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"
72 #include "rx_internal.h"
74 #include "rx_globals.h"
76 #define AFSOP_STOP_RXCALLBACK 210 /* Stop CALLBACK process */
77 #define AFSOP_STOP_AFS 211 /* Stop AFS process */
78 #define AFSOP_STOP_BKG 212 /* Stop BKG process */
80 extern afs_int32 afs_termState;
82 #include "sys/lockl.h"
83 #include "sys/lock_def.h"
84 #endif /* AFS_AIX41_ENV */
85 # include "rxgen_consts.h"
87 # include <sys/types.h>
94 # include <afs/afsutil.h>
95 # include <WINNT\afsreg.h>
97 # include <sys/socket.h>
98 # include <sys/file.h>
100 # include <sys/stat.h>
101 # include <netinet/in.h>
102 # include <sys/time.h>
104 # include "rx_internal.h"
106 # include "rx_user.h"
107 # include "rx_clock.h"
108 # include "rx_queue.h"
109 # include "rx_globals.h"
110 # include "rx_trace.h"
111 # include <afs/rxgen_consts.h>
115 #ifdef AFS_PTHREAD_ENV
117 int (*registerProgram) (pid_t, char *) = 0;
118 int (*swapNameProgram) (pid_t, const char *, char *) = 0;
121 int (*registerProgram) (PROCESS, char *) = 0;
122 int (*swapNameProgram) (PROCESS, const char *, char *) = 0;
126 /* Local static routines */
127 static void rxi_DestroyConnectionNoLock(register struct rx_connection *conn);
128 #ifdef RX_ENABLE_LOCKS
129 static void rxi_SetAcksInTransmitQueue(register struct rx_call *call);
132 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
134 afs_int32 rxi_start_aborted; /* rxi_start awoke after rxi_Send in error. */
135 afs_int32 rxi_start_in_error;
137 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
140 * rxi_rpc_peer_stat_cnt counts the total number of peer stat structures
141 * currently allocated within rx. This number is used to allocate the
142 * memory required to return the statistics when queried.
145 static unsigned int rxi_rpc_peer_stat_cnt;
148 * rxi_rpc_process_stat_cnt counts the total number of local process stat
149 * structures currently allocated within rx. The number is used to allocate
150 * the memory required to return the statistics when queried.
153 static unsigned int rxi_rpc_process_stat_cnt;
155 #if !defined(offsetof)
156 #include <stddef.h> /* for definition of offsetof() */
159 #ifdef AFS_PTHREAD_ENV
163 * Use procedural initialization of mutexes/condition variables
167 extern pthread_mutex_t rx_stats_mutex;
168 extern pthread_mutex_t des_init_mutex;
169 extern pthread_mutex_t des_random_mutex;
170 extern pthread_mutex_t rx_clock_mutex;
171 extern pthread_mutex_t rxi_connCacheMutex;
172 extern pthread_mutex_t rx_event_mutex;
173 extern pthread_mutex_t osi_malloc_mutex;
174 extern pthread_mutex_t event_handler_mutex;
175 extern pthread_mutex_t listener_mutex;
176 extern pthread_mutex_t rx_if_init_mutex;
177 extern pthread_mutex_t rx_if_mutex;
178 extern pthread_mutex_t rxkad_client_uid_mutex;
179 extern pthread_mutex_t rxkad_random_mutex;
181 extern pthread_cond_t rx_event_handler_cond;
182 extern pthread_cond_t rx_listener_cond;
184 static pthread_mutex_t epoch_mutex;
185 static pthread_mutex_t rx_init_mutex;
186 static pthread_mutex_t rx_debug_mutex;
187 static pthread_mutex_t rx_rpc_stats;
190 rxi_InitPthread(void)
192 assert(pthread_mutex_init(&rx_clock_mutex, (const pthread_mutexattr_t *)0)
194 assert(pthread_mutex_init(&rx_stats_mutex, (const pthread_mutexattr_t *)0)
196 assert(pthread_mutex_init
197 (&rxi_connCacheMutex, (const pthread_mutexattr_t *)0) == 0);
198 assert(pthread_mutex_init(&rx_init_mutex, (const pthread_mutexattr_t *)0)
200 assert(pthread_mutex_init(&epoch_mutex, (const pthread_mutexattr_t *)0) ==
202 assert(pthread_mutex_init(&rx_event_mutex, (const pthread_mutexattr_t *)0)
204 assert(pthread_mutex_init(&des_init_mutex, (const pthread_mutexattr_t *)0)
206 assert(pthread_mutex_init
207 (&des_random_mutex, (const pthread_mutexattr_t *)0) == 0);
208 assert(pthread_mutex_init
209 (&osi_malloc_mutex, (const pthread_mutexattr_t *)0) == 0);
210 assert(pthread_mutex_init
211 (&event_handler_mutex, (const pthread_mutexattr_t *)0) == 0);
212 assert(pthread_mutex_init(&listener_mutex, (const pthread_mutexattr_t *)0)
214 assert(pthread_mutex_init
215 (&rx_if_init_mutex, (const pthread_mutexattr_t *)0) == 0);
216 assert(pthread_mutex_init(&rx_if_mutex, (const pthread_mutexattr_t *)0) ==
218 assert(pthread_mutex_init
219 (&rxkad_client_uid_mutex, (const pthread_mutexattr_t *)0) == 0);
220 assert(pthread_mutex_init
221 (&rxkad_random_mutex, (const pthread_mutexattr_t *)0) == 0);
222 assert(pthread_mutex_init(&rx_debug_mutex, (const pthread_mutexattr_t *)0)
225 assert(pthread_cond_init
226 (&rx_event_handler_cond, (const pthread_condattr_t *)0) == 0);
227 assert(pthread_cond_init(&rx_listener_cond, (const pthread_condattr_t *)0)
229 assert(pthread_key_create(&rx_thread_id_key, NULL) == 0);
230 assert(pthread_key_create(&rx_ts_info_key, NULL) == 0);
232 rxkad_global_stats_init();
234 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
235 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
236 #ifdef RX_ENABLE_LOCKS
239 #endif /* RX_LOCKS_DB */
240 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
241 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
243 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
245 RWLOCK_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
247 RWLOCK_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
249 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
250 MUTEX_INIT(&rxi_keyCreate_lock, "rxi_keyCreate_lock", MUTEX_DEFAULT, 0);
251 #endif /* RX_ENABLE_LOCKS */
254 pthread_once_t rx_once_init = PTHREAD_ONCE_INIT;
255 #define INIT_PTHREAD_LOCKS \
256 assert(pthread_once(&rx_once_init, rxi_InitPthread)==0)
258 * The rx_stats_mutex mutex protects the following global variables:
263 * rxi_lowConnRefCount
264 * rxi_lowPeerRefCount
273 #define INIT_PTHREAD_LOCKS
277 /* Variables for handling the minProcs implementation. availProcs gives the
278 * number of threads available in the pool at this moment (not counting dudes
279 * executing right now). totalMin gives the total number of procs required
280 * for handling all minProcs requests. minDeficit is a dynamic variable
281 * tracking the # of procs required to satisfy all of the remaining minProcs
283 * For fine grain locking to work, the quota check and the reservation of
284 * a server thread has to come while rxi_availProcs and rxi_minDeficit
285 * are locked. To this end, the code has been modified under #ifdef
286 * RX_ENABLE_LOCKS so that quota checks and reservation occur at the
287 * same time. A new function, ReturnToServerPool() returns the allocation.
289 * A call can be on several queue's (but only one at a time). When
290 * rxi_ResetCall wants to remove the call from a queue, it has to ensure
291 * that no one else is touching the queue. To this end, we store the address
292 * of the queue lock in the call structure (under the call lock) when we
293 * put the call on a queue, and we clear the call_queue_lock when the
294 * call is removed from a queue (once the call lock has been obtained).
295 * This allows rxi_ResetCall to safely synchronize with others wishing
296 * to manipulate the queue.
299 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
300 static afs_kmutex_t rx_rpc_stats;
301 void rxi_StartUnlocked(struct rxevent *event, void *call,
302 void *arg1, int istack);
305 /* We keep a "last conn pointer" in rxi_FindConnection. The odds are
306 ** pretty good that the next packet coming in is from the same connection
307 ** as the last packet, since we're send multiple packets in a transmit window.
309 struct rx_connection *rxLastConn = 0;
311 #ifdef RX_ENABLE_LOCKS
312 /* The locking hierarchy for rx fine grain locking is composed of these
315 * rx_connHashTable_lock - synchronizes conn creation, rx_connHashTable access
316 * conn_call_lock - used to synchonize rx_EndCall and rx_NewCall
317 * call->lock - locks call data fields.
318 * These are independent of each other:
319 * rx_freeCallQueue_lock
324 * serverQueueEntry->lock
326 * rx_peerHashTable_lock - locked under rx_connHashTable_lock
327 * peer->lock - locks peer data fields.
328 * conn_data_lock - that more than one thread is not updating a conn data
329 * field at the same time.
337 * Do we need a lock to protect the peer field in the conn structure?
338 * conn->peer was previously a constant for all intents and so has no
339 * lock protecting this field. The multihomed client delta introduced
340 * a RX code change : change the peer field in the connection structure
341 * to that remote inetrface from which the last packet for this
342 * connection was sent out. This may become an issue if further changes
345 #define SET_CALL_QUEUE_LOCK(C, L) (C)->call_queue_lock = (L)
346 #define CLEAR_CALL_QUEUE_LOCK(C) (C)->call_queue_lock = NULL
348 /* rxdb_fileID is used to identify the lock location, along with line#. */
349 static int rxdb_fileID = RXDB_FILE_RX;
350 #endif /* RX_LOCKS_DB */
351 #else /* RX_ENABLE_LOCKS */
352 #define SET_CALL_QUEUE_LOCK(C, L)
353 #define CLEAR_CALL_QUEUE_LOCK(C)
354 #endif /* RX_ENABLE_LOCKS */
355 struct rx_serverQueueEntry *rx_waitForPacket = 0;
356 struct rx_serverQueueEntry *rx_waitingForPacket = 0;
358 /* ------------Exported Interfaces------------- */
360 /* This function allows rxkad to set the epoch to a suitably random number
361 * which rx_NewConnection will use in the future. The principle purpose is to
362 * get rxnull connections to use the same epoch as the rxkad connections do, at
363 * least once the first rxkad connection is established. This is important now
364 * that the host/port addresses aren't used in FindConnection: the uniqueness
365 * of epoch/cid matters and the start time won't do. */
367 #ifdef AFS_PTHREAD_ENV
369 * This mutex protects the following global variables:
373 #define LOCK_EPOCH assert(pthread_mutex_lock(&epoch_mutex)==0)
374 #define UNLOCK_EPOCH assert(pthread_mutex_unlock(&epoch_mutex)==0)
378 #endif /* AFS_PTHREAD_ENV */
381 rx_SetEpoch(afs_uint32 epoch)
388 /* Initialize rx. A port number may be mentioned, in which case this
389 * becomes the default port number for any service installed later.
390 * If 0 is provided for the port number, a random port will be chosen
391 * by the kernel. Whether this will ever overlap anything in
392 * /etc/services is anybody's guess... Returns 0 on success, -1 on
397 int rxinit_status = 1;
398 #ifdef AFS_PTHREAD_ENV
400 * This mutex protects the following global variables:
404 #define LOCK_RX_INIT assert(pthread_mutex_lock(&rx_init_mutex)==0)
405 #define UNLOCK_RX_INIT assert(pthread_mutex_unlock(&rx_init_mutex)==0)
408 #define UNLOCK_RX_INIT
412 rx_InitHost(u_int host, u_int port)
419 char *htable, *ptable;
426 if (rxinit_status == 0) {
427 tmp_status = rxinit_status;
429 return tmp_status; /* Already started; return previous error code. */
435 if (afs_winsockInit() < 0)
441 * Initialize anything necessary to provide a non-premptive threading
444 rxi_InitializeThreadSupport();
447 /* Allocate and initialize a socket for client and perhaps server
450 rx_socket = rxi_GetHostUDPSocket(host, (u_short) port);
451 if (rx_socket == OSI_NULLSOCKET) {
455 #if defined(RX_ENABLE_LOCKS) && defined(KERNEL)
458 #endif /* RX_LOCKS_DB */
459 MUTEX_INIT(&rx_stats_mutex, "rx_stats_mutex", MUTEX_DEFAULT, 0);
460 MUTEX_INIT(&rx_rpc_stats, "rx_rpc_stats", MUTEX_DEFAULT, 0);
461 MUTEX_INIT(&rx_freePktQ_lock, "rx_freePktQ_lock", MUTEX_DEFAULT, 0);
462 MUTEX_INIT(&freeSQEList_lock, "freeSQEList lock", MUTEX_DEFAULT, 0);
463 MUTEX_INIT(&rx_freeCallQueue_lock, "rx_freeCallQueue_lock", MUTEX_DEFAULT,
465 CV_INIT(&rx_waitingForPackets_cv, "rx_waitingForPackets_cv", CV_DEFAULT,
467 RWLOCK_INIT(&rx_peerHashTable_lock, "rx_peerHashTable_lock", MUTEX_DEFAULT,
469 RWLOCK_INIT(&rx_connHashTable_lock, "rx_connHashTable_lock", MUTEX_DEFAULT,
471 MUTEX_INIT(&rx_serverPool_lock, "rx_serverPool_lock", MUTEX_DEFAULT, 0);
472 #if defined(AFS_HPUX110_ENV)
474 rx_sleepLock = alloc_spinlock(LAST_HELD_ORDER - 10, "rx_sleepLock");
475 #endif /* AFS_HPUX110_ENV */
476 #endif /* RX_ENABLE_LOCKS && KERNEL */
479 rx_connDeadTime = 12;
480 rx_tranquil = 0; /* reset flag */
481 memset((char *)&rx_stats, 0, sizeof(struct rx_statistics));
483 osi_Alloc(rx_hashTableSize * sizeof(struct rx_connection *));
484 PIN(htable, rx_hashTableSize * sizeof(struct rx_connection *)); /* XXXXX */
485 memset(htable, 0, rx_hashTableSize * sizeof(struct rx_connection *));
486 ptable = (char *)osi_Alloc(rx_hashTableSize * sizeof(struct rx_peer *));
487 PIN(ptable, rx_hashTableSize * sizeof(struct rx_peer *)); /* XXXXX */
488 memset(ptable, 0, rx_hashTableSize * sizeof(struct rx_peer *));
490 /* Malloc up a bunch of packets & buffers */
492 queue_Init(&rx_freePacketQueue);
493 rxi_NeedMorePackets = FALSE;
494 #ifdef RX_ENABLE_TSFPQ
495 rx_nPackets = 0; /* in TSFPQ version, rx_nPackets is managed by rxi_MorePackets* */
496 rxi_MorePacketsTSFPQ(rx_extraPackets + RX_MAX_QUOTA + 2, RX_TS_FPQ_FLUSH_GLOBAL, 0);
497 #else /* RX_ENABLE_TSFPQ */
498 rx_nPackets = rx_extraPackets + RX_MAX_QUOTA + 2; /* fudge */
499 rxi_MorePackets(rx_nPackets);
500 #endif /* RX_ENABLE_TSFPQ */
507 #if defined(AFS_NT40_ENV) && !defined(AFS_PTHREAD_ENV)
508 tv.tv_sec = clock_now.sec;
509 tv.tv_usec = clock_now.usec;
510 srand((unsigned int)tv.tv_usec);
517 #if defined(KERNEL) && !defined(UKERNEL)
518 /* Really, this should never happen in a real kernel */
521 struct sockaddr_in addr;
523 int addrlen = sizeof(addr);
525 socklen_t addrlen = sizeof(addr);
527 if (getsockname((int)rx_socket, (struct sockaddr *)&addr, &addrlen)) {
531 rx_port = addr.sin_port;
534 rx_stats.minRtt.sec = 9999999;
536 rx_SetEpoch(tv.tv_sec | 0x80000000);
538 rx_SetEpoch(tv.tv_sec); /* Start time of this package, rxkad
539 * will provide a randomer value. */
541 rx_MutexAdd(rxi_dataQuota, rx_extraQuota, rx_stats_mutex); /* + extra pkts caller asked to rsrv */
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(afs_uint32 shost, u_short sport, u_short sservice,
777 struct rx_securityClass *securityObject,
778 int serviceSecurityIndex)
781 afs_int32 cix, nclones;
782 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 RWLOCK_WRLOCK(&rx_connHashTable_lock);
793 * allocate the connection and all of its clones.
794 * clones are flagged as such and have their
795 * parent set to the 0th connection object.
797 for (nclones = rx_max_clones_per_connection,
801 ++cix, ptconn = tconn) {
803 tconn = rxi_AllocConnection();
804 tconn->cid = (rx_nextCid += RX_MAXCALLS);
805 tconn->type = RX_CLIENT_CONNECTION;
806 tconn->epoch = rx_epoch;
807 tconn->peer = rxi_FindPeer(shost, sport, 0, 1);
808 tconn->serviceId = sservice;
809 tconn->securityObject = securityObject;
810 tconn->securityData = (void *) 0;
811 tconn->securityIndex = serviceSecurityIndex;
812 tconn->ackRate = RX_FAST_ACK_RATE;
813 tconn->nSpecific = 0;
814 tconn->specific = NULL;
815 tconn->challengeEvent = NULL;
816 tconn->delayedAbortEvent = NULL;
817 tconn->abortCount = 0;
820 for (i = 0; i < RX_MAXCALLS; i++) {
821 tconn->twind[i] = rx_initSendWindow;
822 tconn->rwind[i] = rx_initReceiveWindow;
827 conn->nclones = nclones;
829 conn->next_clone = 0;
830 rx_SetConnDeadTime(conn, rx_connDeadTime);
833 tconn->flags |= RX_CLONED_CONNECTION;
834 tconn->parent = conn;
835 ptconn->next_clone = tconn;
836 tconn->secondsUntilDead = 0;
837 tconn->secondsUntilPing = 0;
840 /* generic connection setup */
841 #ifdef RX_ENABLE_LOCKS
842 MUTEX_INIT(&tconn->conn_call_lock, "conn call lock", MUTEX_DEFAULT,
844 MUTEX_INIT(&tconn->conn_data_lock, "conn data lock", MUTEX_DEFAULT,
846 CV_INIT(&tconn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
848 RXS_NewConnection(securityObject, tconn);
850 CONN_HASH(shost, sport, tconn->cid, tconn->epoch,
851 RX_CLIENT_CONNECTION);
852 tconn->refCount++; /* no lock required since only this thread knows */
853 tconn->next = rx_connHashTable[hashindex];
854 rx_connHashTable[hashindex] = tconn;
855 rx_MutexIncrement(rx_stats.nClientConns, rx_stats_mutex);
858 RWLOCK_UNLOCK(&rx_connHashTable_lock);
864 rx_SetConnDeadTime(struct rx_connection *conn, int seconds)
866 /* The idea is to set the dead time to a value that allows several
867 * keepalives to be dropped without timing out the connection. */
868 struct rx_connection *tconn =
869 (rx_IsClonedConn(conn)) ? conn->parent : conn;
871 tconn->secondsUntilDead = MAX(seconds, 6);
872 tconn->secondsUntilPing = rx_ConnSecondsUntilDead(tconn) / 6;
875 int rxi_lowPeerRefCount = 0;
876 int rxi_lowConnRefCount = 0;
879 * Cleanup a connection that was destroyed in rxi_DestroyConnectioNoLock.
880 * NOTE: must not be called with rx_connHashTable_lock held.
883 rxi_CleanupConnection(struct rx_connection *conn)
885 /* Notify the service exporter, if requested, that this connection
886 * is being destroyed */
887 if (conn->type == RX_SERVER_CONNECTION && conn->service->destroyConnProc)
888 (*conn->service->destroyConnProc) (conn);
890 /* Notify the security module that this connection is being destroyed */
891 RXS_DestroyConnection(conn->securityObject, conn);
893 /* If this is the last connection using the rx_peer struct, set its
894 * idle time to now. rxi_ReapConnections will reap it if it's still
895 * idle (refCount == 0) after rx_idlePeerTime (60 seconds) have passed.
897 RWLOCK_WRLOCK(&rx_peerHashTable_lock);
898 if (conn->peer->refCount < 2) {
899 conn->peer->idleWhen = clock_Sec();
900 if (conn->peer->refCount < 1) {
901 conn->peer->refCount = 1;
902 rx_MutexIncrement(rxi_lowPeerRefCount, rx_stats_mutex);
905 conn->peer->refCount--;
906 RWLOCK_UNLOCK(&rx_peerHashTable_lock);
908 if (conn->type == RX_SERVER_CONNECTION)
909 rx_MutexDecrement(rx_stats.nServerConns, rx_stats_mutex);
911 rx_MutexDecrement(rx_stats.nClientConns, rx_stats_mutex);
913 if (conn->specific) {
915 for (i = 0; i < conn->nSpecific; i++) {
916 if (conn->specific[i] && rxi_keyCreate_destructor[i])
917 (*rxi_keyCreate_destructor[i]) (conn->specific[i]);
918 conn->specific[i] = NULL;
920 free(conn->specific);
922 conn->specific = NULL;
926 MUTEX_DESTROY(&conn->conn_call_lock);
927 MUTEX_DESTROY(&conn->conn_data_lock);
928 CV_DESTROY(&conn->conn_call_cv);
930 rxi_FreeConnection(conn);
933 /* Destroy the specified connection */
935 rxi_DestroyConnection(register struct rx_connection *conn)
937 register struct rx_connection *tconn, *dtconn;
939 RWLOCK_WRLOCK(&rx_connHashTable_lock);
941 /* destroy any clones that might exist */
942 if (!rx_IsClonedConn(conn)) {
943 tconn = conn->next_clone;
944 conn->next_clone = 0; /* once */
948 tconn = tconn->next_clone;
949 rxi_DestroyConnectionNoLock(dtconn);
951 * if destroyed dtconn will be the head of
952 * rx_connCleanup_list. Remove it and clean
953 * it up now as no one else is holding a
956 if (dtconn == rx_connCleanup_list) {
957 rx_connCleanup_list = rx_connCleanup_list->next;
958 MUTEX_EXIT(&rx_connHashTable_lock);
959 /* rxi_CleanupConnection will free dtconn */
960 rxi_CleanupConnection(dtconn);
961 MUTEX_ENTER(&rx_connHashTable_lock);
967 rxi_DestroyConnectionNoLock(conn);
968 /* conn should be at the head of the cleanup list */
969 if (conn == rx_connCleanup_list) {
970 rx_connCleanup_list = rx_connCleanup_list->next;
971 RWLOCK_UNLOCK(&rx_connHashTable_lock);
972 rxi_CleanupConnection(conn);
974 #ifdef RX_ENABLE_LOCKS
976 RWLOCK_UNLOCK(&rx_connHashTable_lock);
978 #endif /* RX_ENABLE_LOCKS */
982 rxi_DestroyConnectionNoLock(register struct rx_connection *conn)
984 register struct rx_connection **conn_ptr;
985 register int havecalls = 0;
986 struct rx_packet *packet;
993 MUTEX_ENTER(&conn->conn_data_lock);
994 if (conn->refCount > 0)
997 rx_MutexIncrement(rxi_lowConnRefCount, rx_stats_mutex);
1000 if ((conn->refCount > 0) || (conn->flags & RX_CONN_BUSY)) {
1001 /* Busy; wait till the last guy before proceeding */
1002 MUTEX_EXIT(&conn->conn_data_lock);
1007 /* If the client previously called rx_NewCall, but it is still
1008 * waiting, treat this as a running call, and wait to destroy the
1009 * connection later when the call completes. */
1010 if ((conn->type == RX_CLIENT_CONNECTION)
1011 && (conn->flags & RX_CONN_MAKECALL_WAITING)) {
1012 conn->flags |= RX_CONN_DESTROY_ME;
1013 MUTEX_EXIT(&conn->conn_data_lock);
1017 MUTEX_EXIT(&conn->conn_data_lock);
1019 /* Check for extant references to this connection */
1020 for (i = 0; i < RX_MAXCALLS; i++) {
1021 register struct rx_call *call = conn->call[i];
1024 if (conn->type == RX_CLIENT_CONNECTION) {
1025 MUTEX_ENTER(&call->lock);
1026 if (call->delayedAckEvent) {
1027 /* Push the final acknowledgment out now--there
1028 * won't be a subsequent call to acknowledge the
1029 * last reply packets */
1030 rxevent_Cancel(call->delayedAckEvent, call,
1031 RX_CALL_REFCOUNT_DELAY);
1032 if (call->state == RX_STATE_PRECALL
1033 || call->state == RX_STATE_ACTIVE) {
1034 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1036 rxi_AckAll(NULL, call, 0);
1039 MUTEX_EXIT(&call->lock);
1043 #ifdef RX_ENABLE_LOCKS
1045 if (MUTEX_TRYENTER(&conn->conn_data_lock)) {
1046 MUTEX_EXIT(&conn->conn_data_lock);
1048 /* Someone is accessing a packet right now. */
1052 #endif /* RX_ENABLE_LOCKS */
1055 /* Don't destroy the connection if there are any call
1056 * structures still in use */
1057 rx_MutexOr(conn->flags, RX_CONN_DESTROY_ME, conn->conn_data_lock);
1062 if (conn->delayedAbortEvent) {
1063 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
1064 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
1066 MUTEX_ENTER(&conn->conn_data_lock);
1067 rxi_SendConnectionAbort(conn, packet, 0, 1);
1068 MUTEX_EXIT(&conn->conn_data_lock);
1069 rxi_FreePacket(packet);
1073 /* Remove from connection hash table before proceeding */
1075 &rx_connHashTable[CONN_HASH
1076 (peer->host, peer->port, conn->cid, conn->epoch,
1078 for (; *conn_ptr; conn_ptr = &(*conn_ptr)->next) {
1079 if (*conn_ptr == conn) {
1080 *conn_ptr = conn->next;
1084 /* if the conn that we are destroying was the last connection, then we
1085 * clear rxLastConn as well */
1086 if (rxLastConn == conn)
1089 /* Make sure the connection is completely reset before deleting it. */
1090 /* get rid of pending events that could zap us later */
1091 if (conn->challengeEvent)
1092 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
1093 if (conn->checkReachEvent)
1094 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
1096 /* Add the connection to the list of destroyed connections that
1097 * need to be cleaned up. This is necessary to avoid deadlocks
1098 * in the routines we call to inform others that this connection is
1099 * being destroyed. */
1100 conn->next = rx_connCleanup_list;
1101 rx_connCleanup_list = conn;
1104 /* Externally available version */
1106 rx_DestroyConnection(register struct rx_connection *conn)
1111 rxi_DestroyConnection(conn);
1116 rx_GetConnection(register struct rx_connection *conn)
1121 rx_MutexIncrement(conn->refCount, conn->conn_data_lock);
1125 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1126 /* Wait for the transmit queue to no longer be busy.
1127 * requires the call->lock to be held */
1128 static void rxi_WaitforTQBusy(struct rx_call *call) {
1129 while (call->flags & RX_CALL_TQ_BUSY) {
1130 call->flags |= RX_CALL_TQ_WAIT;
1132 #ifdef RX_ENABLE_LOCKS
1133 osirx_AssertMine(&call->lock, "rxi_WaitforTQ lock");
1134 CV_WAIT(&call->cv_tq, &call->lock);
1135 #else /* RX_ENABLE_LOCKS */
1136 osi_rxSleep(&call->tq);
1137 #endif /* RX_ENABLE_LOCKS */
1139 if (call->tqWaiters == 0) {
1140 call->flags &= ~RX_CALL_TQ_WAIT;
1146 /* Start a new rx remote procedure call, on the specified connection.
1147 * If wait is set to 1, wait for a free call channel; otherwise return
1148 * 0. Maxtime gives the maximum number of seconds this call may take,
1149 * after rx_NewCall returns. After this time interval, a call to any
1150 * of rx_SendData, rx_ReadData, etc. will fail with RX_CALL_TIMEOUT.
1151 * For fine grain locking, we hold the conn_call_lock in order to
1152 * to ensure that we don't get signalle after we found a call in an active
1153 * state and before we go to sleep.
1156 rx_NewCall(register struct rx_connection *conn)
1159 register struct rx_call *call;
1160 register struct rx_connection *tconn;
1161 struct clock queueTime;
1165 dpf(("rx_NewCall(conn %x)\n", conn));
1168 clock_GetTime(&queueTime);
1169 MUTEX_ENTER(&conn->conn_call_lock);
1172 * Check if there are others waiting for a new call.
1173 * If so, let them go first to avoid starving them.
1174 * This is a fairly simple scheme, and might not be
1175 * a complete solution for large numbers of waiters.
1177 * makeCallWaiters keeps track of the number of
1178 * threads waiting to make calls and the
1179 * RX_CONN_MAKECALL_WAITING flag bit is used to
1180 * indicate that there are indeed calls waiting.
1181 * The flag is set when the waiter is incremented.
1182 * It is only cleared in rx_EndCall when
1183 * makeCallWaiters is 0. This prevents us from
1184 * accidently destroying the connection while it
1185 * is potentially about to be used.
1187 MUTEX_ENTER(&conn->conn_data_lock);
1188 if (conn->makeCallWaiters) {
1189 conn->flags |= RX_CONN_MAKECALL_WAITING;
1190 conn->makeCallWaiters++;
1191 MUTEX_EXIT(&conn->conn_data_lock);
1193 #ifdef RX_ENABLE_LOCKS
1194 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1198 rx_MutexDecrement(conn->makeCallWaiters, conn->conn_data_lock);
1200 MUTEX_EXIT(&conn->conn_data_lock);
1203 /* search for next free call on this connection or
1204 * its clones, if any */
1206 register struct rx_connection *tconn;
1208 for (tconn = conn; tconn; tconn = tconn->next_clone) {
1209 for (i = 0; i < RX_MAXCALLS; i++) {
1210 call = tconn->call[i];
1212 MUTEX_ENTER(&call->lock);
1213 if (call->state == RX_STATE_DALLY) {
1214 rxi_ResetCall(call, 0);
1215 (*call->callNumber)++;
1218 MUTEX_EXIT(&call->lock);
1220 call = rxi_NewCall(tconn, i);
1223 } /* for i < RX_MAXCALLS */
1227 * to be here, all available calls for this connection (and all
1228 * of its clones) must be in use
1231 MUTEX_ENTER(&conn->conn_data_lock);
1232 conn->flags |= RX_CONN_MAKECALL_WAITING;
1233 conn->makeCallWaiters++;
1234 MUTEX_EXIT(&conn->conn_data_lock);
1236 #ifdef RX_ENABLE_LOCKS
1237 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1241 rx_MutexDecrement(conn->makeCallWaiters, conn->conn_data_lock);
1246 * Wake up anyone else who might be giving us a chance to
1247 * run (see code above that avoids resource starvation).
1249 #ifdef RX_ENABLE_LOCKS
1250 CV_BROADCAST(&conn->conn_call_cv);
1255 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1257 /* Client is initially in send mode */
1258 call->state = RX_STATE_ACTIVE;
1259 call->error = rx_ConnError(conn);
1261 call->mode = RX_MODE_ERROR;
1263 call->mode = RX_MODE_SENDING;
1265 /* remember start time for call in case we have hard dead time limit */
1266 call->queueTime = queueTime;
1267 clock_GetTime(&call->startTime);
1268 hzero(call->bytesSent);
1269 hzero(call->bytesRcvd);
1271 /* Turn on busy protocol. */
1272 rxi_KeepAliveOn(call);
1274 MUTEX_EXIT(&call->lock);
1275 MUTEX_EXIT(&conn->conn_call_lock);
1278 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1279 /* Now, if TQ wasn't cleared earlier, do it now. */
1280 MUTEX_ENTER(&call->lock);
1281 rxi_WaitforTQBusy(call);
1282 if (call->flags & RX_CALL_TQ_CLEARME) {
1283 rxi_ClearTransmitQueue(call, 1);
1284 /*queue_Init(&call->tq);*/
1286 MUTEX_EXIT(&call->lock);
1287 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1289 dpf(("rx_NewCall(call %x)\n", call));
1294 rxi_HasActiveCalls(register struct rx_connection *aconn)
1297 register struct rx_call *tcall;
1301 for (i = 0; i < RX_MAXCALLS; i++) {
1302 if ((tcall = aconn->call[i])) {
1303 if ((tcall->state == RX_STATE_ACTIVE)
1304 || (tcall->state == RX_STATE_PRECALL)) {
1315 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1316 register afs_int32 * aint32s)
1319 register struct rx_call *tcall;
1323 for (i = 0; i < RX_MAXCALLS; i++) {
1324 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1325 aint32s[i] = aconn->callNumber[i] + 1;
1327 aint32s[i] = aconn->callNumber[i];
1334 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1335 register afs_int32 * aint32s)
1338 register struct rx_call *tcall;
1342 for (i = 0; i < RX_MAXCALLS; i++) {
1343 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1344 aconn->callNumber[i] = aint32s[i] - 1;
1346 aconn->callNumber[i] = aint32s[i];
1352 /* Advertise a new service. A service is named locally by a UDP port
1353 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1356 char *serviceName; Name for identification purposes (e.g. the
1357 service name might be used for probing for
1360 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1361 char *serviceName, struct rx_securityClass **securityObjects,
1362 int nSecurityObjects,
1363 afs_int32(*serviceProc) (struct rx_call * acall))
1365 osi_socket socket = OSI_NULLSOCKET;
1366 register struct rx_service *tservice;
1372 if (serviceId == 0) {
1374 "rx_NewService: service id for service %s is not non-zero.\n",
1381 "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",
1389 tservice = rxi_AllocService();
1391 for (i = 0; i < RX_MAX_SERVICES; i++) {
1392 register struct rx_service *service = rx_services[i];
1394 if (port == service->servicePort && host == service->serviceHost) {
1395 if (service->serviceId == serviceId) {
1396 /* The identical service has already been
1397 * installed; if the caller was intending to
1398 * change the security classes used by this
1399 * service, he/she loses. */
1401 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1402 serviceName, serviceId, service->serviceName);
1404 rxi_FreeService(tservice);
1407 /* Different service, same port: re-use the socket
1408 * which is bound to the same port */
1409 socket = service->socket;
1412 if (socket == OSI_NULLSOCKET) {
1413 /* If we don't already have a socket (from another
1414 * service on same port) get a new one */
1415 socket = rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
1416 if (socket == OSI_NULLSOCKET) {
1418 rxi_FreeService(tservice);
1423 service->socket = socket;
1424 service->serviceHost = host;
1425 service->servicePort = port;
1426 service->serviceId = serviceId;
1427 service->serviceName = serviceName;
1428 service->nSecurityObjects = nSecurityObjects;
1429 service->securityObjects = securityObjects;
1430 service->minProcs = 0;
1431 service->maxProcs = 1;
1432 service->idleDeadTime = 60;
1433 service->idleDeadErr = 0;
1434 service->connDeadTime = rx_connDeadTime;
1435 service->executeRequestProc = serviceProc;
1436 service->checkReach = 0;
1437 rx_services[i] = service; /* not visible until now */
1443 rxi_FreeService(tservice);
1444 (osi_Msg "rx_NewService: cannot support > %d services\n",
1449 /* Set configuration options for all of a service's security objects */
1452 rx_SetSecurityConfiguration(struct rx_service *service,
1453 rx_securityConfigVariables type,
1457 for (i = 0; i<service->nSecurityObjects; i++) {
1458 if (service->securityObjects[i]) {
1459 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1467 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1468 struct rx_securityClass **securityObjects, int nSecurityObjects,
1469 afs_int32(*serviceProc) (struct rx_call * acall))
1471 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1474 /* Generic request processing loop. This routine should be called
1475 * by the implementation dependent rx_ServerProc. If socketp is
1476 * non-null, it will be set to the file descriptor that this thread
1477 * is now listening on. If socketp is null, this routine will never
1480 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1482 register struct rx_call *call;
1483 register afs_int32 code;
1484 register struct rx_service *tservice = NULL;
1491 call = rx_GetCall(threadID, tservice, socketp);
1492 if (socketp && *socketp != OSI_NULLSOCKET) {
1493 /* We are now a listener thread */
1498 /* if server is restarting( typically smooth shutdown) then do not
1499 * allow any new calls.
1502 if (rx_tranquil && (call != NULL)) {
1506 MUTEX_ENTER(&call->lock);
1508 rxi_CallError(call, RX_RESTARTING);
1509 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1511 MUTEX_EXIT(&call->lock);
1515 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1516 #ifdef RX_ENABLE_LOCKS
1518 #endif /* RX_ENABLE_LOCKS */
1519 afs_termState = AFSOP_STOP_AFS;
1520 afs_osi_Wakeup(&afs_termState);
1521 #ifdef RX_ENABLE_LOCKS
1523 #endif /* RX_ENABLE_LOCKS */
1528 tservice = call->conn->service;
1530 if (tservice->beforeProc)
1531 (*tservice->beforeProc) (call);
1533 code = call->conn->service->executeRequestProc(call);
1535 if (tservice->afterProc)
1536 (*tservice->afterProc) (call, code);
1538 rx_EndCall(call, code);
1539 rx_MutexIncrement(rxi_nCalls, rx_stats_mutex);
1545 rx_WakeupServerProcs(void)
1547 struct rx_serverQueueEntry *np, *tqp;
1551 MUTEX_ENTER(&rx_serverPool_lock);
1553 #ifdef RX_ENABLE_LOCKS
1554 if (rx_waitForPacket)
1555 CV_BROADCAST(&rx_waitForPacket->cv);
1556 #else /* RX_ENABLE_LOCKS */
1557 if (rx_waitForPacket)
1558 osi_rxWakeup(rx_waitForPacket);
1559 #endif /* RX_ENABLE_LOCKS */
1560 MUTEX_ENTER(&freeSQEList_lock);
1561 for (np = rx_FreeSQEList; np; np = tqp) {
1562 tqp = *(struct rx_serverQueueEntry **)np;
1563 #ifdef RX_ENABLE_LOCKS
1564 CV_BROADCAST(&np->cv);
1565 #else /* RX_ENABLE_LOCKS */
1567 #endif /* RX_ENABLE_LOCKS */
1569 MUTEX_EXIT(&freeSQEList_lock);
1570 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1571 #ifdef RX_ENABLE_LOCKS
1572 CV_BROADCAST(&np->cv);
1573 #else /* RX_ENABLE_LOCKS */
1575 #endif /* RX_ENABLE_LOCKS */
1577 MUTEX_EXIT(&rx_serverPool_lock);
1582 * One thing that seems to happen is that all the server threads get
1583 * tied up on some empty or slow call, and then a whole bunch of calls
1584 * arrive at once, using up the packet pool, so now there are more
1585 * empty calls. The most critical resources here are server threads
1586 * and the free packet pool. The "doreclaim" code seems to help in
1587 * general. I think that eventually we arrive in this state: there
1588 * are lots of pending calls which do have all their packets present,
1589 * so they won't be reclaimed, are multi-packet calls, so they won't
1590 * be scheduled until later, and thus are tying up most of the free
1591 * packet pool for a very long time.
1593 * 1. schedule multi-packet calls if all the packets are present.
1594 * Probably CPU-bound operation, useful to return packets to pool.
1595 * Do what if there is a full window, but the last packet isn't here?
1596 * 3. preserve one thread which *only* runs "best" calls, otherwise
1597 * it sleeps and waits for that type of call.
1598 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1599 * the current dataquota business is badly broken. The quota isn't adjusted
1600 * to reflect how many packets are presently queued for a running call.
1601 * So, when we schedule a queued call with a full window of packets queued
1602 * up for it, that *should* free up a window full of packets for other 2d-class
1603 * calls to be able to use from the packet pool. But it doesn't.
1605 * NB. Most of the time, this code doesn't run -- since idle server threads
1606 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1607 * as a new call arrives.
1609 /* Sleep until a call arrives. Returns a pointer to the call, ready
1610 * for an rx_Read. */
1611 #ifdef RX_ENABLE_LOCKS
1613 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1615 struct rx_serverQueueEntry *sq;
1616 register struct rx_call *call = (struct rx_call *)0;
1617 struct rx_service *service = NULL;
1620 MUTEX_ENTER(&freeSQEList_lock);
1622 if ((sq = rx_FreeSQEList)) {
1623 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1624 MUTEX_EXIT(&freeSQEList_lock);
1625 } else { /* otherwise allocate a new one and return that */
1626 MUTEX_EXIT(&freeSQEList_lock);
1627 sq = (struct rx_serverQueueEntry *)
1628 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1629 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1630 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1633 MUTEX_ENTER(&rx_serverPool_lock);
1634 if (cur_service != NULL) {
1635 ReturnToServerPool(cur_service);
1638 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1639 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1641 /* Scan for eligible incoming calls. A call is not eligible
1642 * if the maximum number of calls for its service type are
1643 * already executing */
1644 /* One thread will process calls FCFS (to prevent starvation),
1645 * while the other threads may run ahead looking for calls which
1646 * have all their input data available immediately. This helps
1647 * keep threads from blocking, waiting for data from the client. */
1648 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1649 service = tcall->conn->service;
1650 if (!QuotaOK(service)) {
1653 if (tno == rxi_fcfs_thread_num
1654 || !tcall->queue_item_header.next) {
1655 /* If we're the fcfs thread , then we'll just use
1656 * this call. If we haven't been able to find an optimal
1657 * choice, and we're at the end of the list, then use a
1658 * 2d choice if one has been identified. Otherwise... */
1659 call = (choice2 ? choice2 : tcall);
1660 service = call->conn->service;
1661 } else if (!queue_IsEmpty(&tcall->rq)) {
1662 struct rx_packet *rp;
1663 rp = queue_First(&tcall->rq, rx_packet);
1664 if (rp->header.seq == 1) {
1666 || (rp->header.flags & RX_LAST_PACKET)) {
1668 } else if (rxi_2dchoice && !choice2
1669 && !(tcall->flags & RX_CALL_CLEARED)
1670 && (tcall->rprev > rxi_HardAckRate)) {
1679 ReturnToServerPool(service);
1686 MUTEX_EXIT(&rx_serverPool_lock);
1687 MUTEX_ENTER(&call->lock);
1689 if (call->flags & RX_CALL_WAIT_PROC) {
1690 call->flags &= ~RX_CALL_WAIT_PROC;
1691 rx_MutexDecrement(rx_nWaiting, rx_stats_mutex);
1694 if (call->state != RX_STATE_PRECALL || call->error) {
1695 MUTEX_EXIT(&call->lock);
1696 MUTEX_ENTER(&rx_serverPool_lock);
1697 ReturnToServerPool(service);
1702 if (queue_IsEmpty(&call->rq)
1703 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1704 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1706 CLEAR_CALL_QUEUE_LOCK(call);
1709 /* If there are no eligible incoming calls, add this process
1710 * to the idle server queue, to wait for one */
1714 *socketp = OSI_NULLSOCKET;
1716 sq->socketp = socketp;
1717 queue_Append(&rx_idleServerQueue, sq);
1718 #ifndef AFS_AIX41_ENV
1719 rx_waitForPacket = sq;
1721 rx_waitingForPacket = sq;
1722 #endif /* AFS_AIX41_ENV */
1724 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1726 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1727 MUTEX_EXIT(&rx_serverPool_lock);
1728 return (struct rx_call *)0;
1731 } while (!(call = sq->newcall)
1732 && !(socketp && *socketp != OSI_NULLSOCKET));
1733 MUTEX_EXIT(&rx_serverPool_lock);
1735 MUTEX_ENTER(&call->lock);
1741 MUTEX_ENTER(&freeSQEList_lock);
1742 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1743 rx_FreeSQEList = sq;
1744 MUTEX_EXIT(&freeSQEList_lock);
1747 clock_GetTime(&call->startTime);
1748 call->state = RX_STATE_ACTIVE;
1749 call->mode = RX_MODE_RECEIVING;
1750 #ifdef RX_KERNEL_TRACE
1751 if (ICL_SETACTIVE(afs_iclSetp)) {
1752 int glockOwner = ISAFS_GLOCK();
1755 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1756 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1763 rxi_calltrace(RX_CALL_START, call);
1764 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1765 call->conn->service->servicePort, call->conn->service->serviceId,
1768 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1769 MUTEX_EXIT(&call->lock);
1771 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1776 #else /* RX_ENABLE_LOCKS */
1778 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1780 struct rx_serverQueueEntry *sq;
1781 register struct rx_call *call = (struct rx_call *)0, *choice2;
1782 struct rx_service *service = NULL;
1786 MUTEX_ENTER(&freeSQEList_lock);
1788 if ((sq = rx_FreeSQEList)) {
1789 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1790 MUTEX_EXIT(&freeSQEList_lock);
1791 } else { /* otherwise allocate a new one and return that */
1792 MUTEX_EXIT(&freeSQEList_lock);
1793 sq = (struct rx_serverQueueEntry *)
1794 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1795 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1796 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1798 MUTEX_ENTER(&sq->lock);
1800 if (cur_service != NULL) {
1801 cur_service->nRequestsRunning--;
1802 if (cur_service->nRequestsRunning < cur_service->minProcs)
1806 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1807 register struct rx_call *tcall, *ncall;
1808 /* Scan for eligible incoming calls. A call is not eligible
1809 * if the maximum number of calls for its service type are
1810 * already executing */
1811 /* One thread will process calls FCFS (to prevent starvation),
1812 * while the other threads may run ahead looking for calls which
1813 * have all their input data available immediately. This helps
1814 * keep threads from blocking, waiting for data from the client. */
1815 choice2 = (struct rx_call *)0;
1816 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1817 service = tcall->conn->service;
1818 if (QuotaOK(service)) {
1819 if (tno == rxi_fcfs_thread_num
1820 || !tcall->queue_item_header.next) {
1821 /* If we're the fcfs thread, then we'll just use
1822 * this call. If we haven't been able to find an optimal
1823 * choice, and we're at the end of the list, then use a
1824 * 2d choice if one has been identified. Otherwise... */
1825 call = (choice2 ? choice2 : tcall);
1826 service = call->conn->service;
1827 } else if (!queue_IsEmpty(&tcall->rq)) {
1828 struct rx_packet *rp;
1829 rp = queue_First(&tcall->rq, rx_packet);
1830 if (rp->header.seq == 1
1832 || (rp->header.flags & RX_LAST_PACKET))) {
1834 } else if (rxi_2dchoice && !choice2
1835 && !(tcall->flags & RX_CALL_CLEARED)
1836 && (tcall->rprev > rxi_HardAckRate)) {
1849 /* we can't schedule a call if there's no data!!! */
1850 /* send an ack if there's no data, if we're missing the
1851 * first packet, or we're missing something between first
1852 * and last -- there's a "hole" in the incoming data. */
1853 if (queue_IsEmpty(&call->rq)
1854 || queue_First(&call->rq, rx_packet)->header.seq != 1
1855 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1856 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1858 call->flags &= (~RX_CALL_WAIT_PROC);
1859 service->nRequestsRunning++;
1860 /* just started call in minProcs pool, need fewer to maintain
1862 if (service->nRequestsRunning <= service->minProcs)
1866 /* MUTEX_EXIT(&call->lock); */
1868 /* If there are no eligible incoming calls, add this process
1869 * to the idle server queue, to wait for one */
1872 *socketp = OSI_NULLSOCKET;
1874 sq->socketp = socketp;
1875 queue_Append(&rx_idleServerQueue, sq);
1879 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1881 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1882 return (struct rx_call *)0;
1885 } while (!(call = sq->newcall)
1886 && !(socketp && *socketp != OSI_NULLSOCKET));
1888 MUTEX_EXIT(&sq->lock);
1890 MUTEX_ENTER(&freeSQEList_lock);
1891 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1892 rx_FreeSQEList = sq;
1893 MUTEX_EXIT(&freeSQEList_lock);
1896 clock_GetTime(&call->startTime);
1897 call->state = RX_STATE_ACTIVE;
1898 call->mode = RX_MODE_RECEIVING;
1899 #ifdef RX_KERNEL_TRACE
1900 if (ICL_SETACTIVE(afs_iclSetp)) {
1901 int glockOwner = ISAFS_GLOCK();
1904 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1905 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1912 rxi_calltrace(RX_CALL_START, call);
1913 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1914 call->conn->service->servicePort, call->conn->service->serviceId,
1917 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1924 #endif /* RX_ENABLE_LOCKS */
1928 /* Establish a procedure to be called when a packet arrives for a
1929 * call. This routine will be called at most once after each call,
1930 * and will also be called if there is an error condition on the or
1931 * the call is complete. Used by multi rx to build a selection
1932 * function which determines which of several calls is likely to be a
1933 * good one to read from.
1934 * NOTE: the way this is currently implemented it is probably only a
1935 * good idea to (1) use it immediately after a newcall (clients only)
1936 * and (2) only use it once. Other uses currently void your warranty
1939 rx_SetArrivalProc(register struct rx_call *call,
1940 register void (*proc) (register struct rx_call * call,
1942 register int index),
1943 register void * handle, register int arg)
1945 call->arrivalProc = proc;
1946 call->arrivalProcHandle = handle;
1947 call->arrivalProcArg = arg;
1950 /* Call is finished (possibly prematurely). Return rc to the peer, if
1951 * appropriate, and return the final error code from the conversation
1955 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1957 register struct rx_connection *conn = call->conn;
1958 register struct rx_service *service;
1964 dpf(("rx_EndCall(call %x rc %d error %d abortCode %d)\n", call, rc, call->error, call->abortCode));
1967 MUTEX_ENTER(&call->lock);
1969 if (rc == 0 && call->error == 0) {
1970 call->abortCode = 0;
1971 call->abortCount = 0;
1974 call->arrivalProc = (void (*)())0;
1975 if (rc && call->error == 0) {
1976 rxi_CallError(call, rc);
1977 /* Send an abort message to the peer if this error code has
1978 * only just been set. If it was set previously, assume the
1979 * peer has already been sent the error code or will request it
1981 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1983 if (conn->type == RX_SERVER_CONNECTION) {
1984 /* Make sure reply or at least dummy reply is sent */
1985 if (call->mode == RX_MODE_RECEIVING) {
1986 rxi_WriteProc(call, 0, 0);
1988 if (call->mode == RX_MODE_SENDING) {
1989 rxi_FlushWrite(call);
1991 service = conn->service;
1992 rxi_calltrace(RX_CALL_END, call);
1993 /* Call goes to hold state until reply packets are acknowledged */
1994 if (call->tfirst + call->nSoftAcked < call->tnext) {
1995 call->state = RX_STATE_HOLD;
1997 call->state = RX_STATE_DALLY;
1998 rxi_ClearTransmitQueue(call, 0);
1999 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
2000 rxevent_Cancel(call->keepAliveEvent, call,
2001 RX_CALL_REFCOUNT_ALIVE);
2003 } else { /* Client connection */
2005 /* Make sure server receives input packets, in the case where
2006 * no reply arguments are expected */
2007 if ((call->mode == RX_MODE_SENDING)
2008 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2009 (void)rxi_ReadProc(call, &dummy, 1);
2012 /* If we had an outstanding delayed ack, be nice to the server
2013 * and force-send it now.
2015 if (call->delayedAckEvent) {
2016 rxevent_Cancel(call->delayedAckEvent, call,
2017 RX_CALL_REFCOUNT_DELAY);
2018 call->delayedAckEvent = NULL;
2019 rxi_SendDelayedAck(NULL, call, NULL);
2022 /* We need to release the call lock since it's lower than the
2023 * conn_call_lock and we don't want to hold the conn_call_lock
2024 * over the rx_ReadProc call. The conn_call_lock needs to be held
2025 * here for the case where rx_NewCall is perusing the calls on
2026 * the connection structure. We don't want to signal until
2027 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2028 * have checked this call, found it active and by the time it
2029 * goes to sleep, will have missed the signal.
2031 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
2032 * there are threads waiting to use the conn object.
2034 MUTEX_EXIT(&call->lock);
2035 MUTEX_ENTER(&conn->conn_call_lock);
2036 MUTEX_ENTER(&call->lock);
2037 MUTEX_ENTER(&conn->conn_data_lock);
2038 conn->flags |= RX_CONN_BUSY;
2039 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2040 if (conn->makeCallWaiters == 0)
2041 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
2042 MUTEX_EXIT(&conn->conn_data_lock);
2043 #ifdef RX_ENABLE_LOCKS
2044 CV_BROADCAST(&conn->conn_call_cv);
2049 #ifdef RX_ENABLE_LOCKS
2051 MUTEX_EXIT(&conn->conn_data_lock);
2053 #endif /* RX_ENABLE_LOCKS */
2054 call->state = RX_STATE_DALLY;
2056 error = call->error;
2058 /* currentPacket, nLeft, and NFree must be zeroed here, because
2059 * ResetCall cannot: ResetCall may be called at splnet(), in the
2060 * kernel version, and may interrupt the macros rx_Read or
2061 * rx_Write, which run at normal priority for efficiency. */
2062 if (call->currentPacket) {
2063 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2064 rxi_FreePacket(call->currentPacket);
2065 call->currentPacket = (struct rx_packet *)0;
2068 call->nLeft = call->nFree = call->curlen = 0;
2070 /* Free any packets from the last call to ReadvProc/WritevProc */
2071 rxi_FreePackets(0, &call->iovq);
2073 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2074 MUTEX_EXIT(&call->lock);
2075 if (conn->type == RX_CLIENT_CONNECTION) {
2076 MUTEX_EXIT(&conn->conn_call_lock);
2077 conn->flags &= ~RX_CONN_BUSY;
2081 * Map errors to the local host's errno.h format.
2083 error = ntoh_syserr_conv(error);
2087 #if !defined(KERNEL)
2089 /* Call this routine when shutting down a server or client (especially
2090 * clients). This will allow Rx to gracefully garbage collect server
2091 * connections, and reduce the number of retries that a server might
2092 * make to a dead client.
2093 * This is not quite right, since some calls may still be ongoing and
2094 * we can't lock them to destroy them. */
2098 register struct rx_connection **conn_ptr, **conn_end;
2102 if (rxinit_status == 1) {
2104 return; /* Already shutdown. */
2106 rxi_DeleteCachedConnections();
2107 if (rx_connHashTable) {
2108 RWLOCK_WRLOCK(&rx_connHashTable_lock);
2109 for (conn_ptr = &rx_connHashTable[0], conn_end =
2110 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2112 struct rx_connection *conn, *next;
2113 for (conn = *conn_ptr; conn; conn = next) {
2115 if (conn->type == RX_CLIENT_CONNECTION) {
2116 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2118 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2119 #ifdef RX_ENABLE_LOCKS
2120 rxi_DestroyConnectionNoLock(conn);
2121 #else /* RX_ENABLE_LOCKS */
2122 rxi_DestroyConnection(conn);
2123 #endif /* RX_ENABLE_LOCKS */
2127 #ifdef RX_ENABLE_LOCKS
2128 while (rx_connCleanup_list) {
2129 struct rx_connection *conn;
2130 conn = rx_connCleanup_list;
2131 rx_connCleanup_list = rx_connCleanup_list->next;
2132 RWLOCK_UNLOCK(&rx_connHashTable_lock);
2133 rxi_CleanupConnection(conn);
2134 RWLOCK_WRLOCK(&rx_connHashTable_lock);
2136 RWLOCK_UNLOCK(&rx_connHashTable_lock);
2137 #endif /* RX_ENABLE_LOCKS */
2142 afs_winsockCleanup();
2150 /* if we wakeup packet waiter too often, can get in loop with two
2151 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2153 rxi_PacketsUnWait(void)
2155 if (!rx_waitingForPackets) {
2159 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2160 return; /* still over quota */
2163 rx_waitingForPackets = 0;
2164 #ifdef RX_ENABLE_LOCKS
2165 CV_BROADCAST(&rx_waitingForPackets_cv);
2167 osi_rxWakeup(&rx_waitingForPackets);
2173 /* ------------------Internal interfaces------------------------- */
2175 /* Return this process's service structure for the
2176 * specified socket and service */
2178 rxi_FindService(register osi_socket socket, register u_short serviceId)
2180 register struct rx_service **sp;
2181 for (sp = &rx_services[0]; *sp; sp++) {
2182 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2188 /* Allocate a call structure, for the indicated channel of the
2189 * supplied connection. The mode and state of the call must be set by
2190 * the caller. Returns the call with mutex locked. */
2192 rxi_NewCall(register struct rx_connection *conn, register int channel)
2194 register struct rx_call *call;
2195 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2196 register struct rx_call *cp; /* Call pointer temp */
2197 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2198 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2200 dpf(("rxi_NewCall(conn %x, channel %d)\n", conn, channel));
2202 /* Grab an existing call structure, or allocate a new one.
2203 * Existing call structures are assumed to have been left reset by
2205 MUTEX_ENTER(&rx_freeCallQueue_lock);
2207 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2209 * EXCEPT that the TQ might not yet be cleared out.
2210 * Skip over those with in-use TQs.
2213 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2214 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2220 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2221 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2222 call = queue_First(&rx_freeCallQueue, rx_call);
2223 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2225 rx_MutexDecrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2226 MUTEX_EXIT(&rx_freeCallQueue_lock);
2227 MUTEX_ENTER(&call->lock);
2228 CLEAR_CALL_QUEUE_LOCK(call);
2229 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2230 /* Now, if TQ wasn't cleared earlier, do it now. */
2231 if (call->flags & RX_CALL_TQ_CLEARME) {
2232 rxi_ClearTransmitQueue(call, 1);
2233 /*queue_Init(&call->tq);*/
2235 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2236 /* Bind the call to its connection structure */
2238 rxi_ResetCall(call, 1);
2241 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2242 rx_MutexIncrement(rx_stats.nCallStructs, rx_stats_mutex);
2243 MUTEX_EXIT(&rx_freeCallQueue_lock);
2244 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2245 MUTEX_ENTER(&call->lock);
2246 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2247 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2248 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2250 /* Initialize once-only items */
2251 queue_Init(&call->tq);
2252 queue_Init(&call->rq);
2253 queue_Init(&call->iovq);
2254 /* Bind the call to its connection structure (prereq for reset) */
2256 rxi_ResetCall(call, 1);
2258 call->channel = channel;
2259 call->callNumber = &conn->callNumber[channel];
2260 call->rwind = conn->rwind[channel];
2261 call->twind = conn->twind[channel];
2262 /* Note that the next expected call number is retained (in
2263 * conn->callNumber[i]), even if we reallocate the call structure
2265 conn->call[channel] = call;
2266 /* if the channel's never been used (== 0), we should start at 1, otherwise
2267 * the call number is valid from the last time this channel was used */
2268 if (*call->callNumber == 0)
2269 *call->callNumber = 1;
2274 /* A call has been inactive long enough that so we can throw away
2275 * state, including the call structure, which is placed on the call
2277 * Call is locked upon entry.
2278 * haveCTLock set if called from rxi_ReapConnections
2280 #ifdef RX_ENABLE_LOCKS
2282 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2283 #else /* RX_ENABLE_LOCKS */
2285 rxi_FreeCall(register struct rx_call *call)
2286 #endif /* RX_ENABLE_LOCKS */
2288 register int channel = call->channel;
2289 register struct rx_connection *conn = call->conn;
2292 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2293 (*call->callNumber)++;
2294 rxi_ResetCall(call, 0);
2295 call->conn->call[channel] = (struct rx_call *)0;
2297 MUTEX_ENTER(&rx_freeCallQueue_lock);
2298 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2299 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2300 /* A call may be free even though its transmit queue is still in use.
2301 * Since we search the call list from head to tail, put busy calls at
2302 * the head of the list, and idle calls at the tail.
2304 if (call->flags & RX_CALL_TQ_BUSY)
2305 queue_Prepend(&rx_freeCallQueue, call);
2307 queue_Append(&rx_freeCallQueue, call);
2308 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2309 queue_Append(&rx_freeCallQueue, call);
2310 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2311 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2312 MUTEX_EXIT(&rx_freeCallQueue_lock);
2314 /* Destroy the connection if it was previously slated for
2315 * destruction, i.e. the Rx client code previously called
2316 * rx_DestroyConnection (client connections), or
2317 * rxi_ReapConnections called the same routine (server
2318 * connections). Only do this, however, if there are no
2319 * outstanding calls. Note that for fine grain locking, there appears
2320 * to be a deadlock in that rxi_FreeCall has a call locked and
2321 * DestroyConnectionNoLock locks each call in the conn. But note a
2322 * few lines up where we have removed this call from the conn.
2323 * If someone else destroys a connection, they either have no
2324 * call lock held or are going through this section of code.
2326 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2327 rx_MutexIncrement(conn->refCount, conn->conn_data_lock);
2328 #ifdef RX_ENABLE_LOCKS
2330 rxi_DestroyConnectionNoLock(conn);
2332 rxi_DestroyConnection(conn);
2333 #else /* RX_ENABLE_LOCKS */
2334 rxi_DestroyConnection(conn);
2335 #endif /* RX_ENABLE_LOCKS */
2339 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2341 rxi_Alloc(register size_t size)
2345 rx_MutexAdd1Increment2(rxi_Allocsize, (afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2348 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2349 afs_osi_Alloc_NoSleep(size);
2354 osi_Panic("rxi_Alloc error");
2360 rxi_Free(void *addr, register size_t size)
2362 rx_MutexAdd1Decrement2(rxi_Allocsize, -(afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2363 osi_Free(addr, size);
2367 rxi_SetPeerMtu(register afs_uint32 host, register afs_uint32 port, int mtu)
2369 struct rx_peer **peer_ptr, **peer_end;
2372 RWLOCK_RDLOCK(&rx_peerHashTable_lock);
2374 for (peer_ptr = &rx_peerHashTable[0], peer_end =
2375 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
2377 struct rx_peer *peer, *next;
2378 for (peer = *peer_ptr; peer; peer = next) {
2380 if (host == peer->host) {
2381 MUTEX_ENTER(&peer->peer_lock);
2382 peer->ifMTU=MIN(mtu, peer->ifMTU);
2383 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2384 MUTEX_EXIT(&peer->peer_lock);
2389 struct rx_peer *peer;
2390 hashIndex = PEER_HASH(host, port);
2391 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2392 if ((peer->host == host) && (peer->port == port)) {
2393 MUTEX_ENTER(&peer->peer_lock);
2394 peer->ifMTU=MIN(mtu, peer->ifMTU);
2395 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2396 MUTEX_EXIT(&peer->peer_lock);
2400 RWLOCK_UNLOCK(&rx_peerHashTable_lock);
2403 /* Find the peer process represented by the supplied (host,port)
2404 * combination. If there is no appropriate active peer structure, a
2405 * new one will be allocated and initialized
2406 * The origPeer, if set, is a pointer to a peer structure on which the
2407 * refcount will be be decremented. This is used to replace the peer
2408 * structure hanging off a connection structure */
2410 rxi_FindPeer(register afs_uint32 host, register u_short port,
2411 struct rx_peer *origPeer, int create)
2413 register struct rx_peer *pp;
2415 hashIndex = PEER_HASH(host, port);
2416 RWLOCK_RDLOCK(&rx_peerHashTable_lock);
2417 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2418 if ((pp->host == host) && (pp->port == port))
2423 pp = rxi_AllocPeer(); /* This bzero's *pp */
2424 pp->host = host; /* set here or in InitPeerParams is zero */
2426 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2427 queue_Init(&pp->congestionQueue);
2428 queue_Init(&pp->rpcStats);
2429 RWLOCK_UPLOCK(&rx_peerHashTable_lock);
2430 pp->next = rx_peerHashTable[hashIndex];
2431 rx_peerHashTable[hashIndex] = pp;
2432 rxi_InitPeerParams(pp);
2433 rx_MutexIncrement(rx_stats.nPeerStructs, rx_stats_mutex);
2440 origPeer->refCount--;
2441 RWLOCK_UNLOCK(&rx_peerHashTable_lock);
2446 /* Find the connection at (host, port) started at epoch, and with the
2447 * given connection id. Creates the server connection if necessary.
2448 * The type specifies whether a client connection or a server
2449 * connection is desired. In both cases, (host, port) specify the
2450 * peer's (host, pair) pair. Client connections are not made
2451 * automatically by this routine. The parameter socket gives the
2452 * socket descriptor on which the packet was received. This is used,
2453 * in the case of server connections, to check that *new* connections
2454 * come via a valid (port, serviceId). Finally, the securityIndex
2455 * parameter must match the existing index for the connection. If a
2456 * server connection is created, it will be created using the supplied
2457 * index, if the index is valid for this service */
2458 struct rx_connection *
2459 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2460 register u_short port, u_short serviceId, afs_uint32 cid,
2461 afs_uint32 epoch, int type, u_int securityIndex)
2463 int hashindex, flag, i;
2464 register struct rx_connection *conn;
2465 hashindex = CONN_HASH(host, port, cid, epoch, type);
2466 RWLOCK_RDLOCK(&rx_connHashTable_lock);
2467 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2468 rx_connHashTable[hashindex],
2471 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2472 && (epoch == conn->epoch)) {
2473 register struct rx_peer *pp = conn->peer;
2474 if (securityIndex != conn->securityIndex) {
2475 /* this isn't supposed to happen, but someone could forge a packet
2476 * like this, and there seems to be some CM bug that makes this
2477 * happen from time to time -- in which case, the fileserver
2479 RWLOCK_UNLOCK(&rx_connHashTable_lock);
2480 return (struct rx_connection *)0;
2482 if (pp->host == host && pp->port == port)
2484 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2486 /* So what happens when it's a callback connection? */
2487 if ( /*type == RX_CLIENT_CONNECTION && */
2488 (conn->epoch & 0x80000000))
2492 /* the connection rxLastConn that was used the last time is not the
2493 ** one we are looking for now. Hence, start searching in the hash */
2495 conn = rx_connHashTable[hashindex];
2500 struct rx_service *service;
2501 if (type == RX_CLIENT_CONNECTION) {
2502 RWLOCK_UNLOCK(&rx_connHashTable_lock);
2503 return (struct rx_connection *)0;
2505 service = rxi_FindService(socket, serviceId);
2506 if (!service || (securityIndex >= service->nSecurityObjects)
2507 || (service->securityObjects[securityIndex] == 0)) {
2508 RWLOCK_UNLOCK(&rx_connHashTable_lock);
2509 return (struct rx_connection *)0;
2511 conn = rxi_AllocConnection(); /* This bzero's the connection */
2512 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2513 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2514 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2515 conn->peer = rxi_FindPeer(host, port, 0, 1);
2516 conn->type = RX_SERVER_CONNECTION;
2517 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2518 conn->epoch = epoch;
2519 /* conn->serial = conn->lastSerial = 0; */
2520 /* conn->timeout = 0; */
2521 conn->ackRate = RX_FAST_ACK_RATE;
2522 conn->service = service;
2523 conn->serviceId = serviceId;
2524 conn->securityIndex = securityIndex;
2525 conn->securityObject = service->securityObjects[securityIndex];
2526 conn->nSpecific = 0;
2527 conn->specific = NULL;
2528 rx_SetConnDeadTime(conn, service->connDeadTime);
2529 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2530 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2531 for (i = 0; i < RX_MAXCALLS; i++) {
2532 conn->twind[i] = rx_initSendWindow;
2533 conn->rwind[i] = rx_initReceiveWindow;
2535 RWLOCK_UPLOCK(&rx_connHashTable_lock);
2536 conn->next = rx_connHashTable[hashindex];
2537 rx_connHashTable[hashindex] = conn;
2538 conn->cid = cid & RX_CIDMASK;
2539 /* Notify security object of the new connection */
2540 RXS_NewConnection(conn->securityObject, conn);
2541 /* XXXX Connection timeout? */
2542 if (service->newConnProc)
2543 (*service->newConnProc) (conn);
2544 rx_MutexIncrement(rx_stats.nServerConns, rx_stats_mutex);
2547 rx_MutexIncrement(conn->refCount, conn->conn_data_lock);
2549 rxLastConn = conn; /* store this connection as the last conn used */
2550 RWLOCK_UNLOCK(&rx_connHashTable_lock);
2554 /* There are two packet tracing routines available for testing and monitoring
2555 * Rx. One is called just after every packet is received and the other is
2556 * called just before every packet is sent. Received packets, have had their
2557 * headers decoded, and packets to be sent have not yet had their headers
2558 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2559 * containing the network address. Both can be modified. The return value, if
2560 * non-zero, indicates that the packet should be dropped. */
2562 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
2563 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
2565 /* A packet has been received off the interface. Np is the packet, socket is
2566 * the socket number it was received from (useful in determining which service
2567 * this packet corresponds to), and (host, port) reflect the host,port of the
2568 * sender. This call returns the packet to the caller if it is finished with
2569 * it, rather than de-allocating it, just as a small performance hack */
2572 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2573 afs_uint32 host, u_short port, int *tnop,
2574 struct rx_call **newcallp)
2576 register struct rx_call *call;
2577 register struct rx_connection *conn;
2579 afs_uint32 currentCallNumber;
2585 struct rx_packet *tnp;
2588 /* We don't print out the packet until now because (1) the time may not be
2589 * accurate enough until now in the lwp implementation (rx_Listener only gets
2590 * the time after the packet is read) and (2) from a protocol point of view,
2591 * this is the first time the packet has been seen */
2592 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2593 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2594 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2595 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2596 np->header.epoch, np->header.cid, np->header.callNumber,
2597 np->header.seq, np->header.flags, np));
2600 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2601 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2604 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2605 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2608 /* If an input tracer function is defined, call it with the packet and
2609 * network address. Note this function may modify its arguments. */
2610 if (rx_justReceived) {
2611 struct sockaddr_in addr;
2613 addr.sin_family = AF_INET;
2614 addr.sin_port = port;
2615 addr.sin_addr.s_addr = host;
2616 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2617 addr.sin_len = sizeof(addr);
2618 #endif /* AFS_OSF_ENV */
2619 drop = (*rx_justReceived) (np, &addr);
2620 /* drop packet if return value is non-zero */
2623 port = addr.sin_port; /* in case fcn changed addr */
2624 host = addr.sin_addr.s_addr;
2628 /* If packet was not sent by the client, then *we* must be the client */
2629 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2630 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2632 /* Find the connection (or fabricate one, if we're the server & if
2633 * necessary) associated with this packet */
2635 rxi_FindConnection(socket, host, port, np->header.serviceId,
2636 np->header.cid, np->header.epoch, type,
2637 np->header.securityIndex);
2640 /* If no connection found or fabricated, just ignore the packet.
2641 * (An argument could be made for sending an abort packet for
2646 MUTEX_ENTER(&conn->conn_data_lock);
2647 if (conn->maxSerial < np->header.serial)
2648 conn->maxSerial = np->header.serial;
2649 MUTEX_EXIT(&conn->conn_data_lock);
2651 /* If the connection is in an error state, send an abort packet and ignore
2652 * the incoming packet */
2653 if (rx_ConnError(conn)) {
2654 /* Don't respond to an abort packet--we don't want loops! */
2655 MUTEX_ENTER(&conn->conn_data_lock);
2656 if (np->header.type != RX_PACKET_TYPE_ABORT)
2657 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2659 MUTEX_EXIT(&conn->conn_data_lock);
2663 /* Check for connection-only requests (i.e. not call specific). */
2664 if (np->header.callNumber == 0) {
2665 switch (np->header.type) {
2666 case RX_PACKET_TYPE_ABORT: {
2667 /* What if the supplied error is zero? */
2668 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2669 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2670 rxi_ConnectionError(conn, errcode);
2671 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2674 case RX_PACKET_TYPE_CHALLENGE:
2675 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2676 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2678 case RX_PACKET_TYPE_RESPONSE:
2679 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2680 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2682 case RX_PACKET_TYPE_PARAMS:
2683 case RX_PACKET_TYPE_PARAMS + 1:
2684 case RX_PACKET_TYPE_PARAMS + 2:
2685 /* ignore these packet types for now */
2686 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2691 /* Should not reach here, unless the peer is broken: send an
2693 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2694 MUTEX_ENTER(&conn->conn_data_lock);
2695 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2697 MUTEX_EXIT(&conn->conn_data_lock);
2702 channel = np->header.cid & RX_CHANNELMASK;
2703 call = conn->call[channel];
2704 #ifdef RX_ENABLE_LOCKS
2706 MUTEX_ENTER(&call->lock);
2707 /* Test to see if call struct is still attached to conn. */
2708 if (call != conn->call[channel]) {
2710 MUTEX_EXIT(&call->lock);
2711 if (type == RX_SERVER_CONNECTION) {
2712 call = conn->call[channel];
2713 /* If we started with no call attached and there is one now,
2714 * another thread is also running this routine and has gotten
2715 * the connection channel. We should drop this packet in the tests
2716 * below. If there was a call on this connection and it's now
2717 * gone, then we'll be making a new call below.
2718 * If there was previously a call and it's now different then
2719 * the old call was freed and another thread running this routine
2720 * has created a call on this channel. One of these two threads
2721 * has a packet for the old call and the code below handles those
2725 MUTEX_ENTER(&call->lock);
2727 /* This packet can't be for this call. If the new call address is
2728 * 0 then no call is running on this channel. If there is a call
2729 * then, since this is a client connection we're getting data for
2730 * it must be for the previous call.
2732 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2733 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2738 currentCallNumber = conn->callNumber[channel];
2740 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2741 if (np->header.callNumber < currentCallNumber) {
2742 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2743 #ifdef RX_ENABLE_LOCKS
2745 MUTEX_EXIT(&call->lock);
2747 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2751 MUTEX_ENTER(&conn->conn_call_lock);
2752 call = rxi_NewCall(conn, channel);
2753 MUTEX_EXIT(&conn->conn_call_lock);
2754 *call->callNumber = np->header.callNumber;
2755 if (np->header.callNumber == 0)
2756 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));
2758 call->state = RX_STATE_PRECALL;
2759 clock_GetTime(&call->queueTime);
2760 hzero(call->bytesSent);
2761 hzero(call->bytesRcvd);
2763 * If the number of queued calls exceeds the overload
2764 * threshold then abort this call.
2766 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2767 struct rx_packet *tp;
2769 rxi_CallError(call, rx_BusyError);
2770 tp = rxi_SendCallAbort(call, np, 1, 0);
2771 MUTEX_EXIT(&call->lock);
2772 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2773 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2776 rxi_KeepAliveOn(call);
2777 } else if (np->header.callNumber != currentCallNumber) {
2778 /* Wait until the transmit queue is idle before deciding
2779 * whether to reset the current call. Chances are that the
2780 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2783 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2784 while ((call->state == RX_STATE_ACTIVE)
2785 && (call->flags & RX_CALL_TQ_BUSY)) {
2786 call->flags |= RX_CALL_TQ_WAIT;
2788 #ifdef RX_ENABLE_LOCKS
2789 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2790 CV_WAIT(&call->cv_tq, &call->lock);
2791 #else /* RX_ENABLE_LOCKS */
2792 osi_rxSleep(&call->tq);
2793 #endif /* RX_ENABLE_LOCKS */
2795 if (call->tqWaiters == 0)
2796 call->flags &= ~RX_CALL_TQ_WAIT;
2798 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2799 /* If the new call cannot be taken right now send a busy and set
2800 * the error condition in this call, so that it terminates as
2801 * quickly as possible */
2802 if (call->state == RX_STATE_ACTIVE) {
2803 struct rx_packet *tp;
2805 rxi_CallError(call, RX_CALL_DEAD);
2806 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2808 MUTEX_EXIT(&call->lock);
2809 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2812 rxi_ResetCall(call, 0);
2813 *call->callNumber = np->header.callNumber;
2814 if (np->header.callNumber == 0)
2815 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));
2817 call->state = RX_STATE_PRECALL;
2818 clock_GetTime(&call->queueTime);
2819 hzero(call->bytesSent);
2820 hzero(call->bytesRcvd);
2822 * If the number of queued calls exceeds the overload
2823 * threshold then abort this call.
2825 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2826 struct rx_packet *tp;
2828 rxi_CallError(call, rx_BusyError);
2829 tp = rxi_SendCallAbort(call, np, 1, 0);
2830 MUTEX_EXIT(&call->lock);
2831 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2832 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2835 rxi_KeepAliveOn(call);
2837 /* Continuing call; do nothing here. */
2839 } else { /* we're the client */
2840 /* Ignore all incoming acknowledgements for calls in DALLY state */
2841 if (call && (call->state == RX_STATE_DALLY)
2842 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2843 rx_MutexIncrement(rx_stats.ignorePacketDally, rx_stats_mutex);
2844 #ifdef RX_ENABLE_LOCKS
2846 MUTEX_EXIT(&call->lock);
2849 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2853 /* Ignore anything that's not relevant to the current call. If there
2854 * isn't a current call, then no packet is relevant. */
2855 if (!call || (np->header.callNumber != currentCallNumber)) {
2856 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2857 #ifdef RX_ENABLE_LOCKS
2859 MUTEX_EXIT(&call->lock);
2862 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2865 /* If the service security object index stamped in the packet does not
2866 * match the connection's security index, ignore the packet */
2867 if (np->header.securityIndex != conn->securityIndex) {
2868 #ifdef RX_ENABLE_LOCKS
2869 MUTEX_EXIT(&call->lock);
2871 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2875 /* If we're receiving the response, then all transmit packets are
2876 * implicitly acknowledged. Get rid of them. */
2877 if (np->header.type == RX_PACKET_TYPE_DATA) {
2878 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2879 /* XXX Hack. Because we must release the global rx lock when
2880 * sending packets (osi_NetSend) we drop all acks while we're
2881 * traversing the tq in rxi_Start sending packets out because
2882 * packets may move to the freePacketQueue as result of being here!
2883 * So we drop these packets until we're safely out of the
2884 * traversing. Really ugly!
2885 * For fine grain RX locking, we set the acked field in the
2886 * packets and let rxi_Start remove them from the transmit queue.
2888 if (call->flags & RX_CALL_TQ_BUSY) {
2889 #ifdef RX_ENABLE_LOCKS
2890 rxi_SetAcksInTransmitQueue(call);
2893 return np; /* xmitting; drop packet */
2896 rxi_ClearTransmitQueue(call, 0);
2898 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2899 rxi_ClearTransmitQueue(call, 0);
2900 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2902 if (np->header.type == RX_PACKET_TYPE_ACK) {
2903 /* now check to see if this is an ack packet acknowledging that the
2904 * server actually *lost* some hard-acked data. If this happens we
2905 * ignore this packet, as it may indicate that the server restarted in
2906 * the middle of a call. It is also possible that this is an old ack
2907 * packet. We don't abort the connection in this case, because this
2908 * *might* just be an old ack packet. The right way to detect a server
2909 * restart in the midst of a call is to notice that the server epoch
2911 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2912 * XXX unacknowledged. I think that this is off-by-one, but
2913 * XXX I don't dare change it just yet, since it will
2914 * XXX interact badly with the server-restart detection
2915 * XXX code in receiveackpacket. */
2916 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2917 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2918 MUTEX_EXIT(&call->lock);
2919 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2923 } /* else not a data packet */
2926 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2927 /* Set remote user defined status from packet */
2928 call->remoteStatus = np->header.userStatus;
2930 /* Note the gap between the expected next packet and the actual
2931 * packet that arrived, when the new packet has a smaller serial number
2932 * than expected. Rioses frequently reorder packets all by themselves,
2933 * so this will be quite important with very large window sizes.
2934 * Skew is checked against 0 here to avoid any dependence on the type of
2935 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2937 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2938 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2939 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2941 MUTEX_ENTER(&conn->conn_data_lock);
2942 skew = conn->lastSerial - np->header.serial;
2943 conn->lastSerial = np->header.serial;
2944 MUTEX_EXIT(&conn->conn_data_lock);
2946 register struct rx_peer *peer;
2948 if (skew > peer->inPacketSkew) {
2949 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2951 peer->inPacketSkew = skew;
2955 /* Now do packet type-specific processing */
2956 switch (np->header.type) {
2957 case RX_PACKET_TYPE_DATA:
2958 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2961 case RX_PACKET_TYPE_ACK:
2962 /* Respond immediately to ack packets requesting acknowledgement
2964 if (np->header.flags & RX_REQUEST_ACK) {
2966 (void)rxi_SendCallAbort(call, 0, 1, 0);
2968 (void)rxi_SendAck(call, 0, np->header.serial,
2969 RX_ACK_PING_RESPONSE, 1);
2971 np = rxi_ReceiveAckPacket(call, np, 1);
2973 case RX_PACKET_TYPE_ABORT: {
2974 /* An abort packet: reset the call, passing the error up to the user. */
2975 /* What if error is zero? */
2976 /* What if the error is -1? the application will treat it as a timeout. */
2977 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
2978 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
2979 rxi_CallError(call, errdata);
2980 MUTEX_EXIT(&call->lock);
2981 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2982 return np; /* xmitting; drop packet */
2984 case RX_PACKET_TYPE_BUSY:
2987 case RX_PACKET_TYPE_ACKALL:
2988 /* All packets acknowledged, so we can drop all packets previously
2989 * readied for sending */
2990 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2991 /* XXX Hack. We because we can't release the global rx lock when
2992 * sending packets (osi_NetSend) we drop all ack pkts while we're
2993 * traversing the tq in rxi_Start sending packets out because
2994 * packets may move to the freePacketQueue as result of being
2995 * here! So we drop these packets until we're safely out of the
2996 * traversing. Really ugly!
2997 * For fine grain RX locking, we set the acked field in the packets
2998 * and let rxi_Start remove the packets from the transmit queue.
3000 if (call->flags & RX_CALL_TQ_BUSY) {
3001 #ifdef RX_ENABLE_LOCKS
3002 rxi_SetAcksInTransmitQueue(call);
3004 #else /* RX_ENABLE_LOCKS */
3005 MUTEX_EXIT(&call->lock);
3006 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
3007 return np; /* xmitting; drop packet */
3008 #endif /* RX_ENABLE_LOCKS */
3010 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3011 rxi_ClearTransmitQueue(call, 0);
3012 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
3015 /* Should not reach here, unless the peer is broken: send an abort
3017 rxi_CallError(call, RX_PROTOCOL_ERROR);
3018 np = rxi_SendCallAbort(call, np, 1, 0);
3021 /* Note when this last legitimate packet was received, for keep-alive
3022 * processing. Note, we delay getting the time until now in the hope that
3023 * the packet will be delivered to the user before any get time is required
3024 * (if not, then the time won't actually be re-evaluated here). */
3025 call->lastReceiveTime = clock_Sec();
3026 MUTEX_EXIT(&call->lock);
3027 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
3031 /* return true if this is an "interesting" connection from the point of view
3032 of someone trying to debug the system */
3034 rxi_IsConnInteresting(struct rx_connection *aconn)
3037 register struct rx_call *tcall;
3039 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3041 for (i = 0; i < RX_MAXCALLS; i++) {
3042 tcall = aconn->call[i];
3044 if ((tcall->state == RX_STATE_PRECALL)
3045 || (tcall->state == RX_STATE_ACTIVE))
3047 if ((tcall->mode == RX_MODE_SENDING)
3048 || (tcall->mode == RX_MODE_RECEIVING))
3056 /* if this is one of the last few packets AND it wouldn't be used by the
3057 receiving call to immediately satisfy a read request, then drop it on
3058 the floor, since accepting it might prevent a lock-holding thread from
3059 making progress in its reading. If a call has been cleared while in
3060 the precall state then ignore all subsequent packets until the call
3061 is assigned to a thread. */
3064 TooLow(struct rx_packet *ap, struct rx_call *acall)
3067 MUTEX_ENTER(&rx_stats_mutex);
3068 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3069 && (acall->state == RX_STATE_PRECALL))
3070 || ((rx_nFreePackets < rxi_dataQuota + 2)
3071 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3072 && (acall->flags & RX_CALL_READER_WAIT)))) {
3075 MUTEX_EXIT(&rx_stats_mutex);
3081 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
3083 struct rx_connection *conn = arg1;
3084 struct rx_call *acall = arg2;
3085 struct rx_call *call = acall;
3086 struct clock when, now;
3089 MUTEX_ENTER(&conn->conn_data_lock);
3090 conn->checkReachEvent = NULL;
3091 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3094 MUTEX_EXIT(&conn->conn_data_lock);
3098 MUTEX_ENTER(&conn->conn_call_lock);
3099 MUTEX_ENTER(&conn->conn_data_lock);
3100 for (i = 0; i < RX_MAXCALLS; i++) {
3101 struct rx_call *tc = conn->call[i];
3102 if (tc && tc->state == RX_STATE_PRECALL) {
3108 /* Indicate that rxi_CheckReachEvent is no longer running by
3109 * clearing the flag. Must be atomic under conn_data_lock to
3110 * avoid a new call slipping by: rxi_CheckConnReach holds
3111 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3113 conn->flags &= ~RX_CONN_ATTACHWAIT;
3114 MUTEX_EXIT(&conn->conn_data_lock);
3115 MUTEX_EXIT(&conn->conn_call_lock);
3120 MUTEX_ENTER(&call->lock);
3121 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3123 MUTEX_EXIT(&call->lock);
3125 clock_GetTime(&now);
3127 when.sec += RX_CHECKREACH_TIMEOUT;
3128 MUTEX_ENTER(&conn->conn_data_lock);
3129 if (!conn->checkReachEvent) {
3131 conn->checkReachEvent =
3132 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3135 MUTEX_EXIT(&conn->conn_data_lock);
3141 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3143 struct rx_service *service = conn->service;
3144 struct rx_peer *peer = conn->peer;
3145 afs_uint32 now, lastReach;
3147 if (service->checkReach == 0)
3151 MUTEX_ENTER(&peer->peer_lock);
3152 lastReach = peer->lastReachTime;
3153 MUTEX_EXIT(&peer->peer_lock);
3154 if (now - lastReach < RX_CHECKREACH_TTL)
3157 MUTEX_ENTER(&conn->conn_data_lock);
3158 if (conn->flags & RX_CONN_ATTACHWAIT) {
3159 MUTEX_EXIT(&conn->conn_data_lock);
3162 conn->flags |= RX_CONN_ATTACHWAIT;
3163 MUTEX_EXIT(&conn->conn_data_lock);
3164 if (!conn->checkReachEvent)
3165 rxi_CheckReachEvent(NULL, conn, call);
3170 /* try to attach call, if authentication is complete */
3172 TryAttach(register struct rx_call *acall, register osi_socket socket,
3173 register int *tnop, register struct rx_call **newcallp,
3176 struct rx_connection *conn = acall->conn;
3178 if (conn->type == RX_SERVER_CONNECTION
3179 && acall->state == RX_STATE_PRECALL) {
3180 /* Don't attach until we have any req'd. authentication. */
3181 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3182 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3183 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3184 /* Note: this does not necessarily succeed; there
3185 * may not any proc available
3188 rxi_ChallengeOn(acall->conn);
3193 /* A data packet has been received off the interface. This packet is
3194 * appropriate to the call (the call is in the right state, etc.). This
3195 * routine can return a packet to the caller, for re-use */
3198 rxi_ReceiveDataPacket(register struct rx_call *call,
3199 register struct rx_packet *np, int istack,
3200 osi_socket socket, afs_uint32 host, u_short port,
3201 int *tnop, struct rx_call **newcallp)
3203 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3207 afs_uint32 seq, serial, flags;
3209 struct rx_packet *tnp;
3210 struct clock when, now;
3211 rx_MutexIncrement(rx_stats.dataPacketsRead, rx_stats_mutex);
3214 /* If there are no packet buffers, drop this new packet, unless we can find
3215 * packet buffers from inactive calls */
3217 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3218 MUTEX_ENTER(&rx_freePktQ_lock);
3219 rxi_NeedMorePackets = TRUE;
3220 MUTEX_EXIT(&rx_freePktQ_lock);
3221 rx_MutexIncrement(rx_stats.noPacketBuffersOnRead, rx_stats_mutex);
3222 call->rprev = np->header.serial;
3223 rxi_calltrace(RX_TRACE_DROP, call);
3224 dpf(("packet %x dropped on receipt - quota problems", np));
3226 rxi_ClearReceiveQueue(call);
3227 clock_GetTime(&now);
3229 clock_Add(&when, &rx_softAckDelay);
3230 if (!call->delayedAckEvent
3231 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3232 rxevent_Cancel(call->delayedAckEvent, call,
3233 RX_CALL_REFCOUNT_DELAY);
3234 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3235 call->delayedAckEvent =
3236 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3238 /* we've damaged this call already, might as well do it in. */
3244 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3245 * packet is one of several packets transmitted as a single
3246 * datagram. Do not send any soft or hard acks until all packets
3247 * in a jumbogram have been processed. Send negative acks right away.
3249 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3250 /* tnp is non-null when there are more packets in the
3251 * current jumbo gram */
3258 seq = np->header.seq;
3259 serial = np->header.serial;
3260 flags = np->header.flags;
3262 /* If the call is in an error state, send an abort message */
3264 return rxi_SendCallAbort(call, np, istack, 0);
3266 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3267 * AFS 3.5 jumbogram. */
3268 if (flags & RX_JUMBO_PACKET) {
3269 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3274 if (np->header.spare != 0) {
3275 MUTEX_ENTER(&call->conn->conn_data_lock);
3276 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3277 MUTEX_EXIT(&call->conn->conn_data_lock);
3280 /* The usual case is that this is the expected next packet */
3281 if (seq == call->rnext) {
3283 /* Check to make sure it is not a duplicate of one already queued */
3284 if (queue_IsNotEmpty(&call->rq)
3285 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3286 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3287 dpf(("packet %x dropped on receipt - duplicate", np));
3288 rxevent_Cancel(call->delayedAckEvent, call,
3289 RX_CALL_REFCOUNT_DELAY);
3290 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3296 /* It's the next packet. Stick it on the receive queue
3297 * for this call. Set newPackets to make sure we wake
3298 * the reader once all packets have been processed */
3299 np->flags |= RX_PKTFLAG_RQ;
3300 queue_Prepend(&call->rq, np);
3302 np = NULL; /* We can't use this anymore */
3305 /* If an ack is requested then set a flag to make sure we
3306 * send an acknowledgement for this packet */
3307 if (flags & RX_REQUEST_ACK) {
3308 ackNeeded = RX_ACK_REQUESTED;
3311 /* Keep track of whether we have received the last packet */
3312 if (flags & RX_LAST_PACKET) {
3313 call->flags |= RX_CALL_HAVE_LAST;
3317 /* Check whether we have all of the packets for this call */
3318 if (call->flags & RX_CALL_HAVE_LAST) {
3319 afs_uint32 tseq; /* temporary sequence number */
3320 struct rx_packet *tp; /* Temporary packet pointer */
3321 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3323 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3324 if (tseq != tp->header.seq)
3326 if (tp->header.flags & RX_LAST_PACKET) {
3327 call->flags |= RX_CALL_RECEIVE_DONE;
3334 /* Provide asynchronous notification for those who want it
3335 * (e.g. multi rx) */
3336 if (call->arrivalProc) {
3337 (*call->arrivalProc) (call, call->arrivalProcHandle,
3338 call->arrivalProcArg);
3339 call->arrivalProc = (void (*)())0;
3342 /* Update last packet received */
3345 /* If there is no server process serving this call, grab
3346 * one, if available. We only need to do this once. If a
3347 * server thread is available, this thread becomes a server
3348 * thread and the server thread becomes a listener thread. */
3350 TryAttach(call, socket, tnop, newcallp, 0);
3353 /* This is not the expected next packet. */
3355 /* Determine whether this is a new or old packet, and if it's
3356 * a new one, whether it fits into the current receive window.
3357 * Also figure out whether the packet was delivered in sequence.
3358 * We use the prev variable to determine whether the new packet
3359 * is the successor of its immediate predecessor in the
3360 * receive queue, and the missing flag to determine whether
3361 * any of this packets predecessors are missing. */
3363 afs_uint32 prev; /* "Previous packet" sequence number */
3364 struct rx_packet *tp; /* Temporary packet pointer */
3365 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3366 int missing; /* Are any predecessors missing? */
3368 /* If the new packet's sequence number has been sent to the
3369 * application already, then this is a duplicate */
3370 if (seq < call->rnext) {
3371 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3372 rxevent_Cancel(call->delayedAckEvent, call,
3373 RX_CALL_REFCOUNT_DELAY);
3374 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3380 /* If the sequence number is greater than what can be
3381 * accomodated by the current window, then send a negative
3382 * acknowledge and drop the packet */
3383 if ((call->rnext + call->rwind) <= seq) {
3384 rxevent_Cancel(call->delayedAckEvent, call,
3385 RX_CALL_REFCOUNT_DELAY);
3386 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3393 /* Look for the packet in the queue of old received packets */
3394 for (prev = call->rnext - 1, missing =
3395 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3396 /*Check for duplicate packet */
3397 if (seq == tp->header.seq) {
3398 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3399 rxevent_Cancel(call->delayedAckEvent, call,
3400 RX_CALL_REFCOUNT_DELAY);
3401 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3407 /* If we find a higher sequence packet, break out and
3408 * insert the new packet here. */
3409 if (seq < tp->header.seq)
3411 /* Check for missing packet */
3412 if (tp->header.seq != prev + 1) {
3416 prev = tp->header.seq;
3419 /* Keep track of whether we have received the last packet. */
3420 if (flags & RX_LAST_PACKET) {
3421 call->flags |= RX_CALL_HAVE_LAST;
3424 /* It's within the window: add it to the the receive queue.
3425 * tp is left by the previous loop either pointing at the
3426 * packet before which to insert the new packet, or at the
3427 * queue head if the queue is empty or the packet should be
3429 queue_InsertBefore(tp, np);
3433 /* Check whether we have all of the packets for this call */
3434 if ((call->flags & RX_CALL_HAVE_LAST)
3435 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3436 afs_uint32 tseq; /* temporary sequence number */
3439 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3440 if (tseq != tp->header.seq)
3442 if (tp->header.flags & RX_LAST_PACKET) {
3443 call->flags |= RX_CALL_RECEIVE_DONE;
3450 /* We need to send an ack of the packet is out of sequence,
3451 * or if an ack was requested by the peer. */
3452 if (seq != prev + 1 || missing) {
3453 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3454 } else if (flags & RX_REQUEST_ACK) {
3455 ackNeeded = RX_ACK_REQUESTED;
3458 /* Acknowledge the last packet for each call */
3459 if (flags & RX_LAST_PACKET) {
3470 * If the receiver is waiting for an iovec, fill the iovec
3471 * using the data from the receive queue */
3472 if (call->flags & RX_CALL_IOVEC_WAIT) {
3473 didHardAck = rxi_FillReadVec(call, serial);
3474 /* the call may have been aborted */
3483 /* Wakeup the reader if any */
3484 if ((call->flags & RX_CALL_READER_WAIT)
3485 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3486 || (call->iovNext >= call->iovMax)
3487 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3488 call->flags &= ~RX_CALL_READER_WAIT;
3489 #ifdef RX_ENABLE_LOCKS
3490 CV_BROADCAST(&call->cv_rq);
3492 osi_rxWakeup(&call->rq);
3498 * Send an ack when requested by the peer, or once every
3499 * rxi_SoftAckRate packets until the last packet has been
3500 * received. Always send a soft ack for the last packet in
3501 * the server's reply. */
3503 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3504 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3505 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3506 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3507 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3508 } else if (call->nSoftAcks) {
3509 clock_GetTime(&now);
3511 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3512 clock_Add(&when, &rx_lastAckDelay);
3514 clock_Add(&when, &rx_softAckDelay);
3516 if (!call->delayedAckEvent
3517 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3518 rxevent_Cancel(call->delayedAckEvent, call,
3519 RX_CALL_REFCOUNT_DELAY);
3520 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3521 call->delayedAckEvent =
3522 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3524 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3525 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3532 static void rxi_ComputeRate();
3536 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3538 struct rx_peer *peer = conn->peer;
3540 MUTEX_ENTER(&peer->peer_lock);
3541 peer->lastReachTime = clock_Sec();
3542 MUTEX_EXIT(&peer->peer_lock);
3544 MUTEX_ENTER(&conn->conn_data_lock);
3545 if (conn->flags & RX_CONN_ATTACHWAIT) {
3548 conn->flags &= ~RX_CONN_ATTACHWAIT;
3549 MUTEX_EXIT(&conn->conn_data_lock);
3551 for (i = 0; i < RX_MAXCALLS; i++) {
3552 struct rx_call *call = conn->call[i];
3555 MUTEX_ENTER(&call->lock);
3556 /* tnop can be null if newcallp is null */
3557 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3559 MUTEX_EXIT(&call->lock);
3563 MUTEX_EXIT(&conn->conn_data_lock);
3567 rx_ack_reason(int reason)
3570 case RX_ACK_REQUESTED:
3572 case RX_ACK_DUPLICATE:
3574 case RX_ACK_OUT_OF_SEQUENCE:
3576 case RX_ACK_EXCEEDS_WINDOW:
3578 case RX_ACK_NOSPACE:
3582 case RX_ACK_PING_RESPONSE:
3594 /* rxi_ComputePeerNetStats
3596 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3597 * estimates (like RTT and throughput) based on ack packets. Caller
3598 * must ensure that the packet in question is the right one (i.e.
3599 * serial number matches).
3602 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3603 struct rx_ackPacket *ap, struct rx_packet *np)
3605 struct rx_peer *peer = call->conn->peer;
3607 /* Use RTT if not delayed by client. */
3608 if (ap->reason != RX_ACK_DELAY)
3609 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3611 rxi_ComputeRate(peer, call, p, np, ap->reason);
3615 /* The real smarts of the whole thing. */
3617 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3620 struct rx_ackPacket *ap;
3622 register struct rx_packet *tp;
3623 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3624 register struct rx_connection *conn = call->conn;
3625 struct rx_peer *peer = conn->peer;
3628 /* because there are CM's that are bogus, sending weird values for this. */
3629 afs_uint32 skew = 0;
3634 int newAckCount = 0;
3635 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3636 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3638 rx_MutexIncrement(rx_stats.ackPacketsRead, rx_stats_mutex);
3639 ap = (struct rx_ackPacket *)rx_DataOf(np);
3640 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3642 return np; /* truncated ack packet */
3644 /* depends on ack packet struct */
3645 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3646 first = ntohl(ap->firstPacket);
3647 serial = ntohl(ap->serial);
3648 /* temporarily disabled -- needs to degrade over time
3649 * skew = ntohs(ap->maxSkew); */
3651 /* Ignore ack packets received out of order */
3652 if (first < call->tfirst) {
3656 if (np->header.flags & RX_SLOW_START_OK) {
3657 call->flags |= RX_CALL_SLOW_START_OK;
3660 if (ap->reason == RX_ACK_PING_RESPONSE)
3661 rxi_UpdatePeerReach(conn, call);
3665 if (rxdebug_active) {
3669 len = _snprintf(msg, sizeof(msg),
3670 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3671 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3672 ntohl(ap->serial), ntohl(ap->previousPacket),
3673 (unsigned int)np->header.seq, (unsigned int)skew,
3674 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3678 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3679 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3683 OutputDebugString(msg);
3685 #else /* AFS_NT40_ENV */
3688 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3689 ap->reason, ntohl(ap->previousPacket),
3690 (unsigned int)np->header.seq, (unsigned int)serial,
3691 (unsigned int)skew, ntohl(ap->firstPacket));
3694 for (offset = 0; offset < nAcks; offset++)
3695 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3700 #endif /* AFS_NT40_ENV */
3703 /* Update the outgoing packet skew value to the latest value of
3704 * the peer's incoming packet skew value. The ack packet, of
3705 * course, could arrive out of order, but that won't affect things
3707 MUTEX_ENTER(&peer->peer_lock);
3708 peer->outPacketSkew = skew;
3710 /* Check for packets that no longer need to be transmitted, and
3711 * discard them. This only applies to packets positively
3712 * acknowledged as having been sent to the peer's upper level.
3713 * All other packets must be retained. So only packets with
3714 * sequence numbers < ap->firstPacket are candidates. */
3715 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3716 if (tp->header.seq >= first)
3718 call->tfirst = tp->header.seq + 1;
3720 && (tp->header.serial == serial || tp->firstSerial == serial))
3721 rxi_ComputePeerNetStats(call, tp, ap, np);
3722 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3725 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3726 /* XXX Hack. Because we have to release the global rx lock when sending
3727 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3728 * in rxi_Start sending packets out because packets may move to the
3729 * freePacketQueue as result of being here! So we drop these packets until
3730 * we're safely out of the traversing. Really ugly!
3731 * To make it even uglier, if we're using fine grain locking, we can
3732 * set the ack bits in the packets and have rxi_Start remove the packets
3733 * when it's done transmitting.
3735 if (call->flags & RX_CALL_TQ_BUSY) {
3736 #ifdef RX_ENABLE_LOCKS
3737 tp->flags |= RX_PKTFLAG_ACKED;
3738 call->flags |= RX_CALL_TQ_SOME_ACKED;
3739 #else /* RX_ENABLE_LOCKS */
3741 #endif /* RX_ENABLE_LOCKS */
3743 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3746 tp->flags &= ~RX_PKTFLAG_TQ;
3747 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3752 /* Give rate detector a chance to respond to ping requests */
3753 if (ap->reason == RX_ACK_PING_RESPONSE) {
3754 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3758 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3760 /* Now go through explicit acks/nacks and record the results in
3761 * the waiting packets. These are packets that can't be released
3762 * yet, even with a positive acknowledge. This positive
3763 * acknowledge only means the packet has been received by the
3764 * peer, not that it will be retained long enough to be sent to
3765 * the peer's upper level. In addition, reset the transmit timers
3766 * of any missing packets (those packets that must be missing
3767 * because this packet was out of sequence) */
3769 call->nSoftAcked = 0;
3770 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3771 /* Update round trip time if the ack was stimulated on receipt
3773 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3774 #ifdef RX_ENABLE_LOCKS
3775 if (tp->header.seq >= first)
3776 #endif /* RX_ENABLE_LOCKS */
3777 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3779 && (tp->header.serial == serial || tp->firstSerial == serial))
3780 rxi_ComputePeerNetStats(call, tp, ap, np);
3782 /* Set the acknowledge flag per packet based on the
3783 * information in the ack packet. An acknowlegded packet can
3784 * be downgraded when the server has discarded a packet it
3785 * soacked previously, or when an ack packet is received
3786 * out of sequence. */
3787 if (tp->header.seq < first) {
3788 /* Implicit ack information */
3789 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3792 tp->flags |= RX_PKTFLAG_ACKED;
3793 } else if (tp->header.seq < first + nAcks) {
3794 /* Explicit ack information: set it in the packet appropriately */
3795 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3796 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3798 tp->flags |= RX_PKTFLAG_ACKED;
3805 } else /* RX_ACK_TYPE_NACK */ {
3806 tp->flags &= ~RX_PKTFLAG_ACKED;
3810 tp->flags &= ~RX_PKTFLAG_ACKED;
3814 /* If packet isn't yet acked, and it has been transmitted at least
3815 * once, reset retransmit time using latest timeout
3816 * ie, this should readjust the retransmit timer for all outstanding
3817 * packets... So we don't just retransmit when we should know better*/
3819 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3820 tp->retryTime = tp->timeSent;
3821 clock_Add(&tp->retryTime, &peer->timeout);
3822 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3823 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3827 /* If the window has been extended by this acknowledge packet,
3828 * then wakeup a sender waiting in alloc for window space, or try
3829 * sending packets now, if he's been sitting on packets due to
3830 * lack of window space */
3831 if (call->tnext < (call->tfirst + call->twind)) {
3832 #ifdef RX_ENABLE_LOCKS
3833 CV_SIGNAL(&call->cv_twind);
3835 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3836 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3837 osi_rxWakeup(&call->twind);
3840 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3841 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3845 /* if the ack packet has a receivelen field hanging off it,
3846 * update our state */
3847 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3850 /* If the ack packet has a "recommended" size that is less than
3851 * what I am using now, reduce my size to match */
3852 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3853 (int)sizeof(afs_int32), &tSize);
3854 tSize = (afs_uint32) ntohl(tSize);
3855 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3857 /* Get the maximum packet size to send to this peer */
3858 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
3860 tSize = (afs_uint32) ntohl(tSize);
3861 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3862 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3864 /* sanity check - peer might have restarted with different params.
3865 * If peer says "send less", dammit, send less... Peer should never
3866 * be unable to accept packets of the size that prior AFS versions would
3867 * send without asking. */
3868 if (peer->maxMTU != tSize) {
3869 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
3871 peer->maxMTU = tSize;
3872 peer->MTU = MIN(tSize, peer->MTU);
3873 call->MTU = MIN(call->MTU, tSize);
3876 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3879 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3880 (int)sizeof(afs_int32), &tSize);
3881 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3882 if (tSize < call->twind) { /* smaller than our send */
3883 call->twind = tSize; /* window, we must send less... */
3884 call->ssthresh = MIN(call->twind, call->ssthresh);
3885 call->conn->twind[call->channel] = call->twind;
3888 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3889 * network MTU confused with the loopback MTU. Calculate the
3890 * maximum MTU here for use in the slow start code below.
3892 maxMTU = peer->maxMTU;
3893 /* Did peer restart with older RX version? */
3894 if (peer->maxDgramPackets > 1) {
3895 peer->maxDgramPackets = 1;
3897 } else if (np->length >=
3898 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3901 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3902 sizeof(afs_int32), &tSize);
3903 tSize = (afs_uint32) ntohl(tSize);
3905 * As of AFS 3.5 we set the send window to match the receive window.
3907 if (tSize < call->twind) {
3908 call->twind = tSize;
3909 call->conn->twind[call->channel] = call->twind;
3910 call->ssthresh = MIN(call->twind, call->ssthresh);
3911 } else if (tSize > call->twind) {
3912 call->twind = tSize;
3913 call->conn->twind[call->channel] = call->twind;
3917 * As of AFS 3.5, a jumbogram is more than one fixed size
3918 * packet transmitted in a single UDP datagram. If the remote
3919 * MTU is smaller than our local MTU then never send a datagram
3920 * larger than the natural MTU.
3923 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3924 sizeof(afs_int32), &tSize);
3925 maxDgramPackets = (afs_uint32) ntohl(tSize);
3926 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3927 maxDgramPackets = MIN(maxDgramPackets, peer->ifDgramPackets);
3928 if (peer->natMTU < peer->ifMTU)
3929 maxDgramPackets = MIN(maxDgramPackets, rxi_AdjustDgramPackets(1, peer->natMTU));
3930 if (maxDgramPackets > 1) {
3931 peer->maxDgramPackets = maxDgramPackets;
3932 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3934 peer->maxDgramPackets = 1;
3935 call->MTU = peer->natMTU;
3937 } else if (peer->maxDgramPackets > 1) {
3938 /* Restarted with lower version of RX */
3939 peer->maxDgramPackets = 1;
3941 } else if (peer->maxDgramPackets > 1
3942 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3943 /* Restarted with lower version of RX */
3944 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3945 peer->natMTU = OLD_MAX_PACKET_SIZE;
3946 peer->MTU = OLD_MAX_PACKET_SIZE;
3947 peer->maxDgramPackets = 1;
3948 peer->nDgramPackets = 1;
3950 call->MTU = OLD_MAX_PACKET_SIZE;
3955 * Calculate how many datagrams were successfully received after
3956 * the first missing packet and adjust the negative ack counter
3961 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3962 if (call->nNacks < nNacked) {
3963 call->nNacks = nNacked;
3966 call->nAcks += newAckCount;
3970 if (call->flags & RX_CALL_FAST_RECOVER) {
3972 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
3974 call->flags &= ~RX_CALL_FAST_RECOVER;
3975 call->cwind = call->nextCwind;
3976 call->nextCwind = 0;
3979 call->nCwindAcks = 0;
3980 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
3981 /* Three negative acks in a row trigger congestion recovery */
3982 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3983 MUTEX_EXIT(&peer->peer_lock);
3984 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
3985 /* someone else is waiting to start recovery */
3988 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
3989 rxi_WaitforTQBusy(call);
3990 MUTEX_ENTER(&peer->peer_lock);
3991 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3992 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
3993 call->flags |= RX_CALL_FAST_RECOVER;
3994 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
3996 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
3997 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
3998 call->nextCwind = call->ssthresh;
4001 peer->MTU = call->MTU;
4002 peer->cwind = call->nextCwind;
4003 peer->nDgramPackets = call->nDgramPackets;
4005 call->congestSeq = peer->congestSeq;
4006 /* Reset the resend times on the packets that were nacked
4007 * so we will retransmit as soon as the window permits*/
4008 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4010 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4011 clock_Zero(&tp->retryTime);
4013 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4018 /* If cwind is smaller than ssthresh, then increase
4019 * the window one packet for each ack we receive (exponential
4021 * If cwind is greater than or equal to ssthresh then increase
4022 * the congestion window by one packet for each cwind acks we
4023 * receive (linear growth). */
4024 if (call->cwind < call->ssthresh) {
4026 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4027 call->nCwindAcks = 0;
4029 call->nCwindAcks += newAckCount;
4030 if (call->nCwindAcks >= call->cwind) {
4031 call->nCwindAcks = 0;
4032 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4036 * If we have received several acknowledgements in a row then
4037 * it is time to increase the size of our datagrams
4039 if ((int)call->nAcks > rx_nDgramThreshold) {
4040 if (peer->maxDgramPackets > 1) {
4041 if (call->nDgramPackets < peer->maxDgramPackets) {
4042 call->nDgramPackets++;
4044 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4045 } else if (call->MTU < peer->maxMTU) {
4046 call->MTU += peer->natMTU;
4047 call->MTU = MIN(call->MTU, peer->maxMTU);
4053 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4055 /* Servers need to hold the call until all response packets have
4056 * been acknowledged. Soft acks are good enough since clients
4057 * are not allowed to clear their receive queues. */
4058 if (call->state == RX_STATE_HOLD
4059 && call->tfirst + call->nSoftAcked >= call->tnext) {
4060 call->state = RX_STATE_DALLY;
4061 rxi_ClearTransmitQueue(call, 0);
4062 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4063 } else if (!queue_IsEmpty(&call->tq)) {
4064 rxi_Start(0, call, 0, istack);
4069 /* Received a response to a challenge packet */
4071 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
4072 register struct rx_packet *np, int istack)
4076 /* Ignore the packet if we're the client */
4077 if (conn->type == RX_CLIENT_CONNECTION)
4080 /* If already authenticated, ignore the packet (it's probably a retry) */
4081 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4084 /* Otherwise, have the security object evaluate the response packet */
4085 error = RXS_CheckResponse(conn->securityObject, conn, np);
4087 /* If the response is invalid, reset the connection, sending
4088 * an abort to the peer */
4092 rxi_ConnectionError(conn, error);
4093 MUTEX_ENTER(&conn->conn_data_lock);
4094 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4095 MUTEX_EXIT(&conn->conn_data_lock);
4098 /* If the response is valid, any calls waiting to attach
4099 * servers can now do so */
4102 for (i = 0; i < RX_MAXCALLS; i++) {
4103 struct rx_call *call = conn->call[i];
4105 MUTEX_ENTER(&call->lock);
4106 if (call->state == RX_STATE_PRECALL)
4107 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4108 /* tnop can be null if newcallp is null */
4109 MUTEX_EXIT(&call->lock);
4113 /* Update the peer reachability information, just in case
4114 * some calls went into attach-wait while we were waiting
4115 * for authentication..
4117 rxi_UpdatePeerReach(conn, NULL);
4122 /* A client has received an authentication challenge: the security
4123 * object is asked to cough up a respectable response packet to send
4124 * back to the server. The server is responsible for retrying the
4125 * challenge if it fails to get a response. */
4128 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
4129 register struct rx_packet *np, int istack)
4133 /* Ignore the challenge if we're the server */
4134 if (conn->type == RX_SERVER_CONNECTION)
4137 /* Ignore the challenge if the connection is otherwise idle; someone's
4138 * trying to use us as an oracle. */
4139 if (!rxi_HasActiveCalls(conn))
4142 /* Send the security object the challenge packet. It is expected to fill
4143 * in the response. */
4144 error = RXS_GetResponse(conn->securityObject, conn, np);
4146 /* If the security object is unable to return a valid response, reset the
4147 * connection and send an abort to the peer. Otherwise send the response
4148 * packet to the peer connection. */
4150 rxi_ConnectionError(conn, error);
4151 MUTEX_ENTER(&conn->conn_data_lock);
4152 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4153 MUTEX_EXIT(&conn->conn_data_lock);
4155 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4156 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4162 /* Find an available server process to service the current request in
4163 * the given call structure. If one isn't available, queue up this
4164 * call so it eventually gets one */
4166 rxi_AttachServerProc(register struct rx_call *call,
4167 register osi_socket socket, register int *tnop,
4168 register struct rx_call **newcallp)
4170 register struct rx_serverQueueEntry *sq;
4171 register struct rx_service *service = call->conn->service;
4172 register int haveQuota = 0;
4174 /* May already be attached */
4175 if (call->state == RX_STATE_ACTIVE)
4178 MUTEX_ENTER(&rx_serverPool_lock);
4180 haveQuota = QuotaOK(service);
4181 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4182 /* If there are no processes available to service this call,
4183 * put the call on the incoming call queue (unless it's
4184 * already on the queue).
4186 #ifdef RX_ENABLE_LOCKS
4188 ReturnToServerPool(service);
4189 #endif /* RX_ENABLE_LOCKS */
4191 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4192 call->flags |= RX_CALL_WAIT_PROC;
4193 MUTEX_ENTER(&rx_stats_mutex);
4196 MUTEX_EXIT(&rx_stats_mutex);
4197 rxi_calltrace(RX_CALL_ARRIVAL, call);
4198 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4199 queue_Append(&rx_incomingCallQueue, call);
4202 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4204 /* If hot threads are enabled, and both newcallp and sq->socketp
4205 * are non-null, then this thread will process the call, and the
4206 * idle server thread will start listening on this threads socket.
4209 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4212 *sq->socketp = socket;
4213 clock_GetTime(&call->startTime);
4214 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4218 if (call->flags & RX_CALL_WAIT_PROC) {
4219 /* Conservative: I don't think this should happen */
4220 call->flags &= ~RX_CALL_WAIT_PROC;
4221 if (queue_IsOnQueue(call)) {
4223 rx_MutexDecrement(rx_nWaiting, rx_stats_mutex);
4226 call->state = RX_STATE_ACTIVE;
4227 call->mode = RX_MODE_RECEIVING;
4228 #ifdef RX_KERNEL_TRACE
4230 int glockOwner = ISAFS_GLOCK();
4233 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4234 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4240 if (call->flags & RX_CALL_CLEARED) {
4241 /* send an ack now to start the packet flow up again */
4242 call->flags &= ~RX_CALL_CLEARED;
4243 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4245 #ifdef RX_ENABLE_LOCKS
4248 service->nRequestsRunning++;
4249 if (service->nRequestsRunning <= service->minProcs)
4255 MUTEX_EXIT(&rx_serverPool_lock);
4258 /* Delay the sending of an acknowledge event for a short while, while
4259 * a new call is being prepared (in the case of a client) or a reply
4260 * is being prepared (in the case of a server). Rather than sending
4261 * an ack packet, an ACKALL packet is sent. */
4263 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4265 #ifdef RX_ENABLE_LOCKS
4267 MUTEX_ENTER(&call->lock);
4268 call->delayedAckEvent = NULL;
4269 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4271 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4272 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4274 MUTEX_EXIT(&call->lock);
4275 #else /* RX_ENABLE_LOCKS */
4277 call->delayedAckEvent = NULL;
4278 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4279 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4280 #endif /* RX_ENABLE_LOCKS */
4284 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused)
4286 struct rx_call *call = arg1;
4287 #ifdef RX_ENABLE_LOCKS
4289 MUTEX_ENTER(&call->lock);
4290 if (event == call->delayedAckEvent)
4291 call->delayedAckEvent = NULL;
4292 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4294 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4296 MUTEX_EXIT(&call->lock);
4297 #else /* RX_ENABLE_LOCKS */
4299 call->delayedAckEvent = NULL;
4300 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4301 #endif /* RX_ENABLE_LOCKS */
4305 #ifdef RX_ENABLE_LOCKS
4306 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4307 * clearing them out.
4310 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4312 register struct rx_packet *p, *tp;
4315 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4316 p->flags |= RX_PKTFLAG_ACKED;
4320 call->flags |= RX_CALL_TQ_CLEARME;
4321 call->flags |= RX_CALL_TQ_SOME_ACKED;
4324 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4325 call->tfirst = call->tnext;
4326 call->nSoftAcked = 0;
4328 if (call->flags & RX_CALL_FAST_RECOVER) {
4329 call->flags &= ~RX_CALL_FAST_RECOVER;
4330 call->cwind = call->nextCwind;
4331 call->nextCwind = 0;
4334 CV_SIGNAL(&call->cv_twind);
4336 #endif /* RX_ENABLE_LOCKS */
4338 /* Clear out the transmit queue for the current call (all packets have
4339 * been received by peer) */
4341 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4343 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4344 register struct rx_packet *p, *tp;
4346 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4348 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4349 p->flags |= RX_PKTFLAG_ACKED;
4353 call->flags |= RX_CALL_TQ_CLEARME;
4354 call->flags |= RX_CALL_TQ_SOME_ACKED;
4357 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4358 rxi_FreePackets(0, &call->tq);
4359 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4360 call->flags &= ~RX_CALL_TQ_CLEARME;
4362 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4364 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4365 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4366 call->nSoftAcked = 0;
4368 if (call->flags & RX_CALL_FAST_RECOVER) {
4369 call->flags &= ~RX_CALL_FAST_RECOVER;
4370 call->cwind = call->nextCwind;
4372 #ifdef RX_ENABLE_LOCKS
4373 CV_SIGNAL(&call->cv_twind);
4375 osi_rxWakeup(&call->twind);
4380 rxi_ClearReceiveQueue(register struct rx_call *call)
4382 if (queue_IsNotEmpty(&call->rq)) {
4383 rx_packetReclaims += rxi_FreePackets(0, &call->rq);
4384 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4386 if (call->state == RX_STATE_PRECALL) {
4387 call->flags |= RX_CALL_CLEARED;
4391 /* Send an abort packet for the specified call */
4393 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4394 int istack, int force)
4397 struct clock when, now;
4402 /* Clients should never delay abort messages */
4403 if (rx_IsClientConn(call->conn))
4406 if (call->abortCode != call->error) {
4407 call->abortCode = call->error;
4408 call->abortCount = 0;
4411 if (force || rxi_callAbortThreshhold == 0
4412 || call->abortCount < rxi_callAbortThreshhold) {
4413 if (call->delayedAbortEvent) {
4414 rxevent_Cancel(call->delayedAbortEvent, call,
4415 RX_CALL_REFCOUNT_ABORT);
4417 error = htonl(call->error);
4420 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4421 (char *)&error, sizeof(error), istack);
4422 } else if (!call->delayedAbortEvent) {
4423 clock_GetTime(&now);
4425 clock_Addmsec(&when, rxi_callAbortDelay);
4426 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4427 call->delayedAbortEvent =
4428 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4433 /* Send an abort packet for the specified connection. Packet is an
4434 * optional pointer to a packet that can be used to send the abort.
4435 * Once the number of abort messages reaches the threshhold, an
4436 * event is scheduled to send the abort. Setting the force flag
4437 * overrides sending delayed abort messages.
4439 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4440 * to send the abort packet.
4443 rxi_SendConnectionAbort(register struct rx_connection *conn,
4444 struct rx_packet *packet, int istack, int force)
4447 struct clock when, now;
4449 if (!rx_ConnError(conn))
4452 /* Clients should never delay abort messages */
4453 if (rx_IsClientConn(conn))
4456 if (force || rxi_connAbortThreshhold == 0
4457 || conn->abortCount < rxi_connAbortThreshhold) {
4458 if (conn->delayedAbortEvent) {
4459 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4461 error = htonl(rx_ConnError(conn));
4463 MUTEX_EXIT(&conn->conn_data_lock);
4465 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4466 RX_PACKET_TYPE_ABORT, (char *)&error,
4467 sizeof(error), istack);
4468 MUTEX_ENTER(&conn->conn_data_lock);
4469 } else if (!conn->delayedAbortEvent) {
4470 clock_GetTime(&now);
4472 clock_Addmsec(&when, rxi_connAbortDelay);
4473 conn->delayedAbortEvent =
4474 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4480 * Associate an error all of the calls owned by a connection. Called
4481 * with error non-zero. This is only for really fatal things, like
4482 * bad authentication responses. The connection itself is set in
4483 * error at this point, so that future packets received will be
4487 rxi_ConnectionError(register struct rx_connection *conn,
4488 register afs_int32 error)
4492 struct rx_connection *tconn;
4494 dpf(("rxi_ConnectionError conn %x error %d", conn, error));
4496 MUTEX_ENTER(&conn->conn_data_lock);
4497 if (conn->challengeEvent)
4498 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4499 if (conn->checkReachEvent) {
4500 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4501 conn->checkReachEvent = 0;
4502 conn->flags &= ~RX_CONN_ATTACHWAIT;
4505 MUTEX_EXIT(&conn->conn_data_lock);
4507 for ( tconn = rx_IsClonedConn(conn) ? conn->parent : conn;
4509 tconn = tconn->next_clone) {
4510 for (i = 0; i < RX_MAXCALLS; i++) {
4511 struct rx_call *call = tconn->call[i];
4513 MUTEX_ENTER(&call->lock);
4514 rxi_CallError(call, error);
4515 MUTEX_EXIT(&call->lock);
4519 rx_SetConnError(conn, error);
4520 rx_MutexIncrement(rx_stats.fatalErrors, rx_stats_mutex);
4525 rxi_CallError(register struct rx_call *call, afs_int32 error)
4527 dpf(("rxi_CallError call %x error %d call->error %d", call, error, call->error));
4529 error = call->error;
4531 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4532 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4533 rxi_ResetCall(call, 0);
4536 rxi_ResetCall(call, 0);
4538 call->error = error;
4539 call->mode = RX_MODE_ERROR;
4542 /* Reset various fields in a call structure, and wakeup waiting
4543 * processes. Some fields aren't changed: state & mode are not
4544 * touched (these must be set by the caller), and bufptr, nLeft, and
4545 * nFree are not reset, since these fields are manipulated by
4546 * unprotected macros, and may only be reset by non-interrupting code.
4549 /* this code requires that call->conn be set properly as a pre-condition. */
4550 #endif /* ADAPT_WINDOW */
4553 rxi_ResetCall(register struct rx_call *call, register int newcall)
4556 register struct rx_peer *peer;
4557 struct rx_packet *packet;
4559 dpf(("rxi_ResetCall(call %x, newcall %d)\n", call, newcall));
4561 /* Notify anyone who is waiting for asynchronous packet arrival */
4562 if (call->arrivalProc) {
4563 (*call->arrivalProc) (call, call->arrivalProcHandle,
4564 call->arrivalProcArg);
4565 call->arrivalProc = (void (*)())0;
4568 if (call->delayedAbortEvent) {
4569 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4570 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4572 rxi_SendCallAbort(call, packet, 0, 1);
4573 rxi_FreePacket(packet);
4578 * Update the peer with the congestion information in this call
4579 * so other calls on this connection can pick up where this call
4580 * left off. If the congestion sequence numbers don't match then
4581 * another call experienced a retransmission.
4583 peer = call->conn->peer;
4584 MUTEX_ENTER(&peer->peer_lock);
4586 if (call->congestSeq == peer->congestSeq) {
4587 peer->cwind = MAX(peer->cwind, call->cwind);
4588 peer->MTU = MAX(peer->MTU, call->MTU);
4589 peer->nDgramPackets =
4590 MAX(peer->nDgramPackets, call->nDgramPackets);
4593 call->abortCode = 0;
4594 call->abortCount = 0;
4596 if (peer->maxDgramPackets > 1) {
4597 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4599 call->MTU = peer->MTU;
4601 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4602 call->ssthresh = rx_maxSendWindow;
4603 call->nDgramPackets = peer->nDgramPackets;
4604 call->congestSeq = peer->congestSeq;
4605 MUTEX_EXIT(&peer->peer_lock);
4607 flags = call->flags;
4608 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4609 if (flags & RX_CALL_TQ_BUSY) {
4610 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4611 call->flags |= (flags & RX_CALL_TQ_WAIT);
4613 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4615 rxi_ClearTransmitQueue(call, 1);
4616 /* why init the queue if you just emptied it? queue_Init(&call->tq); */
4617 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4618 dpf(("rcall %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4621 while (call->tqWaiters) {
4622 #ifdef RX_ENABLE_LOCKS
4623 CV_BROADCAST(&call->cv_tq);
4624 #else /* RX_ENABLE_LOCKS */
4625 osi_rxWakeup(&call->tq);
4626 #endif /* RX_ENABLE_LOCKS */
4631 rxi_ClearReceiveQueue(call);
4632 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
4634 if (call->currentPacket) {
4635 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
4636 rxi_FreePacket(call->currentPacket);
4637 call->currentPacket = (struct rx_packet *)0;
4639 call->curlen = call->nLeft = call->nFree = 0;
4641 rxi_FreePackets(0, &call->iovq);
4644 call->twind = call->conn->twind[call->channel];
4645 call->rwind = call->conn->rwind[call->channel];
4646 call->nSoftAcked = 0;
4647 call->nextCwind = 0;
4650 call->nCwindAcks = 0;
4651 call->nSoftAcks = 0;
4652 call->nHardAcks = 0;
4654 call->tfirst = call->rnext = call->tnext = 1;
4656 call->lastAcked = 0;
4657 call->localStatus = call->remoteStatus = 0;
4659 if (flags & RX_CALL_READER_WAIT) {
4660 #ifdef RX_ENABLE_LOCKS
4661 CV_BROADCAST(&call->cv_rq);
4663 osi_rxWakeup(&call->rq);
4666 if (flags & RX_CALL_WAIT_PACKETS) {
4667 MUTEX_ENTER(&rx_freePktQ_lock);
4668 rxi_PacketsUnWait(); /* XXX */
4669 MUTEX_EXIT(&rx_freePktQ_lock);
4671 #ifdef RX_ENABLE_LOCKS
4672 CV_SIGNAL(&call->cv_twind);
4674 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4675 osi_rxWakeup(&call->twind);
4678 #ifdef RX_ENABLE_LOCKS
4679 /* The following ensures that we don't mess with any queue while some
4680 * other thread might also be doing so. The call_queue_lock field is
4681 * is only modified under the call lock. If the call is in the process
4682 * of being removed from a queue, the call is not locked until the
4683 * the queue lock is dropped and only then is the call_queue_lock field
4684 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4685 * Note that any other routine which removes a call from a queue has to
4686 * obtain the queue lock before examing the queue and removing the call.
4688 if (call->call_queue_lock) {
4689 MUTEX_ENTER(call->call_queue_lock);
4690 if (queue_IsOnQueue(call)) {
4692 if (flags & RX_CALL_WAIT_PROC) {
4693 rx_MutexDecrement(rx_nWaiting, rx_stats_mutex);
4696 MUTEX_EXIT(call->call_queue_lock);
4697 CLEAR_CALL_QUEUE_LOCK(call);
4699 #else /* RX_ENABLE_LOCKS */
4700 if (queue_IsOnQueue(call)) {
4702 if (flags & RX_CALL_WAIT_PROC)
4705 #endif /* RX_ENABLE_LOCKS */
4707 rxi_KeepAliveOff(call);
4708 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4711 /* Send an acknowledge for the indicated packet (seq,serial) of the
4712 * indicated call, for the indicated reason (reason). This
4713 * acknowledge will specifically acknowledge receiving the packet, and
4714 * will also specify which other packets for this call have been
4715 * received. This routine returns the packet that was used to the
4716 * caller. The caller is responsible for freeing it or re-using it.
4717 * This acknowledgement also returns the highest sequence number
4718 * actually read out by the higher level to the sender; the sender
4719 * promises to keep around packets that have not been read by the
4720 * higher level yet (unless, of course, the sender decides to abort
4721 * the call altogether). Any of p, seq, serial, pflags, or reason may
4722 * be set to zero without ill effect. That is, if they are zero, they
4723 * will not convey any information.
4724 * NOW there is a trailer field, after the ack where it will safely be
4725 * ignored by mundanes, which indicates the maximum size packet this
4726 * host can swallow. */
4728 register struct rx_packet *optionalPacket; use to send ack (or null)
4729 int seq; Sequence number of the packet we are acking
4730 int serial; Serial number of the packet
4731 int pflags; Flags field from packet header
4732 int reason; Reason an acknowledge was prompted
4736 rxi_SendAck(register struct rx_call *call,
4737 register struct rx_packet *optionalPacket, int serial, int reason,
4740 struct rx_ackPacket *ap;
4741 register struct rx_packet *rqp;
4742 register struct rx_packet *nxp; /* For queue_Scan */
4743 register struct rx_packet *p;
4746 #ifdef RX_ENABLE_TSFPQ
4747 struct rx_ts_info_t * rx_ts_info;
4751 * Open the receive window once a thread starts reading packets
4753 if (call->rnext > 1) {
4754 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
4757 call->nHardAcks = 0;
4758 call->nSoftAcks = 0;
4759 if (call->rnext > call->lastAcked)
4760 call->lastAcked = call->rnext;
4764 rx_computelen(p, p->length); /* reset length, you never know */
4765 } /* where that's been... */
4766 #ifdef RX_ENABLE_TSFPQ
4768 RX_TS_INFO_GET(rx_ts_info);
4769 if ((p = rx_ts_info->local_special_packet)) {
4770 rx_computelen(p, p->length);
4771 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4772 rx_ts_info->local_special_packet = p;
4773 } else { /* We won't send the ack, but don't panic. */
4774 return optionalPacket;
4778 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4779 /* We won't send the ack, but don't panic. */
4780 return optionalPacket;
4785 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4788 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4789 #ifndef RX_ENABLE_TSFPQ
4790 if (!optionalPacket)
4793 return optionalPacket;
4795 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4796 if (rx_Contiguous(p) < templ) {
4797 #ifndef RX_ENABLE_TSFPQ
4798 if (!optionalPacket)
4801 return optionalPacket;
4806 /* MTUXXX failing to send an ack is very serious. We should */
4807 /* try as hard as possible to send even a partial ack; it's */
4808 /* better than nothing. */
4809 ap = (struct rx_ackPacket *)rx_DataOf(p);
4810 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4811 ap->reason = reason;
4813 /* The skew computation used to be bogus, I think it's better now. */
4814 /* We should start paying attention to skew. XXX */
4815 ap->serial = htonl(serial);
4816 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4818 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4819 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4821 /* No fear of running out of ack packet here because there can only be at most
4822 * one window full of unacknowledged packets. The window size must be constrained
4823 * to be less than the maximum ack size, of course. Also, an ack should always
4824 * fit into a single packet -- it should not ever be fragmented. */
4825 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4826 if (!rqp || !call->rq.next
4827 || (rqp->header.seq > (call->rnext + call->rwind))) {
4828 #ifndef RX_ENABLE_TSFPQ
4829 if (!optionalPacket)
4832 rxi_CallError(call, RX_CALL_DEAD);
4833 return optionalPacket;
4836 while (rqp->header.seq > call->rnext + offset)
4837 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4838 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4840 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4841 #ifndef RX_ENABLE_TSFPQ
4842 if (!optionalPacket)
4845 rxi_CallError(call, RX_CALL_DEAD);
4846 return optionalPacket;
4851 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4853 /* these are new for AFS 3.3 */
4854 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4855 templ = htonl(templ);
4856 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4857 templ = htonl(call->conn->peer->ifMTU);
4858 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4859 sizeof(afs_int32), &templ);
4861 /* new for AFS 3.4 */
4862 templ = htonl(call->rwind);
4863 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4864 sizeof(afs_int32), &templ);
4866 /* new for AFS 3.5 */
4867 templ = htonl(call->conn->peer->ifDgramPackets);
4868 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4869 sizeof(afs_int32), &templ);
4871 p->header.serviceId = call->conn->serviceId;
4872 p->header.cid = (call->conn->cid | call->channel);
4873 p->header.callNumber = *call->callNumber;
4875 p->header.securityIndex = call->conn->securityIndex;
4876 p->header.epoch = call->conn->epoch;
4877 p->header.type = RX_PACKET_TYPE_ACK;
4878 p->header.flags = RX_SLOW_START_OK;
4879 if (reason == RX_ACK_PING) {
4880 p->header.flags |= RX_REQUEST_ACK;
4882 clock_GetTime(&call->pingRequestTime);
4885 if (call->conn->type == RX_CLIENT_CONNECTION)
4886 p->header.flags |= RX_CLIENT_INITIATED;
4890 if (rxdebug_active) {
4894 len = _snprintf(msg, sizeof(msg),
4895 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4896 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4897 ntohl(ap->serial), ntohl(ap->previousPacket),
4898 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
4899 ap->nAcks, ntohs(ap->bufferSpace) );
4903 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
4904 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4908 OutputDebugString(msg);
4910 #else /* AFS_NT40_ENV */
4912 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
4913 ap->reason, ntohl(ap->previousPacket),
4914 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4916 for (offset = 0; offset < ap->nAcks; offset++)
4917 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4922 #endif /* AFS_NT40_ENV */
4925 register int i, nbytes = p->length;
4927 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4928 if (nbytes <= p->wirevec[i].iov_len) {
4929 register int savelen, saven;
4931 savelen = p->wirevec[i].iov_len;
4933 p->wirevec[i].iov_len = nbytes;
4935 rxi_Send(call, p, istack);
4936 p->wirevec[i].iov_len = savelen;
4940 nbytes -= p->wirevec[i].iov_len;
4943 rx_MutexIncrement(rx_stats.ackPacketsSent, rx_stats_mutex);
4944 #ifndef RX_ENABLE_TSFPQ
4945 if (!optionalPacket)
4948 return optionalPacket; /* Return packet for re-use by caller */
4951 /* Send all of the packets in the list in single datagram */
4953 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
4954 int istack, int moreFlag, struct clock *now,
4955 struct clock *retryTime, int resending)
4960 struct rx_connection *conn = call->conn;
4961 struct rx_peer *peer = conn->peer;
4963 MUTEX_ENTER(&peer->peer_lock);
4966 peer->reSends += len;
4967 rx_MutexIncrement(rx_stats.dataPacketsSent, rx_stats_mutex);
4968 MUTEX_EXIT(&peer->peer_lock);
4970 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
4974 /* Set the packet flags and schedule the resend events */
4975 /* Only request an ack for the last packet in the list */
4976 for (i = 0; i < len; i++) {
4977 list[i]->retryTime = *retryTime;
4978 if (list[i]->header.serial) {
4979 /* Exponentially backoff retry times */
4980 if (list[i]->backoff < MAXBACKOFF) {
4981 /* so it can't stay == 0 */
4982 list[i]->backoff = (list[i]->backoff << 1) + 1;
4985 clock_Addmsec(&(list[i]->retryTime),
4986 ((afs_uint32) list[i]->backoff) << 8);
4989 /* Wait a little extra for the ack on the last packet */
4990 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
4991 clock_Addmsec(&(list[i]->retryTime), 400);
4994 /* Record the time sent */
4995 list[i]->timeSent = *now;
4997 /* Ask for an ack on retransmitted packets, on every other packet
4998 * if the peer doesn't support slow start. Ask for an ack on every
4999 * packet until the congestion window reaches the ack rate. */
5000 if (list[i]->header.serial) {
5002 rx_MutexIncrement(rx_stats.dataPacketsReSent, rx_stats_mutex);
5004 /* improved RTO calculation- not Karn */
5005 list[i]->firstSent = *now;
5006 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5007 || (!(call->flags & RX_CALL_SLOW_START_OK)
5008 && (list[i]->header.seq & 1)))) {
5013 MUTEX_ENTER(&peer->peer_lock);
5017 rx_MutexIncrement(rx_stats.dataPacketsSent, rx_stats_mutex);
5018 MUTEX_EXIT(&peer->peer_lock);
5020 /* Tag this packet as not being the last in this group,
5021 * for the receiver's benefit */
5022 if (i < len - 1 || moreFlag) {
5023 list[i]->header.flags |= RX_MORE_PACKETS;
5026 /* Install the new retransmit time for the packet, and
5027 * record the time sent */
5028 list[i]->timeSent = *now;
5032 list[len - 1]->header.flags |= RX_REQUEST_ACK;
5035 /* Since we're about to send a data packet to the peer, it's
5036 * safe to nuke any scheduled end-of-packets ack */
5037 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5039 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5040 MUTEX_EXIT(&call->lock);
5042 rxi_SendPacketList(call, conn, list, len, istack);
5044 rxi_SendPacket(call, conn, list[0], istack);
5046 MUTEX_ENTER(&call->lock);
5047 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5049 /* Update last send time for this call (for keep-alive
5050 * processing), and for the connection (so that we can discover
5051 * idle connections) */
5052 call->lastSendData = conn->lastSendTime = call->lastSendTime = clock_Sec();
5055 /* When sending packets we need to follow these rules:
5056 * 1. Never send more than maxDgramPackets in a jumbogram.
5057 * 2. Never send a packet with more than two iovecs in a jumbogram.
5058 * 3. Never send a retransmitted packet in a jumbogram.
5059 * 4. Never send more than cwind/4 packets in a jumbogram
5060 * We always keep the last list we should have sent so we
5061 * can set the RX_MORE_PACKETS flags correctly.
5064 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5065 int istack, struct clock *now, struct clock *retryTime,
5068 int i, cnt, lastCnt = 0;
5069 struct rx_packet **listP, **lastP = 0;
5070 struct rx_peer *peer = call->conn->peer;
5071 int morePackets = 0;
5073 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
5074 /* Does the current packet force us to flush the current list? */
5076 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5077 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5079 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
5081 /* If the call enters an error state stop sending, or if
5082 * we entered congestion recovery mode, stop sending */
5083 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5091 /* Add the current packet to the list if it hasn't been acked.
5092 * Otherwise adjust the list pointer to skip the current packet. */
5093 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5095 /* Do we need to flush the list? */
5096 if (cnt >= (int)peer->maxDgramPackets
5097 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5098 || list[i]->header.serial
5099 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5101 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5102 retryTime, resending);
5103 /* If the call enters an error state stop sending, or if
5104 * we entered congestion recovery mode, stop sending */
5106 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5111 listP = &list[i + 1];
5116 osi_Panic("rxi_SendList error");
5118 listP = &list[i + 1];
5122 /* Send the whole list when the call is in receive mode, when
5123 * the call is in eof mode, when we are in fast recovery mode,
5124 * and when we have the last packet */
5125 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5126 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5127 || (call->flags & RX_CALL_FAST_RECOVER)) {
5128 /* Check for the case where the current list contains
5129 * an acked packet. Since we always send retransmissions
5130 * in a separate packet, we only need to check the first
5131 * packet in the list */
5132 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5136 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5137 retryTime, resending);
5138 /* If the call enters an error state stop sending, or if
5139 * we entered congestion recovery mode, stop sending */
5140 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5144 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5147 } else if (lastCnt > 0) {
5148 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5153 #ifdef RX_ENABLE_LOCKS
5154 /* Call rxi_Start, below, but with the call lock held. */
5156 rxi_StartUnlocked(struct rxevent *event,
5157 void *arg0, void *arg1, int istack)
5159 struct rx_call *call = arg0;
5161 MUTEX_ENTER(&call->lock);
5162 rxi_Start(event, call, arg1, istack);
5163 MUTEX_EXIT(&call->lock);
5165 #endif /* RX_ENABLE_LOCKS */
5167 /* This routine is called when new packets are readied for
5168 * transmission and when retransmission may be necessary, or when the
5169 * transmission window or burst count are favourable. This should be
5170 * better optimized for new packets, the usual case, now that we've
5171 * got rid of queues of send packets. XXXXXXXXXXX */
5173 rxi_Start(struct rxevent *event,
5174 void *arg0, void *arg1, int istack)
5176 struct rx_call *call = arg0;
5178 struct rx_packet *p;
5179 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
5180 struct rx_peer *peer = call->conn->peer;
5181 struct clock now, usenow, retryTime;
5185 struct rx_packet **xmitList;
5188 /* If rxi_Start is being called as a result of a resend event,
5189 * then make sure that the event pointer is removed from the call
5190 * structure, since there is no longer a per-call retransmission
5192 if (event && event == call->resendEvent) {
5193 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5194 call->resendEvent = NULL;
5196 if (queue_IsEmpty(&call->tq)) {
5200 /* Timeouts trigger congestion recovery */
5201 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5202 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5203 /* someone else is waiting to start recovery */
5206 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5207 rxi_WaitforTQBusy(call);
5208 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5209 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5210 call->flags |= RX_CALL_FAST_RECOVER;
5211 if (peer->maxDgramPackets > 1) {
5212 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5214 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5216 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5217 call->nDgramPackets = 1;
5219 call->nextCwind = 1;
5222 MUTEX_ENTER(&peer->peer_lock);
5223 peer->MTU = call->MTU;
5224 peer->cwind = call->cwind;
5225 peer->nDgramPackets = 1;
5227 call->congestSeq = peer->congestSeq;
5228 MUTEX_EXIT(&peer->peer_lock);
5229 /* Clear retry times on packets. Otherwise, it's possible for
5230 * some packets in the queue to force resends at rates faster
5231 * than recovery rates.
5233 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5234 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5235 clock_Zero(&p->retryTime);
5240 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5241 rx_MutexIncrement(rx_tq_debug.rxi_start_in_error, rx_stats_mutex);
5246 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5247 /* Get clock to compute the re-transmit time for any packets
5248 * in this burst. Note, if we back off, it's reasonable to
5249 * back off all of the packets in the same manner, even if
5250 * some of them have been retransmitted more times than more
5252 * Do a dance to avoid blocking after setting now. */
5253 clock_Zero(&retryTime);
5254 MUTEX_ENTER(&peer->peer_lock);
5255 clock_Add(&retryTime, &peer->timeout);
5256 MUTEX_EXIT(&peer->peer_lock);
5257 clock_GetTime(&now);
5258 clock_Add(&retryTime, &now);
5260 /* Send (or resend) any packets that need it, subject to
5261 * window restrictions and congestion burst control
5262 * restrictions. Ask for an ack on the last packet sent in
5263 * this burst. For now, we're relying upon the window being
5264 * considerably bigger than the largest number of packets that
5265 * are typically sent at once by one initial call to
5266 * rxi_Start. This is probably bogus (perhaps we should ask
5267 * for an ack when we're half way through the current
5268 * window?). Also, for non file transfer applications, this
5269 * may end up asking for an ack for every packet. Bogus. XXXX
5272 * But check whether we're here recursively, and let the other guy
5275 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5276 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5277 call->flags |= RX_CALL_TQ_BUSY;
5279 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5281 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5282 call->flags &= ~RX_CALL_NEED_START;
5283 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5285 maxXmitPackets = MIN(call->twind, call->cwind);
5286 xmitList = (struct rx_packet **)
5287 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
5288 /* XXXX else we must drop any mtx we hold */
5289 afs_osi_Alloc_NoSleep(maxXmitPackets * sizeof(struct rx_packet *));
5291 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5293 if (xmitList == NULL)
5294 osi_Panic("rxi_Start, failed to allocate xmit list");
5295 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5296 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5297 /* We shouldn't be sending packets if a thread is waiting
5298 * to initiate congestion recovery */
5302 && (call->flags & RX_CALL_FAST_RECOVER)) {
5303 /* Only send one packet during fast recovery */
5306 if ((p->flags & RX_PKTFLAG_FREE)
5307 || (!queue_IsEnd(&call->tq, nxp)
5308 && (nxp->flags & RX_PKTFLAG_FREE))
5309 || (p == (struct rx_packet *)&rx_freePacketQueue)
5310 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5311 osi_Panic("rxi_Start: xmit queue clobbered");
5313 if (p->flags & RX_PKTFLAG_ACKED) {
5314 /* Since we may block, don't trust this */
5315 usenow.sec = usenow.usec = 0;
5316 rx_MutexIncrement(rx_stats.ignoreAckedPacket, rx_stats_mutex);
5317 continue; /* Ignore this packet if it has been acknowledged */
5320 /* Turn off all flags except these ones, which are the same
5321 * on each transmission */
5322 p->header.flags &= RX_PRESET_FLAGS;
5324 if (p->header.seq >=
5325 call->tfirst + MIN((int)call->twind,
5326 (int)(call->nSoftAcked +
5328 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5329 /* Note: if we're waiting for more window space, we can
5330 * still send retransmits; hence we don't return here, but
5331 * break out to schedule a retransmit event */
5332 dpf(("call %d waiting for window",
5333 *(call->callNumber)));
5337 /* Transmit the packet if it needs to be sent. */
5338 if (!clock_Lt(&now, &p->retryTime)) {
5339 if (nXmitPackets == maxXmitPackets) {
5340 rxi_SendXmitList(call, xmitList, nXmitPackets,
5341 istack, &now, &retryTime,
5343 osi_Free(xmitList, maxXmitPackets *
5344 sizeof(struct rx_packet *));
5347 xmitList[nXmitPackets++] = p;
5351 /* xmitList now hold pointers to all of the packets that are
5352 * ready to send. Now we loop to send the packets */
5353 if (nXmitPackets > 0) {
5354 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5355 &now, &retryTime, resending);
5358 maxXmitPackets * sizeof(struct rx_packet *));
5360 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5362 * TQ references no longer protected by this flag; they must remain
5363 * protected by the global lock.
5365 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5366 call->flags &= ~RX_CALL_TQ_BUSY;
5367 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5368 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5369 #ifdef RX_ENABLE_LOCKS
5370 osirx_AssertMine(&call->lock, "rxi_Start start");
5371 CV_BROADCAST(&call->cv_tq);
5372 #else /* RX_ENABLE_LOCKS */
5373 osi_rxWakeup(&call->tq);
5374 #endif /* RX_ENABLE_LOCKS */
5379 /* We went into the error state while sending packets. Now is
5380 * the time to reset the call. This will also inform the using
5381 * process that the call is in an error state.
5383 rx_MutexIncrement(rx_tq_debug.rxi_start_aborted, rx_stats_mutex);
5384 call->flags &= ~RX_CALL_TQ_BUSY;
5385 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5386 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5387 #ifdef RX_ENABLE_LOCKS
5388 osirx_AssertMine(&call->lock, "rxi_Start middle");
5389 CV_BROADCAST(&call->cv_tq);
5390 #else /* RX_ENABLE_LOCKS */
5391 osi_rxWakeup(&call->tq);
5392 #endif /* RX_ENABLE_LOCKS */
5394 rxi_CallError(call, call->error);
5397 #ifdef RX_ENABLE_LOCKS
5398 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5399 register int missing;
5400 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5401 /* Some packets have received acks. If they all have, we can clear
5402 * the transmit queue.
5405 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5406 if (p->header.seq < call->tfirst
5407 && (p->flags & RX_PKTFLAG_ACKED)) {
5409 p->flags &= ~RX_PKTFLAG_TQ;
5415 call->flags |= RX_CALL_TQ_CLEARME;
5417 #endif /* RX_ENABLE_LOCKS */
5418 /* Don't bother doing retransmits if the TQ is cleared. */
5419 if (call->flags & RX_CALL_TQ_CLEARME) {
5420 rxi_ClearTransmitQueue(call, 1);
5422 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5425 /* Always post a resend event, if there is anything in the
5426 * queue, and resend is possible. There should be at least
5427 * one unacknowledged packet in the queue ... otherwise none
5428 * of these packets should be on the queue in the first place.
5430 if (call->resendEvent) {
5431 /* Cancel the existing event and post a new one */
5432 rxevent_Cancel(call->resendEvent, call,
5433 RX_CALL_REFCOUNT_RESEND);
5436 /* The retry time is the retry time on the first unacknowledged
5437 * packet inside the current window */
5439 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5440 /* Don't set timers for packets outside the window */
5441 if (p->header.seq >= call->tfirst + call->twind) {
5445 if (!(p->flags & RX_PKTFLAG_ACKED)
5446 && !clock_IsZero(&p->retryTime)) {
5448 retryTime = p->retryTime;
5453 /* Post a new event to re-run rxi_Start when retries may be needed */
5454 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5455 #ifdef RX_ENABLE_LOCKS
5456 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5458 rxevent_PostNow2(&retryTime, &usenow,
5460 (void *)call, 0, istack);
5461 #else /* RX_ENABLE_LOCKS */
5463 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5464 (void *)call, 0, istack);
5465 #endif /* RX_ENABLE_LOCKS */
5468 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5469 } while (call->flags & RX_CALL_NEED_START);
5471 * TQ references no longer protected by this flag; they must remain
5472 * protected by the global lock.
5474 call->flags &= ~RX_CALL_TQ_BUSY;
5475 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5476 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5477 #ifdef RX_ENABLE_LOCKS
5478 osirx_AssertMine(&call->lock, "rxi_Start end");
5479 CV_BROADCAST(&call->cv_tq);
5480 #else /* RX_ENABLE_LOCKS */
5481 osi_rxWakeup(&call->tq);
5482 #endif /* RX_ENABLE_LOCKS */
5485 call->flags |= RX_CALL_NEED_START;
5487 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5489 if (call->resendEvent) {
5490 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5495 /* Also adjusts the keep alive parameters for the call, to reflect
5496 * that we have just sent a packet (so keep alives aren't sent
5499 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5502 register struct rx_connection *conn = call->conn;
5504 /* Stamp each packet with the user supplied status */
5505 p->header.userStatus = call->localStatus;
5507 /* Allow the security object controlling this call's security to
5508 * make any last-minute changes to the packet */
5509 RXS_SendPacket(conn->securityObject, call, p);
5511 /* Since we're about to send SOME sort of packet to the peer, it's
5512 * safe to nuke any scheduled end-of-packets ack */
5513 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5515 /* Actually send the packet, filling in more connection-specific fields */
5516 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5517 MUTEX_EXIT(&call->lock);
5518 rxi_SendPacket(call, conn, p, istack);
5519 MUTEX_ENTER(&call->lock);
5520 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5522 /* Update last send time for this call (for keep-alive
5523 * processing), and for the connection (so that we can discover
5524 * idle connections) */
5525 conn->lastSendTime = call->lastSendTime = clock_Sec();
5526 /* Don't count keepalives here, so idleness can be tracked. */
5527 if ((p->header.type != RX_PACKET_TYPE_ACK) || (((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING))
5528 call->lastSendData = call->lastSendTime;
5532 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5533 * that things are fine. Also called periodically to guarantee that nothing
5534 * falls through the cracks (e.g. (error + dally) connections have keepalive
5535 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5537 * haveCTLock Set if calling from rxi_ReapConnections
5539 #ifdef RX_ENABLE_LOCKS
5541 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5542 #else /* RX_ENABLE_LOCKS */
5544 rxi_CheckCall(register struct rx_call *call)
5545 #endif /* RX_ENABLE_LOCKS */
5547 register struct rx_connection *conn = call->conn;
5549 afs_uint32 deadTime;
5551 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5552 if (call->flags & RX_CALL_TQ_BUSY) {
5553 /* Call is active and will be reset by rxi_Start if it's
5554 * in an error state.
5559 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5561 (((afs_uint32) rx_ConnSecondsUntilDead(conn) << 10) +
5562 ((afs_uint32) conn->peer->rtt >> 3) +
5563 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5565 /* These are computed to the second (+- 1 second). But that's
5566 * good enough for these values, which should be a significant
5567 * number of seconds. */
5568 if (now > (call->lastReceiveTime + deadTime)) {
5569 if (call->state == RX_STATE_ACTIVE) {
5571 #if defined(KERNEL) && defined(AFS_SUN57_ENV)
5573 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5574 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
5575 ip_stack_t *ipst = ns->netstack_ip;
5577 ire = ire_cache_lookup(call->conn->peer->host
5578 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
5580 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
5582 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5589 if (ire && ire->ire_max_frag > 0)
5590 rxi_SetPeerMtu(call->conn->peer->host, 0, ire->ire_max_frag);
5591 #if defined(GLOBAL_NETSTACKID)
5595 #endif /* ADAPT_PMTU */
5596 rxi_CallError(call, RX_CALL_DEAD);
5599 #ifdef RX_ENABLE_LOCKS
5600 /* Cancel pending events */
5601 rxevent_Cancel(call->delayedAckEvent, call,
5602 RX_CALL_REFCOUNT_DELAY);
5603 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5604 rxevent_Cancel(call->keepAliveEvent, call,
5605 RX_CALL_REFCOUNT_ALIVE);
5606 if (call->refCount == 0) {
5607 rxi_FreeCall(call, haveCTLock);
5611 #else /* RX_ENABLE_LOCKS */
5614 #endif /* RX_ENABLE_LOCKS */
5616 /* Non-active calls are destroyed if they are not responding
5617 * to pings; active calls are simply flagged in error, so the
5618 * attached process can die reasonably gracefully. */
5620 /* see if we have a non-activity timeout */
5621 if (call->startWait && rx_ConnIdleDeadTime(conn)
5622 && ((call->startWait + rx_ConnIdleDeadTime(conn)) < now)) {
5623 if (call->state == RX_STATE_ACTIVE) {
5624 rxi_CallError(call, RX_CALL_TIMEOUT);
5628 if (call->lastSendData && rx_ConnIdleDeadTime(conn)
5629 && (rx_ConnIdleDeadErr(conn) != 0)
5630 && ((call->lastSendData + rx_ConnIdleDeadTime(conn)) < now)) {
5631 if (call->state == RX_STATE_ACTIVE) {
5632 rxi_CallError(call, conn->idleDeadErr);
5636 /* see if we have a hard timeout */
5637 if (rx_ConnHardDeadTime(conn)
5638 && (now > (rx_ConnHardDeadTime(conn) + call->startTime.sec))) {
5639 if (call->state == RX_STATE_ACTIVE)
5640 rxi_CallError(call, RX_CALL_TIMEOUT);
5647 /* When a call is in progress, this routine is called occasionally to
5648 * make sure that some traffic has arrived (or been sent to) the peer.
5649 * If nothing has arrived in a reasonable amount of time, the call is
5650 * declared dead; if nothing has been sent for a while, we send a
5651 * keep-alive packet (if we're actually trying to keep the call alive)
5654 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
5656 struct rx_call *call = arg1;
5657 struct rx_connection *conn;
5660 MUTEX_ENTER(&call->lock);
5661 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5662 if (event == call->keepAliveEvent)
5663 call->keepAliveEvent = NULL;
5666 #ifdef RX_ENABLE_LOCKS
5667 if (rxi_CheckCall(call, 0)) {
5668 MUTEX_EXIT(&call->lock);
5671 #else /* RX_ENABLE_LOCKS */
5672 if (rxi_CheckCall(call))
5674 #endif /* RX_ENABLE_LOCKS */
5676 /* Don't try to keep alive dallying calls */
5677 if (call->state == RX_STATE_DALLY) {
5678 MUTEX_EXIT(&call->lock);
5683 if ((now - call->lastSendTime) > rx_ConnSecondsUntilPing(conn)) {
5684 /* Don't try to send keepalives if there is unacknowledged data */
5685 /* the rexmit code should be good enough, this little hack
5686 * doesn't quite work XXX */
5687 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5689 rxi_ScheduleKeepAliveEvent(call);
5690 MUTEX_EXIT(&call->lock);
5695 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5697 if (!call->keepAliveEvent) {
5698 struct clock when, now;
5699 clock_GetTime(&now);
5701 when.sec += rx_ConnSecondsUntilPing(call->conn);
5702 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5703 call->keepAliveEvent =
5704 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
5708 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5710 rxi_KeepAliveOn(register struct rx_call *call)
5712 /* Pretend last packet received was received now--i.e. if another
5713 * packet isn't received within the keep alive time, then the call
5714 * will die; Initialize last send time to the current time--even
5715 * if a packet hasn't been sent yet. This will guarantee that a
5716 * keep-alive is sent within the ping time */
5717 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5718 rxi_ScheduleKeepAliveEvent(call);
5721 /* This routine is called to send connection abort messages
5722 * that have been delayed to throttle looping clients. */
5724 rxi_SendDelayedConnAbort(struct rxevent *event,
5725 void *arg1, void *unused)
5727 struct rx_connection *conn = arg1;
5730 struct rx_packet *packet;
5732 MUTEX_ENTER(&conn->conn_data_lock);
5733 conn->delayedAbortEvent = NULL;
5734 error = htonl(rx_ConnError(conn));
5736 MUTEX_EXIT(&conn->conn_data_lock);
5737 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5740 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5741 RX_PACKET_TYPE_ABORT, (char *)&error,
5743 rxi_FreePacket(packet);
5747 /* This routine is called to send call abort messages
5748 * that have been delayed to throttle looping clients. */
5750 rxi_SendDelayedCallAbort(struct rxevent *event,
5751 void *arg1, void *dummy)
5753 struct rx_call *call = arg1;
5756 struct rx_packet *packet;
5758 MUTEX_ENTER(&call->lock);
5759 call->delayedAbortEvent = NULL;
5760 error = htonl(call->error);
5762 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5765 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5766 (char *)&error, sizeof(error), 0);
5767 rxi_FreePacket(packet);
5769 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5770 MUTEX_EXIT(&call->lock);
5773 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5774 * seconds) to ask the client to authenticate itself. The routine
5775 * issues a challenge to the client, which is obtained from the
5776 * security object associated with the connection */
5778 rxi_ChallengeEvent(struct rxevent *event,
5779 void *arg0, void *arg1, int tries)
5781 struct rx_connection *conn = arg0;
5783 conn->challengeEvent = NULL;
5784 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5785 register struct rx_packet *packet;
5786 struct clock when, now;
5789 /* We've failed to authenticate for too long.
5790 * Reset any calls waiting for authentication;
5791 * they are all in RX_STATE_PRECALL.
5795 MUTEX_ENTER(&conn->conn_call_lock);
5796 for (i = 0; i < RX_MAXCALLS; i++) {
5797 struct rx_call *call = conn->call[i];
5799 MUTEX_ENTER(&call->lock);
5800 if (call->state == RX_STATE_PRECALL) {
5801 rxi_CallError(call, RX_CALL_DEAD);
5802 rxi_SendCallAbort(call, NULL, 0, 0);
5804 MUTEX_EXIT(&call->lock);
5807 MUTEX_EXIT(&conn->conn_call_lock);
5811 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5813 /* If there's no packet available, do this later. */
5814 RXS_GetChallenge(conn->securityObject, conn, packet);
5815 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5816 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5817 rxi_FreePacket(packet);
5819 clock_GetTime(&now);
5821 when.sec += RX_CHALLENGE_TIMEOUT;
5822 conn->challengeEvent =
5823 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
5828 /* Call this routine to start requesting the client to authenticate
5829 * itself. This will continue until authentication is established,
5830 * the call times out, or an invalid response is returned. The
5831 * security object associated with the connection is asked to create
5832 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5833 * defined earlier. */
5835 rxi_ChallengeOn(register struct rx_connection *conn)
5837 if (!conn->challengeEvent) {
5838 RXS_CreateChallenge(conn->securityObject, conn);
5839 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5844 /* Compute round trip time of the packet provided, in *rttp.
5847 /* rxi_ComputeRoundTripTime is called with peer locked. */
5848 /* sentp and/or peer may be null */
5850 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5851 register struct clock *sentp,
5852 register struct rx_peer *peer)
5854 struct clock thisRtt, *rttp = &thisRtt;
5856 register int rtt_timeout;
5858 clock_GetTime(rttp);
5860 if (clock_Lt(rttp, sentp)) {
5862 return; /* somebody set the clock back, don't count this time. */
5864 clock_Sub(rttp, sentp);
5865 MUTEX_ENTER(&rx_stats_mutex);
5866 if (clock_Lt(rttp, &rx_stats.minRtt))
5867 rx_stats.minRtt = *rttp;
5868 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5869 if (rttp->sec > 60) {
5870 MUTEX_EXIT(&rx_stats_mutex);
5871 return; /* somebody set the clock ahead */
5873 rx_stats.maxRtt = *rttp;
5875 clock_Add(&rx_stats.totalRtt, rttp);
5876 rx_stats.nRttSamples++;
5877 MUTEX_EXIT(&rx_stats_mutex);
5879 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5881 /* Apply VanJacobson round-trip estimations */
5886 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5887 * srtt is stored as fixed point with 3 bits after the binary
5888 * point (i.e., scaled by 8). The following magic is
5889 * equivalent to the smoothing algorithm in rfc793 with an
5890 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5891 * srtt*8 = srtt*8 + rtt - srtt
5892 * srtt = srtt + rtt/8 - srtt/8
5895 delta = MSEC(rttp) - (peer->rtt >> 3);
5899 * We accumulate a smoothed rtt variance (actually, a smoothed
5900 * mean difference), then set the retransmit timer to smoothed
5901 * rtt + 4 times the smoothed variance (was 2x in van's original
5902 * paper, but 4x works better for me, and apparently for him as
5904 * rttvar is stored as
5905 * fixed point with 2 bits after the binary point (scaled by
5906 * 4). The following is equivalent to rfc793 smoothing with
5907 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5908 * replaces rfc793's wired-in beta.
5909 * dev*4 = dev*4 + (|actual - expected| - dev)
5915 delta -= (peer->rtt_dev >> 2);
5916 peer->rtt_dev += delta;
5918 /* I don't have a stored RTT so I start with this value. Since I'm
5919 * probably just starting a call, and will be pushing more data down
5920 * this, I expect congestion to increase rapidly. So I fudge a
5921 * little, and I set deviance to half the rtt. In practice,
5922 * deviance tends to approach something a little less than
5923 * half the smoothed rtt. */
5924 peer->rtt = (MSEC(rttp) << 3) + 8;
5925 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5927 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5928 * the other of these connections is usually in a user process, and can
5929 * be switched and/or swapped out. So on fast, reliable networks, the
5930 * timeout would otherwise be too short.
5932 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5933 clock_Zero(&(peer->timeout));
5934 clock_Addmsec(&(peer->timeout), rtt_timeout);
5936 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)));
5940 /* Find all server connections that have not been active for a long time, and
5943 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2)
5945 struct clock now, when;
5946 clock_GetTime(&now);
5948 /* Find server connection structures that haven't been used for
5949 * greater than rx_idleConnectionTime */
5951 struct rx_connection **conn_ptr, **conn_end;
5952 int i, havecalls = 0;
5953 RWLOCK_WRLOCK(&rx_connHashTable_lock);
5954 for (conn_ptr = &rx_connHashTable[0], conn_end =
5955 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
5957 struct rx_connection *conn, *next;
5958 struct rx_call *call;
5962 for (conn = *conn_ptr; conn; conn = next) {
5963 /* XXX -- Shouldn't the connection be locked? */
5966 for (i = 0; i < RX_MAXCALLS; i++) {
5967 call = conn->call[i];
5970 MUTEX_ENTER(&call->lock);
5971 #ifdef RX_ENABLE_LOCKS
5972 result = rxi_CheckCall(call, 1);
5973 #else /* RX_ENABLE_LOCKS */
5974 result = rxi_CheckCall(call);
5975 #endif /* RX_ENABLE_LOCKS */
5976 MUTEX_EXIT(&call->lock);
5978 /* If CheckCall freed the call, it might
5979 * have destroyed the connection as well,
5980 * which screws up the linked lists.
5986 if (conn->type == RX_SERVER_CONNECTION) {
5987 /* This only actually destroys the connection if
5988 * there are no outstanding calls */
5989 MUTEX_ENTER(&conn->conn_data_lock);
5990 if (!havecalls && !conn->refCount
5991 && ((conn->lastSendTime + rx_idleConnectionTime) <
5993 conn->refCount++; /* it will be decr in rx_DestroyConn */
5994 MUTEX_EXIT(&conn->conn_data_lock);
5995 #ifdef RX_ENABLE_LOCKS
5996 rxi_DestroyConnectionNoLock(conn);
5997 #else /* RX_ENABLE_LOCKS */
5998 rxi_DestroyConnection(conn);
5999 #endif /* RX_ENABLE_LOCKS */
6001 #ifdef RX_ENABLE_LOCKS
6003 MUTEX_EXIT(&conn->conn_data_lock);
6005 #endif /* RX_ENABLE_LOCKS */
6009 #ifdef RX_ENABLE_LOCKS
6010 while (rx_connCleanup_list) {
6011 struct rx_connection *conn;
6012 conn = rx_connCleanup_list;
6013 rx_connCleanup_list = rx_connCleanup_list->next;
6014 RWLOCK_UNLOCK(&rx_connHashTable_lock);
6015 rxi_CleanupConnection(conn);
6016 RWLOCK_WRLOCK(&rx_connHashTable_lock);
6018 RWLOCK_UNLOCK(&rx_connHashTable_lock);
6019 #endif /* RX_ENABLE_LOCKS */
6022 /* Find any peer structures that haven't been used (haven't had an
6023 * associated connection) for greater than rx_idlePeerTime */
6025 struct rx_peer **peer_ptr, **peer_end;
6027 MUTEX_ENTER(&rx_rpc_stats);
6028 RWLOCK_RDLOCK(&rx_peerHashTable_lock);
6029 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6030 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6032 struct rx_peer *peer, *next, *prev;
6033 for (prev = peer = *peer_ptr; peer; peer = next) {
6035 code = MUTEX_TRYENTER(&peer->peer_lock);
6036 if ((code) && (peer->refCount == 0)
6037 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6038 rx_interface_stat_p rpc_stat, nrpc_stat;
6040 MUTEX_EXIT(&peer->peer_lock);
6041 MUTEX_DESTROY(&peer->peer_lock);
6043 (&peer->rpcStats, rpc_stat, nrpc_stat,
6044 rx_interface_stat)) {
6045 unsigned int num_funcs;
6048 queue_Remove(&rpc_stat->queue_header);
6049 queue_Remove(&rpc_stat->all_peers);
6050 num_funcs = rpc_stat->stats[0].func_total;
6052 sizeof(rx_interface_stat_t) +
6053 rpc_stat->stats[0].func_total *
6054 sizeof(rx_function_entry_v1_t);
6056 rxi_Free(rpc_stat, space);
6057 rxi_rpc_peer_stat_cnt -= num_funcs;
6059 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6060 RWLOCK_UPLOCK(&rx_peerHashTable_lock);
6061 if (peer == *peer_ptr) {
6069 MUTEX_EXIT(&peer->peer_lock);
6075 RWLOCK_UNLOCK(&rx_peerHashTable_lock);
6076 MUTEX_EXIT(&rx_rpc_stats);
6079 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
6080 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
6081 * GC, just below. Really, we shouldn't have to keep moving packets from
6082 * one place to another, but instead ought to always know if we can
6083 * afford to hold onto a packet in its particular use. */
6084 MUTEX_ENTER(&rx_freePktQ_lock);
6085 if (rx_waitingForPackets) {
6086 rx_waitingForPackets = 0;
6087 #ifdef RX_ENABLE_LOCKS
6088 CV_BROADCAST(&rx_waitingForPackets_cv);
6090 osi_rxWakeup(&rx_waitingForPackets);
6093 MUTEX_EXIT(&rx_freePktQ_lock);
6096 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
6097 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
6101 /* rxs_Release - This isn't strictly necessary but, since the macro name from
6102 * rx.h is sort of strange this is better. This is called with a security
6103 * object before it is discarded. Each connection using a security object has
6104 * its own refcount to the object so it won't actually be freed until the last
6105 * connection is destroyed.
6107 * This is the only rxs module call. A hold could also be written but no one
6111 rxs_Release(struct rx_securityClass *aobj)
6113 return RXS_Close(aobj);
6117 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
6118 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
6119 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
6120 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
6122 /* Adjust our estimate of the transmission rate to this peer, given
6123 * that the packet p was just acked. We can adjust peer->timeout and
6124 * call->twind. Pragmatically, this is called
6125 * only with packets of maximal length.
6126 * Called with peer and call locked.
6130 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
6131 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
6133 afs_int32 xferSize, xferMs;
6134 register afs_int32 minTime;
6137 /* Count down packets */
6138 if (peer->rateFlag > 0)
6140 /* Do nothing until we're enabled */
6141 if (peer->rateFlag != 0)
6146 /* Count only when the ack seems legitimate */
6147 switch (ackReason) {
6148 case RX_ACK_REQUESTED:
6150 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6154 case RX_ACK_PING_RESPONSE:
6155 if (p) /* want the response to ping-request, not data send */
6157 clock_GetTime(&newTO);
6158 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6159 clock_Sub(&newTO, &call->pingRequestTime);
6160 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6164 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6171 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));
6173 /* Track only packets that are big enough. */
6174 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6178 /* absorb RTT data (in milliseconds) for these big packets */
6179 if (peer->smRtt == 0) {
6180 peer->smRtt = xferMs;
6182 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6187 if (peer->countDown) {
6191 peer->countDown = 10; /* recalculate only every so often */
6193 /* In practice, we can measure only the RTT for full packets,
6194 * because of the way Rx acks the data that it receives. (If it's
6195 * smaller than a full packet, it often gets implicitly acked
6196 * either by the call response (from a server) or by the next call
6197 * (from a client), and either case confuses transmission times
6198 * with processing times.) Therefore, replace the above
6199 * more-sophisticated processing with a simpler version, where the
6200 * smoothed RTT is kept for full-size packets, and the time to
6201 * transmit a windowful of full-size packets is simply RTT *
6202 * windowSize. Again, we take two steps:
6203 - ensure the timeout is large enough for a single packet's RTT;
6204 - ensure that the window is small enough to fit in the desired timeout.*/
6206 /* First, the timeout check. */
6207 minTime = peer->smRtt;
6208 /* Get a reasonable estimate for a timeout period */
6210 newTO.sec = minTime / 1000;
6211 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6213 /* Increase the timeout period so that we can always do at least
6214 * one packet exchange */
6215 if (clock_Gt(&newTO, &peer->timeout)) {
6217 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));
6219 peer->timeout = newTO;
6222 /* Now, get an estimate for the transmit window size. */
6223 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6224 /* Now, convert to the number of full packets that could fit in a
6225 * reasonable fraction of that interval */
6226 minTime /= (peer->smRtt << 1);
6227 xferSize = minTime; /* (make a copy) */
6229 /* Now clamp the size to reasonable bounds. */
6232 else if (minTime > rx_Window)
6233 minTime = rx_Window;
6234 /* if (minTime != peer->maxWindow) {
6235 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6236 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6237 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6239 peer->maxWindow = minTime;
6240 elide... call->twind = minTime;
6244 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6245 * Discern this by calculating the timeout necessary for rx_Window
6247 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6248 /* calculate estimate for transmission interval in milliseconds */
6249 minTime = rx_Window * peer->smRtt;
6250 if (minTime < 1000) {
6251 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
6252 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6253 peer->timeout.usec, peer->smRtt, peer->packetSize));
6255 newTO.sec = 0; /* cut back on timeout by half a second */
6256 newTO.usec = 500000;
6257 clock_Sub(&peer->timeout, &newTO);
6262 } /* end of rxi_ComputeRate */
6263 #endif /* ADAPT_WINDOW */
6271 #define TRACE_OPTION_DEBUGLOG 4
6279 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6280 0, KEY_QUERY_VALUE, &parmKey);
6281 if (code != ERROR_SUCCESS)
6284 dummyLen = sizeof(TraceOption);
6285 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6286 (BYTE *) &TraceOption, &dummyLen);
6287 if (code == ERROR_SUCCESS) {
6288 rxdebug_active = (TraceOption & TRACE_OPTION_DEBUGLOG) ? 1 : 0;
6290 RegCloseKey (parmKey);
6291 #endif /* AFS_NT40_ENV */
6296 rx_DebugOnOff(int on)
6298 rxdebug_active = on;
6300 #endif /* AFS_NT40_ENV */
6303 /* Don't call this debugging routine directly; use dpf */
6305 rxi_DebugPrint(char *format, ...)
6313 va_start(ap, format);
6315 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6318 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
6320 if (msg[len-1] != '\n') {
6324 OutputDebugString(msg);
6331 va_start(ap, format);
6333 clock_GetTime(&now);
6334 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
6335 (unsigned int)now.usec / 1000);
6336 vfprintf(rx_Log, format, ap);
6343 * This function is used to process the rx_stats structure that is local
6344 * to a process as well as an rx_stats structure received from a remote
6345 * process (via rxdebug). Therefore, it needs to do minimal version
6349 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
6350 afs_int32 freePackets, char version)
6354 if (size != sizeof(struct rx_statistics)) {
6356 "Unexpected size of stats structure: was %d, expected %lud\n",
6357 size, sizeof(struct rx_statistics));
6360 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6363 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6364 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6365 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6366 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6367 s->specialPktAllocFailures);
6369 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6370 s->receivePktAllocFailures, s->sendPktAllocFailures,
6371 s->specialPktAllocFailures);
6375 " greedy %d, " "bogusReads %d (last from host %x), "
6376 "noPackets %d, " "noBuffers %d, " "selects %d, "
6377 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6378 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6379 s->selects, s->sendSelects);
6381 fprintf(file, " packets read: ");
6382 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6383 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6385 fprintf(file, "\n");
6388 " other read counters: data %d, " "ack %d, " "dup %d "
6389 "spurious %d " "dally %d\n", s->dataPacketsRead,
6390 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6391 s->ignorePacketDally);
6393 fprintf(file, " packets sent: ");
6394 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6395 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6397 fprintf(file, "\n");
6400 " other send counters: ack %d, " "data %d (not resends), "
6401 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6402 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6403 s->dataPacketsPushed, s->ignoreAckedPacket);
6406 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6407 s->netSendFailures, (int)s->fatalErrors);
6409 if (s->nRttSamples) {
6410 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6411 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6413 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6414 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6418 " %d server connections, " "%d client connections, "
6419 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6420 s->nServerConns, s->nClientConns, s->nPeerStructs,
6421 s->nCallStructs, s->nFreeCallStructs);
6423 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6424 fprintf(file, " %d clock updates\n", clock_nUpdates);
6429 /* for backward compatibility */
6431 rx_PrintStats(FILE * file)
6433 MUTEX_ENTER(&rx_stats_mutex);
6434 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6436 MUTEX_EXIT(&rx_stats_mutex);
6440 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6442 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6443 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6444 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6447 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6448 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6449 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6452 " Packet size %d, " "max in packet skew %d, "
6453 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6454 (int)peer->outPacketSkew);
6457 #ifdef AFS_PTHREAD_ENV
6459 * This mutex protects the following static variables:
6463 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6464 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6466 #define LOCK_RX_DEBUG
6467 #define UNLOCK_RX_DEBUG
6468 #endif /* AFS_PTHREAD_ENV */
6471 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6472 u_char type, void *inputData, size_t inputLength,
6473 void *outputData, size_t outputLength)
6475 static afs_int32 counter = 100;
6476 time_t waitTime, waitCount, startTime;
6477 struct rx_header theader;
6479 register afs_int32 code;
6480 struct timeval tv_now, tv_wake, tv_delta;
6481 struct sockaddr_in taddr, faddr;
6490 startTime = time(0);
6496 tp = &tbuffer[sizeof(struct rx_header)];
6497 taddr.sin_family = AF_INET;
6498 taddr.sin_port = remotePort;
6499 taddr.sin_addr.s_addr = remoteAddr;
6500 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6501 taddr.sin_len = sizeof(struct sockaddr_in);
6504 memset(&theader, 0, sizeof(theader));
6505 theader.epoch = htonl(999);
6507 theader.callNumber = htonl(counter);
6510 theader.type = type;
6511 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6512 theader.serviceId = 0;
6514 memcpy(tbuffer, &theader, sizeof(theader));
6515 memcpy(tp, inputData, inputLength);
6517 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6518 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6520 /* see if there's a packet available */
6521 gettimeofday(&tv_wake,0);
6522 tv_wake.tv_sec += waitTime;
6525 FD_SET(socket, &imask);
6526 tv_delta.tv_sec = tv_wake.tv_sec;
6527 tv_delta.tv_usec = tv_wake.tv_usec;
6528 gettimeofday(&tv_now, 0);
6530 if (tv_delta.tv_usec < tv_now.tv_usec) {
6532 tv_delta.tv_usec += 1000000;
6535 tv_delta.tv_usec -= tv_now.tv_usec;
6537 if (tv_delta.tv_sec < tv_now.tv_sec) {
6541 tv_delta.tv_sec -= tv_now.tv_sec;
6543 code = select(socket + 1, &imask, 0, 0, &tv_delta);
6544 if (code == 1 && FD_ISSET(socket, &imask)) {
6545 /* now receive a packet */
6546 faddrLen = sizeof(struct sockaddr_in);
6548 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6549 (struct sockaddr *)&faddr, &faddrLen);
6552 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6553 if (counter == ntohl(theader.callNumber))
6561 /* see if we've timed out */
6569 code -= sizeof(struct rx_header);
6570 if (code > outputLength)
6571 code = outputLength;
6572 memcpy(outputData, tp, code);
6577 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6578 afs_uint16 remotePort, struct rx_debugStats * stat,
6579 afs_uint32 * supportedValues)
6581 struct rx_debugIn in;
6584 *supportedValues = 0;
6585 in.type = htonl(RX_DEBUGI_GETSTATS);
6588 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6589 &in, sizeof(in), stat, sizeof(*stat));
6592 * If the call was successful, fixup the version and indicate
6593 * what contents of the stat structure are valid.
6594 * Also do net to host conversion of fields here.
6598 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6599 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6601 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6602 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6604 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6605 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6607 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6608 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6610 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6611 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6613 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6614 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6616 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6617 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6619 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6620 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6622 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
6623 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
6625 stat->nFreePackets = ntohl(stat->nFreePackets);
6626 stat->packetReclaims = ntohl(stat->packetReclaims);
6627 stat->callsExecuted = ntohl(stat->callsExecuted);
6628 stat->nWaiting = ntohl(stat->nWaiting);
6629 stat->idleThreads = ntohl(stat->idleThreads);
6630 stat->nWaited = ntohl(stat->nWaited);
6631 stat->nPackets = ntohl(stat->nPackets);
6638 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6639 afs_uint16 remotePort, struct rx_statistics * stat,
6640 afs_uint32 * supportedValues)
6642 struct rx_debugIn in;
6644 afs_int32 *lp = (afs_int32 *) stat;
6648 * supportedValues is currently unused, but added to allow future
6649 * versioning of this function.
6652 *supportedValues = 0;
6653 in.type = htonl(RX_DEBUGI_RXSTATS);
6655 memset(stat, 0, sizeof(*stat));
6657 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6658 &in, sizeof(in), stat, sizeof(*stat));
6663 * Do net to host conversion here
6666 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6675 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6676 afs_uint16 remotePort, size_t version_length,
6680 return MakeDebugCall(socket, remoteAddr, remotePort,
6681 RX_PACKET_TYPE_VERSION, a, 1, version,
6686 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6687 afs_uint16 remotePort, afs_int32 * nextConnection,
6688 int allConnections, afs_uint32 debugSupportedValues,
6689 struct rx_debugConn * conn,
6690 afs_uint32 * supportedValues)
6692 struct rx_debugIn in;
6697 * supportedValues is currently unused, but added to allow future
6698 * versioning of this function.
6701 *supportedValues = 0;
6702 if (allConnections) {
6703 in.type = htonl(RX_DEBUGI_GETALLCONN);
6705 in.type = htonl(RX_DEBUGI_GETCONN);
6707 in.index = htonl(*nextConnection);
6708 memset(conn, 0, sizeof(*conn));
6710 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6711 &in, sizeof(in), conn, sizeof(*conn));
6714 *nextConnection += 1;
6717 * Convert old connection format to new structure.
6720 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6721 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6722 #define MOVEvL(a) (conn->a = vL->a)
6724 /* any old or unrecognized version... */
6725 for (i = 0; i < RX_MAXCALLS; i++) {
6726 MOVEvL(callState[i]);
6727 MOVEvL(callMode[i]);
6728 MOVEvL(callFlags[i]);
6729 MOVEvL(callOther[i]);
6731 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6732 MOVEvL(secStats.type);
6733 MOVEvL(secStats.level);
6734 MOVEvL(secStats.flags);
6735 MOVEvL(secStats.expires);
6736 MOVEvL(secStats.packetsReceived);
6737 MOVEvL(secStats.packetsSent);
6738 MOVEvL(secStats.bytesReceived);
6739 MOVEvL(secStats.bytesSent);
6744 * Do net to host conversion here
6746 * I don't convert host or port since we are most likely
6747 * going to want these in NBO.
6749 conn->cid = ntohl(conn->cid);
6750 conn->serial = ntohl(conn->serial);
6751 for (i = 0; i < RX_MAXCALLS; i++) {
6752 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6754 rx_SetConnError(conn, ntohl(rx_ConnError(conn)));
6755 conn->secStats.flags = ntohl(conn->secStats.flags);
6756 conn->secStats.expires = ntohl(conn->secStats.expires);
6757 conn->secStats.packetsReceived =
6758 ntohl(conn->secStats.packetsReceived);
6759 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6760 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6761 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6762 conn->epoch = ntohl(conn->epoch);
6763 conn->natMTU = ntohl(conn->natMTU);
6770 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6771 afs_uint16 remotePort, afs_int32 * nextPeer,
6772 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6773 afs_uint32 * supportedValues)
6775 struct rx_debugIn in;
6779 * supportedValues is currently unused, but added to allow future
6780 * versioning of this function.
6783 *supportedValues = 0;
6784 in.type = htonl(RX_DEBUGI_GETPEER);
6785 in.index = htonl(*nextPeer);
6786 memset(peer, 0, sizeof(*peer));
6788 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6789 &in, sizeof(in), peer, sizeof(*peer));
6795 * Do net to host conversion here
6797 * I don't convert host or port since we are most likely
6798 * going to want these in NBO.
6800 peer->ifMTU = ntohs(peer->ifMTU);
6801 peer->idleWhen = ntohl(peer->idleWhen);
6802 peer->refCount = ntohs(peer->refCount);
6803 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6804 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6805 peer->rtt = ntohl(peer->rtt);
6806 peer->rtt_dev = ntohl(peer->rtt_dev);
6807 peer->timeout.sec = ntohl(peer->timeout.sec);
6808 peer->timeout.usec = ntohl(peer->timeout.usec);
6809 peer->nSent = ntohl(peer->nSent);
6810 peer->reSends = ntohl(peer->reSends);
6811 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6812 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6813 peer->rateFlag = ntohl(peer->rateFlag);
6814 peer->natMTU = ntohs(peer->natMTU);
6815 peer->maxMTU = ntohs(peer->maxMTU);
6816 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6817 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6818 peer->MTU = ntohs(peer->MTU);
6819 peer->cwind = ntohs(peer->cwind);
6820 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6821 peer->congestSeq = ntohs(peer->congestSeq);
6822 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6823 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6824 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6825 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6830 #endif /* RXDEBUG */
6835 struct rx_serverQueueEntry *np;
6838 register struct rx_call *call;
6839 register struct rx_serverQueueEntry *sq;
6843 if (rxinit_status == 1) {
6845 return; /* Already shutdown. */
6849 #ifndef AFS_PTHREAD_ENV
6850 FD_ZERO(&rx_selectMask);
6851 #endif /* AFS_PTHREAD_ENV */
6852 rxi_dataQuota = RX_MAX_QUOTA;
6853 #ifndef AFS_PTHREAD_ENV
6855 #endif /* AFS_PTHREAD_ENV */
6858 #ifndef AFS_PTHREAD_ENV
6859 #ifndef AFS_USE_GETTIMEOFDAY
6861 #endif /* AFS_USE_GETTIMEOFDAY */
6862 #endif /* AFS_PTHREAD_ENV */
6864 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6865 call = queue_First(&rx_freeCallQueue, rx_call);
6867 rxi_Free(call, sizeof(struct rx_call));
6870 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6871 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6877 struct rx_peer **peer_ptr, **peer_end;
6878 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6879 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6881 struct rx_peer *peer, *next;
6882 for (peer = *peer_ptr; peer; peer = next) {
6883 rx_interface_stat_p rpc_stat, nrpc_stat;
6886 (&peer->rpcStats, rpc_stat, nrpc_stat,
6887 rx_interface_stat)) {
6891 queue_Remove(&rpc_stat->queue_header);
6892 queue_Remove(&rpc_stat->all_peers);
6893 num_funcs = rpc_stat->stats[0].func_total;
6895 sizeof(rx_interface_stat_t) +
6896 rpc_stat->stats[0].func_total *
6897 sizeof(rx_function_entry_v1_t);
6899 rxi_Free(rpc_stat, space);
6900 rx_MutexAdd(rxi_rpc_peer_stat_cnt, -num_funcs, rx_rpc_stats);
6904 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6908 for (i = 0; i < RX_MAX_SERVICES; i++) {
6910 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6912 for (i = 0; i < rx_hashTableSize; i++) {
6913 register struct rx_connection *tc, *ntc;
6914 RWLOCK_RDLOCK(&rx_connHashTable_lock);
6915 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6917 for (j = 0; j < RX_MAXCALLS; j++) {
6919 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6922 rxi_Free(tc, sizeof(*tc));
6924 RWLOCK_UNLOCK(&rx_connHashTable_lock);
6927 MUTEX_ENTER(&freeSQEList_lock);
6929 while ((np = rx_FreeSQEList)) {
6930 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6931 MUTEX_DESTROY(&np->lock);
6932 rxi_Free(np, sizeof(*np));
6935 MUTEX_EXIT(&freeSQEList_lock);
6936 MUTEX_DESTROY(&freeSQEList_lock);
6937 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6938 RWLOCK_DESTROY(&rx_connHashTable_lock);
6939 RWLOCK_DESTROY(&rx_peerHashTable_lock);
6940 MUTEX_DESTROY(&rx_serverPool_lock);
6942 osi_Free(rx_connHashTable,
6943 rx_hashTableSize * sizeof(struct rx_connection *));
6944 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6946 UNPIN(rx_connHashTable,
6947 rx_hashTableSize * sizeof(struct rx_connection *));
6948 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
6950 rxi_FreeAllPackets();
6952 MUTEX_ENTER(&rx_stats_mutex);
6953 rxi_dataQuota = RX_MAX_QUOTA;
6954 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
6955 MUTEX_EXIT(&rx_stats_mutex);
6961 #ifdef RX_ENABLE_LOCKS
6963 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
6965 if (!MUTEX_ISMINE(lockaddr))
6966 osi_Panic("Lock not held: %s", msg);
6968 #endif /* RX_ENABLE_LOCKS */
6973 * Routines to implement connection specific data.
6977 rx_KeyCreate(rx_destructor_t rtn)
6980 MUTEX_ENTER(&rxi_keyCreate_lock);
6981 key = rxi_keyCreate_counter++;
6982 rxi_keyCreate_destructor = (rx_destructor_t *)
6983 realloc((void *)rxi_keyCreate_destructor,
6984 (key + 1) * sizeof(rx_destructor_t));
6985 rxi_keyCreate_destructor[key] = rtn;
6986 MUTEX_EXIT(&rxi_keyCreate_lock);
6991 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
6994 struct rx_connection *tconn =
6995 (rx_IsClonedConn(conn)) ? conn->parent : conn;
6997 MUTEX_ENTER(&tconn->conn_data_lock);
6998 if (!tconn->specific) {
6999 tconn->specific = (void **)malloc((key + 1) * sizeof(void *));
7000 for (i = 0; i < key; i++)
7001 tconn->specific[i] = NULL;
7002 tconn->nSpecific = key + 1;
7003 tconn->specific[key] = ptr;
7004 } else if (key >= tconn->nSpecific) {
7005 tconn->specific = (void **)
7006 realloc(tconn->specific, (key + 1) * sizeof(void *));
7007 for (i = tconn->nSpecific; i < key; i++)
7008 tconn->specific[i] = NULL;
7009 tconn->nSpecific = key + 1;
7010 tconn->specific[key] = ptr;
7012 if (tconn->specific[key] && rxi_keyCreate_destructor[key])
7013 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7014 tconn->specific[key] = ptr;
7016 MUTEX_EXIT(&tconn->conn_data_lock);
7020 rx_GetSpecific(struct rx_connection *conn, int key)
7023 struct rx_connection *tconn =
7024 (rx_IsClonedConn(conn)) ? conn->parent : conn;
7026 MUTEX_ENTER(&tconn->conn_data_lock);
7027 if (key >= tconn->nSpecific)
7030 ptr = tconn->specific[key];
7031 MUTEX_EXIT(&tconn->conn_data_lock);
7035 #endif /* !KERNEL */
7038 * processStats is a queue used to store the statistics for the local
7039 * process. Its contents are similar to the contents of the rpcStats
7040 * queue on a rx_peer structure, but the actual data stored within
7041 * this queue contains totals across the lifetime of the process (assuming
7042 * the stats have not been reset) - unlike the per peer structures
7043 * which can come and go based upon the peer lifetime.
7046 static struct rx_queue processStats = { &processStats, &processStats };
7049 * peerStats is a queue used to store the statistics for all peer structs.
7050 * Its contents are the union of all the peer rpcStats queues.
7053 static struct rx_queue peerStats = { &peerStats, &peerStats };
7056 * rxi_monitor_processStats is used to turn process wide stat collection
7060 static int rxi_monitor_processStats = 0;
7063 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
7066 static int rxi_monitor_peerStats = 0;
7069 * rxi_AddRpcStat - given all of the information for a particular rpc
7070 * call, create (if needed) and update the stat totals for the rpc.
7074 * IN stats - the queue of stats that will be updated with the new value
7076 * IN rxInterface - a unique number that identifies the rpc interface
7078 * IN currentFunc - the index of the function being invoked
7080 * IN totalFunc - the total number of functions in this interface
7082 * IN queueTime - the amount of time this function waited for a thread
7084 * IN execTime - the amount of time this function invocation took to execute
7086 * IN bytesSent - the number bytes sent by this invocation
7088 * IN bytesRcvd - the number bytes received by this invocation
7090 * IN isServer - if true, this invocation was made to a server
7092 * IN remoteHost - the ip address of the remote host
7094 * IN remotePort - the port of the remote host
7096 * IN addToPeerList - if != 0, add newly created stat to the global peer list
7098 * INOUT counter - if a new stats structure is allocated, the counter will
7099 * be updated with the new number of allocated stat structures
7107 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
7108 afs_uint32 currentFunc, afs_uint32 totalFunc,
7109 struct clock *queueTime, struct clock *execTime,
7110 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
7111 afs_uint32 remoteHost, afs_uint32 remotePort,
7112 int addToPeerList, unsigned int *counter)
7115 rx_interface_stat_p rpc_stat, nrpc_stat;
7118 * See if there's already a structure for this interface
7121 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7122 if ((rpc_stat->stats[0].interfaceId == rxInterface)
7123 && (rpc_stat->stats[0].remote_is_server == isServer))
7128 * Didn't find a match so allocate a new structure and add it to the
7132 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
7133 || (rpc_stat->stats[0].interfaceId != rxInterface)
7134 || (rpc_stat->stats[0].remote_is_server != isServer)) {
7139 sizeof(rx_interface_stat_t) +
7140 totalFunc * sizeof(rx_function_entry_v1_t);
7142 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
7143 if (rpc_stat == NULL) {
7147 *counter += totalFunc;
7148 for (i = 0; i < totalFunc; i++) {
7149 rpc_stat->stats[i].remote_peer = remoteHost;
7150 rpc_stat->stats[i].remote_port = remotePort;
7151 rpc_stat->stats[i].remote_is_server = isServer;
7152 rpc_stat->stats[i].interfaceId = rxInterface;
7153 rpc_stat->stats[i].func_total = totalFunc;
7154 rpc_stat->stats[i].func_index = i;
7155 hzero(rpc_stat->stats[i].invocations);
7156 hzero(rpc_stat->stats[i].bytes_sent);
7157 hzero(rpc_stat->stats[i].bytes_rcvd);
7158 rpc_stat->stats[i].queue_time_sum.sec = 0;
7159 rpc_stat->stats[i].queue_time_sum.usec = 0;
7160 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7161 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7162 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7163 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7164 rpc_stat->stats[i].queue_time_max.sec = 0;
7165 rpc_stat->stats[i].queue_time_max.usec = 0;
7166 rpc_stat->stats[i].execution_time_sum.sec = 0;
7167 rpc_stat->stats[i].execution_time_sum.usec = 0;
7168 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7169 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7170 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7171 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7172 rpc_stat->stats[i].execution_time_max.sec = 0;
7173 rpc_stat->stats[i].execution_time_max.usec = 0;
7175 queue_Prepend(stats, rpc_stat);
7176 if (addToPeerList) {
7177 queue_Prepend(&peerStats, &rpc_stat->all_peers);
7182 * Increment the stats for this function
7185 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
7186 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
7187 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
7188 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
7189 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7190 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7191 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7193 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7194 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7196 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7197 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7199 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7200 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7202 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7203 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7211 * rx_IncrementTimeAndCount - increment the times and count for a particular
7216 * IN peer - the peer who invoked the rpc
7218 * IN rxInterface - a unique number that identifies the rpc interface
7220 * IN currentFunc - the index of the function being invoked
7222 * IN totalFunc - the total number of functions in this interface
7224 * IN queueTime - the amount of time this function waited for a thread
7226 * IN execTime - the amount of time this function invocation took to execute
7228 * IN bytesSent - the number bytes sent by this invocation
7230 * IN bytesRcvd - the number bytes received by this invocation
7232 * IN isServer - if true, this invocation was made to a server
7240 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7241 afs_uint32 currentFunc, afs_uint32 totalFunc,
7242 struct clock *queueTime, struct clock *execTime,
7243 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7247 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
7250 MUTEX_ENTER(&rx_rpc_stats);
7251 MUTEX_ENTER(&peer->peer_lock);
7253 if (rxi_monitor_peerStats) {
7254 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7255 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7256 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7259 if (rxi_monitor_processStats) {
7260 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7261 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7262 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7265 MUTEX_EXIT(&peer->peer_lock);
7266 MUTEX_EXIT(&rx_rpc_stats);
7271 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7275 * IN callerVersion - the rpc stat version of the caller.
7277 * IN count - the number of entries to marshall.
7279 * IN stats - pointer to stats to be marshalled.
7281 * OUT ptr - Where to store the marshalled data.
7288 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7289 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7295 * We only support the first version
7297 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7298 *(ptr++) = stats->remote_peer;
7299 *(ptr++) = stats->remote_port;
7300 *(ptr++) = stats->remote_is_server;
7301 *(ptr++) = stats->interfaceId;
7302 *(ptr++) = stats->func_total;
7303 *(ptr++) = stats->func_index;
7304 *(ptr++) = hgethi(stats->invocations);
7305 *(ptr++) = hgetlo(stats->invocations);
7306 *(ptr++) = hgethi(stats->bytes_sent);
7307 *(ptr++) = hgetlo(stats->bytes_sent);
7308 *(ptr++) = hgethi(stats->bytes_rcvd);
7309 *(ptr++) = hgetlo(stats->bytes_rcvd);
7310 *(ptr++) = stats->queue_time_sum.sec;
7311 *(ptr++) = stats->queue_time_sum.usec;
7312 *(ptr++) = stats->queue_time_sum_sqr.sec;
7313 *(ptr++) = stats->queue_time_sum_sqr.usec;
7314 *(ptr++) = stats->queue_time_min.sec;
7315 *(ptr++) = stats->queue_time_min.usec;
7316 *(ptr++) = stats->queue_time_max.sec;
7317 *(ptr++) = stats->queue_time_max.usec;
7318 *(ptr++) = stats->execution_time_sum.sec;
7319 *(ptr++) = stats->execution_time_sum.usec;
7320 *(ptr++) = stats->execution_time_sum_sqr.sec;
7321 *(ptr++) = stats->execution_time_sum_sqr.usec;
7322 *(ptr++) = stats->execution_time_min.sec;
7323 *(ptr++) = stats->execution_time_min.usec;
7324 *(ptr++) = stats->execution_time_max.sec;
7325 *(ptr++) = stats->execution_time_max.usec;
7331 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7336 * IN callerVersion - the rpc stat version of the caller
7338 * OUT myVersion - the rpc stat version of this function
7340 * OUT clock_sec - local time seconds
7342 * OUT clock_usec - local time microseconds
7344 * OUT allocSize - the number of bytes allocated to contain stats
7346 * OUT statCount - the number stats retrieved from this process.
7348 * OUT stats - the actual stats retrieved from this process.
7352 * Returns void. If successful, stats will != NULL.
7356 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7357 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7358 size_t * allocSize, afs_uint32 * statCount,
7359 afs_uint32 ** stats)
7369 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7372 * Check to see if stats are enabled
7375 MUTEX_ENTER(&rx_rpc_stats);
7376 if (!rxi_monitor_processStats) {
7377 MUTEX_EXIT(&rx_rpc_stats);
7381 clock_GetTime(&now);
7382 *clock_sec = now.sec;
7383 *clock_usec = now.usec;
7386 * Allocate the space based upon the caller version
7388 * If the client is at an older version than we are,
7389 * we return the statistic data in the older data format, but
7390 * we still return our version number so the client knows we
7391 * are maintaining more data than it can retrieve.
7394 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7395 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7396 *statCount = rxi_rpc_process_stat_cnt;
7399 * This can't happen yet, but in the future version changes
7400 * can be handled by adding additional code here
7404 if (space > (size_t) 0) {
7406 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7409 rx_interface_stat_p rpc_stat, nrpc_stat;
7413 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7415 * Copy the data based upon the caller version
7417 rx_MarshallProcessRPCStats(callerVersion,
7418 rpc_stat->stats[0].func_total,
7419 rpc_stat->stats, &ptr);
7425 MUTEX_EXIT(&rx_rpc_stats);
7430 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7434 * IN callerVersion - the rpc stat version of the caller
7436 * OUT myVersion - the rpc stat version of this function
7438 * OUT clock_sec - local time seconds
7440 * OUT clock_usec - local time microseconds
7442 * OUT allocSize - the number of bytes allocated to contain stats
7444 * OUT statCount - the number of stats retrieved from the individual
7447 * OUT stats - the actual stats retrieved from the individual peer structures.
7451 * Returns void. If successful, stats will != NULL.
7455 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7456 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7457 size_t * allocSize, afs_uint32 * statCount,
7458 afs_uint32 ** stats)
7468 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7471 * Check to see if stats are enabled
7474 MUTEX_ENTER(&rx_rpc_stats);
7475 if (!rxi_monitor_peerStats) {
7476 MUTEX_EXIT(&rx_rpc_stats);
7480 clock_GetTime(&now);
7481 *clock_sec = now.sec;
7482 *clock_usec = now.usec;
7485 * Allocate the space based upon the caller version
7487 * If the client is at an older version than we are,
7488 * we return the statistic data in the older data format, but
7489 * we still return our version number so the client knows we
7490 * are maintaining more data than it can retrieve.
7493 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7494 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7495 *statCount = rxi_rpc_peer_stat_cnt;
7498 * This can't happen yet, but in the future version changes
7499 * can be handled by adding additional code here
7503 if (space > (size_t) 0) {
7505 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7508 rx_interface_stat_p rpc_stat, nrpc_stat;
7512 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7514 * We have to fix the offset of rpc_stat since we are
7515 * keeping this structure on two rx_queues. The rx_queue
7516 * package assumes that the rx_queue member is the first
7517 * member of the structure. That is, rx_queue assumes that
7518 * any one item is only on one queue at a time. We are
7519 * breaking that assumption and so we have to do a little
7520 * math to fix our pointers.
7523 fix_offset = (char *)rpc_stat;
7524 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7525 rpc_stat = (rx_interface_stat_p) fix_offset;
7528 * Copy the data based upon the caller version
7530 rx_MarshallProcessRPCStats(callerVersion,
7531 rpc_stat->stats[0].func_total,
7532 rpc_stat->stats, &ptr);
7538 MUTEX_EXIT(&rx_rpc_stats);
7543 * rx_FreeRPCStats - free memory allocated by
7544 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7548 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7549 * rx_RetrievePeerRPCStats
7551 * IN allocSize - the number of bytes in stats.
7559 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7561 rxi_Free(stats, allocSize);
7565 * rx_queryProcessRPCStats - see if process rpc stat collection is
7566 * currently enabled.
7572 * Returns 0 if stats are not enabled != 0 otherwise
7576 rx_queryProcessRPCStats(void)
7579 MUTEX_ENTER(&rx_rpc_stats);
7580 rc = rxi_monitor_processStats;
7581 MUTEX_EXIT(&rx_rpc_stats);
7586 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7592 * Returns 0 if stats are not enabled != 0 otherwise
7596 rx_queryPeerRPCStats(void)
7599 MUTEX_ENTER(&rx_rpc_stats);
7600 rc = rxi_monitor_peerStats;
7601 MUTEX_EXIT(&rx_rpc_stats);
7606 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7616 rx_enableProcessRPCStats(void)
7618 MUTEX_ENTER(&rx_rpc_stats);
7619 rx_enable_stats = 1;
7620 rxi_monitor_processStats = 1;
7621 MUTEX_EXIT(&rx_rpc_stats);
7625 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7635 rx_enablePeerRPCStats(void)
7637 MUTEX_ENTER(&rx_rpc_stats);
7638 rx_enable_stats = 1;
7639 rxi_monitor_peerStats = 1;
7640 MUTEX_EXIT(&rx_rpc_stats);
7644 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7654 rx_disableProcessRPCStats(void)
7656 rx_interface_stat_p rpc_stat, nrpc_stat;
7659 MUTEX_ENTER(&rx_rpc_stats);
7662 * Turn off process statistics and if peer stats is also off, turn
7666 rxi_monitor_processStats = 0;
7667 if (rxi_monitor_peerStats == 0) {
7668 rx_enable_stats = 0;
7671 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7672 unsigned int num_funcs = 0;
7675 queue_Remove(rpc_stat);
7676 num_funcs = rpc_stat->stats[0].func_total;
7678 sizeof(rx_interface_stat_t) +
7679 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7681 rxi_Free(rpc_stat, space);
7682 rxi_rpc_process_stat_cnt -= num_funcs;
7684 MUTEX_EXIT(&rx_rpc_stats);
7688 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7698 rx_disablePeerRPCStats(void)
7700 struct rx_peer **peer_ptr, **peer_end;
7703 MUTEX_ENTER(&rx_rpc_stats);
7706 * Turn off peer statistics and if process stats is also off, turn
7710 rxi_monitor_peerStats = 0;
7711 if (rxi_monitor_processStats == 0) {
7712 rx_enable_stats = 0;
7715 RWLOCK_RDLOCK(&rx_peerHashTable_lock);
7716 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7717 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7719 struct rx_peer *peer, *next, *prev;
7720 for (prev = peer = *peer_ptr; peer; peer = next) {
7722 code = MUTEX_TRYENTER(&peer->peer_lock);
7724 rx_interface_stat_p rpc_stat, nrpc_stat;
7727 (&peer->rpcStats, rpc_stat, nrpc_stat,
7728 rx_interface_stat)) {
7729 unsigned int num_funcs = 0;
7732 queue_Remove(&rpc_stat->queue_header);
7733 queue_Remove(&rpc_stat->all_peers);
7734 num_funcs = rpc_stat->stats[0].func_total;
7736 sizeof(rx_interface_stat_t) +
7737 rpc_stat->stats[0].func_total *
7738 sizeof(rx_function_entry_v1_t);
7740 rxi_Free(rpc_stat, space);
7741 rxi_rpc_peer_stat_cnt -= num_funcs;
7743 MUTEX_EXIT(&peer->peer_lock);
7744 if (prev == *peer_ptr) {
7754 RWLOCK_UNLOCK(&rx_peerHashTable_lock);
7755 MUTEX_EXIT(&rx_rpc_stats);
7759 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7764 * IN clearFlag - flag indicating which stats to clear
7772 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7774 rx_interface_stat_p rpc_stat, nrpc_stat;
7776 MUTEX_ENTER(&rx_rpc_stats);
7778 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7779 unsigned int num_funcs = 0, i;
7780 num_funcs = rpc_stat->stats[0].func_total;
7781 for (i = 0; i < num_funcs; i++) {
7782 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7783 hzero(rpc_stat->stats[i].invocations);
7785 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7786 hzero(rpc_stat->stats[i].bytes_sent);
7788 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7789 hzero(rpc_stat->stats[i].bytes_rcvd);
7791 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7792 rpc_stat->stats[i].queue_time_sum.sec = 0;
7793 rpc_stat->stats[i].queue_time_sum.usec = 0;
7795 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7796 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7797 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7799 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7800 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7801 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7803 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7804 rpc_stat->stats[i].queue_time_max.sec = 0;
7805 rpc_stat->stats[i].queue_time_max.usec = 0;
7807 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7808 rpc_stat->stats[i].execution_time_sum.sec = 0;
7809 rpc_stat->stats[i].execution_time_sum.usec = 0;
7811 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7812 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7813 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7815 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7816 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7817 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7819 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7820 rpc_stat->stats[i].execution_time_max.sec = 0;
7821 rpc_stat->stats[i].execution_time_max.usec = 0;
7826 MUTEX_EXIT(&rx_rpc_stats);
7830 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7835 * IN clearFlag - flag indicating which stats to clear
7843 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7845 rx_interface_stat_p rpc_stat, nrpc_stat;
7847 MUTEX_ENTER(&rx_rpc_stats);
7849 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7850 unsigned int num_funcs = 0, i;
7853 * We have to fix the offset of rpc_stat since we are
7854 * keeping this structure on two rx_queues. The rx_queue
7855 * package assumes that the rx_queue member is the first
7856 * member of the structure. That is, rx_queue assumes that
7857 * any one item is only on one queue at a time. We are
7858 * breaking that assumption and so we have to do a little
7859 * math to fix our pointers.
7862 fix_offset = (char *)rpc_stat;
7863 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7864 rpc_stat = (rx_interface_stat_p) fix_offset;
7866 num_funcs = rpc_stat->stats[0].func_total;
7867 for (i = 0; i < num_funcs; i++) {
7868 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7869 hzero(rpc_stat->stats[i].invocations);
7871 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7872 hzero(rpc_stat->stats[i].bytes_sent);
7874 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7875 hzero(rpc_stat->stats[i].bytes_rcvd);
7877 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7878 rpc_stat->stats[i].queue_time_sum.sec = 0;
7879 rpc_stat->stats[i].queue_time_sum.usec = 0;
7881 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7882 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7883 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7885 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7886 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7887 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7889 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7890 rpc_stat->stats[i].queue_time_max.sec = 0;
7891 rpc_stat->stats[i].queue_time_max.usec = 0;
7893 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7894 rpc_stat->stats[i].execution_time_sum.sec = 0;
7895 rpc_stat->stats[i].execution_time_sum.usec = 0;
7897 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7898 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7899 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7901 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7902 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7903 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7905 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7906 rpc_stat->stats[i].execution_time_max.sec = 0;
7907 rpc_stat->stats[i].execution_time_max.usec = 0;
7912 MUTEX_EXIT(&rx_rpc_stats);
7916 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7917 * is authorized to enable/disable/clear RX statistics.
7919 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7922 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7924 rxi_rxstat_userok = proc;
7928 rx_RxStatUserOk(struct rx_call *call)
7930 if (!rxi_rxstat_userok)
7932 return rxi_rxstat_userok(call);
7937 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
7938 * function in the MSVC runtime DLL (msvcrt.dll).
7940 * Note: the system serializes calls to this function.
7943 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
7944 DWORD reason, /* reason function is being called */
7945 LPVOID reserved) /* reserved for future use */
7948 case DLL_PROCESS_ATTACH:
7949 /* library is being attached to a process */
7953 case DLL_PROCESS_DETACH: