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
115 int (*registerProgram) (pid_t, char *) = 0;
116 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
119 int (*registerProgram) (PROCESS, char *) = 0;
120 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
124 /* Local static routines */
125 static void rxi_DestroyConnectionNoLock(register struct rx_connection *conn);
126 #ifdef RX_ENABLE_LOCKS
127 static void rxi_SetAcksInTransmitQueue(register struct rx_call *call);
130 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
132 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
133 afs_int32 rxi_start_in_error;
135 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
138 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
139 * currently allocated within rx. This number is used to allocate the
140 * memory required to return the statistics when queried.
143 static unsigned int rxi_rpc_peer_stat_cnt;
146 * rxi_rpc_process_stat_cnt counts the total number of local process stat
147 * structures currently allocated within rx. The number is used to allocate
148 * the memory required to return the statistics when queried.
151 static unsigned int rxi_rpc_process_stat_cnt;
153 #if !defined(offsetof)
154 #include <stddef.h> /* for definition of offsetof() */
157 #ifdef AFS_PTHREAD_ENV
161 * Use procedural initialization of mutexes/condition variables
165 extern pthread_mutex_t rx_stats_mutex;
166 extern pthread_mutex_t des_init_mutex;
167 extern pthread_mutex_t des_random_mutex;
168 extern pthread_mutex_t rx_clock_mutex;
169 extern pthread_mutex_t rxi_connCacheMutex;
170 extern pthread_mutex_t rx_event_mutex;
171 extern pthread_mutex_t osi_malloc_mutex;
172 extern pthread_mutex_t event_handler_mutex;
173 extern pthread_mutex_t listener_mutex;
174 extern pthread_mutex_t rx_if_init_mutex;
175 extern pthread_mutex_t rx_if_mutex;
176 extern pthread_mutex_t rxkad_client_uid_mutex;
177 extern pthread_mutex_t rxkad_random_mutex;
179 extern pthread_cond_t rx_event_handler_cond;
180 extern pthread_cond_t rx_listener_cond;
182 static pthread_mutex_t epoch_mutex;
183 static pthread_mutex_t rx_init_mutex;
184 static pthread_mutex_t rx_debug_mutex;
185 static pthread_mutex_t rx_rpc_stats;
188 rxi_InitPthread(void)
190 assert(pthread_mutex_init(&rx_clock_mutex, (const pthread_mutexattr_t *)0)
192 assert(pthread_mutex_init(&rx_stats_mutex, (const pthread_mutexattr_t *)0)
194 assert(pthread_mutex_init
195 (&rxi_connCacheMutex, (const pthread_mutexattr_t *)0) == 0);
196 assert(pthread_mutex_init(&rx_init_mutex, (const pthread_mutexattr_t *)0)
198 assert(pthread_mutex_init(&epoch_mutex, (const pthread_mutexattr_t *)0) ==
200 assert(pthread_mutex_init(&rx_event_mutex, (const pthread_mutexattr_t *)0)
202 assert(pthread_mutex_init(&des_init_mutex, (const pthread_mutexattr_t *)0)
204 assert(pthread_mutex_init
205 (&des_random_mutex, (const pthread_mutexattr_t *)0) == 0);
206 assert(pthread_mutex_init
207 (&osi_malloc_mutex, (const pthread_mutexattr_t *)0) == 0);
208 assert(pthread_mutex_init
209 (&event_handler_mutex, (const pthread_mutexattr_t *)0) == 0);
210 assert(pthread_mutex_init(&listener_mutex, (const pthread_mutexattr_t *)0)
212 assert(pthread_mutex_init
213 (&rx_if_init_mutex, (const pthread_mutexattr_t *)0) == 0);
214 assert(pthread_mutex_init(&rx_if_mutex, (const pthread_mutexattr_t *)0) ==
216 assert(pthread_mutex_init
217 (&rxkad_client_uid_mutex, (const pthread_mutexattr_t *)0) == 0);
218 assert(pthread_mutex_init
219 (&rxkad_random_mutex, (const pthread_mutexattr_t *)0) == 0);
220 assert(pthread_mutex_init(&rx_debug_mutex, (const pthread_mutexattr_t *)0)
223 assert(pthread_cond_init
224 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
225 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
227 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
228 assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
230 rxkad_global_stats_init();
232 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
233 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
234 #ifdef RX_ENABLE_LOCKS
237 #endif /* RX_LOCKS_DB */
238 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
239 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
241 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
243 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
245 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
247 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
248 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
249 #endif /* RX_ENABLE_LOCKS */
252 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
253 #define INIT_PTHREAD_LOCKS \
254 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
256 * The rx_stats_mutex mutex protects the following global variables:
261 * rxi_lowConnRefCount
262 * rxi_lowPeerRefCount
271 #define INIT_PTHREAD_LOCKS
275 /* Variables for handling the minProcs implementation. availProcs gives the
276 * number of threads available in the pool at this moment (not counting dudes
277 * executing right now). totalMin gives the total number of procs required
278 * for handling all minProcs requests. minDeficit is a dynamic variable
279 * tracking the # of procs required to satisfy all of the remaining minProcs
281 * For fine grain locking to work, the quota check and the reservation of
282 * a server thread has to come while rxi_availProcs and rxi_minDeficit
283 * are locked. To this end, the code has been modified under #ifdef
284 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
285 * same time. A new function, ReturnToServerPool() returns the allocation.
287 * A call can be on several queue's (but only one at a time). When
288 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
289 * that no one else is touching the queue. To this end, we store the address
290 * of the queue lock in the call structure (under the call lock) when we
291 * put the call on a queue, and we clear the call_queue_lock when the
292 * call is removed from a queue (once the call lock has been obtained).
293 * This allows rxi_ResetCall to safely synchronize with others wishing
294 * to manipulate the queue.
297 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
298 static afs_kmutex_t rx_rpc_stats;
299 void rxi_StartUnlocked(struct rxevent *event, void *call,
300 void *arg1, int istack);
303 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
304 ** pretty good that the next packet coming in is from the same connection
305 ** as the last packet, since we're send multiple packets in a transmit window.
307 struct rx_connection *rxLastConn = 0;
309 #ifdef RX_ENABLE_LOCKS
310 /* The locking hierarchy for rx fine grain locking is composed of these
313 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
314 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
315 * call->lock - locks call data fields.
316 * These are independent of each other:
317 * rx_freeCallQueue_lock
322 * serverQueueEntry->lock
324 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
325 * peer->lock - locks peer data fields.
326 * conn_data_lock - that more than one thread is not updating a conn data
327 * field at the same time.
335 * Do we need a lock to protect the peer field in the conn structure?
336 * conn->peer was previously a constant for all intents and so has no
337 * lock protecting this field. The multihomed client delta introduced
338 * a RX code change : change the peer field in the connection structure
339 * to that remote inetrface from which the last packet for this
340 * connection was sent out. This may become an issue if further changes
343 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
344 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
346 /* rxdb_fileID is used to identify the lock location, along with line#. */
347 static int rxdb_fileID = RXDB_FILE_RX;
348 #endif /* RX_LOCKS_DB */
349 #else /* RX_ENABLE_LOCKS */
350 #define SET_CALL_QUEUE_LOCK(C, L)
351 #define CLEAR_CALL_QUEUE_LOCK(C)
352 #endif /* RX_ENABLE_LOCKS */
353 struct rx_serverQueueEntry *rx_waitForPacket = 0;
354 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
356 /* ------------Exported Interfaces------------- */
358 /* This function allows rxkad to set the epoch to a suitably random number
359 * which rx_NewConnection will use in the future. The principle purpose is to
360 * get rxnull connections to use the same epoch as the rxkad connections do, at
361 * least once the first rxkad connection is established. This is important now
362 * that the host/port addresses aren't used in FindConnection: the uniqueness
363 * of epoch/cid matters and the start time won't do. */
365 #ifdef AFS_PTHREAD_ENV
367 * This mutex protects the following global variables:
371 #define LOCK_EPOCH assert(pthread_mutex_lock(&epoch_mutex)==0)
372 #define UNLOCK_EPOCH assert(pthread_mutex_unlock(&epoch_mutex)==0)
376 #endif /* AFS_PTHREAD_ENV */
379 rx_SetEpoch(afs_uint32 epoch)
386 /* Initialize rx. A port number may be mentioned, in which case this
387 * becomes the default port number for any service installed later.
388 * If 0 is provided for the port number, a random port will be chosen
389 * by the kernel. Whether this will ever overlap anything in
390 * /etc/services is anybody's guess... Returns 0 on success, -1 on
395 int rxinit_status = 1;
396 #ifdef AFS_PTHREAD_ENV
398 * This mutex protects the following global variables:
402 #define LOCK_RX_INIT assert(pthread_mutex_lock(&rx_init_mutex)==0)
403 #define UNLOCK_RX_INIT assert(pthread_mutex_unlock(&rx_init_mutex)==0)
406 #define UNLOCK_RX_INIT
410 rx_InitHost(u_int host, u_int port)
417 char *htable, *ptable;
424 if (rxinit_status == 0) {
425 tmp_status = rxinit_status;
427 return tmp_status; /* Already started; return previous error code. */
433 if (afs_winsockInit() < 0)
439 * Initialize anything necessary to provide a non-premptive threading
442 rxi_InitializeThreadSupport();
445 /* Allocate and initialize a socket for client and perhaps server
448 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
449 if (rx_socket == OSI_NULLSOCKET) {
453 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
456 #endif /* RX_LOCKS_DB */
457 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
458 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
459 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
460 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
461 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
463 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
465 MUTEX_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
467 MUTEX_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
469 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
470 #if defined(AFS_HPUX110_ENV)
472 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
473 #endif /* AFS_HPUX110_ENV */
474 #endif /* RX_ENABLE_LOCKS && KERNEL */
477 rx_connDeadTime = 12;
478 rx_tranquil = 0; /* reset flag */
479 memset((char *)&rx_stats, 0, sizeof(struct rx_statistics));
481 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
482 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
483 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
484 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
485 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
486 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
488 /* Malloc up a bunch of packets & buffers */
490 queue_Init(&rx_freePacketQueue);
491 rxi_NeedMorePackets = FALSE;
492 #ifdef RX_ENABLE_TSFPQ
493 rx_nPackets = 0; /* in TSFPQ version, rx_nPackets is managed by rxi_MorePackets* */
494 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
495 #else /* RX_ENABLE_TSFPQ */
496 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
497 rxi_MorePackets(rx_nPackets);
498 #endif /* RX_ENABLE_TSFPQ */
505 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
506 tv.tv_sec = clock_now.sec;
507 tv.tv_usec = clock_now.usec;
508 srand((unsigned int)tv.tv_usec);
515 #if defined(KERNEL) && !defined(UKERNEL)
516 /* Really, this should never happen in a real kernel */
519 struct sockaddr_in addr;
521 int addrlen = sizeof(addr);
523 socklen_t addrlen = sizeof(addr);
525 if (getsockname((int)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
529 rx_port = addr.sin_port;
532 rx_stats.minRtt.sec = 9999999;
534 rx_SetEpoch(tv.tv_sec | 0x80000000);
536 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
537 * will provide a randomer value. */
539 MUTEX_ENTER(&rx_stats_mutex);
540 rxi_dataQuota += rx_extraQuota; /* + extra pkts caller asked to rsrv */
541 MUTEX_EXIT(&rx_stats_mutex);
542 /* *Slightly* random start time for the cid. This is just to help
543 * out with the hashing function at the peer */
544 rx_nextCid = ((tv.tv_sec ^ tv.tv_usec) << RX_CIDSHIFT);
545 rx_connHashTable = (struct rx_connection **)htable;
546 rx_peerHashTable = (struct rx_peer **)ptable;
548 rx_lastAckDelay.sec = 0;
549 rx_lastAckDelay.usec = 400000; /* 400 milliseconds */
550 rx_hardAckDelay.sec = 0;
551 rx_hardAckDelay.usec = 100000; /* 100 milliseconds */
552 rx_softAckDelay.sec = 0;
553 rx_softAckDelay.usec = 100000; /* 100 milliseconds */
555 rxevent_Init(20, rxi_ReScheduleEvents);
557 /* Initialize various global queues */
558 queue_Init(&rx_idleServerQueue);
559 queue_Init(&rx_incomingCallQueue);
560 queue_Init(&rx_freeCallQueue);
562 #if defined(AFS_NT40_ENV) && !defined(KERNEL)
563 /* Initialize our list of usable IP addresses. */
567 /* Start listener process (exact function is dependent on the
568 * implementation environment--kernel or user space) */
572 tmp_status = rxinit_status = 0;
580 return rx_InitHost(htonl(INADDR_ANY), port);
583 /* called with unincremented nRequestsRunning to see if it is OK to start
584 * a new thread in this service. Could be "no" for two reasons: over the
585 * max quota, or would prevent others from reaching their min quota.
587 #ifdef RX_ENABLE_LOCKS
588 /* This verion of QuotaOK reserves quota if it's ok while the
589 * rx_serverPool_lock is held. Return quota using ReturnToServerPool().
592 QuotaOK(register struct rx_service *aservice)
594 /* check if over max quota */
595 if (aservice->nRequestsRunning >= aservice->maxProcs) {
599 /* under min quota, we're OK */
600 /* otherwise, can use only if there are enough to allow everyone
601 * to go to their min quota after this guy starts.
603 MUTEX_ENTER(&rx_stats_mutex);
604 if ((aservice->nRequestsRunning < aservice->minProcs)
605 || (rxi_availProcs > rxi_minDeficit)) {
606 aservice->nRequestsRunning++;
607 /* just started call in minProcs pool, need fewer to maintain
609 if (aservice->nRequestsRunning <= aservice->minProcs)
612 MUTEX_EXIT(&rx_stats_mutex);
615 MUTEX_EXIT(&rx_stats_mutex);
621 ReturnToServerPool(register struct rx_service *aservice)
623 aservice->nRequestsRunning--;
624 MUTEX_ENTER(&rx_stats_mutex);
625 if (aservice->nRequestsRunning < aservice->minProcs)
628 MUTEX_EXIT(&rx_stats_mutex);
631 #else /* RX_ENABLE_LOCKS */
633 QuotaOK(register struct rx_service *aservice)
636 /* under min quota, we're OK */
637 if (aservice->nRequestsRunning < aservice->minProcs)
640 /* check if over max quota */
641 if (aservice->nRequestsRunning >= aservice->maxProcs)
644 /* otherwise, can use only if there are enough to allow everyone
645 * to go to their min quota after this guy starts.
647 if (rxi_availProcs > rxi_minDeficit)
651 #endif /* RX_ENABLE_LOCKS */
654 /* Called by rx_StartServer to start up lwp's to service calls.
655 NExistingProcs gives the number of procs already existing, and which
656 therefore needn't be created. */
658 rxi_StartServerProcs(int nExistingProcs)
660 register struct rx_service *service;
665 /* For each service, reserve N processes, where N is the "minimum"
666 * number of processes that MUST be able to execute a request in parallel,
667 * at any time, for that process. Also compute the maximum difference
668 * between any service's maximum number of processes that can run
669 * (i.e. the maximum number that ever will be run, and a guarantee
670 * that this number will run if other services aren't running), and its
671 * minimum number. The result is the extra number of processes that
672 * we need in order to provide the latter guarantee */
673 for (i = 0; i < RX_MAX_SERVICES; i++) {
675 service = rx_services[i];
676 if (service == (struct rx_service *)0)
678 nProcs += service->minProcs;
679 diff = service->maxProcs - service->minProcs;
683 nProcs += maxdiff; /* Extra processes needed to allow max number requested to run in any given service, under good conditions */
684 nProcs -= nExistingProcs; /* Subtract the number of procs that were previously created for use as server procs */
685 for (i = 0; i < nProcs; i++) {
686 rxi_StartServerProc(rx_ServerProc, rx_stackSize);
692 /* This routine is only required on Windows */
694 rx_StartClientThread(void)
696 #ifdef AFS_PTHREAD_ENV
698 pid = pthread_self();
699 #endif /* AFS_PTHREAD_ENV */
701 #endif /* AFS_NT40_ENV */
703 /* This routine must be called if any services are exported. If the
704 * donateMe flag is set, the calling process is donated to the server
707 rx_StartServer(int donateMe)
709 register struct rx_service *service;
715 /* Start server processes, if necessary (exact function is dependent
716 * on the implementation environment--kernel or user space). DonateMe
717 * will be 1 if there is 1 pre-existing proc, i.e. this one. In this
718 * case, one less new proc will be created rx_StartServerProcs.
720 rxi_StartServerProcs(donateMe);
722 /* count up the # of threads in minProcs, and add set the min deficit to
723 * be that value, too.
725 for (i = 0; i < RX_MAX_SERVICES; i++) {
726 service = rx_services[i];
727 if (service == (struct rx_service *)0)
729 MUTEX_ENTER(&rx_stats_mutex);
730 rxi_totalMin += service->minProcs;
731 /* below works even if a thread is running, since minDeficit would
732 * still have been decremented and later re-incremented.
734 rxi_minDeficit += service->minProcs;
735 MUTEX_EXIT(&rx_stats_mutex);
738 /* Turn on reaping of idle server connections */
739 rxi_ReapConnections(NULL, NULL, NULL);
748 #ifdef AFS_PTHREAD_ENV
750 pid = (pid_t) pthread_self();
751 #else /* AFS_PTHREAD_ENV */
753 LWP_CurrentProcess(&pid);
754 #endif /* AFS_PTHREAD_ENV */
756 sprintf(name, "srv_%d", ++nProcs);
758 (*registerProgram) (pid, name);
760 #endif /* AFS_NT40_ENV */
761 rx_ServerProc(NULL); /* Never returns */
763 #ifdef RX_ENABLE_TSFPQ
764 /* no use leaving packets around in this thread's local queue if
765 * it isn't getting donated to the server thread pool.
767 rxi_FlushLocalPacketsTSFPQ();
768 #endif /* RX_ENABLE_TSFPQ */
772 /* Create a new client connection to the specified service, using the
773 * specified security object to implement the security model for this
775 struct rx_connection *
776 rx_NewConnection(register afs_uint32 shost, u_short sport, u_short sservice,
777 register struct rx_securityClass *securityObject,
778 int serviceSecurityIndex)
781 afs_int32 cid, cix, nclones;
782 register struct rx_connection *conn, *tconn, *ptconn;
787 dpf(("rx_NewConnection(host %x, port %u, service %u, securityObject %x, serviceSecurityIndex %d)\n", ntohl(shost), ntohs(sport), sservice, securityObject, serviceSecurityIndex));
790 nclones = rx_max_clones_per_connection;
792 /* Vasilsi said: "NETPRI protects Cid and Alloc", but can this be true in
793 * the case of kmem_alloc? */
796 MUTEX_ENTER(&rx_connHashTable_lock);
798 /* send in the clones */
799 for(cix = 0; cix <= nclones; ++cix) {
802 tconn = rxi_AllocConnection();
803 tconn->type = RX_CLIENT_CONNECTION;
804 tconn->epoch = rx_epoch;
805 tconn->peer = rxi_FindPeer(shost, sport, 0, 1);
806 tconn->serviceId = sservice;
807 tconn->securityObject = securityObject;
808 tconn->securityData = (void *) 0;
809 tconn->securityIndex = serviceSecurityIndex;
810 tconn->ackRate = RX_FAST_ACK_RATE;
811 tconn->nSpecific = 0;
812 tconn->specific = NULL;
813 tconn->challengeEvent = NULL;
814 tconn->delayedAbortEvent = NULL;
815 tconn->abortCount = 0;
817 for (i = 0; i < RX_MAXCALLS; i++) {
818 tconn->twind[i] = rx_initSendWindow;
819 tconn->rwind[i] = rx_initReceiveWindow;
822 tconn->next_clone = 0;
823 tconn->nclones = nclones;
824 rx_SetConnDeadTime(tconn, rx_connDeadTime);
829 tconn->flags |= RX_CLONED_CONNECTION;
830 tconn->parent = conn;
831 ptconn->next_clone = tconn;
834 /* generic connection setup */
835 #ifdef RX_ENABLE_LOCKS
836 MUTEX_INIT(&tconn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
837 MUTEX_INIT(&tconn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
838 CV_INIT(&tconn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
840 cid = (rx_nextCid += RX_MAXCALLS);
842 RXS_NewConnection(securityObject, tconn);
844 CONN_HASH(shost, sport, tconn->cid, tconn->epoch,
845 RX_CLIENT_CONNECTION);
846 tconn->refCount++; /* no lock required since only this thread knows */
847 tconn->next = rx_connHashTable[hashindex];
848 rx_connHashTable[hashindex] = tconn;
849 rx_MutexIncrement(rx_stats.nClientConns, rx_stats_mutex);
852 MUTEX_EXIT(&rx_connHashTable_lock);
858 rx_SetConnDeadTime(register struct rx_connection *conn, register int seconds)
860 /* The idea is to set the dead time to a value that allows several
861 * keepalives to be dropped without timing out the connection. */
862 struct rx_connection *tconn;
865 tconn->secondsUntilDead = MAX(seconds, 6);
866 tconn->secondsUntilPing = tconn->secondsUntilDead / 6;
867 } while(tconn->next_clone && (tconn = tconn->next_clone));
870 int rxi_lowPeerRefCount = 0;
871 int rxi_lowConnRefCount = 0;
874 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
875 * NOTE: must not be called with rx_connHashTable_lock held.
878 rxi_CleanupConnection(struct rx_connection *conn)
880 /* Notify the service exporter, if requested, that this connection
881 * is being destroyed */
882 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
883 (*conn->service->destroyConnProc) (conn);
885 /* Notify the security module that this connection is being destroyed */
886 RXS_DestroyConnection(conn->securityObject, conn);
888 /* If this is the last connection using the rx_peer struct, set its
889 * idle time to now. rxi_ReapConnections will reap it if it's still
890 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
892 MUTEX_ENTER(&rx_peerHashTable_lock);
893 if (conn->peer->refCount < 2) {
894 conn->peer->idleWhen = clock_Sec();
895 if (conn->peer->refCount < 1) {
896 conn->peer->refCount = 1;
897 MUTEX_ENTER(&rx_stats_mutex);
898 rxi_lowPeerRefCount++;
899 MUTEX_EXIT(&rx_stats_mutex);
902 conn->peer->refCount--;
903 MUTEX_EXIT(&rx_peerHashTable_lock);
905 if (conn->type == RX_SERVER_CONNECTION)
906 rx_MutexDecrement(rx_stats.nServerConns, rx_stats_mutex);
908 rx_MutexDecrement(rx_stats.nClientConns, rx_stats_mutex);
910 if (conn->specific) {
912 for (i = 0; i < conn->nSpecific; i++) {
913 if (conn->specific[i] && rxi_keyCreate_destructor[i])
914 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
915 conn->specific[i] = NULL;
917 free(conn->specific);
919 conn->specific = NULL;
923 MUTEX_DESTROY(&conn->conn_call_lock);
924 MUTEX_DESTROY(&conn->conn_data_lock);
925 CV_DESTROY(&conn->conn_call_cv);
927 rxi_FreeConnection(conn);
930 /* Destroy the specified connection */
932 rxi_DestroyConnection(register struct rx_connection *conn)
934 register struct rx_connection *tconn, *dtconn;
936 MUTEX_ENTER(&rx_connHashTable_lock);
938 if(!(conn->flags & RX_CLONED_CONNECTION)) {
939 tconn = conn->next_clone;
940 conn->next_clone = 0; /* once */
944 tconn = tconn->next_clone;
945 rxi_DestroyConnectionNoLock(dtconn);
947 if (dtconn == rx_connCleanup_list) {
948 rx_connCleanup_list = rx_connCleanup_list->next;
949 MUTEX_EXIT(&rx_connHashTable_lock);
950 /* rxi_CleanupConnection will free tconn */
951 rxi_CleanupConnection(dtconn);
952 MUTEX_ENTER(&rx_connHashTable_lock);
959 rxi_DestroyConnectionNoLock(conn);
960 /* conn should be at the head of the cleanup list */
961 if (conn == rx_connCleanup_list) {
962 rx_connCleanup_list = rx_connCleanup_list->next;
963 MUTEX_EXIT(&rx_connHashTable_lock);
964 rxi_CleanupConnection(conn);
966 #ifdef RX_ENABLE_LOCKS
968 MUTEX_EXIT(&rx_connHashTable_lock);
970 #endif /* RX_ENABLE_LOCKS */
974 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
976 register struct rx_connection **conn_ptr;
977 register int havecalls = 0;
978 struct rx_packet *packet;
985 MUTEX_ENTER(&conn->conn_data_lock);
986 if (conn->refCount > 0)
989 MUTEX_ENTER(&rx_stats_mutex);
990 rxi_lowConnRefCount++;
991 MUTEX_EXIT(&rx_stats_mutex);
994 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
995 /* Busy; wait till the last guy before proceeding */
996 MUTEX_EXIT(&conn->conn_data_lock);
1001 /* If the client previously called rx_NewCall, but it is still
1002 * waiting, treat this as a running call, and wait to destroy the
1003 * connection later when the call completes. */
1004 if ((conn->type == RX_CLIENT_CONNECTION)
1005 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
1006 conn->flags |= RX_CONN_DESTROY_ME;
1007 MUTEX_EXIT(&conn->conn_data_lock);
1011 MUTEX_EXIT(&conn->conn_data_lock);
1013 /* Check for extant references to this connection */
1014 for (i = 0; i < RX_MAXCALLS; i++) {
1015 register struct rx_call *call = conn->call[i];
1018 if (conn->type == RX_CLIENT_CONNECTION) {
1019 MUTEX_ENTER(&call->lock);
1020 if (call->delayedAckEvent) {
1021 /* Push the final acknowledgment out now--there
1022 * won't be a subsequent call to acknowledge the
1023 * last reply packets */
1024 rxevent_Cancel(call->delayedAckEvent, call,
1025 RX_CALL_REFCOUNT_DELAY);
1026 if (call->state == RX_STATE_PRECALL
1027 || call->state == RX_STATE_ACTIVE) {
1028 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1030 rxi_AckAll(NULL, call, 0);
1033 MUTEX_EXIT(&call->lock);
1037 #ifdef RX_ENABLE_LOCKS
1039 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1040 MUTEX_EXIT(&conn->conn_data_lock);
1042 /* Someone is accessing a packet right now. */
1046 #endif /* RX_ENABLE_LOCKS */
1049 /* Don't destroy the connection if there are any call
1050 * structures still in use */
1051 MUTEX_ENTER(&conn->conn_data_lock);
1052 conn->flags |= RX_CONN_DESTROY_ME;
1053 MUTEX_EXIT(&conn->conn_data_lock);
1058 if (conn->delayedAbortEvent) {
1059 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
1060 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1062 MUTEX_ENTER(&conn->conn_data_lock);
1063 rxi_SendConnectionAbort(conn, packet, 0, 1);
1064 MUTEX_EXIT(&conn->conn_data_lock);
1065 rxi_FreePacket(packet);
1069 /* Remove from connection hash table before proceeding */
1071 &rx_connHashTable[CONN_HASH
1072 (peer->host, peer->port, conn->cid, conn->epoch,
1074 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1075 if (*conn_ptr == conn) {
1076 *conn_ptr = conn->next;
1080 /* if the conn that we are destroying was the last connection, then we
1081 * clear rxLastConn as well */
1082 if (rxLastConn == conn)
1085 /* Make sure the connection is completely reset before deleting it. */
1086 /* get rid of pending events that could zap us later */
1087 if (conn->challengeEvent)
1088 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1089 if (conn->checkReachEvent)
1090 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1092 /* Add the connection to the list of destroyed connections that
1093 * need to be cleaned up. This is necessary to avoid deadlocks
1094 * in the routines we call to inform others that this connection is
1095 * being destroyed. */
1096 conn->next = rx_connCleanup_list;
1097 rx_connCleanup_list = conn;
1100 /* Externally available version */
1102 rx_DestroyConnection(register struct rx_connection *conn)
1107 rxi_DestroyConnection(conn);
1112 rx_GetConnection(register struct rx_connection *conn)
1117 MUTEX_ENTER(&conn->conn_data_lock);
1119 MUTEX_EXIT(&conn->conn_data_lock);
1123 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1124 /* Wait for the transmit queue to no longer be busy.
1125 * requires the call->lock to be held */
1126 static void rxi_WaitforTQBusy(struct rx_call *call) {
1127 while (call->flags & RX_CALL_TQ_BUSY) {
1128 call->flags |= RX_CALL_TQ_WAIT;
1130 #ifdef RX_ENABLE_LOCKS
1131 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1132 CV_WAIT(&call->cv_tq, &call->lock);
1133 #else /* RX_ENABLE_LOCKS */
1134 osi_rxSleep(&call->tq);
1135 #endif /* RX_ENABLE_LOCKS */
1137 if (call->tqWaiters == 0) {
1138 call->flags &= ~RX_CALL_TQ_WAIT;
1144 /* Start a new rx remote procedure call, on the specified connection.
1145 * If wait is set to 1, wait for a free call channel; otherwise return
1146 * 0. Maxtime gives the maximum number of seconds this call may take,
1147 * after rx_NewCall returns. After this time interval, a call to any
1148 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1149 * For fine grain locking, we hold the conn_call_lock in order to
1150 * to ensure that we don't get signalle after we found a call in an active
1151 * state and before we go to sleep.
1154 rx_NewCall(register struct rx_connection *conn)
1157 register struct rx_call *call;
1158 register struct rx_connection *tconn;
1159 struct clock queueTime;
1163 dpf(("rx_NewCall(conn %x)\n", conn));
1166 clock_GetTime(&queueTime);
1167 MUTEX_ENTER(&conn->conn_call_lock);
1170 * Check if there are others waiting for a new call.
1171 * If so, let them go first to avoid starving them.
1172 * This is a fairly simple scheme, and might not be
1173 * a complete solution for large numbers of waiters.
1175 * makeCallWaiters keeps track of the number of
1176 * threads waiting to make calls and the
1177 * RX_CONN_MAKECALL_WAITING flag bit is used to
1178 * indicate that there are indeed calls waiting.
1179 * The flag is set when the waiter is incremented.
1180 * It is only cleared in rx_EndCall when
1181 * makeCallWaiters is 0. This prevents us from
1182 * accidently destroying the connection while it
1183 * is potentially about to be used.
1185 MUTEX_ENTER(&conn->conn_data_lock);
1186 if (conn->makeCallWaiters) {
1187 conn->flags |= RX_CONN_MAKECALL_WAITING;
1188 conn->makeCallWaiters++;
1189 MUTEX_EXIT(&conn->conn_data_lock);
1191 #ifdef RX_ENABLE_LOCKS
1192 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1196 MUTEX_ENTER(&conn->conn_data_lock);
1197 conn->makeCallWaiters--;
1199 MUTEX_EXIT(&conn->conn_data_lock);
1201 /* search for next free call on this connection or
1202 * its clones, if any */
1206 for (i = 0; i < RX_MAXCALLS; i++) {
1207 call = tconn->call[i];
1209 MUTEX_ENTER(&call->lock);
1210 if (call->state == RX_STATE_DALLY) {
1211 rxi_ResetCall(call, 0);
1212 (*call->callNumber)++;
1215 MUTEX_EXIT(&call->lock);
1217 call = rxi_NewCall(tconn, i);
1220 } /* for i < RX_MAXCALLS */
1221 } while (tconn->next_clone && (tconn = tconn->next_clone));
1225 if (i < RX_MAXCALLS) {
1229 /* to be here, all available calls for this connection (and all
1230 * its clones) must be in use */
1232 MUTEX_ENTER(&conn->conn_data_lock);
1233 conn->flags |= RX_CONN_MAKECALL_WAITING;
1234 conn->makeCallWaiters++;
1235 MUTEX_EXIT(&conn->conn_data_lock);
1237 #ifdef RX_ENABLE_LOCKS
1238 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1242 MUTEX_ENTER(&conn->conn_data_lock);
1243 conn->makeCallWaiters--;
1244 MUTEX_EXIT(&conn->conn_data_lock);
1247 * Wake up anyone else who might be giving us a chance to
1248 * run (see code above that avoids resource starvation).
1250 #ifdef RX_ENABLE_LOCKS
1251 CV_BROADCAST(&conn->conn_call_cv);
1256 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1258 /* Client is initially in send mode */
1259 call->state = RX_STATE_ACTIVE;
1260 call->error = conn->error;
1262 call->mode = RX_MODE_ERROR;
1264 call->mode = RX_MODE_SENDING;
1266 /* remember start time for call in case we have hard dead time limit */
1267 call->queueTime = queueTime;
1268 clock_GetTime(&call->startTime);
1269 hzero(call->bytesSent);
1270 hzero(call->bytesRcvd);
1272 /* Turn on busy protocol. */
1273 rxi_KeepAliveOn(call);
1275 MUTEX_EXIT(&call->lock);
1276 MUTEX_EXIT(&conn->conn_call_lock);
1279 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1280 /* Now, if TQ wasn't cleared earlier, do it now. */
1281 MUTEX_ENTER(&call->lock);
1282 rxi_WaitforTQBusy(call);
1283 if (call->flags & RX_CALL_TQ_CLEARME) {
1284 rxi_ClearTransmitQueue(call, 1);
1285 queue_Init(&call->tq);
1287 MUTEX_EXIT(&call->lock);
1288 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1290 dpf(("rx_NewCall(call %x)\n", call));
1295 rxi_HasActiveCalls(register struct rx_connection *aconn)
1298 register struct rx_call *tcall;
1302 for (i = 0; i < RX_MAXCALLS; i++) {
1303 if ((tcall = aconn->call[i])) {
1304 if ((tcall->state == RX_STATE_ACTIVE)
1305 || (tcall->state == RX_STATE_PRECALL)) {
1316 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1317 register afs_int32 * aint32s)
1320 register struct rx_call *tcall;
1324 for (i = 0; i < RX_MAXCALLS; i++) {
1325 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1326 aint32s[i] = aconn->callNumber[i] + 1;
1328 aint32s[i] = aconn->callNumber[i];
1335 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1336 register afs_int32 * aint32s)
1339 register struct rx_call *tcall;
1343 for (i = 0; i < RX_MAXCALLS; i++) {
1344 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1345 aconn->callNumber[i] = aint32s[i] - 1;
1347 aconn->callNumber[i] = aint32s[i];
1353 /* Advertise a new service. A service is named locally by a UDP port
1354 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1357 char *serviceName; Name for identification purposes (e.g. the
1358 service name might be used for probing for
1361 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1362 char *serviceName, struct rx_securityClass **securityObjects,
1363 int nSecurityObjects,
1364 afs_int32(*serviceProc) (struct rx_call * acall))
1366 osi_socket socket = OSI_NULLSOCKET;
1367 register struct rx_service *tservice;
1373 if (serviceId == 0) {
1375 "rx_NewService: service id for service %s is not non-zero.\n",
1382 "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",
1390 tservice = rxi_AllocService();
1392 for (i = 0; i < RX_MAX_SERVICES; i++) {
1393 register struct rx_service *service = rx_services[i];
1395 if (port == service->servicePort && host == service->serviceHost) {
1396 if (service->serviceId == serviceId) {
1397 /* The identical service has already been
1398 * installed; if the caller was intending to
1399 * change the security classes used by this
1400 * service, he/she loses. */
1402 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1403 serviceName, serviceId, service->serviceName);
1405 rxi_FreeService(tservice);
1408 /* Different service, same port: re-use the socket
1409 * which is bound to the same port */
1410 socket = service->socket;
1413 if (socket == OSI_NULLSOCKET) {
1414 /* If we don't already have a socket (from another
1415 * service on same port) get a new one */
1416 socket = rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
1417 if (socket == OSI_NULLSOCKET) {
1419 rxi_FreeService(tservice);
1424 service->socket = socket;
1425 service->serviceHost = host;
1426 service->servicePort = port;
1427 service->serviceId = serviceId;
1428 service->serviceName = serviceName;
1429 service->nSecurityObjects = nSecurityObjects;
1430 service->securityObjects = securityObjects;
1431 service->minProcs = 0;
1432 service->maxProcs = 1;
1433 service->idleDeadTime = 60;
1434 service->idleDeadErr = 0;
1435 service->connDeadTime = rx_connDeadTime;
1436 service->executeRequestProc = serviceProc;
1437 service->checkReach = 0;
1438 rx_services[i] = service; /* not visible until now */
1444 rxi_FreeService(tservice);
1445 (osi_Msg "rx_NewService: cannot support > %d services\n",
1450 /* Set configuration options for all of a service's security objects */
1453 rx_SetSecurityConfiguration(struct rx_service *service,
1454 rx_securityConfigVariables type,
1458 for (i = 0; i<service->nSecurityObjects; i++) {
1459 if (service->securityObjects[i]) {
1460 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1468 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1469 struct rx_securityClass **securityObjects, int nSecurityObjects,
1470 afs_int32(*serviceProc) (struct rx_call * acall))
1472 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1475 /* Generic request processing loop. This routine should be called
1476 * by the implementation dependent rx_ServerProc. If socketp is
1477 * non-null, it will be set to the file descriptor that this thread
1478 * is now listening on. If socketp is null, this routine will never
1481 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1483 register struct rx_call *call;
1484 register afs_int32 code;
1485 register struct rx_service *tservice = NULL;
1492 call = rx_GetCall(threadID, tservice, socketp);
1493 if (socketp && *socketp != OSI_NULLSOCKET) {
1494 /* We are now a listener thread */
1499 /* if server is restarting( typically smooth shutdown) then do not
1500 * allow any new calls.
1503 if (rx_tranquil && (call != NULL)) {
1507 MUTEX_ENTER(&call->lock);
1509 rxi_CallError(call, RX_RESTARTING);
1510 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1512 MUTEX_EXIT(&call->lock);
1516 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1517 #ifdef RX_ENABLE_LOCKS
1519 #endif /* RX_ENABLE_LOCKS */
1520 afs_termState = AFSOP_STOP_AFS;
1521 afs_osi_Wakeup(&afs_termState);
1522 #ifdef RX_ENABLE_LOCKS
1524 #endif /* RX_ENABLE_LOCKS */
1529 tservice = call->conn->service;
1531 if (tservice->beforeProc)
1532 (*tservice->beforeProc) (call);
1534 code = call->conn->service->executeRequestProc(call);
1536 if (tservice->afterProc)
1537 (*tservice->afterProc) (call, code);
1539 rx_EndCall(call, code);
1540 MUTEX_ENTER(&rx_stats_mutex);
1542 MUTEX_EXIT(&rx_stats_mutex);
1548 rx_WakeupServerProcs(void)
1550 struct rx_serverQueueEntry *np, *tqp;
1554 MUTEX_ENTER(&rx_serverPool_lock);
1556 #ifdef RX_ENABLE_LOCKS
1557 if (rx_waitForPacket)
1558 CV_BROADCAST(&rx_waitForPacket->cv);
1559 #else /* RX_ENABLE_LOCKS */
1560 if (rx_waitForPacket)
1561 osi_rxWakeup(rx_waitForPacket);
1562 #endif /* RX_ENABLE_LOCKS */
1563 MUTEX_ENTER(&freeSQEList_lock);
1564 for (np = rx_FreeSQEList; np; np = tqp) {
1565 tqp = *(struct rx_serverQueueEntry **)np;
1566 #ifdef RX_ENABLE_LOCKS
1567 CV_BROADCAST(&np->cv);
1568 #else /* RX_ENABLE_LOCKS */
1570 #endif /* RX_ENABLE_LOCKS */
1572 MUTEX_EXIT(&freeSQEList_lock);
1573 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1574 #ifdef RX_ENABLE_LOCKS
1575 CV_BROADCAST(&np->cv);
1576 #else /* RX_ENABLE_LOCKS */
1578 #endif /* RX_ENABLE_LOCKS */
1580 MUTEX_EXIT(&rx_serverPool_lock);
1585 * One thing that seems to happen is that all the server threads get
1586 * tied up on some empty or slow call, and then a whole bunch of calls
1587 * arrive at once, using up the packet pool, so now there are more
1588 * empty calls. The most critical resources here are server threads
1589 * and the free packet pool. The "doreclaim" code seems to help in
1590 * general. I think that eventually we arrive in this state: there
1591 * are lots of pending calls which do have all their packets present,
1592 * so they won't be reclaimed, are multi-packet calls, so they won't
1593 * be scheduled until later, and thus are tying up most of the free
1594 * packet pool for a very long time.
1596 * 1. schedule multi-packet calls if all the packets are present.
1597 * Probably CPU-bound operation, useful to return packets to pool.
1598 * Do what if there is a full window, but the last packet isn't here?
1599 * 3. preserve one thread which *only* runs "best" calls, otherwise
1600 * it sleeps and waits for that type of call.
1601 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1602 * the current dataquota business is badly broken. The quota isn't adjusted
1603 * to reflect how many packets are presently queued for a running call.
1604 * So, when we schedule a queued call with a full window of packets queued
1605 * up for it, that *should* free up a window full of packets for other 2d-class
1606 * calls to be able to use from the packet pool. But it doesn't.
1608 * NB. Most of the time, this code doesn't run -- since idle server threads
1609 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1610 * as a new call arrives.
1612 /* Sleep until a call arrives. Returns a pointer to the call, ready
1613 * for an rx_Read. */
1614 #ifdef RX_ENABLE_LOCKS
1616 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1618 struct rx_serverQueueEntry *sq;
1619 register struct rx_call *call = (struct rx_call *)0;
1620 struct rx_service *service = NULL;
1623 MUTEX_ENTER(&freeSQEList_lock);
1625 if ((sq = rx_FreeSQEList)) {
1626 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1627 MUTEX_EXIT(&freeSQEList_lock);
1628 } else { /* otherwise allocate a new one and return that */
1629 MUTEX_EXIT(&freeSQEList_lock);
1630 sq = (struct rx_serverQueueEntry *)
1631 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1632 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1633 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1636 MUTEX_ENTER(&rx_serverPool_lock);
1637 if (cur_service != NULL) {
1638 ReturnToServerPool(cur_service);
1641 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1642 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1644 /* Scan for eligible incoming calls. A call is not eligible
1645 * if the maximum number of calls for its service type are
1646 * already executing */
1647 /* One thread will process calls FCFS (to prevent starvation),
1648 * while the other threads may run ahead looking for calls which
1649 * have all their input data available immediately. This helps
1650 * keep threads from blocking, waiting for data from the client. */
1651 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1652 service = tcall->conn->service;
1653 if (!QuotaOK(service)) {
1656 if (tno == rxi_fcfs_thread_num
1657 || !tcall->queue_item_header.next) {
1658 /* If we're the fcfs thread , then we'll just use
1659 * this call. If we haven't been able to find an optimal
1660 * choice, and we're at the end of the list, then use a
1661 * 2d choice if one has been identified. Otherwise... */
1662 call = (choice2 ? choice2 : tcall);
1663 service = call->conn->service;
1664 } else if (!queue_IsEmpty(&tcall->rq)) {
1665 struct rx_packet *rp;
1666 rp = queue_First(&tcall->rq, rx_packet);
1667 if (rp->header.seq == 1) {
1669 || (rp->header.flags & RX_LAST_PACKET)) {
1671 } else if (rxi_2dchoice && !choice2
1672 && !(tcall->flags & RX_CALL_CLEARED)
1673 && (tcall->rprev > rxi_HardAckRate)) {
1682 ReturnToServerPool(service);
1689 MUTEX_EXIT(&rx_serverPool_lock);
1690 MUTEX_ENTER(&call->lock);
1692 if (call->flags & RX_CALL_WAIT_PROC) {
1693 call->flags &= ~RX_CALL_WAIT_PROC;
1694 MUTEX_ENTER(&rx_stats_mutex);
1696 MUTEX_EXIT(&rx_stats_mutex);
1699 if (call->state != RX_STATE_PRECALL || call->error) {
1700 MUTEX_EXIT(&call->lock);
1701 MUTEX_ENTER(&rx_serverPool_lock);
1702 ReturnToServerPool(service);
1707 if (queue_IsEmpty(&call->rq)
1708 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1709 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1711 CLEAR_CALL_QUEUE_LOCK(call);
1714 /* If there are no eligible incoming calls, add this process
1715 * to the idle server queue, to wait for one */
1719 *socketp = OSI_NULLSOCKET;
1721 sq->socketp = socketp;
1722 queue_Append(&rx_idleServerQueue, sq);
1723 #ifndef AFS_AIX41_ENV
1724 rx_waitForPacket = sq;
1726 rx_waitingForPacket = sq;
1727 #endif /* AFS_AIX41_ENV */
1729 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1731 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1732 MUTEX_EXIT(&rx_serverPool_lock);
1733 return (struct rx_call *)0;
1736 } while (!(call = sq->newcall)
1737 && !(socketp && *socketp != OSI_NULLSOCKET));
1738 MUTEX_EXIT(&rx_serverPool_lock);
1740 MUTEX_ENTER(&call->lock);
1746 MUTEX_ENTER(&freeSQEList_lock);
1747 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1748 rx_FreeSQEList = sq;
1749 MUTEX_EXIT(&freeSQEList_lock);
1752 clock_GetTime(&call->startTime);
1753 call->state = RX_STATE_ACTIVE;
1754 call->mode = RX_MODE_RECEIVING;
1755 #ifdef RX_KERNEL_TRACE
1756 if (ICL_SETACTIVE(afs_iclSetp)) {
1757 int glockOwner = ISAFS_GLOCK();
1760 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1761 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1768 rxi_calltrace(RX_CALL_START, call);
1769 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1770 call->conn->service->servicePort, call->conn->service->serviceId,
1773 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1774 MUTEX_EXIT(&call->lock);
1776 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1781 #else /* RX_ENABLE_LOCKS */
1783 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1785 struct rx_serverQueueEntry *sq;
1786 register struct rx_call *call = (struct rx_call *)0, *choice2;
1787 struct rx_service *service = NULL;
1791 MUTEX_ENTER(&freeSQEList_lock);
1793 if ((sq = rx_FreeSQEList)) {
1794 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1795 MUTEX_EXIT(&freeSQEList_lock);
1796 } else { /* otherwise allocate a new one and return that */
1797 MUTEX_EXIT(&freeSQEList_lock);
1798 sq = (struct rx_serverQueueEntry *)
1799 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1800 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1801 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1803 MUTEX_ENTER(&sq->lock);
1805 if (cur_service != NULL) {
1806 cur_service->nRequestsRunning--;
1807 if (cur_service->nRequestsRunning < cur_service->minProcs)
1811 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1812 register struct rx_call *tcall, *ncall;
1813 /* Scan for eligible incoming calls. A call is not eligible
1814 * if the maximum number of calls for its service type are
1815 * already executing */
1816 /* One thread will process calls FCFS (to prevent starvation),
1817 * while the other threads may run ahead looking for calls which
1818 * have all their input data available immediately. This helps
1819 * keep threads from blocking, waiting for data from the client. */
1820 choice2 = (struct rx_call *)0;
1821 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1822 service = tcall->conn->service;
1823 if (QuotaOK(service)) {
1824 if (tno == rxi_fcfs_thread_num
1825 || !tcall->queue_item_header.next) {
1826 /* If we're the fcfs thread, then we'll just use
1827 * this call. If we haven't been able to find an optimal
1828 * choice, and we're at the end of the list, then use a
1829 * 2d choice if one has been identified. Otherwise... */
1830 call = (choice2 ? choice2 : tcall);
1831 service = call->conn->service;
1832 } else if (!queue_IsEmpty(&tcall->rq)) {
1833 struct rx_packet *rp;
1834 rp = queue_First(&tcall->rq, rx_packet);
1835 if (rp->header.seq == 1
1837 || (rp->header.flags & RX_LAST_PACKET))) {
1839 } else if (rxi_2dchoice && !choice2
1840 && !(tcall->flags & RX_CALL_CLEARED)
1841 && (tcall->rprev > rxi_HardAckRate)) {
1854 /* we can't schedule a call if there's no data!!! */
1855 /* send an ack if there's no data, if we're missing the
1856 * first packet, or we're missing something between first
1857 * and last -- there's a "hole" in the incoming data. */
1858 if (queue_IsEmpty(&call->rq)
1859 || queue_First(&call->rq, rx_packet)->header.seq != 1
1860 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1861 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1863 call->flags &= (~RX_CALL_WAIT_PROC);
1864 service->nRequestsRunning++;
1865 /* just started call in minProcs pool, need fewer to maintain
1867 if (service->nRequestsRunning <= service->minProcs)
1871 /* MUTEX_EXIT(&call->lock); */
1873 /* If there are no eligible incoming calls, add this process
1874 * to the idle server queue, to wait for one */
1877 *socketp = OSI_NULLSOCKET;
1879 sq->socketp = socketp;
1880 queue_Append(&rx_idleServerQueue, sq);
1884 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1886 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1887 return (struct rx_call *)0;
1890 } while (!(call = sq->newcall)
1891 && !(socketp && *socketp != OSI_NULLSOCKET));
1893 MUTEX_EXIT(&sq->lock);
1895 MUTEX_ENTER(&freeSQEList_lock);
1896 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1897 rx_FreeSQEList = sq;
1898 MUTEX_EXIT(&freeSQEList_lock);
1901 clock_GetTime(&call->startTime);
1902 call->state = RX_STATE_ACTIVE;
1903 call->mode = RX_MODE_RECEIVING;
1904 #ifdef RX_KERNEL_TRACE
1905 if (ICL_SETACTIVE(afs_iclSetp)) {
1906 int glockOwner = ISAFS_GLOCK();
1909 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1910 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1917 rxi_calltrace(RX_CALL_START, call);
1918 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1919 call->conn->service->servicePort, call->conn->service->serviceId,
1922 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1929 #endif /* RX_ENABLE_LOCKS */
1933 /* Establish a procedure to be called when a packet arrives for a
1934 * call. This routine will be called at most once after each call,
1935 * and will also be called if there is an error condition on the or
1936 * the call is complete. Used by multi rx to build a selection
1937 * function which determines which of several calls is likely to be a
1938 * good one to read from.
1939 * NOTE: the way this is currently implemented it is probably only a
1940 * good idea to (1) use it immediately after a newcall (clients only)
1941 * and (2) only use it once. Other uses currently void your warranty
1944 rx_SetArrivalProc(register struct rx_call *call,
1945 register void (*proc) (register struct rx_call * call,
1947 register int index),
1948 register void * handle, register int arg)
1950 call->arrivalProc = proc;
1951 call->arrivalProcHandle = handle;
1952 call->arrivalProcArg = arg;
1955 /* Call is finished (possibly prematurely). Return rc to the peer, if
1956 * appropriate, and return the final error code from the conversation
1960 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1962 register struct rx_connection *conn = call->conn;
1963 register struct rx_service *service;
1969 dpf(("rx_EndCall(call %x rc %d error %d abortCode %d)\n", call, rc, call->error, call->abortCode));
1972 MUTEX_ENTER(&call->lock);
1974 if (rc == 0 && call->error == 0) {
1975 call->abortCode = 0;
1976 call->abortCount = 0;
1979 call->arrivalProc = (void (*)())0;
1980 if (rc && call->error == 0) {
1981 rxi_CallError(call, rc);
1982 /* Send an abort message to the peer if this error code has
1983 * only just been set. If it was set previously, assume the
1984 * peer has already been sent the error code or will request it
1986 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1988 if (conn->type == RX_SERVER_CONNECTION) {
1989 /* Make sure reply or at least dummy reply is sent */
1990 if (call->mode == RX_MODE_RECEIVING) {
1991 rxi_WriteProc(call, 0, 0);
1993 if (call->mode == RX_MODE_SENDING) {
1994 rxi_FlushWrite(call);
1996 service = conn->service;
1997 rxi_calltrace(RX_CALL_END, call);
1998 /* Call goes to hold state until reply packets are acknowledged */
1999 if (call->tfirst + call->nSoftAcked < call->tnext) {
2000 call->state = RX_STATE_HOLD;
2002 call->state = RX_STATE_DALLY;
2003 rxi_ClearTransmitQueue(call, 0);
2004 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
2005 rxevent_Cancel(call->keepAliveEvent, call,
2006 RX_CALL_REFCOUNT_ALIVE);
2008 } else { /* Client connection */
2010 /* Make sure server receives input packets, in the case where
2011 * no reply arguments are expected */
2012 if ((call->mode == RX_MODE_SENDING)
2013 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2014 (void)rxi_ReadProc(call, &dummy, 1);
2017 /* If we had an outstanding delayed ack, be nice to the server
2018 * and force-send it now.
2020 if (call->delayedAckEvent) {
2021 rxevent_Cancel(call->delayedAckEvent, call,
2022 RX_CALL_REFCOUNT_DELAY);
2023 call->delayedAckEvent = NULL;
2024 rxi_SendDelayedAck(NULL, call, NULL);
2027 /* We need to release the call lock since it's lower than the
2028 * conn_call_lock and we don't want to hold the conn_call_lock
2029 * over the rx_ReadProc call. The conn_call_lock needs to be held
2030 * here for the case where rx_NewCall is perusing the calls on
2031 * the connection structure. We don't want to signal until
2032 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2033 * have checked this call, found it active and by the time it
2034 * goes to sleep, will have missed the signal.
2036 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
2037 * there are threads waiting to use the conn object.
2039 MUTEX_EXIT(&call->lock);
2040 MUTEX_ENTER(&conn->conn_call_lock);
2041 MUTEX_ENTER(&call->lock);
2042 MUTEX_ENTER(&conn->conn_data_lock);
2043 conn->flags |= RX_CONN_BUSY;
2044 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2045 if (conn->makeCallWaiters == 0)
2046 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
2047 MUTEX_EXIT(&conn->conn_data_lock);
2048 #ifdef RX_ENABLE_LOCKS
2049 CV_BROADCAST(&conn->conn_call_cv);
2054 #ifdef RX_ENABLE_LOCKS
2056 MUTEX_EXIT(&conn->conn_data_lock);
2058 #endif /* RX_ENABLE_LOCKS */
2059 call->state = RX_STATE_DALLY;
2061 error = call->error;
2063 /* currentPacket, nLeft, and NFree must be zeroed here, because
2064 * ResetCall cannot: ResetCall may be called at splnet(), in the
2065 * kernel version, and may interrupt the macros rx_Read or
2066 * rx_Write, which run at normal priority for efficiency. */
2067 if (call->currentPacket) {
2068 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2069 rxi_FreePacket(call->currentPacket);
2070 call->currentPacket = (struct rx_packet *)0;
2073 call->nLeft = call->nFree = call->curlen = 0;
2075 /* Free any packets from the last call to ReadvProc/WritevProc */
2076 rxi_FreePackets(0, &call->iovq);
2078 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2079 MUTEX_EXIT(&call->lock);
2080 if (conn->type == RX_CLIENT_CONNECTION) {
2081 MUTEX_EXIT(&conn->conn_call_lock);
2082 conn->flags &= ~RX_CONN_BUSY;
2086 * Map errors to the local host's errno.h format.
2088 error = ntoh_syserr_conv(error);
2092 #if !defined(KERNEL)
2094 /* Call this routine when shutting down a server or client (especially
2095 * clients). This will allow Rx to gracefully garbage collect server
2096 * connections, and reduce the number of retries that a server might
2097 * make to a dead client.
2098 * This is not quite right, since some calls may still be ongoing and
2099 * we can't lock them to destroy them. */
2103 register struct rx_connection **conn_ptr, **conn_end;
2107 if (rxinit_status == 1) {
2109 return; /* Already shutdown. */
2111 rxi_DeleteCachedConnections();
2112 if (rx_connHashTable) {
2113 MUTEX_ENTER(&rx_connHashTable_lock);
2114 for (conn_ptr = &rx_connHashTable[0], conn_end =
2115 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2117 struct rx_connection *conn, *next;
2118 for (conn = *conn_ptr; conn; conn = next) {
2120 if (conn->type == RX_CLIENT_CONNECTION) {
2121 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2123 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2124 #ifdef RX_ENABLE_LOCKS
2125 rxi_DestroyConnectionNoLock(conn);
2126 #else /* RX_ENABLE_LOCKS */
2127 rxi_DestroyConnection(conn);
2128 #endif /* RX_ENABLE_LOCKS */
2132 #ifdef RX_ENABLE_LOCKS
2133 while (rx_connCleanup_list) {
2134 struct rx_connection *conn;
2135 conn = rx_connCleanup_list;
2136 rx_connCleanup_list = rx_connCleanup_list->next;
2137 MUTEX_EXIT(&rx_connHashTable_lock);
2138 rxi_CleanupConnection(conn);
2139 MUTEX_ENTER(&rx_connHashTable_lock);
2141 MUTEX_EXIT(&rx_connHashTable_lock);
2142 #endif /* RX_ENABLE_LOCKS */
2147 afs_winsockCleanup();
2155 /* if we wakeup packet waiter too often, can get in loop with two
2156 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2158 rxi_PacketsUnWait(void)
2160 if (!rx_waitingForPackets) {
2164 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2165 return; /* still over quota */
2168 rx_waitingForPackets = 0;
2169 #ifdef RX_ENABLE_LOCKS
2170 CV_BROADCAST(&rx_waitingForPackets_cv);
2172 osi_rxWakeup(&rx_waitingForPackets);
2178 /* ------------------Internal interfaces------------------------- */
2180 /* Return this process's service structure for the
2181 * specified socket and service */
2183 rxi_FindService(register osi_socket socket, register u_short serviceId)
2185 register struct rx_service **sp;
2186 for (sp = &rx_services[0]; *sp; sp++) {
2187 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2193 /* Allocate a call structure, for the indicated channel of the
2194 * supplied connection. The mode and state of the call must be set by
2195 * the caller. Returns the call with mutex locked. */
2197 rxi_NewCall(register struct rx_connection *conn, register int channel)
2199 register struct rx_call *call;
2200 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2201 register struct rx_call *cp; /* Call pointer temp */
2202 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2203 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2205 dpf(("rxi_NewCall(conn %x, channel %d)\n", conn, channel));
2207 /* Grab an existing call structure, or allocate a new one.
2208 * Existing call structures are assumed to have been left reset by
2210 MUTEX_ENTER(&rx_freeCallQueue_lock);
2212 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2214 * EXCEPT that the TQ might not yet be cleared out.
2215 * Skip over those with in-use TQs.
2218 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2219 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2225 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2226 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2227 call = queue_First(&rx_freeCallQueue, rx_call);
2228 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2230 rx_MutexDecrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2231 MUTEX_EXIT(&rx_freeCallQueue_lock);
2232 MUTEX_ENTER(&call->lock);
2233 CLEAR_CALL_QUEUE_LOCK(call);
2234 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2235 /* Now, if TQ wasn't cleared earlier, do it now. */
2236 if (call->flags & RX_CALL_TQ_CLEARME) {
2237 rxi_ClearTransmitQueue(call, 1);
2238 queue_Init(&call->tq);
2240 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2241 /* Bind the call to its connection structure */
2243 rxi_ResetCall(call, 1);
2246 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2247 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2248 MUTEX_EXIT(&rx_freeCallQueue_lock);
2249 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2250 MUTEX_ENTER(&call->lock);
2251 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2252 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2253 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2255 /* Initialize once-only items */
2256 queue_Init(&call->tq);
2257 queue_Init(&call->rq);
2258 queue_Init(&call->iovq);
2259 /* Bind the call to its connection structure (prereq for reset) */
2261 rxi_ResetCall(call, 1);
2263 call->channel = channel;
2264 call->callNumber = &conn->callNumber[channel];
2265 call->rwind = conn->rwind[channel];
2266 call->twind = conn->twind[channel];
2267 /* Note that the next expected call number is retained (in
2268 * conn->callNumber[i]), even if we reallocate the call structure
2270 conn->call[channel] = call;
2271 /* if the channel's never been used (== 0), we should start at 1, otherwise
2272 * the call number is valid from the last time this channel was used */
2273 if (*call->callNumber == 0)
2274 *call->callNumber = 1;
2279 /* A call has been inactive long enough that so we can throw away
2280 * state, including the call structure, which is placed on the call
2282 * Call is locked upon entry.
2283 * haveCTLock set if called from rxi_ReapConnections
2285 #ifdef RX_ENABLE_LOCKS
2287 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2288 #else /* RX_ENABLE_LOCKS */
2290 rxi_FreeCall(register struct rx_call *call)
2291 #endif /* RX_ENABLE_LOCKS */
2293 register int channel = call->channel;
2294 register struct rx_connection *conn = call->conn;
2297 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2298 (*call->callNumber)++;
2299 rxi_ResetCall(call, 0);
2300 call->conn->call[channel] = (struct rx_call *)0;
2302 MUTEX_ENTER(&rx_freeCallQueue_lock);
2303 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2304 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2305 /* A call may be free even though its transmit queue is still in use.
2306 * Since we search the call list from head to tail, put busy calls at
2307 * the head of the list, and idle calls at the tail.
2309 if (call->flags & RX_CALL_TQ_BUSY)
2310 queue_Prepend(&rx_freeCallQueue, call);
2312 queue_Append(&rx_freeCallQueue, call);
2313 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2314 queue_Append(&rx_freeCallQueue, call);
2315 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2316 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2317 MUTEX_EXIT(&rx_freeCallQueue_lock);
2319 /* Destroy the connection if it was previously slated for
2320 * destruction, i.e. the Rx client code previously called
2321 * rx_DestroyConnection (client connections), or
2322 * rxi_ReapConnections called the same routine (server
2323 * connections). Only do this, however, if there are no
2324 * outstanding calls. Note that for fine grain locking, there appears
2325 * to be a deadlock in that rxi_FreeCall has a call locked and
2326 * DestroyConnectionNoLock locks each call in the conn. But note a
2327 * few lines up where we have removed this call from the conn.
2328 * If someone else destroys a connection, they either have no
2329 * call lock held or are going through this section of code.
2331 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2332 MUTEX_ENTER(&conn->conn_data_lock);
2334 MUTEX_EXIT(&conn->conn_data_lock);
2335 #ifdef RX_ENABLE_LOCKS
2337 rxi_DestroyConnectionNoLock(conn);
2339 rxi_DestroyConnection(conn);
2340 #else /* RX_ENABLE_LOCKS */
2341 rxi_DestroyConnection(conn);
2342 #endif /* RX_ENABLE_LOCKS */
2346 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2348 rxi_Alloc(register size_t size)
2352 rx_MutexAdd1Increment2(rxi_Allocsize, (afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2355 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2356 afs_osi_Alloc_NoSleep(size);
2361 osi_Panic("rxi_Alloc error");
2367 rxi_Free(void *addr, register size_t size)
2369 rx_MutexAdd1Decrement2(rxi_Allocsize, -(afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2370 osi_Free(addr, size);
2374 rxi_SetPeerMtu(register afs_uint32 host, register afs_uint32 port, int mtu)
2376 struct rx_peer **peer_ptr, **peer_end;
2379 MUTEX_ENTER(&rx_peerHashTable_lock);
2381 for (peer_ptr = &rx_peerHashTable[0], peer_end =
2382 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
2384 struct rx_peer *peer, *next;
2385 for (peer = *peer_ptr; peer; peer = next) {
2387 if (host == peer->host) {
2388 MUTEX_ENTER(&peer->peer_lock);
2389 peer->ifMTU=MIN(mtu, peer->ifMTU);
2390 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2391 MUTEX_EXIT(&peer->peer_lock);
2396 struct rx_peer *peer;
2397 hashIndex = PEER_HASH(host, port);
2398 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2399 if ((peer->host == host) && (peer->port == port)) {
2400 MUTEX_ENTER(&peer->peer_lock);
2401 peer->ifMTU=MIN(mtu, peer->ifMTU);
2402 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2403 MUTEX_EXIT(&peer->peer_lock);
2407 MUTEX_EXIT(&rx_peerHashTable_lock);
2410 /* Find the peer process represented by the supplied (host,port)
2411 * combination. If there is no appropriate active peer structure, a
2412 * new one will be allocated and initialized
2413 * The origPeer, if set, is a pointer to a peer structure on which the
2414 * refcount will be be decremented. This is used to replace the peer
2415 * structure hanging off a connection structure */
2417 rxi_FindPeer(register afs_uint32 host, register u_short port,
2418 struct rx_peer *origPeer, int create)
2420 register struct rx_peer *pp;
2422 hashIndex = PEER_HASH(host, port);
2423 MUTEX_ENTER(&rx_peerHashTable_lock);
2424 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2425 if ((pp->host == host) && (pp->port == port))
2430 pp = rxi_AllocPeer(); /* This bzero's *pp */
2431 pp->host = host; /* set here or in InitPeerParams is zero */
2433 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2434 queue_Init(&pp->congestionQueue);
2435 queue_Init(&pp->rpcStats);
2436 pp->next = rx_peerHashTable[hashIndex];
2437 rx_peerHashTable[hashIndex] = pp;
2438 rxi_InitPeerParams(pp);
2439 rx_MutexIncrement(rx_stats.nPeerStructs, rx_stats_mutex);
2446 origPeer->refCount--;
2447 MUTEX_EXIT(&rx_peerHashTable_lock);
2452 /* Find the connection at (host, port) started at epoch, and with the
2453 * given connection id. Creates the server connection if necessary.
2454 * The type specifies whether a client connection or a server
2455 * connection is desired. In both cases, (host, port) specify the
2456 * peer's (host, pair) pair. Client connections are not made
2457 * automatically by this routine. The parameter socket gives the
2458 * socket descriptor on which the packet was received. This is used,
2459 * in the case of server connections, to check that *new* connections
2460 * come via a valid (port, serviceId). Finally, the securityIndex
2461 * parameter must match the existing index for the connection. If a
2462 * server connection is created, it will be created using the supplied
2463 * index, if the index is valid for this service */
2464 struct rx_connection *
2465 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2466 register u_short port, u_short serviceId, afs_uint32 cid,
2467 afs_uint32 epoch, int type, u_int securityIndex)
2469 int hashindex, flag, i;
2470 register struct rx_connection *conn;
2471 hashindex = CONN_HASH(host, port, cid, epoch, type);
2472 MUTEX_ENTER(&rx_connHashTable_lock);
2473 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2474 rx_connHashTable[hashindex],
2477 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2478 && (epoch == conn->epoch)) {
2479 register struct rx_peer *pp = conn->peer;
2480 if (securityIndex != conn->securityIndex) {
2481 /* this isn't supposed to happen, but someone could forge a packet
2482 * like this, and there seems to be some CM bug that makes this
2483 * happen from time to time -- in which case, the fileserver
2485 MUTEX_EXIT(&rx_connHashTable_lock);
2486 return (struct rx_connection *)0;
2488 if (pp->host == host && pp->port == port)
2490 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2492 /* So what happens when it's a callback connection? */
2493 if ( /*type == RX_CLIENT_CONNECTION && */
2494 (conn->epoch & 0x80000000))
2498 /* the connection rxLastConn that was used the last time is not the
2499 ** one we are looking for now. Hence, start searching in the hash */
2501 conn = rx_connHashTable[hashindex];
2506 struct rx_service *service;
2507 if (type == RX_CLIENT_CONNECTION) {
2508 MUTEX_EXIT(&rx_connHashTable_lock);
2509 return (struct rx_connection *)0;
2511 service = rxi_FindService(socket, serviceId);
2512 if (!service || (securityIndex >= service->nSecurityObjects)
2513 || (service->securityObjects[securityIndex] == 0)) {
2514 MUTEX_EXIT(&rx_connHashTable_lock);
2515 return (struct rx_connection *)0;
2517 conn = rxi_AllocConnection(); /* This bzero's the connection */
2518 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2519 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2520 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2521 conn->next = rx_connHashTable[hashindex];
2522 rx_connHashTable[hashindex] = conn;
2523 conn->peer = rxi_FindPeer(host, port, 0, 1);
2524 conn->type = RX_SERVER_CONNECTION;
2525 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2526 conn->epoch = epoch;
2527 conn->cid = cid & RX_CIDMASK;
2528 /* conn->serial = conn->lastSerial = 0; */
2529 /* conn->timeout = 0; */
2530 conn->ackRate = RX_FAST_ACK_RATE;
2531 conn->service = service;
2532 conn->serviceId = serviceId;
2533 conn->securityIndex = securityIndex;
2534 conn->securityObject = service->securityObjects[securityIndex];
2535 conn->nSpecific = 0;
2536 conn->specific = NULL;
2537 rx_SetConnDeadTime(conn, service->connDeadTime);
2538 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2539 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2540 for (i = 0; i < RX_MAXCALLS; i++) {
2541 conn->twind[i] = rx_initSendWindow;
2542 conn->rwind[i] = rx_initReceiveWindow;
2544 /* Notify security object of the new connection */
2545 RXS_NewConnection(conn->securityObject, conn);
2546 /* XXXX Connection timeout? */
2547 if (service->newConnProc)
2548 (*service->newConnProc) (conn);
2549 rx_MutexIncrement(rx_stats.nServerConns, rx_stats_mutex);
2552 MUTEX_ENTER(&conn->conn_data_lock);
2554 MUTEX_EXIT(&conn->conn_data_lock);
2556 rxLastConn = conn; /* store this connection as the last conn used */
2557 MUTEX_EXIT(&rx_connHashTable_lock);
2561 /* There are two packet tracing routines available for testing and monitoring
2562 * Rx. One is called just after every packet is received and the other is
2563 * called just before every packet is sent. Received packets, have had their
2564 * headers decoded, and packets to be sent have not yet had their headers
2565 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2566 * containing the network address. Both can be modified. The return value, if
2567 * non-zero, indicates that the packet should be dropped. */
2569 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
2570 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
2572 /* A packet has been received off the interface. Np is the packet, socket is
2573 * the socket number it was received from (useful in determining which service
2574 * this packet corresponds to), and (host, port) reflect the host,port of the
2575 * sender. This call returns the packet to the caller if it is finished with
2576 * it, rather than de-allocating it, just as a small performance hack */
2579 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2580 afs_uint32 host, u_short port, int *tnop,
2581 struct rx_call **newcallp)
2583 register struct rx_call *call;
2584 register struct rx_connection *conn;
2586 afs_uint32 currentCallNumber;
2592 struct rx_packet *tnp;
2595 /* We don't print out the packet until now because (1) the time may not be
2596 * accurate enough until now in the lwp implementation (rx_Listener only gets
2597 * the time after the packet is read) and (2) from a protocol point of view,
2598 * this is the first time the packet has been seen */
2599 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2600 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2601 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2602 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2603 np->header.epoch, np->header.cid, np->header.callNumber,
2604 np->header.seq, np->header.flags, np));
2607 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2608 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2611 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2612 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2615 /* If an input tracer function is defined, call it with the packet and
2616 * network address. Note this function may modify its arguments. */
2617 if (rx_justReceived) {
2618 struct sockaddr_in addr;
2620 addr.sin_family = AF_INET;
2621 addr.sin_port = port;
2622 addr.sin_addr.s_addr = host;
2623 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2624 addr.sin_len = sizeof(addr);
2625 #endif /* AFS_OSF_ENV */
2626 drop = (*rx_justReceived) (np, &addr);
2627 /* drop packet if return value is non-zero */
2630 port = addr.sin_port; /* in case fcn changed addr */
2631 host = addr.sin_addr.s_addr;
2635 /* If packet was not sent by the client, then *we* must be the client */
2636 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2637 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2639 /* Find the connection (or fabricate one, if we're the server & if
2640 * necessary) associated with this packet */
2642 rxi_FindConnection(socket, host, port, np->header.serviceId,
2643 np->header.cid, np->header.epoch, type,
2644 np->header.securityIndex);
2647 /* If no connection found or fabricated, just ignore the packet.
2648 * (An argument could be made for sending an abort packet for
2653 MUTEX_ENTER(&conn->conn_data_lock);
2654 if (conn->maxSerial < np->header.serial)
2655 conn->maxSerial = np->header.serial;
2656 MUTEX_EXIT(&conn->conn_data_lock);
2658 /* If the connection is in an error state, send an abort packet and ignore
2659 * the incoming packet */
2661 /* Don't respond to an abort packet--we don't want loops! */
2662 MUTEX_ENTER(&conn->conn_data_lock);
2663 if (np->header.type != RX_PACKET_TYPE_ABORT)
2664 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2666 MUTEX_EXIT(&conn->conn_data_lock);
2670 /* Check for connection-only requests (i.e. not call specific). */
2671 if (np->header.callNumber == 0) {
2672 switch (np->header.type) {
2673 case RX_PACKET_TYPE_ABORT: {
2674 /* What if the supplied error is zero? */
2675 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2676 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2677 rxi_ConnectionError(conn, errcode);
2678 MUTEX_ENTER(&conn->conn_data_lock);
2680 MUTEX_EXIT(&conn->conn_data_lock);
2683 case RX_PACKET_TYPE_CHALLENGE:
2684 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2685 MUTEX_ENTER(&conn->conn_data_lock);
2687 MUTEX_EXIT(&conn->conn_data_lock);
2689 case RX_PACKET_TYPE_RESPONSE:
2690 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2691 MUTEX_ENTER(&conn->conn_data_lock);
2693 MUTEX_EXIT(&conn->conn_data_lock);
2695 case RX_PACKET_TYPE_PARAMS:
2696 case RX_PACKET_TYPE_PARAMS + 1:
2697 case RX_PACKET_TYPE_PARAMS + 2:
2698 /* ignore these packet types for now */
2699 MUTEX_ENTER(&conn->conn_data_lock);
2701 MUTEX_EXIT(&conn->conn_data_lock);
2706 /* Should not reach here, unless the peer is broken: send an
2708 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2709 MUTEX_ENTER(&conn->conn_data_lock);
2710 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2712 MUTEX_EXIT(&conn->conn_data_lock);
2717 channel = np->header.cid & RX_CHANNELMASK;
2718 call = conn->call[channel];
2719 #ifdef RX_ENABLE_LOCKS
2721 MUTEX_ENTER(&call->lock);
2722 /* Test to see if call struct is still attached to conn. */
2723 if (call != conn->call[channel]) {
2725 MUTEX_EXIT(&call->lock);
2726 if (type == RX_SERVER_CONNECTION) {
2727 call = conn->call[channel];
2728 /* If we started with no call attached and there is one now,
2729 * another thread is also running this routine and has gotten
2730 * the connection channel. We should drop this packet in the tests
2731 * below. If there was a call on this connection and it's now
2732 * gone, then we'll be making a new call below.
2733 * If there was previously a call and it's now different then
2734 * the old call was freed and another thread running this routine
2735 * has created a call on this channel. One of these two threads
2736 * has a packet for the old call and the code below handles those
2740 MUTEX_ENTER(&call->lock);
2742 /* This packet can't be for this call. If the new call address is
2743 * 0 then no call is running on this channel. If there is a call
2744 * then, since this is a client connection we're getting data for
2745 * it must be for the previous call.
2747 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2748 MUTEX_ENTER(&conn->conn_data_lock);
2750 MUTEX_EXIT(&conn->conn_data_lock);
2755 currentCallNumber = conn->callNumber[channel];
2757 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2758 if (np->header.callNumber < currentCallNumber) {
2759 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2760 #ifdef RX_ENABLE_LOCKS
2762 MUTEX_EXIT(&call->lock);
2764 MUTEX_ENTER(&conn->conn_data_lock);
2766 MUTEX_EXIT(&conn->conn_data_lock);
2770 MUTEX_ENTER(&conn->conn_call_lock);
2771 call = rxi_NewCall(conn, channel);
2772 MUTEX_EXIT(&conn->conn_call_lock);
2773 *call->callNumber = np->header.callNumber;
2774 if (np->header.callNumber == 0)
2775 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));
2777 call->state = RX_STATE_PRECALL;
2778 clock_GetTime(&call->queueTime);
2779 hzero(call->bytesSent);
2780 hzero(call->bytesRcvd);
2782 * If the number of queued calls exceeds the overload
2783 * threshold then abort this call.
2785 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2786 struct rx_packet *tp;
2788 rxi_CallError(call, rx_BusyError);
2789 tp = rxi_SendCallAbort(call, np, 1, 0);
2790 MUTEX_EXIT(&call->lock);
2791 MUTEX_ENTER(&conn->conn_data_lock);
2793 MUTEX_EXIT(&conn->conn_data_lock);
2794 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2797 rxi_KeepAliveOn(call);
2798 } else if (np->header.callNumber != currentCallNumber) {
2799 /* Wait until the transmit queue is idle before deciding
2800 * whether to reset the current call. Chances are that the
2801 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2804 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2805 while ((call->state == RX_STATE_ACTIVE)
2806 && (call->flags & RX_CALL_TQ_BUSY)) {
2807 call->flags |= RX_CALL_TQ_WAIT;
2809 #ifdef RX_ENABLE_LOCKS
2810 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2811 CV_WAIT(&call->cv_tq, &call->lock);
2812 #else /* RX_ENABLE_LOCKS */
2813 osi_rxSleep(&call->tq);
2814 #endif /* RX_ENABLE_LOCKS */
2816 if (call->tqWaiters == 0)
2817 call->flags &= ~RX_CALL_TQ_WAIT;
2819 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2820 /* If the new call cannot be taken right now send a busy and set
2821 * the error condition in this call, so that it terminates as
2822 * quickly as possible */
2823 if (call->state == RX_STATE_ACTIVE) {
2824 struct rx_packet *tp;
2826 rxi_CallError(call, RX_CALL_DEAD);
2827 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2829 MUTEX_EXIT(&call->lock);
2830 MUTEX_ENTER(&conn->conn_data_lock);
2832 MUTEX_EXIT(&conn->conn_data_lock);
2835 rxi_ResetCall(call, 0);
2836 *call->callNumber = np->header.callNumber;
2837 if (np->header.callNumber == 0)
2838 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));
2840 call->state = RX_STATE_PRECALL;
2841 clock_GetTime(&call->queueTime);
2842 hzero(call->bytesSent);
2843 hzero(call->bytesRcvd);
2845 * If the number of queued calls exceeds the overload
2846 * threshold then abort this call.
2848 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2849 struct rx_packet *tp;
2851 rxi_CallError(call, rx_BusyError);
2852 tp = rxi_SendCallAbort(call, np, 1, 0);
2853 MUTEX_EXIT(&call->lock);
2854 MUTEX_ENTER(&conn->conn_data_lock);
2856 MUTEX_EXIT(&conn->conn_data_lock);
2857 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2860 rxi_KeepAliveOn(call);
2862 /* Continuing call; do nothing here. */
2864 } else { /* we're the client */
2865 /* Ignore all incoming acknowledgements for calls in DALLY state */
2866 if (call && (call->state == RX_STATE_DALLY)
2867 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2868 rx_MutexIncrement(rx_stats.ignorePacketDally, rx_stats_mutex);
2869 #ifdef RX_ENABLE_LOCKS
2871 MUTEX_EXIT(&call->lock);
2874 MUTEX_ENTER(&conn->conn_data_lock);
2876 MUTEX_EXIT(&conn->conn_data_lock);
2880 /* Ignore anything that's not relevant to the current call. If there
2881 * isn't a current call, then no packet is relevant. */
2882 if (!call || (np->header.callNumber != currentCallNumber)) {
2883 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2884 #ifdef RX_ENABLE_LOCKS
2886 MUTEX_EXIT(&call->lock);
2889 MUTEX_ENTER(&conn->conn_data_lock);
2891 MUTEX_EXIT(&conn->conn_data_lock);
2894 /* If the service security object index stamped in the packet does not
2895 * match the connection's security index, ignore the packet */
2896 if (np->header.securityIndex != conn->securityIndex) {
2897 #ifdef RX_ENABLE_LOCKS
2898 MUTEX_EXIT(&call->lock);
2900 MUTEX_ENTER(&conn->conn_data_lock);
2902 MUTEX_EXIT(&conn->conn_data_lock);
2906 /* If we're receiving the response, then all transmit packets are
2907 * implicitly acknowledged. Get rid of them. */
2908 if (np->header.type == RX_PACKET_TYPE_DATA) {
2909 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2910 /* XXX Hack. Because we must release the global rx lock when
2911 * sending packets (osi_NetSend) we drop all acks while we're
2912 * traversing the tq in rxi_Start sending packets out because
2913 * packets may move to the freePacketQueue as result of being here!
2914 * So we drop these packets until we're safely out of the
2915 * traversing. Really ugly!
2916 * For fine grain RX locking, we set the acked field in the
2917 * packets and let rxi_Start remove them from the transmit queue.
2919 if (call->flags & RX_CALL_TQ_BUSY) {
2920 #ifdef RX_ENABLE_LOCKS
2921 rxi_SetAcksInTransmitQueue(call);
2924 return np; /* xmitting; drop packet */
2927 rxi_ClearTransmitQueue(call, 0);
2929 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2930 rxi_ClearTransmitQueue(call, 0);
2931 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2933 if (np->header.type == RX_PACKET_TYPE_ACK) {
2934 /* now check to see if this is an ack packet acknowledging that the
2935 * server actually *lost* some hard-acked data. If this happens we
2936 * ignore this packet, as it may indicate that the server restarted in
2937 * the middle of a call. It is also possible that this is an old ack
2938 * packet. We don't abort the connection in this case, because this
2939 * *might* just be an old ack packet. The right way to detect a server
2940 * restart in the midst of a call is to notice that the server epoch
2942 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2943 * XXX unacknowledged. I think that this is off-by-one, but
2944 * XXX I don't dare change it just yet, since it will
2945 * XXX interact badly with the server-restart detection
2946 * XXX code in receiveackpacket. */
2947 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2948 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2949 MUTEX_EXIT(&call->lock);
2950 MUTEX_ENTER(&conn->conn_data_lock);
2952 MUTEX_EXIT(&conn->conn_data_lock);
2956 } /* else not a data packet */
2959 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2960 /* Set remote user defined status from packet */
2961 call->remoteStatus = np->header.userStatus;
2963 /* Note the gap between the expected next packet and the actual
2964 * packet that arrived, when the new packet has a smaller serial number
2965 * than expected. Rioses frequently reorder packets all by themselves,
2966 * so this will be quite important with very large window sizes.
2967 * Skew is checked against 0 here to avoid any dependence on the type of
2968 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2970 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2971 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2972 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2974 MUTEX_ENTER(&conn->conn_data_lock);
2975 skew = conn->lastSerial - np->header.serial;
2976 conn->lastSerial = np->header.serial;
2977 MUTEX_EXIT(&conn->conn_data_lock);
2979 register struct rx_peer *peer;
2981 if (skew > peer->inPacketSkew) {
2982 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2984 peer->inPacketSkew = skew;
2988 /* Now do packet type-specific processing */
2989 switch (np->header.type) {
2990 case RX_PACKET_TYPE_DATA:
2991 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2994 case RX_PACKET_TYPE_ACK:
2995 /* Respond immediately to ack packets requesting acknowledgement
2997 if (np->header.flags & RX_REQUEST_ACK) {
2999 (void)rxi_SendCallAbort(call, 0, 1, 0);
3001 (void)rxi_SendAck(call, 0, np->header.serial,
3002 RX_ACK_PING_RESPONSE, 1);
3004 np = rxi_ReceiveAckPacket(call, np, 1);
3006 case RX_PACKET_TYPE_ABORT: {
3007 /* An abort packet: reset the call, passing the error up to the user. */
3008 /* What if error is zero? */
3009 /* What if the error is -1? the application will treat it as a timeout. */
3010 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
3011 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
3012 rxi_CallError(call, errdata);
3013 MUTEX_EXIT(&call->lock);
3014 MUTEX_ENTER(&conn->conn_data_lock);
3016 MUTEX_EXIT(&conn->conn_data_lock);
3017 return np; /* xmitting; drop packet */
3019 case RX_PACKET_TYPE_BUSY:
3022 case RX_PACKET_TYPE_ACKALL:
3023 /* All packets acknowledged, so we can drop all packets previously
3024 * readied for sending */
3025 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3026 /* XXX Hack. We because we can't release the global rx lock when
3027 * sending packets (osi_NetSend) we drop all ack pkts while we're
3028 * traversing the tq in rxi_Start sending packets out because
3029 * packets may move to the freePacketQueue as result of being
3030 * here! So we drop these packets until we're safely out of the
3031 * traversing. Really ugly!
3032 * For fine grain RX locking, we set the acked field in the packets
3033 * and let rxi_Start remove the packets from the transmit queue.
3035 if (call->flags & RX_CALL_TQ_BUSY) {
3036 #ifdef RX_ENABLE_LOCKS
3037 rxi_SetAcksInTransmitQueue(call);
3039 #else /* RX_ENABLE_LOCKS */
3040 MUTEX_EXIT(&call->lock);
3041 MUTEX_ENTER(&conn->conn_data_lock);
3043 MUTEX_EXIT(&conn->conn_data_lock);
3044 return np; /* xmitting; drop packet */
3045 #endif /* RX_ENABLE_LOCKS */
3047 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3048 rxi_ClearTransmitQueue(call, 0);
3049 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
3052 /* Should not reach here, unless the peer is broken: send an abort
3054 rxi_CallError(call, RX_PROTOCOL_ERROR);
3055 np = rxi_SendCallAbort(call, np, 1, 0);
3058 /* Note when this last legitimate packet was received, for keep-alive
3059 * processing. Note, we delay getting the time until now in the hope that
3060 * the packet will be delivered to the user before any get time is required
3061 * (if not, then the time won't actually be re-evaluated here). */
3062 call->lastReceiveTime = clock_Sec();
3063 MUTEX_EXIT(&call->lock);
3064 MUTEX_ENTER(&conn->conn_data_lock);
3066 MUTEX_EXIT(&conn->conn_data_lock);
3070 /* return true if this is an "interesting" connection from the point of view
3071 of someone trying to debug the system */
3073 rxi_IsConnInteresting(struct rx_connection *aconn)
3076 register struct rx_call *tcall;
3078 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3080 for (i = 0; i < RX_MAXCALLS; i++) {
3081 tcall = aconn->call[i];
3083 if ((tcall->state == RX_STATE_PRECALL)
3084 || (tcall->state == RX_STATE_ACTIVE))
3086 if ((tcall->mode == RX_MODE_SENDING)
3087 || (tcall->mode == RX_MODE_RECEIVING))
3095 /* if this is one of the last few packets AND it wouldn't be used by the
3096 receiving call to immediately satisfy a read request, then drop it on
3097 the floor, since accepting it might prevent a lock-holding thread from
3098 making progress in its reading. If a call has been cleared while in
3099 the precall state then ignore all subsequent packets until the call
3100 is assigned to a thread. */
3103 TooLow(struct rx_packet *ap, struct rx_call *acall)
3106 MUTEX_ENTER(&rx_stats_mutex);
3107 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3108 && (acall->state == RX_STATE_PRECALL))
3109 || ((rx_nFreePackets < rxi_dataQuota + 2)
3110 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3111 && (acall->flags & RX_CALL_READER_WAIT)))) {
3114 MUTEX_EXIT(&rx_stats_mutex);
3120 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
3122 struct rx_connection *conn = arg1;
3123 struct rx_call *acall = arg2;
3124 struct rx_call *call = acall;
3125 struct clock when, now;
3128 MUTEX_ENTER(&conn->conn_data_lock);
3129 conn->checkReachEvent = NULL;
3130 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3133 MUTEX_EXIT(&conn->conn_data_lock);
3137 MUTEX_ENTER(&conn->conn_call_lock);
3138 MUTEX_ENTER(&conn->conn_data_lock);
3139 for (i = 0; i < RX_MAXCALLS; i++) {
3140 struct rx_call *tc = conn->call[i];
3141 if (tc && tc->state == RX_STATE_PRECALL) {
3147 /* Indicate that rxi_CheckReachEvent is no longer running by
3148 * clearing the flag. Must be atomic under conn_data_lock to
3149 * avoid a new call slipping by: rxi_CheckConnReach holds
3150 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3152 conn->flags &= ~RX_CONN_ATTACHWAIT;
3153 MUTEX_EXIT(&conn->conn_data_lock);
3154 MUTEX_EXIT(&conn->conn_call_lock);
3159 MUTEX_ENTER(&call->lock);
3160 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3162 MUTEX_EXIT(&call->lock);
3164 clock_GetTime(&now);
3166 when.sec += RX_CHECKREACH_TIMEOUT;
3167 MUTEX_ENTER(&conn->conn_data_lock);
3168 if (!conn->checkReachEvent) {
3170 conn->checkReachEvent =
3171 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3174 MUTEX_EXIT(&conn->conn_data_lock);
3180 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3182 struct rx_service *service = conn->service;
3183 struct rx_peer *peer = conn->peer;
3184 afs_uint32 now, lastReach;
3186 if (service->checkReach == 0)
3190 MUTEX_ENTER(&peer->peer_lock);
3191 lastReach = peer->lastReachTime;
3192 MUTEX_EXIT(&peer->peer_lock);
3193 if (now - lastReach < RX_CHECKREACH_TTL)
3196 MUTEX_ENTER(&conn->conn_data_lock);
3197 if (conn->flags & RX_CONN_ATTACHWAIT) {
3198 MUTEX_EXIT(&conn->conn_data_lock);
3201 conn->flags |= RX_CONN_ATTACHWAIT;
3202 MUTEX_EXIT(&conn->conn_data_lock);
3203 if (!conn->checkReachEvent)
3204 rxi_CheckReachEvent(NULL, conn, call);
3209 /* try to attach call, if authentication is complete */
3211 TryAttach(register struct rx_call *acall, register osi_socket socket,
3212 register int *tnop, register struct rx_call **newcallp,
3215 struct rx_connection *conn = acall->conn;
3217 if (conn->type == RX_SERVER_CONNECTION
3218 && acall->state == RX_STATE_PRECALL) {
3219 /* Don't attach until we have any req'd. authentication. */
3220 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3221 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3222 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3223 /* Note: this does not necessarily succeed; there
3224 * may not any proc available
3227 rxi_ChallengeOn(acall->conn);
3232 /* A data packet has been received off the interface. This packet is
3233 * appropriate to the call (the call is in the right state, etc.). This
3234 * routine can return a packet to the caller, for re-use */
3237 rxi_ReceiveDataPacket(register struct rx_call *call,
3238 register struct rx_packet *np, int istack,
3239 osi_socket socket, afs_uint32 host, u_short port,
3240 int *tnop, struct rx_call **newcallp)
3242 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3246 afs_uint32 seq, serial, flags;
3248 struct rx_packet *tnp;
3249 struct clock when, now;
3250 rx_MutexIncrement(rx_stats.dataPacketsRead, rx_stats_mutex);
3253 /* If there are no packet buffers, drop this new packet, unless we can find
3254 * packet buffers from inactive calls */
3256 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3257 MUTEX_ENTER(&rx_freePktQ_lock);
3258 rxi_NeedMorePackets = TRUE;
3259 MUTEX_EXIT(&rx_freePktQ_lock);
3260 rx_MutexIncrement(rx_stats.noPacketBuffersOnRead, rx_stats_mutex);
3261 call->rprev = np->header.serial;
3262 rxi_calltrace(RX_TRACE_DROP, call);
3263 dpf(("packet %x dropped on receipt - quota problems", np));
3265 rxi_ClearReceiveQueue(call);
3266 clock_GetTime(&now);
3268 clock_Add(&when, &rx_softAckDelay);
3269 if (!call->delayedAckEvent
3270 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3271 rxevent_Cancel(call->delayedAckEvent, call,
3272 RX_CALL_REFCOUNT_DELAY);
3273 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3274 call->delayedAckEvent =
3275 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3277 /* we've damaged this call already, might as well do it in. */
3283 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3284 * packet is one of several packets transmitted as a single
3285 * datagram. Do not send any soft or hard acks until all packets
3286 * in a jumbogram have been processed. Send negative acks right away.
3288 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3289 /* tnp is non-null when there are more packets in the
3290 * current jumbo gram */
3297 seq = np->header.seq;
3298 serial = np->header.serial;
3299 flags = np->header.flags;
3301 /* If the call is in an error state, send an abort message */
3303 return rxi_SendCallAbort(call, np, istack, 0);
3305 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3306 * AFS 3.5 jumbogram. */
3307 if (flags & RX_JUMBO_PACKET) {
3308 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3313 if (np->header.spare != 0) {
3314 MUTEX_ENTER(&call->conn->conn_data_lock);
3315 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3316 MUTEX_EXIT(&call->conn->conn_data_lock);
3319 /* The usual case is that this is the expected next packet */
3320 if (seq == call->rnext) {
3322 /* Check to make sure it is not a duplicate of one already queued */
3323 if (queue_IsNotEmpty(&call->rq)
3324 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3325 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3326 dpf(("packet %x dropped on receipt - duplicate", np));
3327 rxevent_Cancel(call->delayedAckEvent, call,
3328 RX_CALL_REFCOUNT_DELAY);
3329 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3335 /* It's the next packet. Stick it on the receive queue
3336 * for this call. Set newPackets to make sure we wake
3337 * the reader once all packets have been processed */
3338 np->flags |= RX_PKTFLAG_RQ;
3339 queue_Prepend(&call->rq, np);
3341 np = NULL; /* We can't use this anymore */
3344 /* If an ack is requested then set a flag to make sure we
3345 * send an acknowledgement for this packet */
3346 if (flags & RX_REQUEST_ACK) {
3347 ackNeeded = RX_ACK_REQUESTED;
3350 /* Keep track of whether we have received the last packet */
3351 if (flags & RX_LAST_PACKET) {
3352 call->flags |= RX_CALL_HAVE_LAST;
3356 /* Check whether we have all of the packets for this call */
3357 if (call->flags & RX_CALL_HAVE_LAST) {
3358 afs_uint32 tseq; /* temporary sequence number */
3359 struct rx_packet *tp; /* Temporary packet pointer */
3360 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3362 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3363 if (tseq != tp->header.seq)
3365 if (tp->header.flags & RX_LAST_PACKET) {
3366 call->flags |= RX_CALL_RECEIVE_DONE;
3373 /* Provide asynchronous notification for those who want it
3374 * (e.g. multi rx) */
3375 if (call->arrivalProc) {
3376 (*call->arrivalProc) (call, call->arrivalProcHandle,
3377 call->arrivalProcArg);
3378 call->arrivalProc = (void (*)())0;
3381 /* Update last packet received */
3384 /* If there is no server process serving this call, grab
3385 * one, if available. We only need to do this once. If a
3386 * server thread is available, this thread becomes a server
3387 * thread and the server thread becomes a listener thread. */
3389 TryAttach(call, socket, tnop, newcallp, 0);
3392 /* This is not the expected next packet. */
3394 /* Determine whether this is a new or old packet, and if it's
3395 * a new one, whether it fits into the current receive window.
3396 * Also figure out whether the packet was delivered in sequence.
3397 * We use the prev variable to determine whether the new packet
3398 * is the successor of its immediate predecessor in the
3399 * receive queue, and the missing flag to determine whether
3400 * any of this packets predecessors are missing. */
3402 afs_uint32 prev; /* "Previous packet" sequence number */
3403 struct rx_packet *tp; /* Temporary packet pointer */
3404 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3405 int missing; /* Are any predecessors missing? */
3407 /* If the new packet's sequence number has been sent to the
3408 * application already, then this is a duplicate */
3409 if (seq < call->rnext) {
3410 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3411 rxevent_Cancel(call->delayedAckEvent, call,
3412 RX_CALL_REFCOUNT_DELAY);
3413 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3419 /* If the sequence number is greater than what can be
3420 * accomodated by the current window, then send a negative
3421 * acknowledge and drop the packet */
3422 if ((call->rnext + call->rwind) <= seq) {
3423 rxevent_Cancel(call->delayedAckEvent, call,
3424 RX_CALL_REFCOUNT_DELAY);
3425 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3432 /* Look for the packet in the queue of old received packets */
3433 for (prev = call->rnext - 1, missing =
3434 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3435 /*Check for duplicate packet */
3436 if (seq == tp->header.seq) {
3437 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3438 rxevent_Cancel(call->delayedAckEvent, call,
3439 RX_CALL_REFCOUNT_DELAY);
3440 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3446 /* If we find a higher sequence packet, break out and
3447 * insert the new packet here. */
3448 if (seq < tp->header.seq)
3450 /* Check for missing packet */
3451 if (tp->header.seq != prev + 1) {
3455 prev = tp->header.seq;
3458 /* Keep track of whether we have received the last packet. */
3459 if (flags & RX_LAST_PACKET) {
3460 call->flags |= RX_CALL_HAVE_LAST;
3463 /* It's within the window: add it to the the receive queue.
3464 * tp is left by the previous loop either pointing at the
3465 * packet before which to insert the new packet, or at the
3466 * queue head if the queue is empty or the packet should be
3468 queue_InsertBefore(tp, np);
3472 /* Check whether we have all of the packets for this call */
3473 if ((call->flags & RX_CALL_HAVE_LAST)
3474 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3475 afs_uint32 tseq; /* temporary sequence number */
3478 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3479 if (tseq != tp->header.seq)
3481 if (tp->header.flags & RX_LAST_PACKET) {
3482 call->flags |= RX_CALL_RECEIVE_DONE;
3489 /* We need to send an ack of the packet is out of sequence,
3490 * or if an ack was requested by the peer. */
3491 if (seq != prev + 1 || missing) {
3492 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3493 } else if (flags & RX_REQUEST_ACK) {
3494 ackNeeded = RX_ACK_REQUESTED;
3497 /* Acknowledge the last packet for each call */
3498 if (flags & RX_LAST_PACKET) {
3509 * If the receiver is waiting for an iovec, fill the iovec
3510 * using the data from the receive queue */
3511 if (call->flags & RX_CALL_IOVEC_WAIT) {
3512 didHardAck = rxi_FillReadVec(call, serial);
3513 /* the call may have been aborted */
3522 /* Wakeup the reader if any */
3523 if ((call->flags & RX_CALL_READER_WAIT)
3524 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3525 || (call->iovNext >= call->iovMax)
3526 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3527 call->flags &= ~RX_CALL_READER_WAIT;
3528 #ifdef RX_ENABLE_LOCKS
3529 CV_BROADCAST(&call->cv_rq);
3531 osi_rxWakeup(&call->rq);
3537 * Send an ack when requested by the peer, or once every
3538 * rxi_SoftAckRate packets until the last packet has been
3539 * received. Always send a soft ack for the last packet in
3540 * the server's reply. */
3542 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3543 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3544 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3545 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3546 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3547 } else if (call->nSoftAcks) {
3548 clock_GetTime(&now);
3550 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3551 clock_Add(&when, &rx_lastAckDelay);
3553 clock_Add(&when, &rx_softAckDelay);
3555 if (!call->delayedAckEvent
3556 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3557 rxevent_Cancel(call->delayedAckEvent, call,
3558 RX_CALL_REFCOUNT_DELAY);
3559 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3560 call->delayedAckEvent =
3561 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3563 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3564 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3571 static void rxi_ComputeRate();
3575 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3577 struct rx_peer *peer = conn->peer;
3579 MUTEX_ENTER(&peer->peer_lock);
3580 peer->lastReachTime = clock_Sec();
3581 MUTEX_EXIT(&peer->peer_lock);
3583 MUTEX_ENTER(&conn->conn_data_lock);
3584 if (conn->flags & RX_CONN_ATTACHWAIT) {
3587 conn->flags &= ~RX_CONN_ATTACHWAIT;
3588 MUTEX_EXIT(&conn->conn_data_lock);
3590 for (i = 0; i < RX_MAXCALLS; i++) {
3591 struct rx_call *call = conn->call[i];
3594 MUTEX_ENTER(&call->lock);
3595 /* tnop can be null if newcallp is null */
3596 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3598 MUTEX_EXIT(&call->lock);
3602 MUTEX_EXIT(&conn->conn_data_lock);
3606 rx_ack_reason(int reason)
3609 case RX_ACK_REQUESTED:
3611 case RX_ACK_DUPLICATE:
3613 case RX_ACK_OUT_OF_SEQUENCE:
3615 case RX_ACK_EXCEEDS_WINDOW:
3617 case RX_ACK_NOSPACE:
3621 case RX_ACK_PING_RESPONSE:
3633 /* rxi_ComputePeerNetStats
3635 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3636 * estimates (like RTT and throughput) based on ack packets. Caller
3637 * must ensure that the packet in question is the right one (i.e.
3638 * serial number matches).
3641 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3642 struct rx_ackPacket *ap, struct rx_packet *np)
3644 struct rx_peer *peer = call->conn->peer;
3646 /* Use RTT if not delayed by client. */
3647 if (ap->reason != RX_ACK_DELAY)
3648 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3650 rxi_ComputeRate(peer, call, p, np, ap->reason);
3654 /* The real smarts of the whole thing. */
3656 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3659 struct rx_ackPacket *ap;
3661 register struct rx_packet *tp;
3662 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3663 register struct rx_connection *conn = call->conn;
3664 struct rx_peer *peer = conn->peer;
3667 /* because there are CM's that are bogus, sending weird values for this. */
3668 afs_uint32 skew = 0;
3673 int newAckCount = 0;
3674 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3675 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3677 rx_MutexIncrement(rx_stats.ackPacketsRead, rx_stats_mutex);
3678 ap = (struct rx_ackPacket *)rx_DataOf(np);
3679 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3681 return np; /* truncated ack packet */
3683 /* depends on ack packet struct */
3684 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3685 first = ntohl(ap->firstPacket);
3686 serial = ntohl(ap->serial);
3687 /* temporarily disabled -- needs to degrade over time
3688 * skew = ntohs(ap->maxSkew); */
3690 /* Ignore ack packets received out of order */
3691 if (first < call->tfirst) {
3695 if (np->header.flags & RX_SLOW_START_OK) {
3696 call->flags |= RX_CALL_SLOW_START_OK;
3699 if (ap->reason == RX_ACK_PING_RESPONSE)
3700 rxi_UpdatePeerReach(conn, call);
3704 if (rxdebug_active) {
3708 len = _snprintf(msg, sizeof(msg),
3709 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3710 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3711 ntohl(ap->serial), ntohl(ap->previousPacket),
3712 (unsigned int)np->header.seq, (unsigned int)skew,
3713 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3717 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3718 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3722 OutputDebugString(msg);
3724 #else /* AFS_NT40_ENV */
3727 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3728 ap->reason, ntohl(ap->previousPacket),
3729 (unsigned int)np->header.seq, (unsigned int)serial,
3730 (unsigned int)skew, ntohl(ap->firstPacket));
3733 for (offset = 0; offset < nAcks; offset++)
3734 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3739 #endif /* AFS_NT40_ENV */
3742 /* Update the outgoing packet skew value to the latest value of
3743 * the peer's incoming packet skew value. The ack packet, of
3744 * course, could arrive out of order, but that won't affect things
3746 MUTEX_ENTER(&peer->peer_lock);
3747 peer->outPacketSkew = skew;
3749 /* Check for packets that no longer need to be transmitted, and
3750 * discard them. This only applies to packets positively
3751 * acknowledged as having been sent to the peer's upper level.
3752 * All other packets must be retained. So only packets with
3753 * sequence numbers < ap->firstPacket are candidates. */
3754 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3755 if (tp->header.seq >= first)
3757 call->tfirst = tp->header.seq + 1;
3759 && (tp->header.serial == serial || tp->firstSerial == serial))
3760 rxi_ComputePeerNetStats(call, tp, ap, np);
3761 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3764 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3765 /* XXX Hack. Because we have to release the global rx lock when sending
3766 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3767 * in rxi_Start sending packets out because packets may move to the
3768 * freePacketQueue as result of being here! So we drop these packets until
3769 * we're safely out of the traversing. Really ugly!
3770 * To make it even uglier, if we're using fine grain locking, we can
3771 * set the ack bits in the packets and have rxi_Start remove the packets
3772 * when it's done transmitting.
3774 if (call->flags & RX_CALL_TQ_BUSY) {
3775 #ifdef RX_ENABLE_LOCKS
3776 tp->flags |= RX_PKTFLAG_ACKED;
3777 call->flags |= RX_CALL_TQ_SOME_ACKED;
3778 #else /* RX_ENABLE_LOCKS */
3780 #endif /* RX_ENABLE_LOCKS */
3782 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3785 tp->flags &= ~RX_PKTFLAG_TQ;
3786 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3791 /* Give rate detector a chance to respond to ping requests */
3792 if (ap->reason == RX_ACK_PING_RESPONSE) {
3793 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3797 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3799 /* Now go through explicit acks/nacks and record the results in
3800 * the waiting packets. These are packets that can't be released
3801 * yet, even with a positive acknowledge. This positive
3802 * acknowledge only means the packet has been received by the
3803 * peer, not that it will be retained long enough to be sent to
3804 * the peer's upper level. In addition, reset the transmit timers
3805 * of any missing packets (those packets that must be missing
3806 * because this packet was out of sequence) */
3808 call->nSoftAcked = 0;
3809 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3810 /* Update round trip time if the ack was stimulated on receipt
3812 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3813 #ifdef RX_ENABLE_LOCKS
3814 if (tp->header.seq >= first)
3815 #endif /* RX_ENABLE_LOCKS */
3816 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3818 && (tp->header.serial == serial || tp->firstSerial == serial))
3819 rxi_ComputePeerNetStats(call, tp, ap, np);
3821 /* Set the acknowledge flag per packet based on the
3822 * information in the ack packet. An acknowlegded packet can
3823 * be downgraded when the server has discarded a packet it
3824 * soacked previously, or when an ack packet is received
3825 * out of sequence. */
3826 if (tp->header.seq < first) {
3827 /* Implicit ack information */
3828 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3831 tp->flags |= RX_PKTFLAG_ACKED;
3832 } else if (tp->header.seq < first + nAcks) {
3833 /* Explicit ack information: set it in the packet appropriately */
3834 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3835 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3837 tp->flags |= RX_PKTFLAG_ACKED;
3844 } else /* RX_ACK_TYPE_NACK */ {
3845 tp->flags &= ~RX_PKTFLAG_ACKED;
3849 tp->flags &= ~RX_PKTFLAG_ACKED;
3853 /* If packet isn't yet acked, and it has been transmitted at least
3854 * once, reset retransmit time using latest timeout
3855 * ie, this should readjust the retransmit timer for all outstanding
3856 * packets... So we don't just retransmit when we should know better*/
3858 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3859 tp->retryTime = tp->timeSent;
3860 clock_Add(&tp->retryTime, &peer->timeout);
3861 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3862 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3866 /* If the window has been extended by this acknowledge packet,
3867 * then wakeup a sender waiting in alloc for window space, or try
3868 * sending packets now, if he's been sitting on packets due to
3869 * lack of window space */
3870 if (call->tnext < (call->tfirst + call->twind)) {
3871 #ifdef RX_ENABLE_LOCKS
3872 CV_SIGNAL(&call->cv_twind);
3874 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3875 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3876 osi_rxWakeup(&call->twind);
3879 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3880 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3884 /* if the ack packet has a receivelen field hanging off it,
3885 * update our state */
3886 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3889 /* If the ack packet has a "recommended" size that is less than
3890 * what I am using now, reduce my size to match */
3891 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3892 (int)sizeof(afs_int32), &tSize);
3893 tSize = (afs_uint32) ntohl(tSize);
3894 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3896 /* Get the maximum packet size to send to this peer */
3897 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
3899 tSize = (afs_uint32) ntohl(tSize);
3900 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3901 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3903 /* sanity check - peer might have restarted with different params.
3904 * If peer says "send less", dammit, send less... Peer should never
3905 * be unable to accept packets of the size that prior AFS versions would
3906 * send without asking. */
3907 if (peer->maxMTU != tSize) {
3908 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
3910 peer->maxMTU = tSize;
3911 peer->MTU = MIN(tSize, peer->MTU);
3912 call->MTU = MIN(call->MTU, tSize);
3915 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3918 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3919 (int)sizeof(afs_int32), &tSize);
3920 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3921 if (tSize < call->twind) { /* smaller than our send */
3922 call->twind = tSize; /* window, we must send less... */
3923 call->ssthresh = MIN(call->twind, call->ssthresh);
3924 call->conn->twind[call->channel] = call->twind;
3927 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3928 * network MTU confused with the loopback MTU. Calculate the
3929 * maximum MTU here for use in the slow start code below.
3931 maxMTU = peer->maxMTU;
3932 /* Did peer restart with older RX version? */
3933 if (peer->maxDgramPackets > 1) {
3934 peer->maxDgramPackets = 1;
3936 } else if (np->length >=
3937 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3940 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3941 sizeof(afs_int32), &tSize);
3942 tSize = (afs_uint32) ntohl(tSize);
3944 * As of AFS 3.5 we set the send window to match the receive window.
3946 if (tSize < call->twind) {
3947 call->twind = tSize;
3948 call->conn->twind[call->channel] = call->twind;
3949 call->ssthresh = MIN(call->twind, call->ssthresh);
3950 } else if (tSize > call->twind) {
3951 call->twind = tSize;
3952 call->conn->twind[call->channel] = call->twind;
3956 * As of AFS 3.5, a jumbogram is more than one fixed size
3957 * packet transmitted in a single UDP datagram. If the remote
3958 * MTU is smaller than our local MTU then never send a datagram
3959 * larger than the natural MTU.
3962 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3963 sizeof(afs_int32), &tSize);
3964 maxDgramPackets = (afs_uint32) ntohl(tSize);
3965 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3966 maxDgramPackets = MIN(maxDgramPackets, peer->ifDgramPackets);
3967 if (peer->natMTU < peer->ifMTU)
3968 maxDgramPackets = MIN(maxDgramPackets, rxi_AdjustDgramPackets(1, peer->natMTU));
3969 if (maxDgramPackets > 1) {
3970 peer->maxDgramPackets = maxDgramPackets;
3971 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3973 peer->maxDgramPackets = 1;
3974 call->MTU = peer->natMTU;
3976 } else if (peer->maxDgramPackets > 1) {
3977 /* Restarted with lower version of RX */
3978 peer->maxDgramPackets = 1;
3980 } else if (peer->maxDgramPackets > 1
3981 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3982 /* Restarted with lower version of RX */
3983 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3984 peer->natMTU = OLD_MAX_PACKET_SIZE;
3985 peer->MTU = OLD_MAX_PACKET_SIZE;
3986 peer->maxDgramPackets = 1;
3987 peer->nDgramPackets = 1;
3989 call->MTU = OLD_MAX_PACKET_SIZE;
3994 * Calculate how many datagrams were successfully received after
3995 * the first missing packet and adjust the negative ack counter
4000 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
4001 if (call->nNacks < nNacked) {
4002 call->nNacks = nNacked;
4005 call->nAcks += newAckCount;
4009 if (call->flags & RX_CALL_FAST_RECOVER) {
4011 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4013 call->flags &= ~RX_CALL_FAST_RECOVER;
4014 call->cwind = call->nextCwind;
4015 call->nextCwind = 0;
4018 call->nCwindAcks = 0;
4019 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4020 /* Three negative acks in a row trigger congestion recovery */
4021 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4022 MUTEX_EXIT(&peer->peer_lock);
4023 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4024 /* someone else is waiting to start recovery */
4027 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4028 rxi_WaitforTQBusy(call);
4029 MUTEX_ENTER(&peer->peer_lock);
4030 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4031 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4032 call->flags |= RX_CALL_FAST_RECOVER;
4033 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4035 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4036 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4037 call->nextCwind = call->ssthresh;
4040 peer->MTU = call->MTU;
4041 peer->cwind = call->nextCwind;
4042 peer->nDgramPackets = call->nDgramPackets;
4044 call->congestSeq = peer->congestSeq;
4045 /* Reset the resend times on the packets that were nacked
4046 * so we will retransmit as soon as the window permits*/
4047 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4049 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4050 clock_Zero(&tp->retryTime);
4052 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4057 /* If cwind is smaller than ssthresh, then increase
4058 * the window one packet for each ack we receive (exponential
4060 * If cwind is greater than or equal to ssthresh then increase
4061 * the congestion window by one packet for each cwind acks we
4062 * receive (linear growth). */
4063 if (call->cwind < call->ssthresh) {
4065 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4066 call->nCwindAcks = 0;
4068 call->nCwindAcks += newAckCount;
4069 if (call->nCwindAcks >= call->cwind) {
4070 call->nCwindAcks = 0;
4071 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4075 * If we have received several acknowledgements in a row then
4076 * it is time to increase the size of our datagrams
4078 if ((int)call->nAcks > rx_nDgramThreshold) {
4079 if (peer->maxDgramPackets > 1) {
4080 if (call->nDgramPackets < peer->maxDgramPackets) {
4081 call->nDgramPackets++;
4083 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4084 } else if (call->MTU < peer->maxMTU) {
4085 call->MTU += peer->natMTU;
4086 call->MTU = MIN(call->MTU, peer->maxMTU);
4092 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4094 /* Servers need to hold the call until all response packets have
4095 * been acknowledged. Soft acks are good enough since clients
4096 * are not allowed to clear their receive queues. */
4097 if (call->state == RX_STATE_HOLD
4098 && call->tfirst + call->nSoftAcked >= call->tnext) {
4099 call->state = RX_STATE_DALLY;
4100 rxi_ClearTransmitQueue(call, 0);
4101 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4102 } else if (!queue_IsEmpty(&call->tq)) {
4103 rxi_Start(0, call, 0, istack);
4108 /* Received a response to a challenge packet */
4110 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
4111 register struct rx_packet *np, int istack)
4115 /* Ignore the packet if we're the client */
4116 if (conn->type == RX_CLIENT_CONNECTION)
4119 /* If already authenticated, ignore the packet (it's probably a retry) */
4120 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4123 /* Otherwise, have the security object evaluate the response packet */
4124 error = RXS_CheckResponse(conn->securityObject, conn, np);
4126 /* If the response is invalid, reset the connection, sending
4127 * an abort to the peer */
4131 rxi_ConnectionError(conn, error);
4132 MUTEX_ENTER(&conn->conn_data_lock);
4133 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4134 MUTEX_EXIT(&conn->conn_data_lock);
4137 /* If the response is valid, any calls waiting to attach
4138 * servers can now do so */
4141 for (i = 0; i < RX_MAXCALLS; i++) {
4142 struct rx_call *call = conn->call[i];
4144 MUTEX_ENTER(&call->lock);
4145 if (call->state == RX_STATE_PRECALL)
4146 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4147 /* tnop can be null if newcallp is null */
4148 MUTEX_EXIT(&call->lock);
4152 /* Update the peer reachability information, just in case
4153 * some calls went into attach-wait while we were waiting
4154 * for authentication..
4156 rxi_UpdatePeerReach(conn, NULL);
4161 /* A client has received an authentication challenge: the security
4162 * object is asked to cough up a respectable response packet to send
4163 * back to the server. The server is responsible for retrying the
4164 * challenge if it fails to get a response. */
4167 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
4168 register struct rx_packet *np, int istack)
4172 /* Ignore the challenge if we're the server */
4173 if (conn->type == RX_SERVER_CONNECTION)
4176 /* Ignore the challenge if the connection is otherwise idle; someone's
4177 * trying to use us as an oracle. */
4178 if (!rxi_HasActiveCalls(conn))
4181 /* Send the security object the challenge packet. It is expected to fill
4182 * in the response. */
4183 error = RXS_GetResponse(conn->securityObject, conn, np);
4185 /* If the security object is unable to return a valid response, reset the
4186 * connection and send an abort to the peer. Otherwise send the response
4187 * packet to the peer connection. */
4189 rxi_ConnectionError(conn, error);
4190 MUTEX_ENTER(&conn->conn_data_lock);
4191 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4192 MUTEX_EXIT(&conn->conn_data_lock);
4194 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4195 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4201 /* Find an available server process to service the current request in
4202 * the given call structure. If one isn't available, queue up this
4203 * call so it eventually gets one */
4205 rxi_AttachServerProc(register struct rx_call *call,
4206 register osi_socket socket, register int *tnop,
4207 register struct rx_call **newcallp)
4209 register struct rx_serverQueueEntry *sq;
4210 register struct rx_service *service = call->conn->service;
4211 register int haveQuota = 0;
4213 /* May already be attached */
4214 if (call->state == RX_STATE_ACTIVE)
4217 MUTEX_ENTER(&rx_serverPool_lock);
4219 haveQuota = QuotaOK(service);
4220 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4221 /* If there are no processes available to service this call,
4222 * put the call on the incoming call queue (unless it's
4223 * already on the queue).
4225 #ifdef RX_ENABLE_LOCKS
4227 ReturnToServerPool(service);
4228 #endif /* RX_ENABLE_LOCKS */
4230 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4231 call->flags |= RX_CALL_WAIT_PROC;
4232 MUTEX_ENTER(&rx_stats_mutex);
4235 MUTEX_EXIT(&rx_stats_mutex);
4236 rxi_calltrace(RX_CALL_ARRIVAL, call);
4237 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4238 queue_Append(&rx_incomingCallQueue, call);
4241 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4243 /* If hot threads are enabled, and both newcallp and sq->socketp
4244 * are non-null, then this thread will process the call, and the
4245 * idle server thread will start listening on this threads socket.
4248 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4251 *sq->socketp = socket;
4252 clock_GetTime(&call->startTime);
4253 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4257 if (call->flags & RX_CALL_WAIT_PROC) {
4258 /* Conservative: I don't think this should happen */
4259 call->flags &= ~RX_CALL_WAIT_PROC;
4260 if (queue_IsOnQueue(call)) {
4262 MUTEX_ENTER(&rx_stats_mutex);
4264 MUTEX_EXIT(&rx_stats_mutex);
4267 call->state = RX_STATE_ACTIVE;
4268 call->mode = RX_MODE_RECEIVING;
4269 #ifdef RX_KERNEL_TRACE
4271 int glockOwner = ISAFS_GLOCK();
4274 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4275 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4281 if (call->flags & RX_CALL_CLEARED) {
4282 /* send an ack now to start the packet flow up again */
4283 call->flags &= ~RX_CALL_CLEARED;
4284 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4286 #ifdef RX_ENABLE_LOCKS
4289 service->nRequestsRunning++;
4290 if (service->nRequestsRunning <= service->minProcs)
4296 MUTEX_EXIT(&rx_serverPool_lock);
4299 /* Delay the sending of an acknowledge event for a short while, while
4300 * a new call is being prepared (in the case of a client) or a reply
4301 * is being prepared (in the case of a server). Rather than sending
4302 * an ack packet, an ACKALL packet is sent. */
4304 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4306 #ifdef RX_ENABLE_LOCKS
4308 MUTEX_ENTER(&call->lock);
4309 call->delayedAckEvent = NULL;
4310 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4312 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4313 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4315 MUTEX_EXIT(&call->lock);
4316 #else /* RX_ENABLE_LOCKS */
4318 call->delayedAckEvent = NULL;
4319 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4320 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4321 #endif /* RX_ENABLE_LOCKS */
4325 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused)
4327 struct rx_call *call = arg1;
4328 #ifdef RX_ENABLE_LOCKS
4330 MUTEX_ENTER(&call->lock);
4331 if (event == call->delayedAckEvent)
4332 call->delayedAckEvent = NULL;
4333 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4335 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4337 MUTEX_EXIT(&call->lock);
4338 #else /* RX_ENABLE_LOCKS */
4340 call->delayedAckEvent = NULL;
4341 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4342 #endif /* RX_ENABLE_LOCKS */
4346 #ifdef RX_ENABLE_LOCKS
4347 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4348 * clearing them out.
4351 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4353 register struct rx_packet *p, *tp;
4356 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4357 p->flags |= RX_PKTFLAG_ACKED;
4361 call->flags |= RX_CALL_TQ_CLEARME;
4362 call->flags |= RX_CALL_TQ_SOME_ACKED;
4365 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4366 call->tfirst = call->tnext;
4367 call->nSoftAcked = 0;
4369 if (call->flags & RX_CALL_FAST_RECOVER) {
4370 call->flags &= ~RX_CALL_FAST_RECOVER;
4371 call->cwind = call->nextCwind;
4372 call->nextCwind = 0;
4375 CV_SIGNAL(&call->cv_twind);
4377 #endif /* RX_ENABLE_LOCKS */
4379 /* Clear out the transmit queue for the current call (all packets have
4380 * been received by peer) */
4382 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4384 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4385 register struct rx_packet *p, *tp;
4387 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4389 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4390 p->flags |= RX_PKTFLAG_ACKED;
4394 call->flags |= RX_CALL_TQ_CLEARME;
4395 call->flags |= RX_CALL_TQ_SOME_ACKED;
4398 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4399 rxi_FreePackets(0, &call->tq);
4400 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4401 call->flags &= ~RX_CALL_TQ_CLEARME;
4403 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4405 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4406 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4407 call->nSoftAcked = 0;
4409 if (call->flags & RX_CALL_FAST_RECOVER) {
4410 call->flags &= ~RX_CALL_FAST_RECOVER;
4411 call->cwind = call->nextCwind;
4413 #ifdef RX_ENABLE_LOCKS
4414 CV_SIGNAL(&call->cv_twind);
4416 osi_rxWakeup(&call->twind);
4421 rxi_ClearReceiveQueue(register struct rx_call *call)
4423 if (queue_IsNotEmpty(&call->rq)) {
4424 rx_packetReclaims += rxi_FreePackets(0, &call->rq);
4425 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4427 if (call->state == RX_STATE_PRECALL) {
4428 call->flags |= RX_CALL_CLEARED;
4432 /* Send an abort packet for the specified call */
4434 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4435 int istack, int force)
4438 struct clock when, now;
4443 /* Clients should never delay abort messages */
4444 if (rx_IsClientConn(call->conn))
4447 if (call->abortCode != call->error) {
4448 call->abortCode = call->error;
4449 call->abortCount = 0;
4452 if (force || rxi_callAbortThreshhold == 0
4453 || call->abortCount < rxi_callAbortThreshhold) {
4454 if (call->delayedAbortEvent) {
4455 rxevent_Cancel(call->delayedAbortEvent, call,
4456 RX_CALL_REFCOUNT_ABORT);
4458 error = htonl(call->error);
4461 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4462 (char *)&error, sizeof(error), istack);
4463 } else if (!call->delayedAbortEvent) {
4464 clock_GetTime(&now);
4466 clock_Addmsec(&when, rxi_callAbortDelay);
4467 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4468 call->delayedAbortEvent =
4469 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4474 /* Send an abort packet for the specified connection. Packet is an
4475 * optional pointer to a packet that can be used to send the abort.
4476 * Once the number of abort messages reaches the threshhold, an
4477 * event is scheduled to send the abort. Setting the force flag
4478 * overrides sending delayed abort messages.
4480 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4481 * to send the abort packet.
4484 rxi_SendConnectionAbort(register struct rx_connection *conn,
4485 struct rx_packet *packet, int istack, int force)
4488 struct clock when, now;
4493 /* Clients should never delay abort messages */
4494 if (rx_IsClientConn(conn))
4497 if (force || rxi_connAbortThreshhold == 0
4498 || conn->abortCount < rxi_connAbortThreshhold) {
4499 if (conn->delayedAbortEvent) {
4500 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4502 error = htonl(conn->error);
4504 MUTEX_EXIT(&conn->conn_data_lock);
4506 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4507 RX_PACKET_TYPE_ABORT, (char *)&error,
4508 sizeof(error), istack);
4509 MUTEX_ENTER(&conn->conn_data_lock);
4510 } else if (!conn->delayedAbortEvent) {
4511 clock_GetTime(&now);
4513 clock_Addmsec(&when, rxi_connAbortDelay);
4514 conn->delayedAbortEvent =
4515 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4520 /* Associate an error all of the calls owned by a connection. Called
4521 * with error non-zero. This is only for really fatal things, like
4522 * bad authentication responses. The connection itself is set in
4523 * error at this point, so that future packets received will be
4526 rxi_ConnectionError(register struct rx_connection *conn,
4527 register afs_int32 error)
4532 dpf(("rxi_ConnectionError conn %x error %d", conn, error));
4534 MUTEX_ENTER(&conn->conn_data_lock);
4535 if (conn->challengeEvent)
4536 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4537 if (conn->checkReachEvent) {
4538 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4539 conn->checkReachEvent = 0;
4540 conn->flags &= ~RX_CONN_ATTACHWAIT;
4543 MUTEX_EXIT(&conn->conn_data_lock);
4544 for (i = 0; i < RX_MAXCALLS; i++) {
4545 struct rx_call *call = conn->call[i];
4547 MUTEX_ENTER(&call->lock);
4548 rxi_CallError(call, error);
4549 MUTEX_EXIT(&call->lock);
4552 conn->error = error;
4553 rx_MutexIncrement(rx_stats.fatalErrors, rx_stats_mutex);
4558 rxi_CallError(register struct rx_call *call, afs_int32 error)
4560 dpf(("rxi_CallError call %x error %d call->error %d", call, error, call->error));
4562 error = call->error;
4564 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4565 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4566 rxi_ResetCall(call, 0);
4569 rxi_ResetCall(call, 0);
4571 call->error = error;
4572 call->mode = RX_MODE_ERROR;
4575 /* Reset various fields in a call structure, and wakeup waiting
4576 * processes. Some fields aren't changed: state & mode are not
4577 * touched (these must be set by the caller), and bufptr, nLeft, and
4578 * nFree are not reset, since these fields are manipulated by
4579 * unprotected macros, and may only be reset by non-interrupting code.
4582 /* this code requires that call->conn be set properly as a pre-condition. */
4583 #endif /* ADAPT_WINDOW */
4586 rxi_ResetCall(register struct rx_call *call, register int newcall)
4589 register struct rx_peer *peer;
4590 struct rx_packet *packet;
4592 dpf(("rxi_ResetCall(call %x, newcall %d)\n", call, newcall));
4594 /* Notify anyone who is waiting for asynchronous packet arrival */
4595 if (call->arrivalProc) {
4596 (*call->arrivalProc) (call, call->arrivalProcHandle,
4597 call->arrivalProcArg);
4598 call->arrivalProc = (void (*)())0;
4601 if (call->delayedAbortEvent) {
4602 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4603 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4605 rxi_SendCallAbort(call, packet, 0, 1);
4606 rxi_FreePacket(packet);
4611 * Update the peer with the congestion information in this call
4612 * so other calls on this connection can pick up where this call
4613 * left off. If the congestion sequence numbers don't match then
4614 * another call experienced a retransmission.
4616 peer = call->conn->peer;
4617 MUTEX_ENTER(&peer->peer_lock);
4619 if (call->congestSeq == peer->congestSeq) {
4620 peer->cwind = MAX(peer->cwind, call->cwind);
4621 peer->MTU = MAX(peer->MTU, call->MTU);
4622 peer->nDgramPackets =
4623 MAX(peer->nDgramPackets, call->nDgramPackets);
4626 call->abortCode = 0;
4627 call->abortCount = 0;
4629 if (peer->maxDgramPackets > 1) {
4630 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4632 call->MTU = peer->MTU;
4634 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4635 call->ssthresh = rx_maxSendWindow;
4636 call->nDgramPackets = peer->nDgramPackets;
4637 call->congestSeq = peer->congestSeq;
4638 MUTEX_EXIT(&peer->peer_lock);
4640 flags = call->flags;
4641 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4642 if (flags & RX_CALL_TQ_BUSY) {
4643 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4644 call->flags |= (flags & RX_CALL_TQ_WAIT);
4646 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4648 rxi_ClearTransmitQueue(call, 1);
4649 /* why init the queue if you just emptied it? queue_Init(&call->tq); */
4650 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4651 dpf(("rcall %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4654 while (call->tqWaiters) {
4655 #ifdef RX_ENABLE_LOCKS
4656 CV_BROADCAST(&call->cv_tq);
4657 #else /* RX_ENABLE_LOCKS */
4658 osi_rxWakeup(&call->tq);
4659 #endif /* RX_ENABLE_LOCKS */
4664 rxi_ClearReceiveQueue(call);
4665 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
4667 if (call->currentPacket) {
4668 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
4669 rxi_FreePacket(call->currentPacket);
4670 call->currentPacket = (struct rx_packet *)0;
4672 call->curlen = call->nLeft = call->nFree = 0;
4674 rxi_FreePackets(0, &call->iovq);
4677 call->twind = call->conn->twind[call->channel];
4678 call->rwind = call->conn->rwind[call->channel];
4679 call->nSoftAcked = 0;
4680 call->nextCwind = 0;
4683 call->nCwindAcks = 0;
4684 call->nSoftAcks = 0;
4685 call->nHardAcks = 0;
4687 call->tfirst = call->rnext = call->tnext = 1;
4689 call->lastAcked = 0;
4690 call->localStatus = call->remoteStatus = 0;
4692 if (flags & RX_CALL_READER_WAIT) {
4693 #ifdef RX_ENABLE_LOCKS
4694 CV_BROADCAST(&call->cv_rq);
4696 osi_rxWakeup(&call->rq);
4699 if (flags & RX_CALL_WAIT_PACKETS) {
4700 MUTEX_ENTER(&rx_freePktQ_lock);
4701 rxi_PacketsUnWait(); /* XXX */
4702 MUTEX_EXIT(&rx_freePktQ_lock);
4704 #ifdef RX_ENABLE_LOCKS
4705 CV_SIGNAL(&call->cv_twind);
4707 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4708 osi_rxWakeup(&call->twind);
4711 #ifdef RX_ENABLE_LOCKS
4712 /* The following ensures that we don't mess with any queue while some
4713 * other thread might also be doing so. The call_queue_lock field is
4714 * is only modified under the call lock. If the call is in the process
4715 * of being removed from a queue, the call is not locked until the
4716 * the queue lock is dropped and only then is the call_queue_lock field
4717 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4718 * Note that any other routine which removes a call from a queue has to
4719 * obtain the queue lock before examing the queue and removing the call.
4721 if (call->call_queue_lock) {
4722 MUTEX_ENTER(call->call_queue_lock);
4723 if (queue_IsOnQueue(call)) {
4725 if (flags & RX_CALL_WAIT_PROC) {
4726 MUTEX_ENTER(&rx_stats_mutex);
4728 MUTEX_EXIT(&rx_stats_mutex);
4731 MUTEX_EXIT(call->call_queue_lock);
4732 CLEAR_CALL_QUEUE_LOCK(call);
4734 #else /* RX_ENABLE_LOCKS */
4735 if (queue_IsOnQueue(call)) {
4737 if (flags & RX_CALL_WAIT_PROC)
4740 #endif /* RX_ENABLE_LOCKS */
4742 rxi_KeepAliveOff(call);
4743 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4746 /* Send an acknowledge for the indicated packet (seq,serial) of the
4747 * indicated call, for the indicated reason (reason). This
4748 * acknowledge will specifically acknowledge receiving the packet, and
4749 * will also specify which other packets for this call have been
4750 * received. This routine returns the packet that was used to the
4751 * caller. The caller is responsible for freeing it or re-using it.
4752 * This acknowledgement also returns the highest sequence number
4753 * actually read out by the higher level to the sender; the sender
4754 * promises to keep around packets that have not been read by the
4755 * higher level yet (unless, of course, the sender decides to abort
4756 * the call altogether). Any of p, seq, serial, pflags, or reason may
4757 * be set to zero without ill effect. That is, if they are zero, they
4758 * will not convey any information.
4759 * NOW there is a trailer field, after the ack where it will safely be
4760 * ignored by mundanes, which indicates the maximum size packet this
4761 * host can swallow. */
4763 register struct rx_packet *optionalPacket; use to send ack (or null)
4764 int seq; Sequence number of the packet we are acking
4765 int serial; Serial number of the packet
4766 int pflags; Flags field from packet header
4767 int reason; Reason an acknowledge was prompted
4771 rxi_SendAck(register struct rx_call *call,
4772 register struct rx_packet *optionalPacket, int serial, int reason,
4775 struct rx_ackPacket *ap;
4776 register struct rx_packet *rqp;
4777 register struct rx_packet *nxp; /* For queue_Scan */
4778 register struct rx_packet *p;
4781 #ifdef RX_ENABLE_TSFPQ
4782 struct rx_ts_info_t * rx_ts_info;
4786 * Open the receive window once a thread starts reading packets
4788 if (call->rnext > 1) {
4789 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
4792 call->nHardAcks = 0;
4793 call->nSoftAcks = 0;
4794 if (call->rnext > call->lastAcked)
4795 call->lastAcked = call->rnext;
4799 rx_computelen(p, p->length); /* reset length, you never know */
4800 } /* where that's been... */
4801 #ifdef RX_ENABLE_TSFPQ
4803 RX_TS_INFO_GET(rx_ts_info);
4804 if ((p = rx_ts_info->local_special_packet)) {
4805 rx_computelen(p, p->length);
4806 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4807 rx_ts_info->local_special_packet = p;
4808 } else { /* We won't send the ack, but don't panic. */
4809 return optionalPacket;
4813 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4814 /* We won't send the ack, but don't panic. */
4815 return optionalPacket;
4820 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4823 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4824 #ifndef RX_ENABLE_TSFPQ
4825 if (!optionalPacket)
4828 return optionalPacket;
4830 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4831 if (rx_Contiguous(p) < templ) {
4832 #ifndef RX_ENABLE_TSFPQ
4833 if (!optionalPacket)
4836 return optionalPacket;
4841 /* MTUXXX failing to send an ack is very serious. We should */
4842 /* try as hard as possible to send even a partial ack; it's */
4843 /* better than nothing. */
4844 ap = (struct rx_ackPacket *)rx_DataOf(p);
4845 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4846 ap->reason = reason;
4848 /* The skew computation used to be bogus, I think it's better now. */
4849 /* We should start paying attention to skew. XXX */
4850 ap->serial = htonl(serial);
4851 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4853 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4854 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4856 /* No fear of running out of ack packet here because there can only be at most
4857 * one window full of unacknowledged packets. The window size must be constrained
4858 * to be less than the maximum ack size, of course. Also, an ack should always
4859 * fit into a single packet -- it should not ever be fragmented. */
4860 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4861 if (!rqp || !call->rq.next
4862 || (rqp->header.seq > (call->rnext + call->rwind))) {
4863 #ifndef RX_ENABLE_TSFPQ
4864 if (!optionalPacket)
4867 rxi_CallError(call, RX_CALL_DEAD);
4868 return optionalPacket;
4871 while (rqp->header.seq > call->rnext + offset)
4872 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4873 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4875 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4876 #ifndef RX_ENABLE_TSFPQ
4877 if (!optionalPacket)
4880 rxi_CallError(call, RX_CALL_DEAD);
4881 return optionalPacket;
4886 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4888 /* these are new for AFS 3.3 */
4889 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4890 templ = htonl(templ);
4891 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4892 templ = htonl(call->conn->peer->ifMTU);
4893 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4894 sizeof(afs_int32), &templ);
4896 /* new for AFS 3.4 */
4897 templ = htonl(call->rwind);
4898 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4899 sizeof(afs_int32), &templ);
4901 /* new for AFS 3.5 */
4902 templ = htonl(call->conn->peer->ifDgramPackets);
4903 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4904 sizeof(afs_int32), &templ);
4906 p->header.serviceId = call->conn->serviceId;
4907 p->header.cid = (call->conn->cid | call->channel);
4908 p->header.callNumber = *call->callNumber;
4910 p->header.securityIndex = call->conn->securityIndex;
4911 p->header.epoch = call->conn->epoch;
4912 p->header.type = RX_PACKET_TYPE_ACK;
4913 p->header.flags = RX_SLOW_START_OK;
4914 if (reason == RX_ACK_PING) {
4915 p->header.flags |= RX_REQUEST_ACK;
4917 clock_GetTime(&call->pingRequestTime);
4920 if (call->conn->type == RX_CLIENT_CONNECTION)
4921 p->header.flags |= RX_CLIENT_INITIATED;
4925 if (rxdebug_active) {
4929 len = _snprintf(msg, sizeof(msg),
4930 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4931 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4932 ntohl(ap->serial), ntohl(ap->previousPacket),
4933 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
4934 ap->nAcks, ntohs(ap->bufferSpace) );
4938 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
4939 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4943 OutputDebugString(msg);
4945 #else /* AFS_NT40_ENV */
4947 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
4948 ap->reason, ntohl(ap->previousPacket),
4949 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4951 for (offset = 0; offset < ap->nAcks; offset++)
4952 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4957 #endif /* AFS_NT40_ENV */
4960 register int i, nbytes = p->length;
4962 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4963 if (nbytes <= p->wirevec[i].iov_len) {
4964 register int savelen, saven;
4966 savelen = p->wirevec[i].iov_len;
4968 p->wirevec[i].iov_len = nbytes;
4970 rxi_Send(call, p, istack);
4971 p->wirevec[i].iov_len = savelen;
4975 nbytes -= p->wirevec[i].iov_len;
4978 rx_MutexIncrement(rx_stats.ackPacketsSent, rx_stats_mutex);
4979 #ifndef RX_ENABLE_TSFPQ
4980 if (!optionalPacket)
4983 return optionalPacket; /* Return packet for re-use by caller */
4986 /* Send all of the packets in the list in single datagram */
4988 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4989 int istack, int moreFlag, struct clock *now,
4990 struct clock *retryTime, int resending)
4995 struct rx_connection *conn = call->conn;
4996 struct rx_peer *peer = conn->peer;
4998 MUTEX_ENTER(&peer->peer_lock);
5001 peer->reSends += len;
5002 rx_MutexIncrement(rx_stats.dataPacketsSent, rx_stats_mutex);
5003 MUTEX_EXIT(&peer->peer_lock);
5005 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
5009 /* Set the packet flags and schedule the resend events */
5010 /* Only request an ack for the last packet in the list */
5011 for (i = 0; i < len; i++) {
5012 list[i]->retryTime = *retryTime;
5013 if (list[i]->header.serial) {
5014 /* Exponentially backoff retry times */
5015 if (list[i]->backoff < MAXBACKOFF) {
5016 /* so it can't stay == 0 */
5017 list[i]->backoff = (list[i]->backoff << 1) + 1;
5020 clock_Addmsec(&(list[i]->retryTime),
5021 ((afs_uint32) list[i]->backoff) << 8);
5024 /* Wait a little extra for the ack on the last packet */
5025 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
5026 clock_Addmsec(&(list[i]->retryTime), 400);
5029 /* Record the time sent */
5030 list[i]->timeSent = *now;
5032 /* Ask for an ack on retransmitted packets, on every other packet
5033 * if the peer doesn't support slow start. Ask for an ack on every
5034 * packet until the congestion window reaches the ack rate. */
5035 if (list[i]->header.serial) {
5037 rx_MutexIncrement(rx_stats.dataPacketsReSent, rx_stats_mutex);
5039 /* improved RTO calculation- not Karn */
5040 list[i]->firstSent = *now;
5041 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5042 || (!(call->flags & RX_CALL_SLOW_START_OK)
5043 && (list[i]->header.seq & 1)))) {
5048 MUTEX_ENTER(&peer->peer_lock);
5052 rx_MutexIncrement(rx_stats.dataPacketsSent, rx_stats_mutex);
5053 MUTEX_EXIT(&peer->peer_lock);
5055 /* Tag this packet as not being the last in this group,
5056 * for the receiver's benefit */
5057 if (i < len - 1 || moreFlag) {
5058 list[i]->header.flags |= RX_MORE_PACKETS;
5061 /* Install the new retransmit time for the packet, and
5062 * record the time sent */
5063 list[i]->timeSent = *now;
5067 list[len - 1]->header.flags |= RX_REQUEST_ACK;
5070 /* Since we're about to send a data packet to the peer, it's
5071 * safe to nuke any scheduled end-of-packets ack */
5072 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5074 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5075 MUTEX_EXIT(&call->lock);
5077 rxi_SendPacketList(call, conn, list, len, istack);
5079 rxi_SendPacket(call, conn, list[0], istack);
5081 MUTEX_ENTER(&call->lock);
5082 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5084 /* Update last send time for this call (for keep-alive
5085 * processing), and for the connection (so that we can discover
5086 * idle connections) */
5087 call->lastSendData = conn->lastSendTime = call->lastSendTime = clock_Sec();
5090 /* When sending packets we need to follow these rules:
5091 * 1. Never send more than maxDgramPackets in a jumbogram.
5092 * 2. Never send a packet with more than two iovecs in a jumbogram.
5093 * 3. Never send a retransmitted packet in a jumbogram.
5094 * 4. Never send more than cwind/4 packets in a jumbogram
5095 * We always keep the last list we should have sent so we
5096 * can set the RX_MORE_PACKETS flags correctly.
5099 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5100 int istack, struct clock *now, struct clock *retryTime,
5103 int i, cnt, lastCnt = 0;
5104 struct rx_packet **listP, **lastP = 0;
5105 struct rx_peer *peer = call->conn->peer;
5106 int morePackets = 0;
5108 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
5109 /* Does the current packet force us to flush the current list? */
5111 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5112 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5114 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
5116 /* If the call enters an error state stop sending, or if
5117 * we entered congestion recovery mode, stop sending */
5118 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5126 /* Add the current packet to the list if it hasn't been acked.
5127 * Otherwise adjust the list pointer to skip the current packet. */
5128 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5130 /* Do we need to flush the list? */
5131 if (cnt >= (int)peer->maxDgramPackets
5132 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5133 || list[i]->header.serial
5134 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5136 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5137 retryTime, resending);
5138 /* If the call enters an error state stop sending, or if
5139 * we entered congestion recovery mode, stop sending */
5141 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5146 listP = &list[i + 1];
5151 osi_Panic("rxi_SendList error");
5153 listP = &list[i + 1];
5157 /* Send the whole list when the call is in receive mode, when
5158 * the call is in eof mode, when we are in fast recovery mode,
5159 * and when we have the last packet */
5160 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5161 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5162 || (call->flags & RX_CALL_FAST_RECOVER)) {
5163 /* Check for the case where the current list contains
5164 * an acked packet. Since we always send retransmissions
5165 * in a separate packet, we only need to check the first
5166 * packet in the list */
5167 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5171 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5172 retryTime, resending);
5173 /* If the call enters an error state stop sending, or if
5174 * we entered congestion recovery mode, stop sending */
5175 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5179 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5182 } else if (lastCnt > 0) {
5183 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5188 #ifdef RX_ENABLE_LOCKS
5189 /* Call rxi_Start, below, but with the call lock held. */
5191 rxi_StartUnlocked(struct rxevent *event,
5192 void *arg0, void *arg1, int istack)
5194 struct rx_call *call = arg0;
5196 MUTEX_ENTER(&call->lock);
5197 rxi_Start(event, call, arg1, istack);
5198 MUTEX_EXIT(&call->lock);
5200 #endif /* RX_ENABLE_LOCKS */
5202 /* This routine is called when new packets are readied for
5203 * transmission and when retransmission may be necessary, or when the
5204 * transmission window or burst count are favourable. This should be
5205 * better optimized for new packets, the usual case, now that we've
5206 * got rid of queues of send packets. XXXXXXXXXXX */
5208 rxi_Start(struct rxevent *event,
5209 void *arg0, void *arg1, int istack)
5211 struct rx_call *call = arg0;
5213 struct rx_packet *p;
5214 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
5215 struct rx_peer *peer = call->conn->peer;
5216 struct clock now, usenow, retryTime;
5220 struct rx_packet **xmitList;
5223 /* If rxi_Start is being called as a result of a resend event,
5224 * then make sure that the event pointer is removed from the call
5225 * structure, since there is no longer a per-call retransmission
5227 if (event && event == call->resendEvent) {
5228 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5229 call->resendEvent = NULL;
5231 if (queue_IsEmpty(&call->tq)) {
5235 /* Timeouts trigger congestion recovery */
5236 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5237 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5238 /* someone else is waiting to start recovery */
5241 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5242 rxi_WaitforTQBusy(call);
5243 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5244 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5245 call->flags |= RX_CALL_FAST_RECOVER;
5246 if (peer->maxDgramPackets > 1) {
5247 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5249 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5251 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5252 call->nDgramPackets = 1;
5254 call->nextCwind = 1;
5257 MUTEX_ENTER(&peer->peer_lock);
5258 peer->MTU = call->MTU;
5259 peer->cwind = call->cwind;
5260 peer->nDgramPackets = 1;
5262 call->congestSeq = peer->congestSeq;
5263 MUTEX_EXIT(&peer->peer_lock);
5264 /* Clear retry times on packets. Otherwise, it's possible for
5265 * some packets in the queue to force resends at rates faster
5266 * than recovery rates.
5268 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5269 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5270 clock_Zero(&p->retryTime);
5275 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5276 rx_MutexIncrement(rx_tq_debug.rxi_start_in_error, rx_stats_mutex);
5281 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5282 /* Get clock to compute the re-transmit time for any packets
5283 * in this burst. Note, if we back off, it's reasonable to
5284 * back off all of the packets in the same manner, even if
5285 * some of them have been retransmitted more times than more
5287 * Do a dance to avoid blocking after setting now. */
5288 clock_Zero(&retryTime);
5289 MUTEX_ENTER(&peer->peer_lock);
5290 clock_Add(&retryTime, &peer->timeout);
5291 MUTEX_EXIT(&peer->peer_lock);
5292 clock_GetTime(&now);
5293 clock_Add(&retryTime, &now);
5295 /* Send (or resend) any packets that need it, subject to
5296 * window restrictions and congestion burst control
5297 * restrictions. Ask for an ack on the last packet sent in
5298 * this burst. For now, we're relying upon the window being
5299 * considerably bigger than the largest number of packets that
5300 * are typically sent at once by one initial call to
5301 * rxi_Start. This is probably bogus (perhaps we should ask
5302 * for an ack when we're half way through the current
5303 * window?). Also, for non file transfer applications, this
5304 * may end up asking for an ack for every packet. Bogus. XXXX
5307 * But check whether we're here recursively, and let the other guy
5310 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5311 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5312 call->flags |= RX_CALL_TQ_BUSY;
5314 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5316 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5317 call->flags &= ~RX_CALL_NEED_START;
5318 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5320 maxXmitPackets = MIN(call->twind, call->cwind);
5321 xmitList = (struct rx_packet **)
5322 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
5323 /* XXXX else we must drop any mtx we hold */
5324 afs_osi_Alloc_NoSleep(maxXmitPackets * sizeof(struct rx_packet *));
5326 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5328 if (xmitList == NULL)
5329 osi_Panic("rxi_Start, failed to allocate xmit list");
5330 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5331 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5332 /* We shouldn't be sending packets if a thread is waiting
5333 * to initiate congestion recovery */
5337 && (call->flags & RX_CALL_FAST_RECOVER)) {
5338 /* Only send one packet during fast recovery */
5341 if ((p->flags & RX_PKTFLAG_FREE)
5342 || (!queue_IsEnd(&call->tq, nxp)
5343 && (nxp->flags & RX_PKTFLAG_FREE))
5344 || (p == (struct rx_packet *)&rx_freePacketQueue)
5345 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5346 osi_Panic("rxi_Start: xmit queue clobbered");
5348 if (p->flags & RX_PKTFLAG_ACKED) {
5349 /* Since we may block, don't trust this */
5350 usenow.sec = usenow.usec = 0;
5351 rx_MutexIncrement(rx_stats.ignoreAckedPacket, rx_stats_mutex);
5352 continue; /* Ignore this packet if it has been acknowledged */
5355 /* Turn off all flags except these ones, which are the same
5356 * on each transmission */
5357 p->header.flags &= RX_PRESET_FLAGS;
5359 if (p->header.seq >=
5360 call->tfirst + MIN((int)call->twind,
5361 (int)(call->nSoftAcked +
5363 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5364 /* Note: if we're waiting for more window space, we can
5365 * still send retransmits; hence we don't return here, but
5366 * break out to schedule a retransmit event */
5367 dpf(("call %d waiting for window",
5368 *(call->callNumber)));
5372 /* Transmit the packet if it needs to be sent. */
5373 if (!clock_Lt(&now, &p->retryTime)) {
5374 if (nXmitPackets == maxXmitPackets) {
5375 rxi_SendXmitList(call, xmitList, nXmitPackets,
5376 istack, &now, &retryTime,
5378 osi_Free(xmitList, maxXmitPackets *
5379 sizeof(struct rx_packet *));
5382 xmitList[nXmitPackets++] = p;
5386 /* xmitList now hold pointers to all of the packets that are
5387 * ready to send. Now we loop to send the packets */
5388 if (nXmitPackets > 0) {
5389 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5390 &now, &retryTime, resending);
5393 maxXmitPackets * sizeof(struct rx_packet *));
5395 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5397 * TQ references no longer protected by this flag; they must remain
5398 * protected by the global lock.
5400 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5401 call->flags &= ~RX_CALL_TQ_BUSY;
5402 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5403 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5404 #ifdef RX_ENABLE_LOCKS
5405 osirx_AssertMine(&call->lock, "rxi_Start start");
5406 CV_BROADCAST(&call->cv_tq);
5407 #else /* RX_ENABLE_LOCKS */
5408 osi_rxWakeup(&call->tq);
5409 #endif /* RX_ENABLE_LOCKS */
5414 /* We went into the error state while sending packets. Now is
5415 * the time to reset the call. This will also inform the using
5416 * process that the call is in an error state.
5418 rx_MutexIncrement(rx_tq_debug.rxi_start_aborted, rx_stats_mutex);
5419 call->flags &= ~RX_CALL_TQ_BUSY;
5420 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5421 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5422 #ifdef RX_ENABLE_LOCKS
5423 osirx_AssertMine(&call->lock, "rxi_Start middle");
5424 CV_BROADCAST(&call->cv_tq);
5425 #else /* RX_ENABLE_LOCKS */
5426 osi_rxWakeup(&call->tq);
5427 #endif /* RX_ENABLE_LOCKS */
5429 rxi_CallError(call, call->error);
5432 #ifdef RX_ENABLE_LOCKS
5433 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5434 register int missing;
5435 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5436 /* Some packets have received acks. If they all have, we can clear
5437 * the transmit queue.
5440 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5441 if (p->header.seq < call->tfirst
5442 && (p->flags & RX_PKTFLAG_ACKED)) {
5444 p->flags &= ~RX_PKTFLAG_TQ;
5450 call->flags |= RX_CALL_TQ_CLEARME;
5452 #endif /* RX_ENABLE_LOCKS */
5453 /* Don't bother doing retransmits if the TQ is cleared. */
5454 if (call->flags & RX_CALL_TQ_CLEARME) {
5455 rxi_ClearTransmitQueue(call, 1);
5457 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5460 /* Always post a resend event, if there is anything in the
5461 * queue, and resend is possible. There should be at least
5462 * one unacknowledged packet in the queue ... otherwise none
5463 * of these packets should be on the queue in the first place.
5465 if (call->resendEvent) {
5466 /* Cancel the existing event and post a new one */
5467 rxevent_Cancel(call->resendEvent, call,
5468 RX_CALL_REFCOUNT_RESEND);
5471 /* The retry time is the retry time on the first unacknowledged
5472 * packet inside the current window */
5474 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5475 /* Don't set timers for packets outside the window */
5476 if (p->header.seq >= call->tfirst + call->twind) {
5480 if (!(p->flags & RX_PKTFLAG_ACKED)
5481 && !clock_IsZero(&p->retryTime)) {
5483 retryTime = p->retryTime;
5488 /* Post a new event to re-run rxi_Start when retries may be needed */
5489 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5490 #ifdef RX_ENABLE_LOCKS
5491 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5493 rxevent_PostNow2(&retryTime, &usenow,
5495 (void *)call, 0, istack);
5496 #else /* RX_ENABLE_LOCKS */
5498 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5499 (void *)call, 0, istack);
5500 #endif /* RX_ENABLE_LOCKS */
5503 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5504 } while (call->flags & RX_CALL_NEED_START);
5506 * TQ references no longer protected by this flag; they must remain
5507 * protected by the global lock.
5509 call->flags &= ~RX_CALL_TQ_BUSY;
5510 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5511 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5512 #ifdef RX_ENABLE_LOCKS
5513 osirx_AssertMine(&call->lock, "rxi_Start end");
5514 CV_BROADCAST(&call->cv_tq);
5515 #else /* RX_ENABLE_LOCKS */
5516 osi_rxWakeup(&call->tq);
5517 #endif /* RX_ENABLE_LOCKS */
5520 call->flags |= RX_CALL_NEED_START;
5522 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5524 if (call->resendEvent) {
5525 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5530 /* Also adjusts the keep alive parameters for the call, to reflect
5531 * that we have just sent a packet (so keep alives aren't sent
5534 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5537 register struct rx_connection *conn = call->conn;
5539 /* Stamp each packet with the user supplied status */
5540 p->header.userStatus = call->localStatus;
5542 /* Allow the security object controlling this call's security to
5543 * make any last-minute changes to the packet */
5544 RXS_SendPacket(conn->securityObject, call, p);
5546 /* Since we're about to send SOME sort of packet to the peer, it's
5547 * safe to nuke any scheduled end-of-packets ack */
5548 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5550 /* Actually send the packet, filling in more connection-specific fields */
5551 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5552 MUTEX_EXIT(&call->lock);
5553 rxi_SendPacket(call, conn, p, istack);
5554 MUTEX_ENTER(&call->lock);
5555 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5557 /* Update last send time for this call (for keep-alive
5558 * processing), and for the connection (so that we can discover
5559 * idle connections) */
5560 conn->lastSendTime = call->lastSendTime = clock_Sec();
5561 /* Don't count keepalives here, so idleness can be tracked. */
5562 if (p->header.type != RX_PACKET_TYPE_ACK)
5563 call->lastSendData = call->lastSendTime;
5567 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5568 * that things are fine. Also called periodically to guarantee that nothing
5569 * falls through the cracks (e.g. (error + dally) connections have keepalive
5570 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5572 * haveCTLock Set if calling from rxi_ReapConnections
5574 #ifdef RX_ENABLE_LOCKS
5576 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5577 #else /* RX_ENABLE_LOCKS */
5579 rxi_CheckCall(register struct rx_call *call)
5580 #endif /* RX_ENABLE_LOCKS */
5582 register struct rx_connection *conn = call->conn;
5584 afs_uint32 deadTime;
5586 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5587 if (call->flags & RX_CALL_TQ_BUSY) {
5588 /* Call is active and will be reset by rxi_Start if it's
5589 * in an error state.
5594 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5596 (((afs_uint32) conn->secondsUntilDead << 10) +
5597 ((afs_uint32) conn->peer->rtt >> 3) +
5598 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5600 /* These are computed to the second (+- 1 second). But that's
5601 * good enough for these values, which should be a significant
5602 * number of seconds. */
5603 if (now > (call->lastReceiveTime + deadTime)) {
5604 if (call->state == RX_STATE_ACTIVE) {
5606 #if defined(KERNEL) && defined(AFS_SUN57_ENV)
5608 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5609 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
5610 ip_stack_t *ipst = ns->netstack_ip;
5612 ire = ire_cache_lookup(call->conn->peer->host
5613 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
5615 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
5617 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5624 if (ire && ire->ire_max_frag > 0)
5625 rxi_SetPeerMtu(call->conn->peer->host, 0, ire->ire_max_frag);
5626 #if defined(GLOBAL_NETSTACKID)
5630 #endif /* ADAPT_PMTU */
5631 rxi_CallError(call, RX_CALL_DEAD);
5634 #ifdef RX_ENABLE_LOCKS
5635 /* Cancel pending events */
5636 rxevent_Cancel(call->delayedAckEvent, call,
5637 RX_CALL_REFCOUNT_DELAY);
5638 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5639 rxevent_Cancel(call->keepAliveEvent, call,
5640 RX_CALL_REFCOUNT_ALIVE);
5641 if (call->refCount == 0) {
5642 rxi_FreeCall(call, haveCTLock);
5646 #else /* RX_ENABLE_LOCKS */
5649 #endif /* RX_ENABLE_LOCKS */
5651 /* Non-active calls are destroyed if they are not responding
5652 * to pings; active calls are simply flagged in error, so the
5653 * attached process can die reasonably gracefully. */
5655 /* see if we have a non-activity timeout */
5656 if (call->startWait && conn->idleDeadTime
5657 && ((call->startWait + conn->idleDeadTime) < now)) {
5658 if (call->state == RX_STATE_ACTIVE) {
5659 rxi_CallError(call, RX_CALL_TIMEOUT);
5663 if (call->lastSendData && conn->idleDeadTime && (conn->idleDeadErr != 0)
5664 && ((call->lastSendData + conn->idleDeadTime) < now)) {
5665 if (call->state == RX_STATE_ACTIVE) {
5666 rxi_CallError(call, conn->idleDeadErr);
5670 /* see if we have a hard timeout */
5671 if (conn->hardDeadTime
5672 && (now > (conn->hardDeadTime + call->startTime.sec))) {
5673 if (call->state == RX_STATE_ACTIVE)
5674 rxi_CallError(call, RX_CALL_TIMEOUT);
5681 /* When a call is in progress, this routine is called occasionally to
5682 * make sure that some traffic has arrived (or been sent to) the peer.
5683 * If nothing has arrived in a reasonable amount of time, the call is
5684 * declared dead; if nothing has been sent for a while, we send a
5685 * keep-alive packet (if we're actually trying to keep the call alive)
5688 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
5690 struct rx_call *call = arg1;
5691 struct rx_connection *conn;
5694 MUTEX_ENTER(&call->lock);
5695 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5696 if (event == call->keepAliveEvent)
5697 call->keepAliveEvent = NULL;
5700 #ifdef RX_ENABLE_LOCKS
5701 if (rxi_CheckCall(call, 0)) {
5702 MUTEX_EXIT(&call->lock);
5705 #else /* RX_ENABLE_LOCKS */
5706 if (rxi_CheckCall(call))
5708 #endif /* RX_ENABLE_LOCKS */
5710 /* Don't try to keep alive dallying calls */
5711 if (call->state == RX_STATE_DALLY) {
5712 MUTEX_EXIT(&call->lock);
5717 if ((now - call->lastSendTime) > conn->secondsUntilPing) {
5718 /* Don't try to send keepalives if there is unacknowledged data */
5719 /* the rexmit code should be good enough, this little hack
5720 * doesn't quite work XXX */
5721 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5723 rxi_ScheduleKeepAliveEvent(call);
5724 MUTEX_EXIT(&call->lock);
5729 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5731 if (!call->keepAliveEvent) {
5732 struct clock when, now;
5733 clock_GetTime(&now);
5735 when.sec += call->conn->secondsUntilPing;
5736 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5737 call->keepAliveEvent =
5738 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
5742 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5744 rxi_KeepAliveOn(register struct rx_call *call)
5746 /* Pretend last packet received was received now--i.e. if another
5747 * packet isn't received within the keep alive time, then the call
5748 * will die; Initialize last send time to the current time--even
5749 * if a packet hasn't been sent yet. This will guarantee that a
5750 * keep-alive is sent within the ping time */
5751 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5752 rxi_ScheduleKeepAliveEvent(call);
5755 /* This routine is called to send connection abort messages
5756 * that have been delayed to throttle looping clients. */
5758 rxi_SendDelayedConnAbort(struct rxevent *event,
5759 void *arg1, void *unused)
5761 struct rx_connection *conn = arg1;
5764 struct rx_packet *packet;
5766 MUTEX_ENTER(&conn->conn_data_lock);
5767 conn->delayedAbortEvent = NULL;
5768 error = htonl(conn->error);
5770 MUTEX_EXIT(&conn->conn_data_lock);
5771 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5774 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5775 RX_PACKET_TYPE_ABORT, (char *)&error,
5777 rxi_FreePacket(packet);
5781 /* This routine is called to send call abort messages
5782 * that have been delayed to throttle looping clients. */
5784 rxi_SendDelayedCallAbort(struct rxevent *event,
5785 void *arg1, void *dummy)
5787 struct rx_call *call = arg1;
5790 struct rx_packet *packet;
5792 MUTEX_ENTER(&call->lock);
5793 call->delayedAbortEvent = NULL;
5794 error = htonl(call->error);
5796 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5799 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5800 (char *)&error, sizeof(error), 0);
5801 rxi_FreePacket(packet);
5803 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5804 MUTEX_EXIT(&call->lock);
5807 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5808 * seconds) to ask the client to authenticate itself. The routine
5809 * issues a challenge to the client, which is obtained from the
5810 * security object associated with the connection */
5812 rxi_ChallengeEvent(struct rxevent *event,
5813 void *arg0, void *arg1, int tries)
5815 struct rx_connection *conn = arg0;
5817 conn->challengeEvent = NULL;
5818 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5819 register struct rx_packet *packet;
5820 struct clock when, now;
5823 /* We've failed to authenticate for too long.
5824 * Reset any calls waiting for authentication;
5825 * they are all in RX_STATE_PRECALL.
5829 MUTEX_ENTER(&conn->conn_call_lock);
5830 for (i = 0; i < RX_MAXCALLS; i++) {
5831 struct rx_call *call = conn->call[i];
5833 MUTEX_ENTER(&call->lock);
5834 if (call->state == RX_STATE_PRECALL) {
5835 rxi_CallError(call, RX_CALL_DEAD);
5836 rxi_SendCallAbort(call, NULL, 0, 0);
5838 MUTEX_EXIT(&call->lock);
5841 MUTEX_EXIT(&conn->conn_call_lock);
5845 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5847 /* If there's no packet available, do this later. */
5848 RXS_GetChallenge(conn->securityObject, conn, packet);
5849 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5850 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5851 rxi_FreePacket(packet);
5853 clock_GetTime(&now);
5855 when.sec += RX_CHALLENGE_TIMEOUT;
5856 conn->challengeEvent =
5857 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
5862 /* Call this routine to start requesting the client to authenticate
5863 * itself. This will continue until authentication is established,
5864 * the call times out, or an invalid response is returned. The
5865 * security object associated with the connection is asked to create
5866 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5867 * defined earlier. */
5869 rxi_ChallengeOn(register struct rx_connection *conn)
5871 if (!conn->challengeEvent) {
5872 RXS_CreateChallenge(conn->securityObject, conn);
5873 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5878 /* Compute round trip time of the packet provided, in *rttp.
5881 /* rxi_ComputeRoundTripTime is called with peer locked. */
5882 /* sentp and/or peer may be null */
5884 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5885 register struct clock *sentp,
5886 register struct rx_peer *peer)
5888 struct clock thisRtt, *rttp = &thisRtt;
5890 register int rtt_timeout;
5892 clock_GetTime(rttp);
5894 if (clock_Lt(rttp, sentp)) {
5896 return; /* somebody set the clock back, don't count this time. */
5898 clock_Sub(rttp, sentp);
5899 MUTEX_ENTER(&rx_stats_mutex);
5900 if (clock_Lt(rttp, &rx_stats.minRtt))
5901 rx_stats.minRtt = *rttp;
5902 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5903 if (rttp->sec > 60) {
5904 MUTEX_EXIT(&rx_stats_mutex);
5905 return; /* somebody set the clock ahead */
5907 rx_stats.maxRtt = *rttp;
5909 clock_Add(&rx_stats.totalRtt, rttp);
5910 rx_stats.nRttSamples++;
5911 MUTEX_EXIT(&rx_stats_mutex);
5913 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5915 /* Apply VanJacobson round-trip estimations */
5920 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5921 * srtt is stored as fixed point with 3 bits after the binary
5922 * point (i.e., scaled by 8). The following magic is
5923 * equivalent to the smoothing algorithm in rfc793 with an
5924 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5925 * srtt*8 = srtt*8 + rtt - srtt
5926 * srtt = srtt + rtt/8 - srtt/8
5929 delta = MSEC(rttp) - (peer->rtt >> 3);
5933 * We accumulate a smoothed rtt variance (actually, a smoothed
5934 * mean difference), then set the retransmit timer to smoothed
5935 * rtt + 4 times the smoothed variance (was 2x in van's original
5936 * paper, but 4x works better for me, and apparently for him as
5938 * rttvar is stored as
5939 * fixed point with 2 bits after the binary point (scaled by
5940 * 4). The following is equivalent to rfc793 smoothing with
5941 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5942 * replaces rfc793's wired-in beta.
5943 * dev*4 = dev*4 + (|actual - expected| - dev)
5949 delta -= (peer->rtt_dev >> 2);
5950 peer->rtt_dev += delta;
5952 /* I don't have a stored RTT so I start with this value. Since I'm
5953 * probably just starting a call, and will be pushing more data down
5954 * this, I expect congestion to increase rapidly. So I fudge a
5955 * little, and I set deviance to half the rtt. In practice,
5956 * deviance tends to approach something a little less than
5957 * half the smoothed rtt. */
5958 peer->rtt = (MSEC(rttp) << 3) + 8;
5959 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5961 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5962 * the other of these connections is usually in a user process, and can
5963 * be switched and/or swapped out. So on fast, reliable networks, the
5964 * timeout would otherwise be too short.
5966 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5967 clock_Zero(&(peer->timeout));
5968 clock_Addmsec(&(peer->timeout), rtt_timeout);
5970 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)));
5974 /* Find all server connections that have not been active for a long time, and
5977 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2)
5979 struct clock now, when;
5980 clock_GetTime(&now);
5982 /* Find server connection structures that haven't been used for
5983 * greater than rx_idleConnectionTime */
5985 struct rx_connection **conn_ptr, **conn_end;
5986 int i, havecalls = 0;
5987 MUTEX_ENTER(&rx_connHashTable_lock);
5988 for (conn_ptr = &rx_connHashTable[0], conn_end =
5989 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5991 struct rx_connection *conn, *next;
5992 struct rx_call *call;
5996 for (conn = *conn_ptr; conn; conn = next) {
5997 /* XXX -- Shouldn't the connection be locked? */
6000 for (i = 0; i < RX_MAXCALLS; i++) {
6001 call = conn->call[i];
6004 MUTEX_ENTER(&call->lock);
6005 #ifdef RX_ENABLE_LOCKS
6006 result = rxi_CheckCall(call, 1);
6007 #else /* RX_ENABLE_LOCKS */
6008 result = rxi_CheckCall(call);
6009 #endif /* RX_ENABLE_LOCKS */
6010 MUTEX_EXIT(&call->lock);
6012 /* If CheckCall freed the call, it might
6013 * have destroyed the connection as well,
6014 * which screws up the linked lists.
6020 if (conn->type == RX_SERVER_CONNECTION) {
6021 /* This only actually destroys the connection if
6022 * there are no outstanding calls */
6023 MUTEX_ENTER(&conn->conn_data_lock);
6024 if (!havecalls && !conn->refCount
6025 && ((conn->lastSendTime + rx_idleConnectionTime) <
6027 conn->refCount++; /* it will be decr in rx_DestroyConn */
6028 MUTEX_EXIT(&conn->conn_data_lock);
6029 #ifdef RX_ENABLE_LOCKS
6030 rxi_DestroyConnectionNoLock(conn);
6031 #else /* RX_ENABLE_LOCKS */
6032 rxi_DestroyConnection(conn);
6033 #endif /* RX_ENABLE_LOCKS */
6035 #ifdef RX_ENABLE_LOCKS
6037 MUTEX_EXIT(&conn->conn_data_lock);
6039 #endif /* RX_ENABLE_LOCKS */
6043 #ifdef RX_ENABLE_LOCKS
6044 while (rx_connCleanup_list) {
6045 struct rx_connection *conn;
6046 conn = rx_connCleanup_list;
6047 rx_connCleanup_list = rx_connCleanup_list->next;
6048 MUTEX_EXIT(&rx_connHashTable_lock);
6049 rxi_CleanupConnection(conn);
6050 MUTEX_ENTER(&rx_connHashTable_lock);
6052 MUTEX_EXIT(&rx_connHashTable_lock);
6053 #endif /* RX_ENABLE_LOCKS */
6056 /* Find any peer structures that haven't been used (haven't had an
6057 * associated connection) for greater than rx_idlePeerTime */
6059 struct rx_peer **peer_ptr, **peer_end;
6061 MUTEX_ENTER(&rx_rpc_stats);
6062 MUTEX_ENTER(&rx_peerHashTable_lock);
6063 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6064 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6066 struct rx_peer *peer, *next, *prev;
6067 for (prev = peer = *peer_ptr; peer; peer = next) {
6069 code = MUTEX_TRYENTER(&peer->peer_lock);
6070 if ((code) && (peer->refCount == 0)
6071 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6072 rx_interface_stat_p rpc_stat, nrpc_stat;
6074 MUTEX_EXIT(&peer->peer_lock);
6075 MUTEX_DESTROY(&peer->peer_lock);
6077 (&peer->rpcStats, rpc_stat, nrpc_stat,
6078 rx_interface_stat)) {
6079 unsigned int num_funcs;
6082 queue_Remove(&rpc_stat->queue_header);
6083 queue_Remove(&rpc_stat->all_peers);
6084 num_funcs = rpc_stat->stats[0].func_total;
6086 sizeof(rx_interface_stat_t) +
6087 rpc_stat->stats[0].func_total *
6088 sizeof(rx_function_entry_v1_t);
6090 rxi_Free(rpc_stat, space);
6091 rxi_rpc_peer_stat_cnt -= num_funcs;
6094 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6095 if (peer == *peer_ptr) {
6102 MUTEX_EXIT(&peer->peer_lock);
6108 MUTEX_EXIT(&rx_peerHashTable_lock);
6109 MUTEX_EXIT(&rx_rpc_stats);
6112 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
6113 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
6114 * GC, just below. Really, we shouldn't have to keep moving packets from
6115 * one place to another, but instead ought to always know if we can
6116 * afford to hold onto a packet in its particular use. */
6117 MUTEX_ENTER(&rx_freePktQ_lock);
6118 if (rx_waitingForPackets) {
6119 rx_waitingForPackets = 0;
6120 #ifdef RX_ENABLE_LOCKS
6121 CV_BROADCAST(&rx_waitingForPackets_cv);
6123 osi_rxWakeup(&rx_waitingForPackets);
6126 MUTEX_EXIT(&rx_freePktQ_lock);
6129 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
6130 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
6134 /* rxs_Release - This isn't strictly necessary but, since the macro name from
6135 * rx.h is sort of strange this is better. This is called with a security
6136 * object before it is discarded. Each connection using a security object has
6137 * its own refcount to the object so it won't actually be freed until the last
6138 * connection is destroyed.
6140 * This is the only rxs module call. A hold could also be written but no one
6144 rxs_Release(struct rx_securityClass *aobj)
6146 return RXS_Close(aobj);
6150 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
6151 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
6152 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
6153 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
6155 /* Adjust our estimate of the transmission rate to this peer, given
6156 * that the packet p was just acked. We can adjust peer->timeout and
6157 * call->twind. Pragmatically, this is called
6158 * only with packets of maximal length.
6159 * Called with peer and call locked.
6163 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
6164 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
6166 afs_int32 xferSize, xferMs;
6167 register afs_int32 minTime;
6170 /* Count down packets */
6171 if (peer->rateFlag > 0)
6173 /* Do nothing until we're enabled */
6174 if (peer->rateFlag != 0)
6179 /* Count only when the ack seems legitimate */
6180 switch (ackReason) {
6181 case RX_ACK_REQUESTED:
6183 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6187 case RX_ACK_PING_RESPONSE:
6188 if (p) /* want the response to ping-request, not data send */
6190 clock_GetTime(&newTO);
6191 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6192 clock_Sub(&newTO, &call->pingRequestTime);
6193 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6197 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6204 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));
6206 /* Track only packets that are big enough. */
6207 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6211 /* absorb RTT data (in milliseconds) for these big packets */
6212 if (peer->smRtt == 0) {
6213 peer->smRtt = xferMs;
6215 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6220 if (peer->countDown) {
6224 peer->countDown = 10; /* recalculate only every so often */
6226 /* In practice, we can measure only the RTT for full packets,
6227 * because of the way Rx acks the data that it receives. (If it's
6228 * smaller than a full packet, it often gets implicitly acked
6229 * either by the call response (from a server) or by the next call
6230 * (from a client), and either case confuses transmission times
6231 * with processing times.) Therefore, replace the above
6232 * more-sophisticated processing with a simpler version, where the
6233 * smoothed RTT is kept for full-size packets, and the time to
6234 * transmit a windowful of full-size packets is simply RTT *
6235 * windowSize. Again, we take two steps:
6236 - ensure the timeout is large enough for a single packet's RTT;
6237 - ensure that the window is small enough to fit in the desired timeout.*/
6239 /* First, the timeout check. */
6240 minTime = peer->smRtt;
6241 /* Get a reasonable estimate for a timeout period */
6243 newTO.sec = minTime / 1000;
6244 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6246 /* Increase the timeout period so that we can always do at least
6247 * one packet exchange */
6248 if (clock_Gt(&newTO, &peer->timeout)) {
6250 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));
6252 peer->timeout = newTO;
6255 /* Now, get an estimate for the transmit window size. */
6256 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6257 /* Now, convert to the number of full packets that could fit in a
6258 * reasonable fraction of that interval */
6259 minTime /= (peer->smRtt << 1);
6260 xferSize = minTime; /* (make a copy) */
6262 /* Now clamp the size to reasonable bounds. */
6265 else if (minTime > rx_Window)
6266 minTime = rx_Window;
6267 /* if (minTime != peer->maxWindow) {
6268 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6269 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6270 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6272 peer->maxWindow = minTime;
6273 elide... call->twind = minTime;
6277 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6278 * Discern this by calculating the timeout necessary for rx_Window
6280 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6281 /* calculate estimate for transmission interval in milliseconds */
6282 minTime = rx_Window * peer->smRtt;
6283 if (minTime < 1000) {
6284 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
6285 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6286 peer->timeout.usec, peer->smRtt, peer->packetSize));
6288 newTO.sec = 0; /* cut back on timeout by half a second */
6289 newTO.usec = 500000;
6290 clock_Sub(&peer->timeout, &newTO);
6295 } /* end of rxi_ComputeRate */
6296 #endif /* ADAPT_WINDOW */
6304 #define TRACE_OPTION_DEBUGLOG 4
6312 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6313 0, KEY_QUERY_VALUE, &parmKey);
6314 if (code != ERROR_SUCCESS)
6317 dummyLen = sizeof(TraceOption);
6318 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6319 (BYTE *) &TraceOption, &dummyLen);
6320 if (code == ERROR_SUCCESS) {
6321 rxdebug_active = (TraceOption & TRACE_OPTION_DEBUGLOG) ? 1 : 0;
6323 RegCloseKey (parmKey);
6324 #endif /* AFS_NT40_ENV */
6329 rx_DebugOnOff(int on)
6331 rxdebug_active = on;
6333 #endif /* AFS_NT40_ENV */
6336 /* Don't call this debugging routine directly; use dpf */
6338 rxi_DebugPrint(char *format, ...)
6345 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6348 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
6350 if (msg[len-1] != '\n') {
6354 OutputDebugString(msg);
6362 va_start(ap, format);
6364 clock_GetTime(&now);
6365 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
6366 (unsigned int)now.usec / 1000);
6367 vfprintf(rx_Log, format, ap);
6374 * This function is used to process the rx_stats structure that is local
6375 * to a process as well as an rx_stats structure received from a remote
6376 * process (via rxdebug). Therefore, it needs to do minimal version
6380 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
6381 afs_int32 freePackets, char version)
6385 if (size != sizeof(struct rx_statistics)) {
6387 "Unexpected size of stats structure: was %d, expected %lud\n",
6388 size, sizeof(struct rx_statistics));
6391 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6394 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6395 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6396 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6397 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6398 s->specialPktAllocFailures);
6400 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6401 s->receivePktAllocFailures, s->sendPktAllocFailures,
6402 s->specialPktAllocFailures);
6406 " greedy %d, " "bogusReads %d (last from host %x), "
6407 "noPackets %d, " "noBuffers %d, " "selects %d, "
6408 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6409 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6410 s->selects, s->sendSelects);
6412 fprintf(file, " packets read: ");
6413 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6414 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6416 fprintf(file, "\n");
6419 " other read counters: data %d, " "ack %d, " "dup %d "
6420 "spurious %d " "dally %d\n", s->dataPacketsRead,
6421 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6422 s->ignorePacketDally);
6424 fprintf(file, " packets sent: ");
6425 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6426 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6428 fprintf(file, "\n");
6431 " other send counters: ack %d, " "data %d (not resends), "
6432 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6433 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6434 s->dataPacketsPushed, s->ignoreAckedPacket);
6437 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6438 s->netSendFailures, (int)s->fatalErrors);
6440 if (s->nRttSamples) {
6441 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6442 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6444 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6445 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6449 " %d server connections, " "%d client connections, "
6450 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6451 s->nServerConns, s->nClientConns, s->nPeerStructs,
6452 s->nCallStructs, s->nFreeCallStructs);
6454 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6455 fprintf(file, " %d clock updates\n", clock_nUpdates);
6460 /* for backward compatibility */
6462 rx_PrintStats(FILE * file)
6464 MUTEX_ENTER(&rx_stats_mutex);
6465 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6467 MUTEX_EXIT(&rx_stats_mutex);
6471 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6473 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6474 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6475 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6478 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6479 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6480 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6483 " Packet size %d, " "max in packet skew %d, "
6484 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6485 (int)peer->outPacketSkew);
6488 #ifdef AFS_PTHREAD_ENV
6490 * This mutex protects the following static variables:
6494 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6495 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6497 #define LOCK_RX_DEBUG
6498 #define UNLOCK_RX_DEBUG
6499 #endif /* AFS_PTHREAD_ENV */
6502 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6503 u_char type, void *inputData, size_t inputLength,
6504 void *outputData, size_t outputLength)
6506 static afs_int32 counter = 100;
6507 time_t waitTime, waitCount, startTime;
6508 struct rx_header theader;
6510 register afs_int32 code;
6511 struct timeval tv_now, tv_wake, tv_delta;
6512 struct sockaddr_in taddr, faddr;
6521 startTime = time(0);
6527 tp = &tbuffer[sizeof(struct rx_header)];
6528 taddr.sin_family = AF_INET;
6529 taddr.sin_port = remotePort;
6530 taddr.sin_addr.s_addr = remoteAddr;
6531 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6532 taddr.sin_len = sizeof(struct sockaddr_in);
6535 memset(&theader, 0, sizeof(theader));
6536 theader.epoch = htonl(999);
6538 theader.callNumber = htonl(counter);
6541 theader.type = type;
6542 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6543 theader.serviceId = 0;
6545 memcpy(tbuffer, &theader, sizeof(theader));
6546 memcpy(tp, inputData, inputLength);
6548 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6549 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6551 /* see if there's a packet available */
6552 gettimeofday(&tv_wake,0);
6553 tv_wake.tv_sec += waitTime;
6556 FD_SET(socket, &imask);
6557 tv_delta.tv_sec = tv_wake.tv_sec;
6558 tv_delta.tv_usec = tv_wake.tv_usec;
6559 gettimeofday(&tv_now, 0);
6561 if (tv_delta.tv_usec < tv_now.tv_usec) {
6563 tv_delta.tv_usec += 1000000;
6566 tv_delta.tv_usec -= tv_now.tv_usec;
6568 if (tv_delta.tv_sec < tv_now.tv_sec) {
6572 tv_delta.tv_sec -= tv_now.tv_sec;
6574 code = select(socket + 1, &imask, 0, 0, &tv_delta);
6575 if (code == 1 && FD_ISSET(socket, &imask)) {
6576 /* now receive a packet */
6577 faddrLen = sizeof(struct sockaddr_in);
6579 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6580 (struct sockaddr *)&faddr, &faddrLen);
6583 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6584 if (counter == ntohl(theader.callNumber))
6592 /* see if we've timed out */
6600 code -= sizeof(struct rx_header);
6601 if (code > outputLength)
6602 code = outputLength;
6603 memcpy(outputData, tp, code);
6608 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6609 afs_uint16 remotePort, struct rx_debugStats * stat,
6610 afs_uint32 * supportedValues)
6612 struct rx_debugIn in;
6613 afs_int32 *lp = (afs_int32 *) stat;
6617 *supportedValues = 0;
6618 in.type = htonl(RX_DEBUGI_GETSTATS);
6621 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6622 &in, sizeof(in), stat, sizeof(*stat));
6625 * If the call was successful, fixup the version and indicate
6626 * what contents of the stat structure are valid.
6627 * Also do net to host conversion of fields here.
6631 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6632 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6634 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6635 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6637 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6638 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6640 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6641 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6643 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6644 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6646 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6647 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6649 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6650 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6652 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6653 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6655 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
6656 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
6658 stat->nFreePackets = ntohl(stat->nFreePackets);
6659 stat->packetReclaims = ntohl(stat->packetReclaims);
6660 stat->callsExecuted = ntohl(stat->callsExecuted);
6661 stat->nWaiting = ntohl(stat->nWaiting);
6662 stat->idleThreads = ntohl(stat->idleThreads);
6663 stat->nWaited = ntohl(stat->nWaited);
6664 stat->nPackets = ntohl(stat->nPackets);
6671 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6672 afs_uint16 remotePort, struct rx_statistics * stat,
6673 afs_uint32 * supportedValues)
6675 struct rx_debugIn in;
6676 afs_int32 *lp = (afs_int32 *) stat;
6681 * supportedValues is currently unused, but added to allow future
6682 * versioning of this function.
6685 *supportedValues = 0;
6686 in.type = htonl(RX_DEBUGI_RXSTATS);
6688 memset(stat, 0, sizeof(*stat));
6690 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6691 &in, sizeof(in), stat, sizeof(*stat));
6696 * Do net to host conversion here
6699 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6708 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6709 afs_uint16 remotePort, size_t version_length,
6713 return MakeDebugCall(socket, remoteAddr, remotePort,
6714 RX_PACKET_TYPE_VERSION, a, 1, version,
6719 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6720 afs_uint16 remotePort, afs_int32 * nextConnection,
6721 int allConnections, afs_uint32 debugSupportedValues,
6722 struct rx_debugConn * conn,
6723 afs_uint32 * supportedValues)
6725 struct rx_debugIn in;
6730 * supportedValues is currently unused, but added to allow future
6731 * versioning of this function.
6734 *supportedValues = 0;
6735 if (allConnections) {
6736 in.type = htonl(RX_DEBUGI_GETALLCONN);
6738 in.type = htonl(RX_DEBUGI_GETCONN);
6740 in.index = htonl(*nextConnection);
6741 memset(conn, 0, sizeof(*conn));
6743 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6744 &in, sizeof(in), conn, sizeof(*conn));
6747 *nextConnection += 1;
6750 * Convert old connection format to new structure.
6753 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6754 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6755 #define MOVEvL(a) (conn->a = vL->a)
6757 /* any old or unrecognized version... */
6758 for (i = 0; i < RX_MAXCALLS; i++) {
6759 MOVEvL(callState[i]);
6760 MOVEvL(callMode[i]);
6761 MOVEvL(callFlags[i]);
6762 MOVEvL(callOther[i]);
6764 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6765 MOVEvL(secStats.type);
6766 MOVEvL(secStats.level);
6767 MOVEvL(secStats.flags);
6768 MOVEvL(secStats.expires);
6769 MOVEvL(secStats.packetsReceived);
6770 MOVEvL(secStats.packetsSent);
6771 MOVEvL(secStats.bytesReceived);
6772 MOVEvL(secStats.bytesSent);
6777 * Do net to host conversion here
6779 * I don't convert host or port since we are most likely
6780 * going to want these in NBO.
6782 conn->cid = ntohl(conn->cid);
6783 conn->serial = ntohl(conn->serial);
6784 for (i = 0; i < RX_MAXCALLS; i++) {
6785 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6787 conn->error = ntohl(conn->error);
6788 conn->secStats.flags = ntohl(conn->secStats.flags);
6789 conn->secStats.expires = ntohl(conn->secStats.expires);
6790 conn->secStats.packetsReceived =
6791 ntohl(conn->secStats.packetsReceived);
6792 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6793 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6794 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6795 conn->epoch = ntohl(conn->epoch);
6796 conn->natMTU = ntohl(conn->natMTU);
6803 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6804 afs_uint16 remotePort, afs_int32 * nextPeer,
6805 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6806 afs_uint32 * supportedValues)
6808 struct rx_debugIn in;
6812 * supportedValues is currently unused, but added to allow future
6813 * versioning of this function.
6816 *supportedValues = 0;
6817 in.type = htonl(RX_DEBUGI_GETPEER);
6818 in.index = htonl(*nextPeer);
6819 memset(peer, 0, sizeof(*peer));
6821 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6822 &in, sizeof(in), peer, sizeof(*peer));
6828 * Do net to host conversion here
6830 * I don't convert host or port since we are most likely
6831 * going to want these in NBO.
6833 peer->ifMTU = ntohs(peer->ifMTU);
6834 peer->idleWhen = ntohl(peer->idleWhen);
6835 peer->refCount = ntohs(peer->refCount);
6836 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6837 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6838 peer->rtt = ntohl(peer->rtt);
6839 peer->rtt_dev = ntohl(peer->rtt_dev);
6840 peer->timeout.sec = ntohl(peer->timeout.sec);
6841 peer->timeout.usec = ntohl(peer->timeout.usec);
6842 peer->nSent = ntohl(peer->nSent);
6843 peer->reSends = ntohl(peer->reSends);
6844 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6845 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6846 peer->rateFlag = ntohl(peer->rateFlag);
6847 peer->natMTU = ntohs(peer->natMTU);
6848 peer->maxMTU = ntohs(peer->maxMTU);
6849 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6850 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6851 peer->MTU = ntohs(peer->MTU);
6852 peer->cwind = ntohs(peer->cwind);
6853 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6854 peer->congestSeq = ntohs(peer->congestSeq);
6855 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6856 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6857 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6858 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6863 #endif /* RXDEBUG */
6868 struct rx_serverQueueEntry *np;
6871 register struct rx_call *call;
6872 register struct rx_serverQueueEntry *sq;
6876 if (rxinit_status == 1) {
6878 return; /* Already shutdown. */
6882 #ifndef AFS_PTHREAD_ENV
6883 FD_ZERO(&rx_selectMask);
6884 #endif /* AFS_PTHREAD_ENV */
6885 rxi_dataQuota = RX_MAX_QUOTA;
6886 #ifndef AFS_PTHREAD_ENV
6888 #endif /* AFS_PTHREAD_ENV */
6891 #ifndef AFS_PTHREAD_ENV
6892 #ifndef AFS_USE_GETTIMEOFDAY
6894 #endif /* AFS_USE_GETTIMEOFDAY */
6895 #endif /* AFS_PTHREAD_ENV */
6897 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6898 call = queue_First(&rx_freeCallQueue, rx_call);
6900 rxi_Free(call, sizeof(struct rx_call));
6903 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6904 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6910 struct rx_peer **peer_ptr, **peer_end;
6911 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6912 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6914 struct rx_peer *peer, *next;
6915 for (peer = *peer_ptr; peer; peer = next) {
6916 rx_interface_stat_p rpc_stat, nrpc_stat;
6919 (&peer->rpcStats, rpc_stat, nrpc_stat,
6920 rx_interface_stat)) {
6921 unsigned int num_funcs;
6924 queue_Remove(&rpc_stat->queue_header);
6925 queue_Remove(&rpc_stat->all_peers);
6926 num_funcs = rpc_stat->stats[0].func_total;
6928 sizeof(rx_interface_stat_t) +
6929 rpc_stat->stats[0].func_total *
6930 sizeof(rx_function_entry_v1_t);
6932 rxi_Free(rpc_stat, space);
6933 MUTEX_ENTER(&rx_rpc_stats);
6934 rxi_rpc_peer_stat_cnt -= num_funcs;
6935 MUTEX_EXIT(&rx_rpc_stats);
6939 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6943 for (i = 0; i < RX_MAX_SERVICES; i++) {
6945 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6947 for (i = 0; i < rx_hashTableSize; i++) {
6948 register struct rx_connection *tc, *ntc;
6949 MUTEX_ENTER(&rx_connHashTable_lock);
6950 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6952 for (j = 0; j < RX_MAXCALLS; j++) {
6954 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6957 rxi_Free(tc, sizeof(*tc));
6959 MUTEX_EXIT(&rx_connHashTable_lock);
6962 MUTEX_ENTER(&freeSQEList_lock);
6964 while ((np = rx_FreeSQEList)) {
6965 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6966 MUTEX_DESTROY(&np->lock);
6967 rxi_Free(np, sizeof(*np));
6970 MUTEX_EXIT(&freeSQEList_lock);
6971 MUTEX_DESTROY(&freeSQEList_lock);
6972 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6973 MUTEX_DESTROY(&rx_connHashTable_lock);
6974 MUTEX_DESTROY(&rx_peerHashTable_lock);
6975 MUTEX_DESTROY(&rx_serverPool_lock);
6977 osi_Free(rx_connHashTable,
6978 rx_hashTableSize * sizeof(struct rx_connection *));
6979 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6981 UNPIN(rx_connHashTable,
6982 rx_hashTableSize * sizeof(struct rx_connection *));
6983 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6985 rxi_FreeAllPackets();
6987 MUTEX_ENTER(&rx_stats_mutex);
6988 rxi_dataQuota = RX_MAX_QUOTA;
6989 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6990 MUTEX_EXIT(&rx_stats_mutex);
6996 #ifdef RX_ENABLE_LOCKS
6998 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7000 if (!MUTEX_ISMINE(lockaddr))
7001 osi_Panic("Lock not held: %s", msg);
7003 #endif /* RX_ENABLE_LOCKS */
7008 * Routines to implement connection specific data.
7012 rx_KeyCreate(rx_destructor_t rtn)
7015 MUTEX_ENTER(&rxi_keyCreate_lock);
7016 key = rxi_keyCreate_counter++;
7017 rxi_keyCreate_destructor = (rx_destructor_t *)
7018 realloc((void *)rxi_keyCreate_destructor,
7019 (key + 1) * sizeof(rx_destructor_t));
7020 rxi_keyCreate_destructor[key] = rtn;
7021 MUTEX_EXIT(&rxi_keyCreate_lock);
7026 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7029 MUTEX_ENTER(&conn->conn_data_lock);
7030 if (!conn->specific) {
7031 conn->specific = (void **)malloc((key + 1) * sizeof(void *));
7032 for (i = 0; i < key; i++)
7033 conn->specific[i] = NULL;
7034 conn->nSpecific = key + 1;
7035 conn->specific[key] = ptr;
7036 } else if (key >= conn->nSpecific) {
7037 conn->specific = (void **)
7038 realloc(conn->specific, (key + 1) * sizeof(void *));
7039 for (i = conn->nSpecific; i < key; i++)
7040 conn->specific[i] = NULL;
7041 conn->nSpecific = key + 1;
7042 conn->specific[key] = ptr;
7044 if (conn->specific[key] && rxi_keyCreate_destructor[key])
7045 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7046 conn->specific[key] = ptr;
7048 MUTEX_EXIT(&conn->conn_data_lock);
7052 rx_GetSpecific(struct rx_connection *conn, int key)
7055 MUTEX_ENTER(&conn->conn_data_lock);
7056 if (key >= conn->nSpecific)
7059 ptr = conn->specific[key];
7060 MUTEX_EXIT(&conn->conn_data_lock);
7064 #endif /* !KERNEL */
7067 * processStats is a queue used to store the statistics for the local
7068 * process. Its contents are similar to the contents of the rpcStats
7069 * queue on a rx_peer structure, but the actual data stored within
7070 * this queue contains totals across the lifetime of the process (assuming
7071 * the stats have not been reset) - unlike the per peer structures
7072 * which can come and go based upon the peer lifetime.
7075 static struct rx_queue processStats = { &processStats, &processStats };
7078 * peerStats is a queue used to store the statistics for all peer structs.
7079 * Its contents are the union of all the peer rpcStats queues.
7082 static struct rx_queue peerStats = { &peerStats, &peerStats };
7085 * rxi_monitor_processStats is used to turn process wide stat collection
7089 static int rxi_monitor_processStats = 0;
7092 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
7095 static int rxi_monitor_peerStats = 0;
7098 * rxi_AddRpcStat - given all of the information for a particular rpc
7099 * call, create (if needed) and update the stat totals for the rpc.
7103 * IN stats - the queue of stats that will be updated with the new value
7105 * IN rxInterface - a unique number that identifies the rpc interface
7107 * IN currentFunc - the index of the function being invoked
7109 * IN totalFunc - the total number of functions in this interface
7111 * IN queueTime - the amount of time this function waited for a thread
7113 * IN execTime - the amount of time this function invocation took to execute
7115 * IN bytesSent - the number bytes sent by this invocation
7117 * IN bytesRcvd - the number bytes received by this invocation
7119 * IN isServer - if true, this invocation was made to a server
7121 * IN remoteHost - the ip address of the remote host
7123 * IN remotePort - the port of the remote host
7125 * IN addToPeerList - if != 0, add newly created stat to the global peer list
7127 * INOUT counter - if a new stats structure is allocated, the counter will
7128 * be updated with the new number of allocated stat structures
7136 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
7137 afs_uint32 currentFunc, afs_uint32 totalFunc,
7138 struct clock *queueTime, struct clock *execTime,
7139 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
7140 afs_uint32 remoteHost, afs_uint32 remotePort,
7141 int addToPeerList, unsigned int *counter)
7144 rx_interface_stat_p rpc_stat, nrpc_stat;
7147 * See if there's already a structure for this interface
7150 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7151 if ((rpc_stat->stats[0].interfaceId == rxInterface)
7152 && (rpc_stat->stats[0].remote_is_server == isServer))
7157 * Didn't find a match so allocate a new structure and add it to the
7161 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
7162 || (rpc_stat->stats[0].interfaceId != rxInterface)
7163 || (rpc_stat->stats[0].remote_is_server != isServer)) {
7168 sizeof(rx_interface_stat_t) +
7169 totalFunc * sizeof(rx_function_entry_v1_t);
7171 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
7172 if (rpc_stat == NULL) {
7176 *counter += totalFunc;
7177 for (i = 0; i < totalFunc; i++) {
7178 rpc_stat->stats[i].remote_peer = remoteHost;
7179 rpc_stat->stats[i].remote_port = remotePort;
7180 rpc_stat->stats[i].remote_is_server = isServer;
7181 rpc_stat->stats[i].interfaceId = rxInterface;
7182 rpc_stat->stats[i].func_total = totalFunc;
7183 rpc_stat->stats[i].func_index = i;
7184 hzero(rpc_stat->stats[i].invocations);
7185 hzero(rpc_stat->stats[i].bytes_sent);
7186 hzero(rpc_stat->stats[i].bytes_rcvd);
7187 rpc_stat->stats[i].queue_time_sum.sec = 0;
7188 rpc_stat->stats[i].queue_time_sum.usec = 0;
7189 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7190 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7191 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7192 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7193 rpc_stat->stats[i].queue_time_max.sec = 0;
7194 rpc_stat->stats[i].queue_time_max.usec = 0;
7195 rpc_stat->stats[i].execution_time_sum.sec = 0;
7196 rpc_stat->stats[i].execution_time_sum.usec = 0;
7197 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7198 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7199 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7200 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7201 rpc_stat->stats[i].execution_time_max.sec = 0;
7202 rpc_stat->stats[i].execution_time_max.usec = 0;
7204 queue_Prepend(stats, rpc_stat);
7205 if (addToPeerList) {
7206 queue_Prepend(&peerStats, &rpc_stat->all_peers);
7211 * Increment the stats for this function
7214 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
7215 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
7216 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
7217 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
7218 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7219 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7220 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7222 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7223 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7225 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7226 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7228 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7229 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7231 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7232 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7240 * rx_IncrementTimeAndCount - increment the times and count for a particular
7245 * IN peer - the peer who invoked the rpc
7247 * IN rxInterface - a unique number that identifies the rpc interface
7249 * IN currentFunc - the index of the function being invoked
7251 * IN totalFunc - the total number of functions in this interface
7253 * IN queueTime - the amount of time this function waited for a thread
7255 * IN execTime - the amount of time this function invocation took to execute
7257 * IN bytesSent - the number bytes sent by this invocation
7259 * IN bytesRcvd - the number bytes received by this invocation
7261 * IN isServer - if true, this invocation was made to a server
7269 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7270 afs_uint32 currentFunc, afs_uint32 totalFunc,
7271 struct clock *queueTime, struct clock *execTime,
7272 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7276 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
7279 MUTEX_ENTER(&rx_rpc_stats);
7280 MUTEX_ENTER(&peer->peer_lock);
7282 if (rxi_monitor_peerStats) {
7283 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7284 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7285 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7288 if (rxi_monitor_processStats) {
7289 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7290 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7291 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7294 MUTEX_EXIT(&peer->peer_lock);
7295 MUTEX_EXIT(&rx_rpc_stats);
7300 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7304 * IN callerVersion - the rpc stat version of the caller.
7306 * IN count - the number of entries to marshall.
7308 * IN stats - pointer to stats to be marshalled.
7310 * OUT ptr - Where to store the marshalled data.
7317 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7318 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7324 * We only support the first version
7326 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7327 *(ptr++) = stats->remote_peer;
7328 *(ptr++) = stats->remote_port;
7329 *(ptr++) = stats->remote_is_server;
7330 *(ptr++) = stats->interfaceId;
7331 *(ptr++) = stats->func_total;
7332 *(ptr++) = stats->func_index;
7333 *(ptr++) = hgethi(stats->invocations);
7334 *(ptr++) = hgetlo(stats->invocations);
7335 *(ptr++) = hgethi(stats->bytes_sent);
7336 *(ptr++) = hgetlo(stats->bytes_sent);
7337 *(ptr++) = hgethi(stats->bytes_rcvd);
7338 *(ptr++) = hgetlo(stats->bytes_rcvd);
7339 *(ptr++) = stats->queue_time_sum.sec;
7340 *(ptr++) = stats->queue_time_sum.usec;
7341 *(ptr++) = stats->queue_time_sum_sqr.sec;
7342 *(ptr++) = stats->queue_time_sum_sqr.usec;
7343 *(ptr++) = stats->queue_time_min.sec;
7344 *(ptr++) = stats->queue_time_min.usec;
7345 *(ptr++) = stats->queue_time_max.sec;
7346 *(ptr++) = stats->queue_time_max.usec;
7347 *(ptr++) = stats->execution_time_sum.sec;
7348 *(ptr++) = stats->execution_time_sum.usec;
7349 *(ptr++) = stats->execution_time_sum_sqr.sec;
7350 *(ptr++) = stats->execution_time_sum_sqr.usec;
7351 *(ptr++) = stats->execution_time_min.sec;
7352 *(ptr++) = stats->execution_time_min.usec;
7353 *(ptr++) = stats->execution_time_max.sec;
7354 *(ptr++) = stats->execution_time_max.usec;
7360 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7365 * IN callerVersion - the rpc stat version of the caller
7367 * OUT myVersion - the rpc stat version of this function
7369 * OUT clock_sec - local time seconds
7371 * OUT clock_usec - local time microseconds
7373 * OUT allocSize - the number of bytes allocated to contain stats
7375 * OUT statCount - the number stats retrieved from this process.
7377 * OUT stats - the actual stats retrieved from this process.
7381 * Returns void. If successful, stats will != NULL.
7385 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7386 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7387 size_t * allocSize, afs_uint32 * statCount,
7388 afs_uint32 ** stats)
7398 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7401 * Check to see if stats are enabled
7404 MUTEX_ENTER(&rx_rpc_stats);
7405 if (!rxi_monitor_processStats) {
7406 MUTEX_EXIT(&rx_rpc_stats);
7410 clock_GetTime(&now);
7411 *clock_sec = now.sec;
7412 *clock_usec = now.usec;
7415 * Allocate the space based upon the caller version
7417 * If the client is at an older version than we are,
7418 * we return the statistic data in the older data format, but
7419 * we still return our version number so the client knows we
7420 * are maintaining more data than it can retrieve.
7423 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7424 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7425 *statCount = rxi_rpc_process_stat_cnt;
7428 * This can't happen yet, but in the future version changes
7429 * can be handled by adding additional code here
7433 if (space > (size_t) 0) {
7435 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7438 rx_interface_stat_p rpc_stat, nrpc_stat;
7442 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7444 * Copy the data based upon the caller version
7446 rx_MarshallProcessRPCStats(callerVersion,
7447 rpc_stat->stats[0].func_total,
7448 rpc_stat->stats, &ptr);
7454 MUTEX_EXIT(&rx_rpc_stats);
7459 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7463 * IN callerVersion - the rpc stat version of the caller
7465 * OUT myVersion - the rpc stat version of this function
7467 * OUT clock_sec - local time seconds
7469 * OUT clock_usec - local time microseconds
7471 * OUT allocSize - the number of bytes allocated to contain stats
7473 * OUT statCount - the number of stats retrieved from the individual
7476 * OUT stats - the actual stats retrieved from the individual peer structures.
7480 * Returns void. If successful, stats will != NULL.
7484 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7485 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7486 size_t * allocSize, afs_uint32 * statCount,
7487 afs_uint32 ** stats)
7497 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7500 * Check to see if stats are enabled
7503 MUTEX_ENTER(&rx_rpc_stats);
7504 if (!rxi_monitor_peerStats) {
7505 MUTEX_EXIT(&rx_rpc_stats);
7509 clock_GetTime(&now);
7510 *clock_sec = now.sec;
7511 *clock_usec = now.usec;
7514 * Allocate the space based upon the caller version
7516 * If the client is at an older version than we are,
7517 * we return the statistic data in the older data format, but
7518 * we still return our version number so the client knows we
7519 * are maintaining more data than it can retrieve.
7522 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7523 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7524 *statCount = rxi_rpc_peer_stat_cnt;
7527 * This can't happen yet, but in the future version changes
7528 * can be handled by adding additional code here
7532 if (space > (size_t) 0) {
7534 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7537 rx_interface_stat_p rpc_stat, nrpc_stat;
7541 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7543 * We have to fix the offset of rpc_stat since we are
7544 * keeping this structure on two rx_queues. The rx_queue
7545 * package assumes that the rx_queue member is the first
7546 * member of the structure. That is, rx_queue assumes that
7547 * any one item is only on one queue at a time. We are
7548 * breaking that assumption and so we have to do a little
7549 * math to fix our pointers.
7552 fix_offset = (char *)rpc_stat;
7553 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7554 rpc_stat = (rx_interface_stat_p) fix_offset;
7557 * Copy the data based upon the caller version
7559 rx_MarshallProcessRPCStats(callerVersion,
7560 rpc_stat->stats[0].func_total,
7561 rpc_stat->stats, &ptr);
7567 MUTEX_EXIT(&rx_rpc_stats);
7572 * rx_FreeRPCStats - free memory allocated by
7573 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7577 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7578 * rx_RetrievePeerRPCStats
7580 * IN allocSize - the number of bytes in stats.
7588 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7590 rxi_Free(stats, allocSize);
7594 * rx_queryProcessRPCStats - see if process rpc stat collection is
7595 * currently enabled.
7601 * Returns 0 if stats are not enabled != 0 otherwise
7605 rx_queryProcessRPCStats(void)
7608 MUTEX_ENTER(&rx_rpc_stats);
7609 rc = rxi_monitor_processStats;
7610 MUTEX_EXIT(&rx_rpc_stats);
7615 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7621 * Returns 0 if stats are not enabled != 0 otherwise
7625 rx_queryPeerRPCStats(void)
7628 MUTEX_ENTER(&rx_rpc_stats);
7629 rc = rxi_monitor_peerStats;
7630 MUTEX_EXIT(&rx_rpc_stats);
7635 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7645 rx_enableProcessRPCStats(void)
7647 MUTEX_ENTER(&rx_rpc_stats);
7648 rx_enable_stats = 1;
7649 rxi_monitor_processStats = 1;
7650 MUTEX_EXIT(&rx_rpc_stats);
7654 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7664 rx_enablePeerRPCStats(void)
7666 MUTEX_ENTER(&rx_rpc_stats);
7667 rx_enable_stats = 1;
7668 rxi_monitor_peerStats = 1;
7669 MUTEX_EXIT(&rx_rpc_stats);
7673 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7683 rx_disableProcessRPCStats(void)
7685 rx_interface_stat_p rpc_stat, nrpc_stat;
7688 MUTEX_ENTER(&rx_rpc_stats);
7691 * Turn off process statistics and if peer stats is also off, turn
7695 rxi_monitor_processStats = 0;
7696 if (rxi_monitor_peerStats == 0) {
7697 rx_enable_stats = 0;
7700 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7701 unsigned int num_funcs = 0;
7704 queue_Remove(rpc_stat);
7705 num_funcs = rpc_stat->stats[0].func_total;
7707 sizeof(rx_interface_stat_t) +
7708 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7710 rxi_Free(rpc_stat, space);
7711 rxi_rpc_process_stat_cnt -= num_funcs;
7713 MUTEX_EXIT(&rx_rpc_stats);
7717 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7727 rx_disablePeerRPCStats(void)
7729 struct rx_peer **peer_ptr, **peer_end;
7732 MUTEX_ENTER(&rx_rpc_stats);
7735 * Turn off peer statistics and if process stats is also off, turn
7739 rxi_monitor_peerStats = 0;
7740 if (rxi_monitor_processStats == 0) {
7741 rx_enable_stats = 0;
7744 MUTEX_ENTER(&rx_peerHashTable_lock);
7745 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7746 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7748 struct rx_peer *peer, *next, *prev;
7749 for (prev = peer = *peer_ptr; peer; peer = next) {
7751 code = MUTEX_TRYENTER(&peer->peer_lock);
7753 rx_interface_stat_p rpc_stat, nrpc_stat;
7756 (&peer->rpcStats, rpc_stat, nrpc_stat,
7757 rx_interface_stat)) {
7758 unsigned int num_funcs = 0;
7761 queue_Remove(&rpc_stat->queue_header);
7762 queue_Remove(&rpc_stat->all_peers);
7763 num_funcs = rpc_stat->stats[0].func_total;
7765 sizeof(rx_interface_stat_t) +
7766 rpc_stat->stats[0].func_total *
7767 sizeof(rx_function_entry_v1_t);
7769 rxi_Free(rpc_stat, space);
7770 rxi_rpc_peer_stat_cnt -= num_funcs;
7772 MUTEX_EXIT(&peer->peer_lock);
7773 if (prev == *peer_ptr) {
7783 MUTEX_EXIT(&rx_peerHashTable_lock);
7784 MUTEX_EXIT(&rx_rpc_stats);
7788 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7793 * IN clearFlag - flag indicating which stats to clear
7801 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7803 rx_interface_stat_p rpc_stat, nrpc_stat;
7805 MUTEX_ENTER(&rx_rpc_stats);
7807 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7808 unsigned int num_funcs = 0, i;
7809 num_funcs = rpc_stat->stats[0].func_total;
7810 for (i = 0; i < num_funcs; i++) {
7811 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7812 hzero(rpc_stat->stats[i].invocations);
7814 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7815 hzero(rpc_stat->stats[i].bytes_sent);
7817 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7818 hzero(rpc_stat->stats[i].bytes_rcvd);
7820 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7821 rpc_stat->stats[i].queue_time_sum.sec = 0;
7822 rpc_stat->stats[i].queue_time_sum.usec = 0;
7824 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7825 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7826 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7828 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7829 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7830 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7832 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7833 rpc_stat->stats[i].queue_time_max.sec = 0;
7834 rpc_stat->stats[i].queue_time_max.usec = 0;
7836 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7837 rpc_stat->stats[i].execution_time_sum.sec = 0;
7838 rpc_stat->stats[i].execution_time_sum.usec = 0;
7840 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7841 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7842 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7844 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7845 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7846 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7848 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7849 rpc_stat->stats[i].execution_time_max.sec = 0;
7850 rpc_stat->stats[i].execution_time_max.usec = 0;
7855 MUTEX_EXIT(&rx_rpc_stats);
7859 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7864 * IN clearFlag - flag indicating which stats to clear
7872 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7874 rx_interface_stat_p rpc_stat, nrpc_stat;
7876 MUTEX_ENTER(&rx_rpc_stats);
7878 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7879 unsigned int num_funcs = 0, i;
7882 * We have to fix the offset of rpc_stat since we are
7883 * keeping this structure on two rx_queues. The rx_queue
7884 * package assumes that the rx_queue member is the first
7885 * member of the structure. That is, rx_queue assumes that
7886 * any one item is only on one queue at a time. We are
7887 * breaking that assumption and so we have to do a little
7888 * math to fix our pointers.
7891 fix_offset = (char *)rpc_stat;
7892 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7893 rpc_stat = (rx_interface_stat_p) fix_offset;
7895 num_funcs = rpc_stat->stats[0].func_total;
7896 for (i = 0; i < num_funcs; i++) {
7897 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7898 hzero(rpc_stat->stats[i].invocations);
7900 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7901 hzero(rpc_stat->stats[i].bytes_sent);
7903 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7904 hzero(rpc_stat->stats[i].bytes_rcvd);
7906 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7907 rpc_stat->stats[i].queue_time_sum.sec = 0;
7908 rpc_stat->stats[i].queue_time_sum.usec = 0;
7910 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7911 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7912 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7914 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7915 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7916 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7918 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7919 rpc_stat->stats[i].queue_time_max.sec = 0;
7920 rpc_stat->stats[i].queue_time_max.usec = 0;
7922 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7923 rpc_stat->stats[i].execution_time_sum.sec = 0;
7924 rpc_stat->stats[i].execution_time_sum.usec = 0;
7926 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7927 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7928 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7930 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7931 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7932 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7934 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7935 rpc_stat->stats[i].execution_time_max.sec = 0;
7936 rpc_stat->stats[i].execution_time_max.usec = 0;
7941 MUTEX_EXIT(&rx_rpc_stats);
7945 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7946 * is authorized to enable/disable/clear RX statistics.
7948 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7951 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7953 rxi_rxstat_userok = proc;
7957 rx_RxStatUserOk(struct rx_call *call)
7959 if (!rxi_rxstat_userok)
7961 return rxi_rxstat_userok(call);
7966 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
7967 * function in the MSVC runtime DLL (msvcrt.dll).
7969 * Note: the system serializes calls to this function.
7972 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
7973 DWORD reason, /* reason function is being called */
7974 LPVOID reserved) /* reserved for future use */
7977 case DLL_PROCESS_ATTACH:
7978 /* library is being attached to a process */
7982 case DLL_PROCESS_DETACH: