desired timeout.*/
6219 /* First, the timeout check. */
6220 minTime = peer->smRtt;
6221 /* Get a reasonable estimate for a timeout period */
6223 newTO.sec = minTime / 1000;
6224 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6226 /* Increase the timeout period so that we can always do at least
6227 * one packet exchange */
6228 if (clock_Gt(&newTO, &peer->timeout)) {
6230 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));
6232 peer->timeout = newTO;
6235 /* Now, get an estimate for the transmit window size. */
6236 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6237 /* Now, convert to the number of full packets that could fit in a
6238 * reasonable fraction of that interval */
6239 minTime /= (peer->smRtt << 1);
6240 xferSize = minTime; /* (make a copy) */
6242 /* Now clamp the size to reasonable bounds. */
6245 else if (minTime > rx_Window)
6246 minTime = rx_Window;
6247 /* if (minTime != peer->maxWindow) {
6248 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6249 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6250 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6252 peer->maxWindow = minTime;
6253 elide... call->twind = minTime;
6257 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6258 * Discern this by calculating the timeout necessary for rx_Window
6260 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6261 /* calculate estimate for transmission interval in milliseconds */
6262 minTime = rx_Window * peer->smRtt;
6263 if (minTime < 1000) {
6264 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
6265 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6266 peer->timeout.usec, peer->smRtt, peer->packetSize));
6268 newTO.sec = 0; /* cut back on timeout by half a second */
6269 newTO.usec = 500000;
6270 clock_Sub(&peer->timeout, &newTO);
6275 } /* end of rxi_ComputeRate */
6276 #endif /* ADAPT_WINDOW */
6284 #define TRACE_OPTION_DEBUGLOG 4
6292 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6293 0, KEY_QUERY_VALUE, &parmKey);
6294 if (code != ERROR_SUCCESS)
6297 dummyLen = sizeof(TraceOption);
6298 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6299 (BYTE *) &TraceOption, &dummyLen);
6300 if (code == ERROR_SUCCESS) {
6301 rxdebug_active = (TraceOption & TRACE_OPTION_DEBUGLOG) ? 1 : 0;
6303 RegCloseKey (parmKey);
6304 #endif /* AFS_NT40_ENV */
6309 rx_DebugOnOff(int on)
6311 rxdebug_active = on;
6313 #endif /* AFS_NT40_ENV */
6316 /* Don't call this debugging routine directly; use dpf */
6318 rxi_DebugPrint(char *format, int a1, int a2, int a3, int a4, int a5, int a6,
6319 int a7, int a8, int a9, int a10, int a11, int a12, int a13,
6327 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6330 len = _snprintf(msg, sizeof(msg)-2,
6331 tformat, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10,
6332 a11, a12, a13, a14, a15);
6334 if (msg[len-1] != '\n') {
6338 OutputDebugString(msg);
6343 clock_GetTime(&now);
6344 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
6345 (unsigned int)now.usec / 1000);
6346 fprintf(rx_Log, format, a1, a2, a3, a4, a5, a6, a7, a8, a9, a10, a11, a12,
6353 * This function is used to process the rx_stats structure that is local
6354 * to a process as well as an rx_stats structure received from a remote
6355 * process (via rxdebug). Therefore, it needs to do minimal version
6359 rx_PrintTheseStats(FILE * file, struct rx_stats *s, int size,
6360 afs_int32 freePackets, char version)
6364 if (size != sizeof(struct rx_stats)) {
6366 "Unexpected size of stats structure: was %d, expected %lud\n",
6367 size, sizeof(struct rx_stats));
6370 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6373 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6374 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6375 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6376 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6377 s->specialPktAllocFailures);
6379 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6380 s->receivePktAllocFailures, s->sendPktAllocFailures,
6381 s->specialPktAllocFailures);
6385 " greedy %d, " "bogusReads %d (last from host %x), "
6386 "noPackets %d, " "noBuffers %d, " "selects %d, "
6387 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6388 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6389 s->selects, s->sendSelects);
6391 fprintf(file, " packets read: ");
6392 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6393 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6395 fprintf(file, "\n");
6398 " other read counters: data %d, " "ack %d, " "dup %d "
6399 "spurious %d " "dally %d\n", s->dataPacketsRead,
6400 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6401 s->ignorePacketDally);
6403 fprintf(file, " packets sent: ");
6404 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6405 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6407 fprintf(file, "\n");
6410 " other send counters: ack %d, " "data %d (not resends), "
6411 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6412 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6413 s->dataPacketsPushed, s->ignoreAckedPacket);
6416 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6417 s->netSendFailures, (int)s->fatalErrors);
6419 if (s->nRttSamples) {
6420 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6421 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6423 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6424 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6428 " %d server connections, " "%d client connections, "
6429 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6430 s->nServerConns, s->nClientConns, s->nPeerStructs,
6431 s->nCallStructs, s->nFreeCallStructs);
6433 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6434 fprintf(file, " %d clock updates\n", clock_nUpdates);
6439 /* for backward compatibility */
6441 rx_PrintStats(FILE * file)
6443 MUTEX_ENTER(&rx_stats_mutex);
6444 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6446 MUTEX_EXIT(&rx_stats_mutex);
6450 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6452 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6453 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6454 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6457 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6458 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6459 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6462 " Packet size %d, " "max in packet skew %d, "
6463 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6464 (int)peer->outPacketSkew);
6467 #ifdef AFS_PTHREAD_ENV
6469 * This mutex protects the following static variables:
6473 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6474 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6476 #define LOCK_RX_DEBUG
6477 #define UNLOCK_RX_DEBUG
6478 #endif /* AFS_PTHREAD_ENV */
6481 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6482 u_char type, void *inputData, size_t inputLength,
6483 void *outputData, size_t outputLength)
6485 static afs_int32 counter = 100;
6486 time_t waitTime, waitCount, startTime;
6487 struct rx_header theader;
6489 register afs_int32 code;
6490 struct timeval tv_now, tv_wake, tv_delta;
6491 struct sockaddr_in taddr, faddr;
6496 startTime = time(0);
6502 tp = &tbuffer[sizeof(struct rx_header)];
6503 taddr.sin_family = AF_INET;
6504 taddr.sin_port = remotePort;
6505 taddr.sin_addr.s_addr = remoteAddr;
6506 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6507 taddr.sin_len = sizeof(struct sockaddr_in);
6510 memset(&theader, 0, sizeof(theader));
6511 theader.epoch = htonl(999);
6513 theader.callNumber = htonl(counter);
6516 theader.type = type;
6517 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6518 theader.serviceId = 0;
6520 memcpy(tbuffer, &theader, sizeof(theader));
6521 memcpy(tp, inputData, inputLength);
6523 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6524 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6526 /* see if there's a packet available */
6527 gettimeofday(&tv_wake,0);
6528 tv_wake.tv_sec += waitTime;
6531 FD_SET(socket, &imask);
6532 tv_delta.tv_sec = tv_wake.tv_sec;
6533 tv_delta.tv_usec = tv_wake.tv_usec;
6534 gettimeofday(&tv_now, 0);
6536 if (tv_delta.tv_usec < tv_now.tv_usec) {
6538 tv_delta.tv_usec += 1000000;
6541 tv_delta.tv_usec -= tv_now.tv_usec;
6543 if (tv_delta.tv_sec < tv_now.tv_sec) {
6547 tv_delta.tv_sec -= tv_now.tv_sec;
6549 code = select(socket + 1, &imask, 0, 0, &tv_delta);
6550 if (code == 1 && FD_ISSET(socket, &imask)) {
6551 /* now receive a packet */
6552 faddrLen = sizeof(struct sockaddr_in);
6554 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6555 (struct sockaddr *)&faddr, &faddrLen);
6558 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6559 if (counter == ntohl(theader.callNumber))
6567 /* see if we've timed out */
6575 code -= sizeof(struct rx_header);
6576 if (code > outputLength)
6577 code = outputLength;
6578 memcpy(outputData, tp, code);
6583 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6584 afs_uint16 remotePort, struct rx_debugStats * stat,
6585 afs_uint32 * supportedValues)
6587 struct rx_debugIn in;
6590 *supportedValues = 0;
6591 in.type = htonl(RX_DEBUGI_GETSTATS);
6594 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6595 &in, sizeof(in), stat, sizeof(*stat));
6598 * If the call was successful, fixup the version and indicate
6599 * what contents of the stat structure are valid.
6600 * Also do net to host conversion of fields here.
6604 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6605 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6607 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6608 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6610 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6611 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6613 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6614 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6616 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6617 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6619 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6620 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6622 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6623 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6625 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6626 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6629 stat->nFreePackets = ntohl(stat->nFreePackets);
6630 stat->packetReclaims = ntohl(stat->packetReclaims);
6631 stat->callsExecuted = ntohl(stat->callsExecuted);
6632 stat->nWaiting = ntohl(stat->nWaiting);
6633 stat->idleThreads = ntohl(stat->idleThreads);
6640 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6641 afs_uint16 remotePort, struct rx_stats * stat,
6642 afs_uint32 * supportedValues)
6644 struct rx_debugIn in;
6645 afs_int32 *lp = (afs_int32 *) stat;
6650 * supportedValues is currently unused, but added to allow future
6651 * versioning of this function.
6654 *supportedValues = 0;
6655 in.type = htonl(RX_DEBUGI_RXSTATS);
6657 memset(stat, 0, sizeof(*stat));
6659 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6660 &in, sizeof(in), stat, sizeof(*stat));
6665 * Do net to host conversion here
6668 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6677 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6678 afs_uint16 remotePort, size_t version_length,
6682 return MakeDebugCall(socket, remoteAddr, remotePort,
6683 RX_PACKET_TYPE_VERSION, a, 1, version,
6688 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6689 afs_uint16 remotePort, afs_int32 * nextConnection,
6690 int allConnections, afs_uint32 debugSupportedValues,
6691 struct rx_debugConn * conn,
6692 afs_uint32 * supportedValues)
6694 struct rx_debugIn in;
6699 * supportedValues is currently unused, but added to allow future
6700 * versioning of this function.
6703 *supportedValues = 0;
6704 if (allConnections) {
6705 in.type = htonl(RX_DEBUGI_GETALLCONN);
6707 in.type = htonl(RX_DEBUGI_GETCONN);
6709 in.index = htonl(*nextConnection);
6710 memset(conn, 0, sizeof(*conn));
6712 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6713 &in, sizeof(in), conn, sizeof(*conn));
6716 *nextConnection += 1;
6719 * Convert old connection format to new structure.
6722 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6723 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6724 #define MOVEvL(a) (conn->a = vL->a)
6726 /* any old or unrecognized version... */
6727 for (i = 0; i < RX_MAXCALLS; i++) {
6728 MOVEvL(callState[i]);
6729 MOVEvL(callMode[i]);
6730 MOVEvL(callFlags[i]);
6731 MOVEvL(callOther[i]);
6733 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6734 MOVEvL(secStats.type);
6735 MOVEvL(secStats.level);
6736 MOVEvL(secStats.flags);
6737 MOVEvL(secStats.expires);
6738 MOVEvL(secStats.packetsReceived);
6739 MOVEvL(secStats.packetsSent);
6740 MOVEvL(secStats.bytesReceived);
6741 MOVEvL(secStats.bytesSent);
6746 * Do net to host conversion here
6748 * I don't convert host or port since we are most likely
6749 * going to want these in NBO.
6751 conn->cid = ntohl(conn->cid);
6752 conn->serial = ntohl(conn->serial);
6753 for (i = 0; i < RX_MAXCALLS; i++) {
6754 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6756 conn->error = ntohl(conn->error);
6757 conn->secStats.flags = ntohl(conn->secStats.flags);
6758 conn->secStats.expires = ntohl(conn->secStats.expires);
6759 conn->secStats.packetsReceived =
6760 ntohl(conn->secStats.packetsReceived);
6761 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6762 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6763 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6764 conn->epoch = ntohl(conn->epoch);
6765 conn->natMTU = ntohl(conn->natMTU);
6772 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6773 afs_uint16 remotePort, afs_int32 * nextPeer,
6774 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6775 afs_uint32 * supportedValues)
6777 struct rx_debugIn in;
6781 * supportedValues is currently unused, but added to allow future
6782 * versioning of this function.
6785 *supportedValues = 0;
6786 in.type = htonl(RX_DEBUGI_GETPEER);
6787 in.index = htonl(*nextPeer);
6788 memset(peer, 0, sizeof(*peer));
6790 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6791 &in, sizeof(in), peer, sizeof(*peer));
6797 * Do net to host conversion here
6799 * I don't convert host or port since we are most likely
6800 * going to want these in NBO.
6802 peer->ifMTU = ntohs(peer->ifMTU);
6803 peer->idleWhen = ntohl(peer->idleWhen);
6804 peer->refCount = ntohs(peer->refCount);
6805 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6806 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6807 peer->rtt = ntohl(peer->rtt);
6808 peer->rtt_dev = ntohl(peer->rtt_dev);
6809 peer->timeout.sec = ntohl(peer->timeout.sec);
6810 peer->timeout.usec = ntohl(peer->timeout.usec);
6811 peer->nSent = ntohl(peer->nSent);
6812 peer->reSends = ntohl(peer->reSends);
6813 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6814 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6815 peer->rateFlag = ntohl(peer->rateFlag);
6816 peer->natMTU = ntohs(peer->natMTU);
6817 peer->maxMTU = ntohs(peer->maxMTU);
6818 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6819 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6820 peer->MTU = ntohs(peer->MTU);
6821 peer->cwind = ntohs(peer->cwind);
6822 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6823 peer->congestSeq = ntohs(peer->congestSeq);
6824 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6825 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6826 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6827 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6832 #endif /* RXDEBUG */
6837 struct rx_serverQueueEntry *np;
6840 register struct rx_call *call;
6841 register struct rx_serverQueueEntry *sq;
6845 if (rxinit_status == 1) {
6847 return; /* Already shutdown. */
6851 #ifndef AFS_PTHREAD_ENV
6852 FD_ZERO(&rx_selectMask);
6853 #endif /* AFS_PTHREAD_ENV */
6854 rxi_dataQuota = RX_MAX_QUOTA;
6855 #ifndef AFS_PTHREAD_ENV
6857 #endif /* AFS_PTHREAD_ENV */
6860 #ifndef AFS_PTHREAD_ENV
6861 #ifndef AFS_USE_GETTIMEOFDAY
6863 #endif /* AFS_USE_GETTIMEOFDAY */
6864 #endif /* AFS_PTHREAD_ENV */
6866 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6867 call = queue_First(&rx_freeCallQueue, rx_call);
6869 rxi_Free(call, sizeof(struct rx_call));
6872 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6873 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6879 struct rx_peer **peer_ptr, **peer_end;
6880 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6881 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6883 struct rx_peer *peer, *next;
6884 for (peer = *peer_ptr; peer; peer = next) {
6885 rx_interface_stat_p rpc_stat, nrpc_stat;
6888 (&peer->rpcStats, rpc_stat, nrpc_stat,
6889 rx_interface_stat)) {
6890 unsigned int num_funcs;
6893 queue_Remove(&rpc_stat->queue_header);
6894 queue_Remove(&rpc_stat->all_peers);
6895 num_funcs = rpc_stat->stats[0].func_total;
6897 sizeof(rx_interface_stat_t) +
6898 rpc_stat->stats[0].func_total *
6899 sizeof(rx_function_entry_v1_t);
6901 rxi_Free(rpc_stat, space);
6902 MUTEX_ENTER(&rx_rpc_stats);
6903 rxi_rpc_peer_stat_cnt -= num_funcs;
6904 MUTEX_EXIT(&rx_rpc_stats);
6908 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6912 for (i = 0; i < RX_MAX_SERVICES; i++) {
6914 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6916 for (i = 0; i < rx_hashTableSize; i++) {
6917 register struct rx_connection *tc, *ntc;
6918 MUTEX_ENTER(&rx_connHashTable_lock);
6919 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6921 for (j = 0; j < RX_MAXCALLS; j++) {
6923 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6926 rxi_Free(tc, sizeof(*tc));
6928 MUTEX_EXIT(&rx_connHashTable_lock);
6931 MUTEX_ENTER(&freeSQEList_lock);
6933 while ((np = rx_FreeSQEList)) {
6934 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6935 MUTEX_DESTROY(&np->lock);
6936 rxi_Free(np, sizeof(*np));
6939 MUTEX_EXIT(&freeSQEList_lock);
6940 MUTEX_DESTROY(&freeSQEList_lock);
6941 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6942 MUTEX_DESTROY(&rx_connHashTable_lock);
6943 MUTEX_DESTROY(&rx_peerHashTable_lock);
6944 MUTEX_DESTROY(&rx_serverPool_lock);
6946 osi_Free(rx_connHashTable,
6947 rx_hashTableSize * sizeof(struct rx_connection *));
6948 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6950 UNPIN(rx_connHashTable,
6951 rx_hashTableSize * sizeof(struct rx_connection *));
6952 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6954 rxi_FreeAllPackets();
6956 MUTEX_ENTER(&rx_stats_mutex);
6957 rxi_dataQuota = RX_MAX_QUOTA;
6958 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6959 MUTEX_EXIT(&rx_stats_mutex);
6965 #ifdef RX_ENABLE_LOCKS
6967 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6969 if (!MUTEX_ISMINE(lockaddr))
6970 osi_Panic("Lock not held: %s", msg);
6972 #endif /* RX_ENABLE_LOCKS */
6977 * Routines to implement connection specific data.
6981 rx_KeyCreate(rx_destructor_t rtn)
6984 MUTEX_ENTER(&rxi_keyCreate_lock);
6985 key = rxi_keyCreate_counter++;
6986 rxi_keyCreate_destructor = (rx_destructor_t *)
6987 realloc((void *)rxi_keyCreate_destructor,
6988 (key + 1) * sizeof(rx_destructor_t));
6989 rxi_keyCreate_destructor[key] = rtn;
6990 MUTEX_EXIT(&rxi_keyCreate_lock);
6995 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6998 MUTEX_ENTER(&conn->conn_data_lock);
6999 if (!conn->specific) {
7000 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
7001 for (i = 0; i < key; i++)
7002 conn->specific[i] = NULL;
7003 conn->nSpecific = key + 1;
7004 conn->specific[key] = ptr;
7005 } else if (key >= conn->nSpecific) {
7006 conn->specific = (void **)
7007 realloc(conn->specific, (key + 1) * sizeof(void *));
7008 for (i = conn->nSpecific; i < key; i++)
7009 conn->specific[i] = NULL;
7010 conn->nSpecific = key + 1;
7011 conn->specific[key] = ptr;
7013 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7014 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7015 conn->specific[key] = ptr;
7017 MUTEX_EXIT(&conn->conn_data_lock);
7021 rx_GetSpecific(struct rx_connection *conn, int key)
7024 MUTEX_ENTER(&conn->conn_data_lock);
7025 if (key >= conn->nSpecific)
7028 ptr = conn->specific[key];
7029 MUTEX_EXIT(&conn->conn_data_lock);
7033 #endif /* !KERNEL */
7036 * processStats is a queue used to store the statistics for the local
7037 * process. Its contents are similar to the contents of the rpcStats
7038 * queue on a rx_peer structure, but the actual data stored within
7039 * this queue contains totals across the lifetime of the process (assuming
7040 * the stats have not been reset) - unlike the per peer structures
7041 * which can come and go based upon the peer lifetime.
7044 static struct rx_queue processStats = { &processStats, &processStats };
7047 * peerStats is a queue used to store the statistics for all peer structs.
7048 * Its contents are the union of all the peer rpcStats queues.
7051 static struct rx_queue peerStats = { &peerStats, &peerStats };
7054 * rxi_monitor_processStats is used to turn process wide stat collection
7058 static int rxi_monitor_processStats = 0;
7061 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
7064 static int rxi_monitor_peerStats = 0;
7067 * rxi_AddRpcStat - given all of the information for a particular rpc
7068 * call, create (if needed) and update the stat totals for the rpc.
7072 * IN stats - the queue of stats that will be updated with the new value
7074 * IN rxInterface - a unique number that identifies the rpc interface
7076 * IN currentFunc - the index of the function being invoked
7078 * IN totalFunc - the total number of functions in this interface
7080 * IN queueTime - the amount of time this function waited for a thread
7082 * IN execTime - the amount of time this function invocation took to execute
7084 * IN bytesSent - the number bytes sent by this invocation
7086 * IN bytesRcvd - the number bytes received by this invocation
7088 * IN isServer - if true, this invocation was made to a server
7090 * IN remoteHost - the ip address of the remote host
7092 * IN remotePort - the port of the remote host
7094 * IN addToPeerList - if != 0, add newly created stat to the global peer list
7096 * INOUT counter - if a new stats structure is allocated, the counter will
7097 * be updated with the new number of allocated stat structures
7105 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
7106 afs_uint32 currentFunc, afs_uint32 totalFunc,
7107 struct clock *queueTime, struct clock *execTime,
7108 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
7109 afs_uint32 remoteHost, afs_uint32 remotePort,
7110 int addToPeerList, unsigned int *counter)
7113 rx_interface_stat_p rpc_stat, nrpc_stat;
7116 * See if there's already a structure for this interface
7119 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7120 if ((rpc_stat->stats[0].interfaceId == rxInterface)
7121 && (rpc_stat->stats[0].remote_is_server == isServer))
7126 * Didn't find a match so allocate a new structure and add it to the
7130 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
7131 || (rpc_stat->stats[0].interfaceId != rxInterface)
7132 || (rpc_stat->stats[0].remote_is_server != isServer)) {
7137 sizeof(rx_interface_stat_t) +
7138 totalFunc * sizeof(rx_function_entry_v1_t);
7140 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
7141 if (rpc_stat == NULL) {
7145 *counter += totalFunc;
7146 for (i = 0; i < totalFunc; i++) {
7147 rpc_stat->stats[i].remote_peer = remoteHost;
7148 rpc_stat->stats[i].remote_port = remotePort;
7149 rpc_stat->stats[i].remote_is_server = isServer;
7150 rpc_stat->stats[i].interfaceId = rxInterface;
7151 rpc_stat->stats[i].func_total = totalFunc;
7152 rpc_stat->stats[i].func_index = i;
7153 hzero(rpc_stat->stats[i].invocations);
7154 hzero(rpc_stat->stats[i].bytes_sent);
7155 hzero(rpc_stat->stats[i].bytes_rcvd);
7156 rpc_stat->stats[i].queue_time_sum.sec = 0;
7157 rpc_stat->stats[i].queue_time_sum.usec = 0;
7158 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7159 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7160 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7161 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7162 rpc_stat->stats[i].queue_time_max.sec = 0;
7163 rpc_stat->stats[i].queue_time_max.usec = 0;
7164 rpc_stat->stats[i].execution_time_sum.sec = 0;
7165 rpc_stat->stats[i].execution_time_sum.usec = 0;
7166 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7167 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7168 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7169 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7170 rpc_stat->stats[i].execution_time_max.sec = 0;
7171 rpc_stat->stats[i].execution_time_max.usec = 0;
7173 queue_Prepend(stats, rpc_stat);
7174 if (addToPeerList) {
7175 queue_Prepend(&peerStats, &rpc_stat->all_peers);
7180 * Increment the stats for this function
7183 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
7184 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
7185 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
7186 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
7187 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7188 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7189 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7191 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7192 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7194 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7195 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7197 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7198 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7200 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7201 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7209 * rx_IncrementTimeAndCount - increment the times and count for a particular
7214 * IN peer - the peer who invoked the rpc
7216 * IN rxInterface - a unique number that identifies the rpc interface
7218 * IN currentFunc - the index of the function being invoked
7220 * IN totalFunc - the total number of functions in this interface
7222 * IN queueTime - the amount of time this function waited for a thread
7224 * IN execTime - the amount of time this function invocation took to execute
7226 * IN bytesSent - the number bytes sent by this invocation
7228 * IN bytesRcvd - the number bytes received by this invocation
7230 * IN isServer - if true, this invocation was made to a server
7238 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7239 afs_uint32 currentFunc, afs_uint32 totalFunc,
7240 struct clock *queueTime, struct clock *execTime,
7241 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7245 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
7248 MUTEX_ENTER(&rx_rpc_stats);
7249 MUTEX_ENTER(&peer->peer_lock);
7251 if (rxi_monitor_peerStats) {
7252 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7253 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7254 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7257 if (rxi_monitor_processStats) {
7258 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7259 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7260 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7263 MUTEX_EXIT(&peer->peer_lock);
7264 MUTEX_EXIT(&rx_rpc_stats);
7269 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7273 * IN callerVersion - the rpc stat version of the caller.
7275 * IN count - the number of entries to marshall.
7277 * IN stats - pointer to stats to be marshalled.
7279 * OUT ptr - Where to store the marshalled data.
7286 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7287 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7293 * We only support the first version
7295 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7296 *(ptr++) = stats->remote_peer;
7297 *(ptr++) = stats->remote_port;
7298 *(ptr++) = stats->remote_is_server;
7299 *(ptr++) = stats->interfaceId;
7300 *(ptr++) = stats->func_total;
7301 *(ptr++) = stats->func_index;
7302 *(ptr++) = hgethi(stats->invocations);
7303 *(ptr++) = hgetlo(stats->invocations);
7304 *(ptr++) = hgethi(stats->bytes_sent);
7305 *(ptr++) = hgetlo(stats->bytes_sent);
7306 *(ptr++) = hgethi(stats->bytes_rcvd);
7307 *(ptr++) = hgetlo(stats->bytes_rcvd);
7308 *(ptr++) = stats->queue_time_sum.sec;
7309 *(ptr++) = stats->queue_time_sum.usec;
7310 *(ptr++) = stats->queue_time_sum_sqr.sec;
7311 *(ptr++) = stats->queue_time_sum_sqr.usec;
7312 *(ptr++) = stats->queue_time_min.sec;
7313 *(ptr++) = stats->queue_time_min.usec;
7314 *(ptr++) = stats->queue_time_max.sec;
7315 *(ptr++) = stats->queue_time_max.usec;
7316 *(ptr++) = stats->execution_time_sum.sec;
7317 *(ptr++) = stats->execution_time_sum.usec;
7318 *(ptr++) = stats->execution_time_sum_sqr.sec;
7319 *(ptr++) = stats->execution_time_sum_sqr.usec;
7320 *(ptr++) = stats->execution_time_min.sec;
7321 *(ptr++) = stats->execution_time_min.usec;
7322 *(ptr++) = stats->execution_time_max.sec;
7323 *(ptr++) = stats->execution_time_max.usec;
7329 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7334 * IN callerVersion - the rpc stat version of the caller
7336 * OUT myVersion - the rpc stat version of this function
7338 * OUT clock_sec - local time seconds
7340 * OUT clock_usec - local time microseconds
7342 * OUT allocSize - the number of bytes allocated to contain stats
7344 * OUT statCount - the number stats retrieved from this process.
7346 * OUT stats - the actual stats retrieved from this process.
7350 * Returns void. If successful, stats will != NULL.
7354 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7355 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7356 size_t * allocSize, afs_uint32 * statCount,
7357 afs_uint32 ** stats)
7367 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7370 * Check to see if stats are enabled
7373 MUTEX_ENTER(&rx_rpc_stats);
7374 if (!rxi_monitor_processStats) {
7375 MUTEX_EXIT(&rx_rpc_stats);
7379 clock_GetTime(&now);
7380 *clock_sec = now.sec;
7381 *clock_usec = now.usec;
7384 * Allocate the space based upon the caller version
7386 * If the client is at an older version than we are,
7387 * we return the statistic data in the older data format, but
7388 * we still return our version number so the client knows we
7389 * are maintaining more data than it can retrieve.
7392 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7393 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7394 *statCount = rxi_rpc_process_stat_cnt;
7397 * This can't happen yet, but in the future version changes
7398 * can be handled by adding additional code here
7402 if (space > (size_t) 0) {
7404 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7407 rx_interface_stat_p rpc_stat, nrpc_stat;
7411 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7413 * Copy the data based upon the caller version
7415 rx_MarshallProcessRPCStats(callerVersion,
7416 rpc_stat->stats[0].func_total,
7417 rpc_stat->stats, &ptr);
7423 MUTEX_EXIT(&rx_rpc_stats);
7428 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7432 * IN callerVersion - the rpc stat version of the caller
7434 * OUT myVersion - the rpc stat version of this function
7436 * OUT clock_sec - local time seconds
7438 * OUT clock_usec - local time microseconds
7440 * OUT allocSize - the number of bytes allocated to contain stats
7442 * OUT statCount - the number of stats retrieved from the individual
7445 * OUT stats - the actual stats retrieved from the individual peer structures.
7449 * Returns void. If successful, stats will != NULL.
7453 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7454 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7455 size_t * allocSize, afs_uint32 * statCount,
7456 afs_uint32 ** stats)
7466 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7469 * Check to see if stats are enabled
7472 MUTEX_ENTER(&rx_rpc_stats);
7473 if (!rxi_monitor_peerStats) {
7474 MUTEX_EXIT(&rx_rpc_stats);
7478 clock_GetTime(&now);
7479 *clock_sec = now.sec;
7480 *clock_usec = now.usec;
7483 * Allocate the space based upon the caller version
7485 * If the client is at an older version than we are,
7486 * we return the statistic data in the older data format, but
7487 * we still return our version number so the client knows we
7488 * are maintaining more data than it can retrieve.
7491 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7492 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7493 *statCount = rxi_rpc_peer_stat_cnt;
7496 * This can't happen yet, but in the future version changes
7497 * can be handled by adding additional code here
7501 if (space > (size_t) 0) {
7503 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7506 rx_interface_stat_p rpc_stat, nrpc_stat;
7510 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7512 * We have to fix the offset of rpc_stat since we are
7513 * keeping this structure on two rx_queues. The rx_queue
7514 * package assumes that the rx_queue member is the first
7515 * member of the structure. That is, rx_queue assumes that
7516 * any one item is only on one queue at a time. We are
7517 * breaking that assumption and so we have to do a little
7518 * math to fix our pointers.
7521 fix_offset = (char *)rpc_stat;
7522 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7523 rpc_stat = (rx_interface_stat_p) fix_offset;
7526 * Copy the data based upon the caller version
7528 rx_MarshallProcessRPCStats(callerVersion,
7529 rpc_stat->stats[0].func_total,
7530 rpc_stat->stats, &ptr);
7536 MUTEX_EXIT(&rx_rpc_stats);
7541 * rx_FreeRPCStats - free memory allocated by
7542 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7546 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7547 * rx_RetrievePeerRPCStats
7549 * IN allocSize - the number of bytes in stats.
7557 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7559 rxi_Free(stats, allocSize);
7563 * rx_queryProcessRPCStats - see if process rpc stat collection is
7564 * currently enabled.
7570 * Returns 0 if stats are not enabled != 0 otherwise
7574 rx_queryProcessRPCStats(void)
7577 MUTEX_ENTER(&rx_rpc_stats);
7578 rc = rxi_monitor_processStats;
7579 MUTEX_EXIT(&rx_rpc_stats);
7584 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7590 * Returns 0 if stats are not enabled != 0 otherwise
7594 rx_queryPeerRPCStats(void)
7597 MUTEX_ENTER(&rx_rpc_stats);
7598 rc = rxi_monitor_peerStats;
7599 MUTEX_EXIT(&rx_rpc_stats);
7604 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7614 rx_enableProcessRPCStats(void)
7616 MUTEX_ENTER(&rx_rpc_stats);
7617 rx_enable_stats = 1;
7618 rxi_monitor_processStats = 1;
7619 MUTEX_EXIT(&rx_rpc_stats);
7623 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7633 rx_enablePeerRPCStats(void)
7635 MUTEX_ENTER(&rx_rpc_stats);
7636 rx_enable_stats = 1;
7637 rxi_monitor_peerStats = 1;
7638 MUTEX_EXIT(&rx_rpc_stats);
7642 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7652 rx_disableProcessRPCStats(void)
7654 rx_interface_stat_p rpc_stat, nrpc_stat;
7657 MUTEX_ENTER(&rx_rpc_stats);
7660 * Turn off process statistics and if peer stats is also off, turn
7664 rxi_monitor_processStats = 0;
7665 if (rxi_monitor_peerStats == 0) {
7666 rx_enable_stats = 0;
7669 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7670 unsigned int num_funcs = 0;
7673 queue_Remove(rpc_stat);
7674 num_funcs = rpc_stat->stats[0].func_total;
7676 sizeof(rx_interface_stat_t) +
7677 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7679 rxi_Free(rpc_stat, space);
7680 rxi_rpc_process_stat_cnt -= num_funcs;
7682 MUTEX_EXIT(&rx_rpc_stats);
7686 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7696 rx_disablePeerRPCStats(void)
7698 struct rx_peer **peer_ptr, **peer_end;
7701 MUTEX_ENTER(&rx_rpc_stats);
7704 * Turn off peer statistics and if process stats is also off, turn
7708 rxi_monitor_peerStats = 0;
7709 if (rxi_monitor_processStats == 0) {
7710 rx_enable_stats = 0;
7713 MUTEX_ENTER(&rx_peerHashTable_lock);
7714 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7715 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7717 struct rx_peer *peer, *next, *prev;
7718 for (prev = peer = *peer_ptr; peer; peer = next) {
7720 code = MUTEX_TRYENTER(&peer->peer_lock);
7722 rx_interface_stat_p rpc_stat, nrpc_stat;
7725 (&peer->rpcStats, rpc_stat, nrpc_stat,
7726 rx_interface_stat)) {
7727 unsigned int num_funcs = 0;
7730 queue_Remove(&rpc_stat->queue_header);
7731 queue_Remove(&rpc_stat->all_peers);
7732 num_funcs = rpc_stat->stats[0].func_total;
7734 sizeof(rx_interface_stat_t) +
7735 rpc_stat->stats[0].func_total *
7736 sizeof(rx_function_entry_v1_t);
7738 rxi_Free(rpc_stat, space);
7739 rxi_rpc_peer_stat_cnt -= num_funcs;
7741 MUTEX_EXIT(&peer->peer_lock);
7742 if (prev == *peer_ptr) {
7752 MUTEX_EXIT(&rx_peerHashTable_lock);
7753 MUTEX_EXIT(&rx_rpc_stats);
7757 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7762 * IN clearFlag - flag indicating which stats to clear
7770 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7772 rx_interface_stat_p rpc_stat, nrpc_stat;
7774 MUTEX_ENTER(&rx_rpc_stats);
7776 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7777 unsigned int num_funcs = 0, i;
7778 num_funcs = rpc_stat->stats[0].func_total;
7779 for (i = 0; i < num_funcs; i++) {
7780 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7781 hzero(rpc_stat->stats[i].invocations);
7783 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7784 hzero(rpc_stat->stats[i].bytes_sent);
7786 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7787 hzero(rpc_stat->stats[i].bytes_rcvd);
7789 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7790 rpc_stat->stats[i].queue_time_sum.sec = 0;
7791 rpc_stat->stats[i].queue_time_sum.usec = 0;
7793 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7794 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7795 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7797 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7798 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7799 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7801 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7802 rpc_stat->stats[i].queue_time_max.sec = 0;
7803 rpc_stat->stats[i].queue_time_max.usec = 0;
7805 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7806 rpc_stat->stats[i].execution_time_sum.sec = 0;
7807 rpc_stat->stats[i].execution_time_sum.usec = 0;
7809 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7810 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7811 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7813 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7814 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7815 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7817 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7818 rpc_stat->stats[i].execution_time_max.sec = 0;
7819 rpc_stat->stats[i].execution_time_max.usec = 0;
7824 MUTEX_EXIT(&rx_rpc_stats);
7828 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7833 * IN clearFlag - flag indicating which stats to clear
7841 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7843 rx_interface_stat_p rpc_stat, nrpc_stat;
7845 MUTEX_ENTER(&rx_rpc_stats);
7847 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7848 unsigned int num_funcs = 0, i;
7851 * We have to fix the offset of rpc_stat since we are
7852 * keeping this structure on two rx_queues. The rx_queue
7853 * package assumes that the rx_queue member is the first
7854 * member of the structure. That is, rx_queue assumes that
7855 * any one item is only on one queue at a time. We are
7856 * breaking that assumption and so we have to do a little
7857 * math to fix our pointers.
7860 fix_offset = (char *)rpc_stat;
7861 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7862 rpc_stat = (rx_interface_stat_p) fix_offset;
7864 num_funcs = rpc_stat->stats[0].func_total;
7865 for (i = 0; i < num_funcs; i++) {
7866 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7867 hzero(rpc_stat->stats[i].invocations);
7869 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7870 hzero(rpc_stat->stats[i].bytes_sent);
7872 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7873 hzero(rpc_stat->stats[i].bytes_rcvd);
7875 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7876 rpc_stat->stats[i].queue_time_sum.sec = 0;
7877 rpc_stat->stats[i].queue_time_sum.usec = 0;
7879 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7880 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7881 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7883 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7884 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7885 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7887 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7888 rpc_stat->stats[i].queue_time_max.sec = 0;
7889 rpc_stat->stats[i].queue_time_max.usec = 0;
7891 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7892 rpc_stat->stats[i].execution_time_sum.sec = 0;
7893 rpc_stat->stats[i].execution_time_sum.usec = 0;
7895 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7896 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7897 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7899 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7900 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7901 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7903 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7904 rpc_stat->stats[i].execution_time_max.sec = 0;
7905 rpc_stat->stats[i].execution_time_max.usec = 0;
7910 MUTEX_EXIT(&rx_rpc_stats);
7914 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7915 * is authorized to enable/disable/clear RX statistics.
7917 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7920 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7922 rxi_rxstat_userok = proc;
7926 rx_RxStatUserOk(struct rx_call *call)
7928 if (!rxi_rxstat_userok)
7930 return rxi_rxstat_userok(call);
7935 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
7936 * function in the MSVC runtime DLL (msvcrt.dll).
7938 * Note: the system serializes calls to this function.
7941 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
7942 DWORD reason, /* reason function is being called */
7943 LPVOID reserved) /* reserved for future use */
7946 case DLL_PROCESS_ATTACH:
7947 /* library is being attached to a process */
7951 case DLL_PROCESS_DETACH: