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 struct clock queueTime;
1164 dpf(("rx_NewCall(conn %x)\n", conn));
1167 clock_GetTime(&queueTime);
1168 MUTEX_ENTER(&conn->conn_call_lock);
1171 * Check if there are others waiting for a new call.
1172 * If so, let them go first to avoid starving them.
1173 * This is a fairly simple scheme, and might not be
1174 * a complete solution for large numbers of waiters.
1176 * makeCallWaiters keeps track of the number of
1177 * threads waiting to make calls and the
1178 * RX_CONN_MAKECALL_WAITING flag bit is used to
1179 * indicate that there are indeed calls waiting.
1180 * The flag is set when the waiter is incremented.
1181 * It is only cleared in rx_EndCall when
1182 * makeCallWaiters is 0. This prevents us from
1183 * accidently destroying the connection while it
1184 * is potentially about to be used.
1186 MUTEX_ENTER(&conn->conn_data_lock);
1187 if (conn->makeCallWaiters) {
1188 conn->flags |= RX_CONN_MAKECALL_WAITING;
1189 conn->makeCallWaiters++;
1190 MUTEX_EXIT(&conn->conn_data_lock);
1192 #ifdef RX_ENABLE_LOCKS
1193 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1197 rx_MutexDecrement(conn->makeCallWaiters, conn->conn_data_lock);
1199 MUTEX_EXIT(&conn->conn_data_lock);
1202 /* search for next free call on this connection or
1203 * its clones, if any */
1205 register struct rx_connection *tconn;
1207 for (tconn = conn; tconn; tconn = tconn->next_clone) {
1208 for (i = 0; i < RX_MAXCALLS; i++) {
1209 call = tconn->call[i];
1211 MUTEX_ENTER(&call->lock);
1212 if (call->state == RX_STATE_DALLY) {
1213 rxi_ResetCall(call, 0);
1214 (*call->callNumber)++;
1217 MUTEX_EXIT(&call->lock);
1219 call = rxi_NewCall(tconn, i);
1222 } /* for i < RX_MAXCALLS */
1226 * to be here, all available calls for this connection (and all
1227 * of its clones) must be in use
1230 MUTEX_ENTER(&conn->conn_data_lock);
1231 conn->flags |= RX_CONN_MAKECALL_WAITING;
1232 conn->makeCallWaiters++;
1233 MUTEX_EXIT(&conn->conn_data_lock);
1235 #ifdef RX_ENABLE_LOCKS
1236 CV_WAIT(&conn->conn_call_cv, &conn->conn_call_lock);
1240 rx_MutexDecrement(conn->makeCallWaiters, conn->conn_data_lock);
1245 * Wake up anyone else who might be giving us a chance to
1246 * run (see code above that avoids resource starvation).
1248 #ifdef RX_ENABLE_LOCKS
1249 CV_BROADCAST(&conn->conn_call_cv);
1254 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1256 /* Client is initially in send mode */
1257 call->state = RX_STATE_ACTIVE;
1258 call->error = rx_ConnError(conn);
1260 call->mode = RX_MODE_ERROR;
1262 call->mode = RX_MODE_SENDING;
1264 /* remember start time for call in case we have hard dead time limit */
1265 call->queueTime = queueTime;
1266 clock_GetTime(&call->startTime);
1267 hzero(call->bytesSent);
1268 hzero(call->bytesRcvd);
1270 /* Turn on busy protocol. */
1271 rxi_KeepAliveOn(call);
1273 MUTEX_EXIT(&call->lock);
1274 MUTEX_EXIT(&conn->conn_call_lock);
1277 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
1278 /* Now, if TQ wasn't cleared earlier, do it now. */
1279 MUTEX_ENTER(&call->lock);
1280 rxi_WaitforTQBusy(call);
1281 if (call->flags & RX_CALL_TQ_CLEARME) {
1282 rxi_ClearTransmitQueue(call, 1);
1283 /*queue_Init(&call->tq);*/
1285 MUTEX_EXIT(&call->lock);
1286 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
1288 dpf(("rx_NewCall(call %x)\n", call));
1293 rxi_HasActiveCalls(register struct rx_connection *aconn)
1296 register struct rx_call *tcall;
1300 for (i = 0; i < RX_MAXCALLS; i++) {
1301 if ((tcall = aconn->call[i])) {
1302 if ((tcall->state == RX_STATE_ACTIVE)
1303 || (tcall->state == RX_STATE_PRECALL)) {
1314 rxi_GetCallNumberVector(register struct rx_connection *aconn,
1315 register afs_int32 * aint32s)
1318 register struct rx_call *tcall;
1322 for (i = 0; i < RX_MAXCALLS; i++) {
1323 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1324 aint32s[i] = aconn->callNumber[i] + 1;
1326 aint32s[i] = aconn->callNumber[i];
1333 rxi_SetCallNumberVector(register struct rx_connection *aconn,
1334 register afs_int32 * aint32s)
1337 register struct rx_call *tcall;
1341 for (i = 0; i < RX_MAXCALLS; i++) {
1342 if ((tcall = aconn->call[i]) && (tcall->state == RX_STATE_DALLY))
1343 aconn->callNumber[i] = aint32s[i] - 1;
1345 aconn->callNumber[i] = aint32s[i];
1351 /* Advertise a new service. A service is named locally by a UDP port
1352 * number plus a 16-bit service id. Returns (struct rx_service *) 0
1355 char *serviceName; Name for identification purposes (e.g. the
1356 service name might be used for probing for
1359 rx_NewServiceHost(afs_uint32 host, u_short port, u_short serviceId,
1360 char *serviceName, struct rx_securityClass **securityObjects,
1361 int nSecurityObjects,
1362 afs_int32(*serviceProc) (struct rx_call * acall))
1364 osi_socket socket = OSI_NULLSOCKET;
1365 register struct rx_service *tservice;
1371 if (serviceId == 0) {
1373 "rx_NewService: service id for service %s is not non-zero.\n",
1380 "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",
1388 tservice = rxi_AllocService();
1390 for (i = 0; i < RX_MAX_SERVICES; i++) {
1391 register struct rx_service *service = rx_services[i];
1393 if (port == service->servicePort && host == service->serviceHost) {
1394 if (service->serviceId == serviceId) {
1395 /* The identical service has already been
1396 * installed; if the caller was intending to
1397 * change the security classes used by this
1398 * service, he/she loses. */
1400 "rx_NewService: tried to install service %s with service id %d, which is already in use for service %s\n",
1401 serviceName, serviceId, service->serviceName);
1403 rxi_FreeService(tservice);
1406 /* Different service, same port: re-use the socket
1407 * which is bound to the same port */
1408 socket = service->socket;
1411 if (socket == OSI_NULLSOCKET) {
1412 /* If we don't already have a socket (from another
1413 * service on same port) get a new one */
1414 socket = rxi_GetHostUDPSocket(htonl(INADDR_ANY), port);
1415 if (socket == OSI_NULLSOCKET) {
1417 rxi_FreeService(tservice);
1422 service->socket = socket;
1423 service->serviceHost = host;
1424 service->servicePort = port;
1425 service->serviceId = serviceId;
1426 service->serviceName = serviceName;
1427 service->nSecurityObjects = nSecurityObjects;
1428 service->securityObjects = securityObjects;
1429 service->minProcs = 0;
1430 service->maxProcs = 1;
1431 service->idleDeadTime = 60;
1432 service->idleDeadErr = 0;
1433 service->connDeadTime = rx_connDeadTime;
1434 service->executeRequestProc = serviceProc;
1435 service->checkReach = 0;
1436 rx_services[i] = service; /* not visible until now */
1442 rxi_FreeService(tservice);
1443 (osi_Msg "rx_NewService: cannot support > %d services\n",
1448 /* Set configuration options for all of a service's security objects */
1451 rx_SetSecurityConfiguration(struct rx_service *service,
1452 rx_securityConfigVariables type,
1456 for (i = 0; i<service->nSecurityObjects; i++) {
1457 if (service->securityObjects[i]) {
1458 RXS_SetConfiguration(service->securityObjects[i], NULL, type,
1466 rx_NewService(u_short port, u_short serviceId, char *serviceName,
1467 struct rx_securityClass **securityObjects, int nSecurityObjects,
1468 afs_int32(*serviceProc) (struct rx_call * acall))
1470 return rx_NewServiceHost(htonl(INADDR_ANY), port, serviceId, serviceName, securityObjects, nSecurityObjects, serviceProc);
1473 /* Generic request processing loop. This routine should be called
1474 * by the implementation dependent rx_ServerProc. If socketp is
1475 * non-null, it will be set to the file descriptor that this thread
1476 * is now listening on. If socketp is null, this routine will never
1479 rxi_ServerProc(int threadID, struct rx_call *newcall, osi_socket * socketp)
1481 register struct rx_call *call;
1482 register afs_int32 code;
1483 register struct rx_service *tservice = NULL;
1490 call = rx_GetCall(threadID, tservice, socketp);
1491 if (socketp && *socketp != OSI_NULLSOCKET) {
1492 /* We are now a listener thread */
1497 /* if server is restarting( typically smooth shutdown) then do not
1498 * allow any new calls.
1501 if (rx_tranquil && (call != NULL)) {
1505 MUTEX_ENTER(&call->lock);
1507 rxi_CallError(call, RX_RESTARTING);
1508 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1510 MUTEX_EXIT(&call->lock);
1514 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1515 #ifdef RX_ENABLE_LOCKS
1517 #endif /* RX_ENABLE_LOCKS */
1518 afs_termState = AFSOP_STOP_AFS;
1519 afs_osi_Wakeup(&afs_termState);
1520 #ifdef RX_ENABLE_LOCKS
1522 #endif /* RX_ENABLE_LOCKS */
1527 tservice = call->conn->service;
1529 if (tservice->beforeProc)
1530 (*tservice->beforeProc) (call);
1532 code = call->conn->service->executeRequestProc(call);
1534 if (tservice->afterProc)
1535 (*tservice->afterProc) (call, code);
1537 rx_EndCall(call, code);
1538 rx_MutexIncrement(rxi_nCalls, rx_stats_mutex);
1544 rx_WakeupServerProcs(void)
1546 struct rx_serverQueueEntry *np, *tqp;
1550 MUTEX_ENTER(&rx_serverPool_lock);
1552 #ifdef RX_ENABLE_LOCKS
1553 if (rx_waitForPacket)
1554 CV_BROADCAST(&rx_waitForPacket->cv);
1555 #else /* RX_ENABLE_LOCKS */
1556 if (rx_waitForPacket)
1557 osi_rxWakeup(rx_waitForPacket);
1558 #endif /* RX_ENABLE_LOCKS */
1559 MUTEX_ENTER(&freeSQEList_lock);
1560 for (np = rx_FreeSQEList; np; np = tqp) {
1561 tqp = *(struct rx_serverQueueEntry **)np;
1562 #ifdef RX_ENABLE_LOCKS
1563 CV_BROADCAST(&np->cv);
1564 #else /* RX_ENABLE_LOCKS */
1566 #endif /* RX_ENABLE_LOCKS */
1568 MUTEX_EXIT(&freeSQEList_lock);
1569 for (queue_Scan(&rx_idleServerQueue, np, tqp, rx_serverQueueEntry)) {
1570 #ifdef RX_ENABLE_LOCKS
1571 CV_BROADCAST(&np->cv);
1572 #else /* RX_ENABLE_LOCKS */
1574 #endif /* RX_ENABLE_LOCKS */
1576 MUTEX_EXIT(&rx_serverPool_lock);
1581 * One thing that seems to happen is that all the server threads get
1582 * tied up on some empty or slow call, and then a whole bunch of calls
1583 * arrive at once, using up the packet pool, so now there are more
1584 * empty calls. The most critical resources here are server threads
1585 * and the free packet pool. The "doreclaim" code seems to help in
1586 * general. I think that eventually we arrive in this state: there
1587 * are lots of pending calls which do have all their packets present,
1588 * so they won't be reclaimed, are multi-packet calls, so they won't
1589 * be scheduled until later, and thus are tying up most of the free
1590 * packet pool for a very long time.
1592 * 1. schedule multi-packet calls if all the packets are present.
1593 * Probably CPU-bound operation, useful to return packets to pool.
1594 * Do what if there is a full window, but the last packet isn't here?
1595 * 3. preserve one thread which *only* runs "best" calls, otherwise
1596 * it sleeps and waits for that type of call.
1597 * 4. Don't necessarily reserve a whole window for each thread. In fact,
1598 * the current dataquota business is badly broken. The quota isn't adjusted
1599 * to reflect how many packets are presently queued for a running call.
1600 * So, when we schedule a queued call with a full window of packets queued
1601 * up for it, that *should* free up a window full of packets for other 2d-class
1602 * calls to be able to use from the packet pool. But it doesn't.
1604 * NB. Most of the time, this code doesn't run -- since idle server threads
1605 * sit on the idle server queue and are assigned by "...ReceivePacket" as soon
1606 * as a new call arrives.
1608 /* Sleep until a call arrives. Returns a pointer to the call, ready
1609 * for an rx_Read. */
1610 #ifdef RX_ENABLE_LOCKS
1612 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1614 struct rx_serverQueueEntry *sq;
1615 register struct rx_call *call = (struct rx_call *)0;
1616 struct rx_service *service = NULL;
1619 MUTEX_ENTER(&freeSQEList_lock);
1621 if ((sq = rx_FreeSQEList)) {
1622 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1623 MUTEX_EXIT(&freeSQEList_lock);
1624 } else { /* otherwise allocate a new one and return that */
1625 MUTEX_EXIT(&freeSQEList_lock);
1626 sq = (struct rx_serverQueueEntry *)
1627 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1628 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1629 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1632 MUTEX_ENTER(&rx_serverPool_lock);
1633 if (cur_service != NULL) {
1634 ReturnToServerPool(cur_service);
1637 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1638 register struct rx_call *tcall, *ncall, *choice2 = NULL;
1640 /* Scan for eligible incoming calls. A call is not eligible
1641 * if the maximum number of calls for its service type are
1642 * already executing */
1643 /* One thread will process calls FCFS (to prevent starvation),
1644 * while the other threads may run ahead looking for calls which
1645 * have all their input data available immediately. This helps
1646 * keep threads from blocking, waiting for data from the client. */
1647 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1648 service = tcall->conn->service;
1649 if (!QuotaOK(service)) {
1652 if (tno == rxi_fcfs_thread_num
1653 || !tcall->queue_item_header.next) {
1654 /* If we're the fcfs thread , then we'll just use
1655 * this call. If we haven't been able to find an optimal
1656 * choice, and we're at the end of the list, then use a
1657 * 2d choice if one has been identified. Otherwise... */
1658 call = (choice2 ? choice2 : tcall);
1659 service = call->conn->service;
1660 } else if (!queue_IsEmpty(&tcall->rq)) {
1661 struct rx_packet *rp;
1662 rp = queue_First(&tcall->rq, rx_packet);
1663 if (rp->header.seq == 1) {
1665 || (rp->header.flags & RX_LAST_PACKET)) {
1667 } else if (rxi_2dchoice && !choice2
1668 && !(tcall->flags & RX_CALL_CLEARED)
1669 && (tcall->rprev > rxi_HardAckRate)) {
1678 ReturnToServerPool(service);
1685 MUTEX_EXIT(&rx_serverPool_lock);
1686 MUTEX_ENTER(&call->lock);
1688 if (call->flags & RX_CALL_WAIT_PROC) {
1689 call->flags &= ~RX_CALL_WAIT_PROC;
1690 rx_MutexDecrement(rx_nWaiting, rx_stats_mutex);
1693 if (call->state != RX_STATE_PRECALL || call->error) {
1694 MUTEX_EXIT(&call->lock);
1695 MUTEX_ENTER(&rx_serverPool_lock);
1696 ReturnToServerPool(service);
1701 if (queue_IsEmpty(&call->rq)
1702 || queue_First(&call->rq, rx_packet)->header.seq != 1)
1703 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1705 CLEAR_CALL_QUEUE_LOCK(call);
1708 /* If there are no eligible incoming calls, add this process
1709 * to the idle server queue, to wait for one */
1713 *socketp = OSI_NULLSOCKET;
1715 sq->socketp = socketp;
1716 queue_Append(&rx_idleServerQueue, sq);
1717 #ifndef AFS_AIX41_ENV
1718 rx_waitForPacket = sq;
1720 rx_waitingForPacket = sq;
1721 #endif /* AFS_AIX41_ENV */
1723 CV_WAIT(&sq->cv, &rx_serverPool_lock);
1725 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1726 MUTEX_EXIT(&rx_serverPool_lock);
1727 return (struct rx_call *)0;
1730 } while (!(call = sq->newcall)
1731 && !(socketp && *socketp != OSI_NULLSOCKET));
1732 MUTEX_EXIT(&rx_serverPool_lock);
1734 MUTEX_ENTER(&call->lock);
1740 MUTEX_ENTER(&freeSQEList_lock);
1741 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1742 rx_FreeSQEList = sq;
1743 MUTEX_EXIT(&freeSQEList_lock);
1746 clock_GetTime(&call->startTime);
1747 call->state = RX_STATE_ACTIVE;
1748 call->mode = RX_MODE_RECEIVING;
1749 #ifdef RX_KERNEL_TRACE
1750 if (ICL_SETACTIVE(afs_iclSetp)) {
1751 int glockOwner = ISAFS_GLOCK();
1754 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1755 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1762 rxi_calltrace(RX_CALL_START, call);
1763 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1764 call->conn->service->servicePort, call->conn->service->serviceId,
1767 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
1768 MUTEX_EXIT(&call->lock);
1770 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1775 #else /* RX_ENABLE_LOCKS */
1777 rx_GetCall(int tno, struct rx_service *cur_service, osi_socket * socketp)
1779 struct rx_serverQueueEntry *sq;
1780 register struct rx_call *call = (struct rx_call *)0, *choice2;
1781 struct rx_service *service = NULL;
1785 MUTEX_ENTER(&freeSQEList_lock);
1787 if ((sq = rx_FreeSQEList)) {
1788 rx_FreeSQEList = *(struct rx_serverQueueEntry **)sq;
1789 MUTEX_EXIT(&freeSQEList_lock);
1790 } else { /* otherwise allocate a new one and return that */
1791 MUTEX_EXIT(&freeSQEList_lock);
1792 sq = (struct rx_serverQueueEntry *)
1793 rxi_Alloc(sizeof(struct rx_serverQueueEntry));
1794 MUTEX_INIT(&sq->lock, "server Queue lock", MUTEX_DEFAULT, 0);
1795 CV_INIT(&sq->cv, "server Queue lock", CV_DEFAULT, 0);
1797 MUTEX_ENTER(&sq->lock);
1799 if (cur_service != NULL) {
1800 cur_service->nRequestsRunning--;
1801 if (cur_service->nRequestsRunning < cur_service->minProcs)
1805 if (queue_IsNotEmpty(&rx_incomingCallQueue)) {
1806 register struct rx_call *tcall, *ncall;
1807 /* Scan for eligible incoming calls. A call is not eligible
1808 * if the maximum number of calls for its service type are
1809 * already executing */
1810 /* One thread will process calls FCFS (to prevent starvation),
1811 * while the other threads may run ahead looking for calls which
1812 * have all their input data available immediately. This helps
1813 * keep threads from blocking, waiting for data from the client. */
1814 choice2 = (struct rx_call *)0;
1815 for (queue_Scan(&rx_incomingCallQueue, tcall, ncall, rx_call)) {
1816 service = tcall->conn->service;
1817 if (QuotaOK(service)) {
1818 if (tno == rxi_fcfs_thread_num
1819 || !tcall->queue_item_header.next) {
1820 /* If we're the fcfs thread, then we'll just use
1821 * this call. If we haven't been able to find an optimal
1822 * choice, and we're at the end of the list, then use a
1823 * 2d choice if one has been identified. Otherwise... */
1824 call = (choice2 ? choice2 : tcall);
1825 service = call->conn->service;
1826 } else if (!queue_IsEmpty(&tcall->rq)) {
1827 struct rx_packet *rp;
1828 rp = queue_First(&tcall->rq, rx_packet);
1829 if (rp->header.seq == 1
1831 || (rp->header.flags & RX_LAST_PACKET))) {
1833 } else if (rxi_2dchoice && !choice2
1834 && !(tcall->flags & RX_CALL_CLEARED)
1835 && (tcall->rprev > rxi_HardAckRate)) {
1848 /* we can't schedule a call if there's no data!!! */
1849 /* send an ack if there's no data, if we're missing the
1850 * first packet, or we're missing something between first
1851 * and last -- there's a "hole" in the incoming data. */
1852 if (queue_IsEmpty(&call->rq)
1853 || queue_First(&call->rq, rx_packet)->header.seq != 1
1854 || call->rprev != queue_Last(&call->rq, rx_packet)->header.seq)
1855 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
1857 call->flags &= (~RX_CALL_WAIT_PROC);
1858 service->nRequestsRunning++;
1859 /* just started call in minProcs pool, need fewer to maintain
1861 if (service->nRequestsRunning <= service->minProcs)
1865 /* MUTEX_EXIT(&call->lock); */
1867 /* If there are no eligible incoming calls, add this process
1868 * to the idle server queue, to wait for one */
1871 *socketp = OSI_NULLSOCKET;
1873 sq->socketp = socketp;
1874 queue_Append(&rx_idleServerQueue, sq);
1878 if (afs_termState == AFSOP_STOP_RXCALLBACK) {
1880 rxi_Free(sq, sizeof(struct rx_serverQueueEntry));
1881 return (struct rx_call *)0;
1884 } while (!(call = sq->newcall)
1885 && !(socketp && *socketp != OSI_NULLSOCKET));
1887 MUTEX_EXIT(&sq->lock);
1889 MUTEX_ENTER(&freeSQEList_lock);
1890 *(struct rx_serverQueueEntry **)sq = rx_FreeSQEList;
1891 rx_FreeSQEList = sq;
1892 MUTEX_EXIT(&freeSQEList_lock);
1895 clock_GetTime(&call->startTime);
1896 call->state = RX_STATE_ACTIVE;
1897 call->mode = RX_MODE_RECEIVING;
1898 #ifdef RX_KERNEL_TRACE
1899 if (ICL_SETACTIVE(afs_iclSetp)) {
1900 int glockOwner = ISAFS_GLOCK();
1903 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
1904 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
1911 rxi_calltrace(RX_CALL_START, call);
1912 dpf(("rx_GetCall(port=%d, service=%d) ==> call %x\n",
1913 call->conn->service->servicePort, call->conn->service->serviceId,
1916 dpf(("rx_GetCall(socketp=0x%x, *socketp=0x%x)\n", socketp, *socketp));
1923 #endif /* RX_ENABLE_LOCKS */
1927 /* Establish a procedure to be called when a packet arrives for a
1928 * call. This routine will be called at most once after each call,
1929 * and will also be called if there is an error condition on the or
1930 * the call is complete. Used by multi rx to build a selection
1931 * function which determines which of several calls is likely to be a
1932 * good one to read from.
1933 * NOTE: the way this is currently implemented it is probably only a
1934 * good idea to (1) use it immediately after a newcall (clients only)
1935 * and (2) only use it once. Other uses currently void your warranty
1938 rx_SetArrivalProc(register struct rx_call *call,
1939 register void (*proc) (register struct rx_call * call,
1941 register int index),
1942 register void * handle, register int arg)
1944 call->arrivalProc = proc;
1945 call->arrivalProcHandle = handle;
1946 call->arrivalProcArg = arg;
1949 /* Call is finished (possibly prematurely). Return rc to the peer, if
1950 * appropriate, and return the final error code from the conversation
1954 rx_EndCall(register struct rx_call *call, afs_int32 rc)
1956 register struct rx_connection *conn = call->conn;
1957 register struct rx_service *service;
1963 dpf(("rx_EndCall(call %x rc %d error %d abortCode %d)\n", call, rc, call->error, call->abortCode));
1966 MUTEX_ENTER(&call->lock);
1968 if (rc == 0 && call->error == 0) {
1969 call->abortCode = 0;
1970 call->abortCount = 0;
1973 call->arrivalProc = (void (*)())0;
1974 if (rc && call->error == 0) {
1975 rxi_CallError(call, rc);
1976 /* Send an abort message to the peer if this error code has
1977 * only just been set. If it was set previously, assume the
1978 * peer has already been sent the error code or will request it
1980 rxi_SendCallAbort(call, (struct rx_packet *)0, 0, 0);
1982 if (conn->type == RX_SERVER_CONNECTION) {
1983 /* Make sure reply or at least dummy reply is sent */
1984 if (call->mode == RX_MODE_RECEIVING) {
1985 rxi_WriteProc(call, 0, 0);
1987 if (call->mode == RX_MODE_SENDING) {
1988 rxi_FlushWrite(call);
1990 service = conn->service;
1991 rxi_calltrace(RX_CALL_END, call);
1992 /* Call goes to hold state until reply packets are acknowledged */
1993 if (call->tfirst + call->nSoftAcked < call->tnext) {
1994 call->state = RX_STATE_HOLD;
1996 call->state = RX_STATE_DALLY;
1997 rxi_ClearTransmitQueue(call, 0);
1998 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
1999 rxevent_Cancel(call->keepAliveEvent, call,
2000 RX_CALL_REFCOUNT_ALIVE);
2002 } else { /* Client connection */
2004 /* Make sure server receives input packets, in the case where
2005 * no reply arguments are expected */
2006 if ((call->mode == RX_MODE_SENDING)
2007 || (call->mode == RX_MODE_RECEIVING && call->rnext == 1)) {
2008 (void)rxi_ReadProc(call, &dummy, 1);
2011 /* If we had an outstanding delayed ack, be nice to the server
2012 * and force-send it now.
2014 if (call->delayedAckEvent) {
2015 rxevent_Cancel(call->delayedAckEvent, call,
2016 RX_CALL_REFCOUNT_DELAY);
2017 call->delayedAckEvent = NULL;
2018 rxi_SendDelayedAck(NULL, call, NULL);
2021 /* We need to release the call lock since it's lower than the
2022 * conn_call_lock and we don't want to hold the conn_call_lock
2023 * over the rx_ReadProc call. The conn_call_lock needs to be held
2024 * here for the case where rx_NewCall is perusing the calls on
2025 * the connection structure. We don't want to signal until
2026 * rx_NewCall is in a stable state. Otherwise, rx_NewCall may
2027 * have checked this call, found it active and by the time it
2028 * goes to sleep, will have missed the signal.
2030 * Do not clear the RX_CONN_MAKECALL_WAITING flag as long as
2031 * there are threads waiting to use the conn object.
2033 MUTEX_EXIT(&call->lock);
2034 MUTEX_ENTER(&conn->conn_call_lock);
2035 MUTEX_ENTER(&call->lock);
2036 MUTEX_ENTER(&conn->conn_data_lock);
2037 conn->flags |= RX_CONN_BUSY;
2038 if (conn->flags & RX_CONN_MAKECALL_WAITING) {
2039 if (conn->makeCallWaiters == 0)
2040 conn->flags &= (~RX_CONN_MAKECALL_WAITING);
2041 MUTEX_EXIT(&conn->conn_data_lock);
2042 #ifdef RX_ENABLE_LOCKS
2043 CV_BROADCAST(&conn->conn_call_cv);
2048 #ifdef RX_ENABLE_LOCKS
2050 MUTEX_EXIT(&conn->conn_data_lock);
2052 #endif /* RX_ENABLE_LOCKS */
2053 call->state = RX_STATE_DALLY;
2055 error = call->error;
2057 /* currentPacket, nLeft, and NFree must be zeroed here, because
2058 * ResetCall cannot: ResetCall may be called at splnet(), in the
2059 * kernel version, and may interrupt the macros rx_Read or
2060 * rx_Write, which run at normal priority for efficiency. */
2061 if (call->currentPacket) {
2062 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
2063 rxi_FreePacket(call->currentPacket);
2064 call->currentPacket = (struct rx_packet *)0;
2067 call->nLeft = call->nFree = call->curlen = 0;
2069 /* Free any packets from the last call to ReadvProc/WritevProc */
2070 #ifdef RXDEBUG_PACKET
2072 #endif /* RXDEBUG_PACKET */
2073 rxi_FreePackets(0, &call->iovq);
2075 CALL_RELE(call, RX_CALL_REFCOUNT_BEGIN);
2076 MUTEX_EXIT(&call->lock);
2077 if (conn->type == RX_CLIENT_CONNECTION) {
2078 MUTEX_EXIT(&conn->conn_call_lock);
2079 conn->flags &= ~RX_CONN_BUSY;
2083 * Map errors to the local host's errno.h format.
2085 error = ntoh_syserr_conv(error);
2089 #if !defined(KERNEL)
2091 /* Call this routine when shutting down a server or client (especially
2092 * clients). This will allow Rx to gracefully garbage collect server
2093 * connections, and reduce the number of retries that a server might
2094 * make to a dead client.
2095 * This is not quite right, since some calls may still be ongoing and
2096 * we can't lock them to destroy them. */
2100 register struct rx_connection **conn_ptr, **conn_end;
2104 if (rxinit_status == 1) {
2106 return; /* Already shutdown. */
2108 rxi_DeleteCachedConnections();
2109 if (rx_connHashTable) {
2110 RWLOCK_WRLOCK(&rx_connHashTable_lock);
2111 for (conn_ptr = &rx_connHashTable[0], conn_end =
2112 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
2114 struct rx_connection *conn, *next;
2115 for (conn = *conn_ptr; conn; conn = next) {
2117 if (conn->type == RX_CLIENT_CONNECTION) {
2118 /* MUTEX_ENTER(&conn->conn_data_lock); when used in kernel */
2120 /* MUTEX_EXIT(&conn->conn_data_lock); when used in kernel */
2121 #ifdef RX_ENABLE_LOCKS
2122 rxi_DestroyConnectionNoLock(conn);
2123 #else /* RX_ENABLE_LOCKS */
2124 rxi_DestroyConnection(conn);
2125 #endif /* RX_ENABLE_LOCKS */
2129 #ifdef RX_ENABLE_LOCKS
2130 while (rx_connCleanup_list) {
2131 struct rx_connection *conn;
2132 conn = rx_connCleanup_list;
2133 rx_connCleanup_list = rx_connCleanup_list->next;
2134 RWLOCK_UNLOCK(&rx_connHashTable_lock);
2135 rxi_CleanupConnection(conn);
2136 RWLOCK_WRLOCK(&rx_connHashTable_lock);
2138 RWLOCK_UNLOCK(&rx_connHashTable_lock);
2139 #endif /* RX_ENABLE_LOCKS */
2144 afs_winsockCleanup();
2152 /* if we wakeup packet waiter too often, can get in loop with two
2153 AllocSendPackets each waking each other up (from ReclaimPacket calls) */
2155 rxi_PacketsUnWait(void)
2157 if (!rx_waitingForPackets) {
2161 if (rxi_OverQuota(RX_PACKET_CLASS_SEND)) {
2162 return; /* still over quota */
2165 rx_waitingForPackets = 0;
2166 #ifdef RX_ENABLE_LOCKS
2167 CV_BROADCAST(&rx_waitingForPackets_cv);
2169 osi_rxWakeup(&rx_waitingForPackets);
2175 /* ------------------Internal interfaces------------------------- */
2177 /* Return this process's service structure for the
2178 * specified socket and service */
2180 rxi_FindService(register osi_socket socket, register u_short serviceId)
2182 register struct rx_service **sp;
2183 for (sp = &rx_services[0]; *sp; sp++) {
2184 if ((*sp)->serviceId == serviceId && (*sp)->socket == socket)
2191 #ifdef KDUMP_RX_LOCK
2192 static struct rx_call_rx_lock *rx_allCallsp = 0;
2194 static struct rx_call *rx_allCallsp = 0;
2198 /* Allocate a call structure, for the indicated channel of the
2199 * supplied connection. The mode and state of the call must be set by
2200 * the caller. Returns the call with mutex locked. */
2202 rxi_NewCall(register struct rx_connection *conn, register int channel)
2204 register struct rx_call *call;
2205 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2206 register struct rx_call *cp; /* Call pointer temp */
2207 register struct rx_call *nxp; /* Next call pointer, for queue_Scan */
2208 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2210 dpf(("rxi_NewCall(conn %x, channel %d)\n", conn, channel));
2212 /* Grab an existing call structure, or allocate a new one.
2213 * Existing call structures are assumed to have been left reset by
2215 MUTEX_ENTER(&rx_freeCallQueue_lock);
2217 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2219 * EXCEPT that the TQ might not yet be cleared out.
2220 * Skip over those with in-use TQs.
2223 for (queue_Scan(&rx_freeCallQueue, cp, nxp, rx_call)) {
2224 if (!(cp->flags & RX_CALL_TQ_BUSY)) {
2230 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2231 if (queue_IsNotEmpty(&rx_freeCallQueue)) {
2232 call = queue_First(&rx_freeCallQueue, rx_call);
2233 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2235 rx_MutexDecrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2236 MUTEX_EXIT(&rx_freeCallQueue_lock);
2237 MUTEX_ENTER(&call->lock);
2238 CLEAR_CALL_QUEUE_LOCK(call);
2239 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2240 /* Now, if TQ wasn't cleared earlier, do it now. */
2241 if (call->flags & RX_CALL_TQ_CLEARME) {
2242 rxi_ClearTransmitQueue(call, 1);
2243 /*queue_Init(&call->tq);*/
2245 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2246 /* Bind the call to its connection structure */
2248 rxi_ResetCall(call, 1);
2251 call = (struct rx_call *)rxi_Alloc(sizeof(struct rx_call));
2252 #ifdef RXDEBUG_PACKET
2253 call->allNextp = rx_allCallsp;
2254 rx_allCallsp = call;
2256 #endif /* RXDEBUG_PACKET */
2257 rx_MutexIncrement(rx_stats.nCallStructs, rx_stats_mutex);
2259 MUTEX_EXIT(&rx_freeCallQueue_lock);
2260 MUTEX_INIT(&call->lock, "call lock", MUTEX_DEFAULT, NULL);
2261 MUTEX_ENTER(&call->lock);
2262 CV_INIT(&call->cv_twind, "call twind", CV_DEFAULT, 0);
2263 CV_INIT(&call->cv_rq, "call rq", CV_DEFAULT, 0);
2264 CV_INIT(&call->cv_tq, "call tq", CV_DEFAULT, 0);
2266 /* Initialize once-only items */
2267 queue_Init(&call->tq);
2268 queue_Init(&call->rq);
2269 queue_Init(&call->iovq);
2270 #ifdef RXDEBUG_PACKET
2271 call->rqc = call->tqc = call->iovqc = 0;
2272 #endif /* RXDEBUG_PACKET */
2273 /* Bind the call to its connection structure (prereq for reset) */
2275 rxi_ResetCall(call, 1);
2277 call->channel = channel;
2278 call->callNumber = &conn->callNumber[channel];
2279 call->rwind = conn->rwind[channel];
2280 call->twind = conn->twind[channel];
2281 /* Note that the next expected call number is retained (in
2282 * conn->callNumber[i]), even if we reallocate the call structure
2284 conn->call[channel] = call;
2285 /* if the channel's never been used (== 0), we should start at 1, otherwise
2286 * the call number is valid from the last time this channel was used */
2287 if (*call->callNumber == 0)
2288 *call->callNumber = 1;
2293 /* A call has been inactive long enough that so we can throw away
2294 * state, including the call structure, which is placed on the call
2296 * Call is locked upon entry.
2297 * haveCTLock set if called from rxi_ReapConnections
2299 #ifdef RX_ENABLE_LOCKS
2301 rxi_FreeCall(register struct rx_call *call, int haveCTLock)
2302 #else /* RX_ENABLE_LOCKS */
2304 rxi_FreeCall(register struct rx_call *call)
2305 #endif /* RX_ENABLE_LOCKS */
2307 register int channel = call->channel;
2308 register struct rx_connection *conn = call->conn;
2311 if (call->state == RX_STATE_DALLY || call->state == RX_STATE_HOLD)
2312 (*call->callNumber)++;
2313 rxi_ResetCall(call, 0);
2314 call->conn->call[channel] = (struct rx_call *)0;
2316 MUTEX_ENTER(&rx_freeCallQueue_lock);
2317 SET_CALL_QUEUE_LOCK(call, &rx_freeCallQueue_lock);
2318 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2319 /* A call may be free even though its transmit queue is still in use.
2320 * Since we search the call list from head to tail, put busy calls at
2321 * the head of the list, and idle calls at the tail.
2323 if (call->flags & RX_CALL_TQ_BUSY)
2324 queue_Prepend(&rx_freeCallQueue, call);
2326 queue_Append(&rx_freeCallQueue, call);
2327 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2328 queue_Append(&rx_freeCallQueue, call);
2329 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2330 rx_MutexIncrement(rx_stats.nFreeCallStructs, rx_stats_mutex);
2331 MUTEX_EXIT(&rx_freeCallQueue_lock);
2333 /* Destroy the connection if it was previously slated for
2334 * destruction, i.e. the Rx client code previously called
2335 * rx_DestroyConnection (client connections), or
2336 * rxi_ReapConnections called the same routine (server
2337 * connections). Only do this, however, if there are no
2338 * outstanding calls. Note that for fine grain locking, there appears
2339 * to be a deadlock in that rxi_FreeCall has a call locked and
2340 * DestroyConnectionNoLock locks each call in the conn. But note a
2341 * few lines up where we have removed this call from the conn.
2342 * If someone else destroys a connection, they either have no
2343 * call lock held or are going through this section of code.
2345 if (conn->flags & RX_CONN_DESTROY_ME && !(conn->flags & RX_CONN_MAKECALL_WAITING)) {
2346 rx_MutexIncrement(conn->refCount, conn->conn_data_lock);
2347 #ifdef RX_ENABLE_LOCKS
2349 rxi_DestroyConnectionNoLock(conn);
2351 rxi_DestroyConnection(conn);
2352 #else /* RX_ENABLE_LOCKS */
2353 rxi_DestroyConnection(conn);
2354 #endif /* RX_ENABLE_LOCKS */
2358 afs_int32 rxi_Alloccnt = 0, rxi_Allocsize = 0;
2360 rxi_Alloc(register size_t size)
2364 rx_MutexAdd1Increment2(rxi_Allocsize, (afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2367 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
2368 afs_osi_Alloc_NoSleep(size);
2373 osi_Panic("rxi_Alloc error");
2379 rxi_Free(void *addr, register size_t size)
2381 rx_MutexAdd1Decrement2(rxi_Allocsize, -(afs_int32)size, rxi_Alloccnt, rx_stats_mutex);
2382 osi_Free(addr, size);
2386 rxi_SetPeerMtu(register afs_uint32 host, register afs_uint32 port, int mtu)
2388 struct rx_peer **peer_ptr, **peer_end;
2391 RWLOCK_RDLOCK(&rx_peerHashTable_lock);
2393 for (peer_ptr = &rx_peerHashTable[0], peer_end =
2394 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
2396 struct rx_peer *peer, *next;
2397 for (peer = *peer_ptr; peer; peer = next) {
2399 if (host == peer->host) {
2400 MUTEX_ENTER(&peer->peer_lock);
2401 peer->ifMTU=MIN(mtu, peer->ifMTU);
2402 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2403 MUTEX_EXIT(&peer->peer_lock);
2408 struct rx_peer *peer;
2409 hashIndex = PEER_HASH(host, port);
2410 for (peer = rx_peerHashTable[hashIndex]; peer; peer = peer->next) {
2411 if ((peer->host == host) && (peer->port == port)) {
2412 MUTEX_ENTER(&peer->peer_lock);
2413 peer->ifMTU=MIN(mtu, peer->ifMTU);
2414 peer->natMTU = rxi_AdjustIfMTU(peer->ifMTU);
2415 MUTEX_EXIT(&peer->peer_lock);
2419 RWLOCK_UNLOCK(&rx_peerHashTable_lock);
2422 /* Find the peer process represented by the supplied (host,port)
2423 * combination. If there is no appropriate active peer structure, a
2424 * new one will be allocated and initialized
2425 * The origPeer, if set, is a pointer to a peer structure on which the
2426 * refcount will be be decremented. This is used to replace the peer
2427 * structure hanging off a connection structure */
2429 rxi_FindPeer(register afs_uint32 host, register u_short port,
2430 struct rx_peer *origPeer, int create)
2432 register struct rx_peer *pp;
2434 hashIndex = PEER_HASH(host, port);
2435 RWLOCK_RDLOCK(&rx_peerHashTable_lock);
2436 for (pp = rx_peerHashTable[hashIndex]; pp; pp = pp->next) {
2437 if ((pp->host == host) && (pp->port == port))
2442 pp = rxi_AllocPeer(); /* This bzero's *pp */
2443 pp->host = host; /* set here or in InitPeerParams is zero */
2445 MUTEX_INIT(&pp->peer_lock, "peer_lock", MUTEX_DEFAULT, 0);
2446 queue_Init(&pp->congestionQueue);
2447 queue_Init(&pp->rpcStats);
2448 RWLOCK_UPLOCK(&rx_peerHashTable_lock);
2449 pp->next = rx_peerHashTable[hashIndex];
2450 rx_peerHashTable[hashIndex] = pp;
2451 rxi_InitPeerParams(pp);
2452 rx_MutexIncrement(rx_stats.nPeerStructs, rx_stats_mutex);
2459 origPeer->refCount--;
2460 RWLOCK_UNLOCK(&rx_peerHashTable_lock);
2465 /* Find the connection at (host, port) started at epoch, and with the
2466 * given connection id. Creates the server connection if necessary.
2467 * The type specifies whether a client connection or a server
2468 * connection is desired. In both cases, (host, port) specify the
2469 * peer's (host, pair) pair. Client connections are not made
2470 * automatically by this routine. The parameter socket gives the
2471 * socket descriptor on which the packet was received. This is used,
2472 * in the case of server connections, to check that *new* connections
2473 * come via a valid (port, serviceId). Finally, the securityIndex
2474 * parameter must match the existing index for the connection. If a
2475 * server connection is created, it will be created using the supplied
2476 * index, if the index is valid for this service */
2477 struct rx_connection *
2478 rxi_FindConnection(osi_socket socket, register afs_int32 host,
2479 register u_short port, u_short serviceId, afs_uint32 cid,
2480 afs_uint32 epoch, int type, u_int securityIndex)
2482 int hashindex, flag, i;
2483 register struct rx_connection *conn;
2484 hashindex = CONN_HASH(host, port, cid, epoch, type);
2485 RWLOCK_RDLOCK(&rx_connHashTable_lock);
2486 rxLastConn ? (conn = rxLastConn, flag = 0) : (conn =
2487 rx_connHashTable[hashindex],
2490 if ((conn->type == type) && ((cid & RX_CIDMASK) == conn->cid)
2491 && (epoch == conn->epoch)) {
2492 register struct rx_peer *pp = conn->peer;
2493 if (securityIndex != conn->securityIndex) {
2494 /* this isn't supposed to happen, but someone could forge a packet
2495 * like this, and there seems to be some CM bug that makes this
2496 * happen from time to time -- in which case, the fileserver
2498 RWLOCK_UNLOCK(&rx_connHashTable_lock);
2499 return (struct rx_connection *)0;
2501 if (pp->host == host && pp->port == port)
2503 if (type == RX_CLIENT_CONNECTION && pp->port == port)
2505 /* So what happens when it's a callback connection? */
2506 if ( /*type == RX_CLIENT_CONNECTION && */
2507 (conn->epoch & 0x80000000))
2511 /* the connection rxLastConn that was used the last time is not the
2512 ** one we are looking for now. Hence, start searching in the hash */
2514 conn = rx_connHashTable[hashindex];
2519 struct rx_service *service;
2520 if (type == RX_CLIENT_CONNECTION) {
2521 RWLOCK_UNLOCK(&rx_connHashTable_lock);
2522 return (struct rx_connection *)0;
2524 service = rxi_FindService(socket, serviceId);
2525 if (!service || (securityIndex >= service->nSecurityObjects)
2526 || (service->securityObjects[securityIndex] == 0)) {
2527 RWLOCK_UNLOCK(&rx_connHashTable_lock);
2528 return (struct rx_connection *)0;
2530 conn = rxi_AllocConnection(); /* This bzero's the connection */
2531 MUTEX_INIT(&conn->conn_call_lock, "conn call lock", MUTEX_DEFAULT, 0);
2532 MUTEX_INIT(&conn->conn_data_lock, "conn data lock", MUTEX_DEFAULT, 0);
2533 CV_INIT(&conn->conn_call_cv, "conn call cv", CV_DEFAULT, 0);
2534 conn->peer = rxi_FindPeer(host, port, 0, 1);
2535 conn->type = RX_SERVER_CONNECTION;
2536 conn->lastSendTime = clock_Sec(); /* don't GC immediately */
2537 conn->epoch = epoch;
2538 /* conn->serial = conn->lastSerial = 0; */
2539 /* conn->timeout = 0; */
2540 conn->ackRate = RX_FAST_ACK_RATE;
2541 conn->service = service;
2542 conn->serviceId = serviceId;
2543 conn->securityIndex = securityIndex;
2544 conn->securityObject = service->securityObjects[securityIndex];
2545 conn->nSpecific = 0;
2546 conn->specific = NULL;
2547 rx_SetConnDeadTime(conn, service->connDeadTime);
2548 rx_SetConnIdleDeadTime(conn, service->idleDeadTime);
2549 rx_SetServerConnIdleDeadErr(conn, service->idleDeadErr);
2550 for (i = 0; i < RX_MAXCALLS; i++) {
2551 conn->twind[i] = rx_initSendWindow;
2552 conn->rwind[i] = rx_initReceiveWindow;
2554 RWLOCK_UPLOCK(&rx_connHashTable_lock);
2555 conn->next = rx_connHashTable[hashindex];
2556 rx_connHashTable[hashindex] = conn;
2557 conn->cid = cid & RX_CIDMASK;
2558 /* Notify security object of the new connection */
2559 RXS_NewConnection(conn->securityObject, conn);
2560 /* XXXX Connection timeout? */
2561 if (service->newConnProc)
2562 (*service->newConnProc) (conn);
2563 rx_MutexIncrement(rx_stats.nServerConns, rx_stats_mutex);
2566 rx_MutexIncrement(conn->refCount, conn->conn_data_lock);
2568 rxLastConn = conn; /* store this connection as the last conn used */
2569 RWLOCK_UNLOCK(&rx_connHashTable_lock);
2573 /* There are two packet tracing routines available for testing and monitoring
2574 * Rx. One is called just after every packet is received and the other is
2575 * called just before every packet is sent. Received packets, have had their
2576 * headers decoded, and packets to be sent have not yet had their headers
2577 * encoded. Both take two parameters: a pointer to the packet and a sockaddr
2578 * containing the network address. Both can be modified. The return value, if
2579 * non-zero, indicates that the packet should be dropped. */
2581 int (*rx_justReceived) (struct rx_packet *, struct sockaddr_in *) = 0;
2582 int (*rx_almostSent) (struct rx_packet *, struct sockaddr_in *) = 0;
2584 /* A packet has been received off the interface. Np is the packet, socket is
2585 * the socket number it was received from (useful in determining which service
2586 * this packet corresponds to), and (host, port) reflect the host,port of the
2587 * sender. This call returns the packet to the caller if it is finished with
2588 * it, rather than de-allocating it, just as a small performance hack */
2591 rxi_ReceivePacket(register struct rx_packet *np, osi_socket socket,
2592 afs_uint32 host, u_short port, int *tnop,
2593 struct rx_call **newcallp)
2595 register struct rx_call *call;
2596 register struct rx_connection *conn;
2598 afs_uint32 currentCallNumber;
2604 struct rx_packet *tnp;
2607 /* We don't print out the packet until now because (1) the time may not be
2608 * accurate enough until now in the lwp implementation (rx_Listener only gets
2609 * the time after the packet is read) and (2) from a protocol point of view,
2610 * this is the first time the packet has been seen */
2611 packetType = (np->header.type > 0 && np->header.type < RX_N_PACKET_TYPES)
2612 ? rx_packetTypes[np->header.type - 1] : "*UNKNOWN*";
2613 dpf(("R %d %s: %x.%d.%d.%d.%d.%d.%d flags %d, packet %x",
2614 np->header.serial, packetType, ntohl(host), ntohs(port), np->header.serviceId,
2615 np->header.epoch, np->header.cid, np->header.callNumber,
2616 np->header.seq, np->header.flags, np));
2619 if (np->header.type == RX_PACKET_TYPE_VERSION) {
2620 return rxi_ReceiveVersionPacket(np, socket, host, port, 1);
2623 if (np->header.type == RX_PACKET_TYPE_DEBUG) {
2624 return rxi_ReceiveDebugPacket(np, socket, host, port, 1);
2627 /* If an input tracer function is defined, call it with the packet and
2628 * network address. Note this function may modify its arguments. */
2629 if (rx_justReceived) {
2630 struct sockaddr_in addr;
2632 addr.sin_family = AF_INET;
2633 addr.sin_port = port;
2634 addr.sin_addr.s_addr = host;
2635 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
2636 addr.sin_len = sizeof(addr);
2637 #endif /* AFS_OSF_ENV */
2638 drop = (*rx_justReceived) (np, &addr);
2639 /* drop packet if return value is non-zero */
2642 port = addr.sin_port; /* in case fcn changed addr */
2643 host = addr.sin_addr.s_addr;
2647 /* If packet was not sent by the client, then *we* must be the client */
2648 type = ((np->header.flags & RX_CLIENT_INITIATED) != RX_CLIENT_INITIATED)
2649 ? RX_CLIENT_CONNECTION : RX_SERVER_CONNECTION;
2651 /* Find the connection (or fabricate one, if we're the server & if
2652 * necessary) associated with this packet */
2654 rxi_FindConnection(socket, host, port, np->header.serviceId,
2655 np->header.cid, np->header.epoch, type,
2656 np->header.securityIndex);
2659 /* If no connection found or fabricated, just ignore the packet.
2660 * (An argument could be made for sending an abort packet for
2665 MUTEX_ENTER(&conn->conn_data_lock);
2666 if (conn->maxSerial < np->header.serial)
2667 conn->maxSerial = np->header.serial;
2668 MUTEX_EXIT(&conn->conn_data_lock);
2670 /* If the connection is in an error state, send an abort packet and ignore
2671 * the incoming packet */
2672 if (rx_ConnError(conn)) {
2673 /* Don't respond to an abort packet--we don't want loops! */
2674 MUTEX_ENTER(&conn->conn_data_lock);
2675 if (np->header.type != RX_PACKET_TYPE_ABORT)
2676 np = rxi_SendConnectionAbort(conn, np, 1, 0);
2678 MUTEX_EXIT(&conn->conn_data_lock);
2682 /* Check for connection-only requests (i.e. not call specific). */
2683 if (np->header.callNumber == 0) {
2684 switch (np->header.type) {
2685 case RX_PACKET_TYPE_ABORT: {
2686 /* What if the supplied error is zero? */
2687 afs_int32 errcode = ntohl(rx_GetInt32(np, 0));
2688 dpf(("rxi_ReceivePacket ABORT rx_GetInt32 = %d", errcode));
2689 rxi_ConnectionError(conn, errcode);
2690 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2693 case RX_PACKET_TYPE_CHALLENGE:
2694 tnp = rxi_ReceiveChallengePacket(conn, np, 1);
2695 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2697 case RX_PACKET_TYPE_RESPONSE:
2698 tnp = rxi_ReceiveResponsePacket(conn, np, 1);
2699 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2701 case RX_PACKET_TYPE_PARAMS:
2702 case RX_PACKET_TYPE_PARAMS + 1:
2703 case RX_PACKET_TYPE_PARAMS + 2:
2704 /* ignore these packet types for now */
2705 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2710 /* Should not reach here, unless the peer is broken: send an
2712 rxi_ConnectionError(conn, RX_PROTOCOL_ERROR);
2713 MUTEX_ENTER(&conn->conn_data_lock);
2714 tnp = rxi_SendConnectionAbort(conn, np, 1, 0);
2716 MUTEX_EXIT(&conn->conn_data_lock);
2721 channel = np->header.cid & RX_CHANNELMASK;
2722 call = conn->call[channel];
2723 #ifdef RX_ENABLE_LOCKS
2725 MUTEX_ENTER(&call->lock);
2726 /* Test to see if call struct is still attached to conn. */
2727 if (call != conn->call[channel]) {
2729 MUTEX_EXIT(&call->lock);
2730 if (type == RX_SERVER_CONNECTION) {
2731 call = conn->call[channel];
2732 /* If we started with no call attached and there is one now,
2733 * another thread is also running this routine and has gotten
2734 * the connection channel. We should drop this packet in the tests
2735 * below. If there was a call on this connection and it's now
2736 * gone, then we'll be making a new call below.
2737 * If there was previously a call and it's now different then
2738 * the old call was freed and another thread running this routine
2739 * has created a call on this channel. One of these two threads
2740 * has a packet for the old call and the code below handles those
2744 MUTEX_ENTER(&call->lock);
2746 /* This packet can't be for this call. If the new call address is
2747 * 0 then no call is running on this channel. If there is a call
2748 * then, since this is a client connection we're getting data for
2749 * it must be for the previous call.
2751 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2752 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2757 currentCallNumber = conn->callNumber[channel];
2759 if (type == RX_SERVER_CONNECTION) { /* We're the server */
2760 if (np->header.callNumber < currentCallNumber) {
2761 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2762 #ifdef RX_ENABLE_LOCKS
2764 MUTEX_EXIT(&call->lock);
2766 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2770 MUTEX_ENTER(&conn->conn_call_lock);
2771 call = rxi_NewCall(conn, channel);
2772 MUTEX_EXIT(&conn->conn_call_lock);
2773 *call->callNumber = np->header.callNumber;
2774 if (np->header.callNumber == 0)
2775 dpf(("RecPacket call 0 %d %s: %x.%u.%u.%u.%u.%u.%u flags %d, packet %lx resend %d.%0.3d len %d", np->header.serial, rx_packetTypes[np->header.type - 1], ntohl(conn->peer->host), ntohs(conn->peer->port), np->header.serial, np->header.epoch, np->header.cid, np->header.callNumber, np->header.seq, np->header.flags, (unsigned long)np, np->retryTime.sec, np->retryTime.usec / 1000, np->length));
2777 call->state = RX_STATE_PRECALL;
2778 clock_GetTime(&call->queueTime);
2779 hzero(call->bytesSent);
2780 hzero(call->bytesRcvd);
2782 * If the number of queued calls exceeds the overload
2783 * threshold then abort this call.
2785 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2786 struct rx_packet *tp;
2788 rxi_CallError(call, rx_BusyError);
2789 tp = rxi_SendCallAbort(call, np, 1, 0);
2790 MUTEX_EXIT(&call->lock);
2791 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2792 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2795 rxi_KeepAliveOn(call);
2796 } else if (np->header.callNumber != currentCallNumber) {
2797 /* Wait until the transmit queue is idle before deciding
2798 * whether to reset the current call. Chances are that the
2799 * call will be in ether DALLY or HOLD state once the TQ_BUSY
2802 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2803 while ((call->state == RX_STATE_ACTIVE)
2804 && (call->flags & RX_CALL_TQ_BUSY)) {
2805 call->flags |= RX_CALL_TQ_WAIT;
2807 #ifdef RX_ENABLE_LOCKS
2808 osirx_AssertMine(&call->lock, "rxi_Start lock3");
2809 CV_WAIT(&call->cv_tq, &call->lock);
2810 #else /* RX_ENABLE_LOCKS */
2811 osi_rxSleep(&call->tq);
2812 #endif /* RX_ENABLE_LOCKS */
2814 if (call->tqWaiters == 0)
2815 call->flags &= ~RX_CALL_TQ_WAIT;
2817 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2818 /* If the new call cannot be taken right now send a busy and set
2819 * the error condition in this call, so that it terminates as
2820 * quickly as possible */
2821 if (call->state == RX_STATE_ACTIVE) {
2822 struct rx_packet *tp;
2824 rxi_CallError(call, RX_CALL_DEAD);
2825 tp = rxi_SendSpecial(call, conn, np, RX_PACKET_TYPE_BUSY,
2827 MUTEX_EXIT(&call->lock);
2828 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2831 rxi_ResetCall(call, 0);
2832 *call->callNumber = np->header.callNumber;
2833 if (np->header.callNumber == 0)
2834 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));
2836 call->state = RX_STATE_PRECALL;
2837 clock_GetTime(&call->queueTime);
2838 hzero(call->bytesSent);
2839 hzero(call->bytesRcvd);
2841 * If the number of queued calls exceeds the overload
2842 * threshold then abort this call.
2844 if ((rx_BusyThreshold > 0) && (rx_nWaiting > rx_BusyThreshold)) {
2845 struct rx_packet *tp;
2847 rxi_CallError(call, rx_BusyError);
2848 tp = rxi_SendCallAbort(call, np, 1, 0);
2849 MUTEX_EXIT(&call->lock);
2850 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2851 rx_MutexIncrement(rx_stats.nBusies, rx_stats_mutex);
2854 rxi_KeepAliveOn(call);
2856 /* Continuing call; do nothing here. */
2858 } else { /* we're the client */
2859 /* Ignore all incoming acknowledgements for calls in DALLY state */
2860 if (call && (call->state == RX_STATE_DALLY)
2861 && (np->header.type == RX_PACKET_TYPE_ACK)) {
2862 rx_MutexIncrement(rx_stats.ignorePacketDally, rx_stats_mutex);
2863 #ifdef RX_ENABLE_LOCKS
2865 MUTEX_EXIT(&call->lock);
2868 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2872 /* Ignore anything that's not relevant to the current call. If there
2873 * isn't a current call, then no packet is relevant. */
2874 if (!call || (np->header.callNumber != currentCallNumber)) {
2875 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2876 #ifdef RX_ENABLE_LOCKS
2878 MUTEX_EXIT(&call->lock);
2881 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2884 /* If the service security object index stamped in the packet does not
2885 * match the connection's security index, ignore the packet */
2886 if (np->header.securityIndex != conn->securityIndex) {
2887 #ifdef RX_ENABLE_LOCKS
2888 MUTEX_EXIT(&call->lock);
2890 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2894 /* If we're receiving the response, then all transmit packets are
2895 * implicitly acknowledged. Get rid of them. */
2896 if (np->header.type == RX_PACKET_TYPE_DATA) {
2897 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
2898 /* XXX Hack. Because we must release the global rx lock when
2899 * sending packets (osi_NetSend) we drop all acks while we're
2900 * traversing the tq in rxi_Start sending packets out because
2901 * packets may move to the freePacketQueue as result of being here!
2902 * So we drop these packets until we're safely out of the
2903 * traversing. Really ugly!
2904 * For fine grain RX locking, we set the acked field in the
2905 * packets and let rxi_Start remove them from the transmit queue.
2907 if (call->flags & RX_CALL_TQ_BUSY) {
2908 #ifdef RX_ENABLE_LOCKS
2909 rxi_SetAcksInTransmitQueue(call);
2912 return np; /* xmitting; drop packet */
2915 rxi_ClearTransmitQueue(call, 0);
2917 #else /* AFS_GLOBAL_RXLOCK_KERNEL */
2918 rxi_ClearTransmitQueue(call, 0);
2919 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
2921 if (np->header.type == RX_PACKET_TYPE_ACK) {
2922 /* now check to see if this is an ack packet acknowledging that the
2923 * server actually *lost* some hard-acked data. If this happens we
2924 * ignore this packet, as it may indicate that the server restarted in
2925 * the middle of a call. It is also possible that this is an old ack
2926 * packet. We don't abort the connection in this case, because this
2927 * *might* just be an old ack packet. The right way to detect a server
2928 * restart in the midst of a call is to notice that the server epoch
2930 /* XXX I'm not sure this is exactly right, since tfirst **IS**
2931 * XXX unacknowledged. I think that this is off-by-one, but
2932 * XXX I don't dare change it just yet, since it will
2933 * XXX interact badly with the server-restart detection
2934 * XXX code in receiveackpacket. */
2935 if (ntohl(rx_GetInt32(np, FIRSTACKOFFSET)) < call->tfirst) {
2936 rx_MutexIncrement(rx_stats.spuriousPacketsRead, rx_stats_mutex);
2937 MUTEX_EXIT(&call->lock);
2938 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
2942 } /* else not a data packet */
2945 osirx_AssertMine(&call->lock, "rxi_ReceivePacket middle");
2946 /* Set remote user defined status from packet */
2947 call->remoteStatus = np->header.userStatus;
2949 /* Note the gap between the expected next packet and the actual
2950 * packet that arrived, when the new packet has a smaller serial number
2951 * than expected. Rioses frequently reorder packets all by themselves,
2952 * so this will be quite important with very large window sizes.
2953 * Skew is checked against 0 here to avoid any dependence on the type of
2954 * inPacketSkew (which may be unsigned). In C, -1 > (unsigned) 0 is always
2956 * The inPacketSkew should be a smoothed running value, not just a maximum. MTUXXX
2957 * see CalculateRoundTripTime for an example of how to keep smoothed values.
2958 * I think using a beta of 1/8 is probably appropriate. 93.04.21
2960 MUTEX_ENTER(&conn->conn_data_lock);
2961 skew = conn->lastSerial - np->header.serial;
2962 conn->lastSerial = np->header.serial;
2963 MUTEX_EXIT(&conn->conn_data_lock);
2965 register struct rx_peer *peer;
2967 if (skew > peer->inPacketSkew) {
2968 dpf(("*** In skew changed from %d to %d\n", peer->inPacketSkew,
2970 peer->inPacketSkew = skew;
2974 /* Now do packet type-specific processing */
2975 switch (np->header.type) {
2976 case RX_PACKET_TYPE_DATA:
2977 np = rxi_ReceiveDataPacket(call, np, 1, socket, host, port, tnop,
2980 case RX_PACKET_TYPE_ACK:
2981 /* Respond immediately to ack packets requesting acknowledgement
2983 if (np->header.flags & RX_REQUEST_ACK) {
2985 (void)rxi_SendCallAbort(call, 0, 1, 0);
2987 (void)rxi_SendAck(call, 0, np->header.serial,
2988 RX_ACK_PING_RESPONSE, 1);
2990 np = rxi_ReceiveAckPacket(call, np, 1);
2992 case RX_PACKET_TYPE_ABORT: {
2993 /* An abort packet: reset the call, passing the error up to the user. */
2994 /* What if error is zero? */
2995 /* What if the error is -1? the application will treat it as a timeout. */
2996 afs_int32 errdata = ntohl(*(afs_int32 *) rx_DataOf(np));
2997 dpf(("rxi_ReceivePacket ABORT rx_DataOf = %d", errdata));
2998 rxi_CallError(call, errdata);
2999 MUTEX_EXIT(&call->lock);
3000 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
3001 return np; /* xmitting; drop packet */
3003 case RX_PACKET_TYPE_BUSY:
3006 case RX_PACKET_TYPE_ACKALL:
3007 /* All packets acknowledged, so we can drop all packets previously
3008 * readied for sending */
3009 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3010 /* XXX Hack. We because we can't release the global rx lock when
3011 * sending packets (osi_NetSend) we drop all ack pkts while we're
3012 * traversing the tq in rxi_Start sending packets out because
3013 * packets may move to the freePacketQueue as result of being
3014 * here! So we drop these packets until we're safely out of the
3015 * traversing. Really ugly!
3016 * For fine grain RX locking, we set the acked field in the packets
3017 * and let rxi_Start remove the packets from the transmit queue.
3019 if (call->flags & RX_CALL_TQ_BUSY) {
3020 #ifdef RX_ENABLE_LOCKS
3021 rxi_SetAcksInTransmitQueue(call);
3023 #else /* RX_ENABLE_LOCKS */
3024 MUTEX_EXIT(&call->lock);
3025 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
3026 return np; /* xmitting; drop packet */
3027 #endif /* RX_ENABLE_LOCKS */
3029 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3030 rxi_ClearTransmitQueue(call, 0);
3031 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
3034 /* Should not reach here, unless the peer is broken: send an abort
3036 rxi_CallError(call, RX_PROTOCOL_ERROR);
3037 np = rxi_SendCallAbort(call, np, 1, 0);
3040 /* Note when this last legitimate packet was received, for keep-alive
3041 * processing. Note, we delay getting the time until now in the hope that
3042 * the packet will be delivered to the user before any get time is required
3043 * (if not, then the time won't actually be re-evaluated here). */
3044 call->lastReceiveTime = clock_Sec();
3045 MUTEX_EXIT(&call->lock);
3046 rx_MutexDecrement(conn->refCount, conn->conn_data_lock);
3050 /* return true if this is an "interesting" connection from the point of view
3051 of someone trying to debug the system */
3053 rxi_IsConnInteresting(struct rx_connection *aconn)
3056 register struct rx_call *tcall;
3058 if (aconn->flags & (RX_CONN_MAKECALL_WAITING | RX_CONN_DESTROY_ME))
3060 for (i = 0; i < RX_MAXCALLS; i++) {
3061 tcall = aconn->call[i];
3063 if ((tcall->state == RX_STATE_PRECALL)
3064 || (tcall->state == RX_STATE_ACTIVE))
3066 if ((tcall->mode == RX_MODE_SENDING)
3067 || (tcall->mode == RX_MODE_RECEIVING))
3075 /* if this is one of the last few packets AND it wouldn't be used by the
3076 receiving call to immediately satisfy a read request, then drop it on
3077 the floor, since accepting it might prevent a lock-holding thread from
3078 making progress in its reading. If a call has been cleared while in
3079 the precall state then ignore all subsequent packets until the call
3080 is assigned to a thread. */
3083 TooLow(struct rx_packet *ap, struct rx_call *acall)
3086 MUTEX_ENTER(&rx_stats_mutex);
3087 if (((ap->header.seq != 1) && (acall->flags & RX_CALL_CLEARED)
3088 && (acall->state == RX_STATE_PRECALL))
3089 || ((rx_nFreePackets < rxi_dataQuota + 2)
3090 && !((ap->header.seq < acall->rnext + rx_initSendWindow)
3091 && (acall->flags & RX_CALL_READER_WAIT)))) {
3094 MUTEX_EXIT(&rx_stats_mutex);
3100 rxi_CheckReachEvent(struct rxevent *event, void *arg1, void *arg2)
3102 struct rx_connection *conn = arg1;
3103 struct rx_call *acall = arg2;
3104 struct rx_call *call = acall;
3105 struct clock when, now;
3108 MUTEX_ENTER(&conn->conn_data_lock);
3109 conn->checkReachEvent = NULL;
3110 waiting = conn->flags & RX_CONN_ATTACHWAIT;
3113 MUTEX_EXIT(&conn->conn_data_lock);
3117 MUTEX_ENTER(&conn->conn_call_lock);
3118 MUTEX_ENTER(&conn->conn_data_lock);
3119 for (i = 0; i < RX_MAXCALLS; i++) {
3120 struct rx_call *tc = conn->call[i];
3121 if (tc && tc->state == RX_STATE_PRECALL) {
3127 /* Indicate that rxi_CheckReachEvent is no longer running by
3128 * clearing the flag. Must be atomic under conn_data_lock to
3129 * avoid a new call slipping by: rxi_CheckConnReach holds
3130 * conn_data_lock while checking RX_CONN_ATTACHWAIT.
3132 conn->flags &= ~RX_CONN_ATTACHWAIT;
3133 MUTEX_EXIT(&conn->conn_data_lock);
3134 MUTEX_EXIT(&conn->conn_call_lock);
3139 MUTEX_ENTER(&call->lock);
3140 rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
3142 MUTEX_EXIT(&call->lock);
3144 clock_GetTime(&now);
3146 when.sec += RX_CHECKREACH_TIMEOUT;
3147 MUTEX_ENTER(&conn->conn_data_lock);
3148 if (!conn->checkReachEvent) {
3150 conn->checkReachEvent =
3151 rxevent_PostNow(&when, &now, rxi_CheckReachEvent, conn,
3154 MUTEX_EXIT(&conn->conn_data_lock);
3160 rxi_CheckConnReach(struct rx_connection *conn, struct rx_call *call)
3162 struct rx_service *service = conn->service;
3163 struct rx_peer *peer = conn->peer;
3164 afs_uint32 now, lastReach;
3166 if (service->checkReach == 0)
3170 MUTEX_ENTER(&peer->peer_lock);
3171 lastReach = peer->lastReachTime;
3172 MUTEX_EXIT(&peer->peer_lock);
3173 if (now - lastReach < RX_CHECKREACH_TTL)
3176 MUTEX_ENTER(&conn->conn_data_lock);
3177 if (conn->flags & RX_CONN_ATTACHWAIT) {
3178 MUTEX_EXIT(&conn->conn_data_lock);
3181 conn->flags |= RX_CONN_ATTACHWAIT;
3182 MUTEX_EXIT(&conn->conn_data_lock);
3183 if (!conn->checkReachEvent)
3184 rxi_CheckReachEvent(NULL, conn, call);
3189 /* try to attach call, if authentication is complete */
3191 TryAttach(register struct rx_call *acall, register osi_socket socket,
3192 register int *tnop, register struct rx_call **newcallp,
3195 struct rx_connection *conn = acall->conn;
3197 if (conn->type == RX_SERVER_CONNECTION
3198 && acall->state == RX_STATE_PRECALL) {
3199 /* Don't attach until we have any req'd. authentication. */
3200 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0) {
3201 if (reachOverride || rxi_CheckConnReach(conn, acall) == 0)
3202 rxi_AttachServerProc(acall, socket, tnop, newcallp);
3203 /* Note: this does not necessarily succeed; there
3204 * may not any proc available
3207 rxi_ChallengeOn(acall->conn);
3212 /* A data packet has been received off the interface. This packet is
3213 * appropriate to the call (the call is in the right state, etc.). This
3214 * routine can return a packet to the caller, for re-use */
3217 rxi_ReceiveDataPacket(register struct rx_call *call,
3218 register struct rx_packet *np, int istack,
3219 osi_socket socket, afs_uint32 host, u_short port,
3220 int *tnop, struct rx_call **newcallp)
3222 int ackNeeded = 0; /* 0 means no, otherwise ack_reason */
3226 afs_uint32 seq, serial, flags;
3228 struct rx_packet *tnp;
3229 struct clock when, now;
3230 rx_MutexIncrement(rx_stats.dataPacketsRead, rx_stats_mutex);
3233 /* If there are no packet buffers, drop this new packet, unless we can find
3234 * packet buffers from inactive calls */
3236 && (rxi_OverQuota(RX_PACKET_CLASS_RECEIVE) || TooLow(np, call))) {
3237 MUTEX_ENTER(&rx_freePktQ_lock);
3238 rxi_NeedMorePackets = TRUE;
3239 MUTEX_EXIT(&rx_freePktQ_lock);
3240 rx_MutexIncrement(rx_stats.noPacketBuffersOnRead, rx_stats_mutex);
3241 call->rprev = np->header.serial;
3242 rxi_calltrace(RX_TRACE_DROP, call);
3243 dpf(("packet %x dropped on receipt - quota problems", np));
3245 rxi_ClearReceiveQueue(call);
3246 clock_GetTime(&now);
3248 clock_Add(&when, &rx_softAckDelay);
3249 if (!call->delayedAckEvent
3250 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3251 rxevent_Cancel(call->delayedAckEvent, call,
3252 RX_CALL_REFCOUNT_DELAY);
3253 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3254 call->delayedAckEvent =
3255 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3257 /* we've damaged this call already, might as well do it in. */
3263 * New in AFS 3.5, if the RX_JUMBO_PACKET flag is set then this
3264 * packet is one of several packets transmitted as a single
3265 * datagram. Do not send any soft or hard acks until all packets
3266 * in a jumbogram have been processed. Send negative acks right away.
3268 for (isFirst = 1, tnp = NULL; isFirst || tnp; isFirst = 0) {
3269 /* tnp is non-null when there are more packets in the
3270 * current jumbo gram */
3277 seq = np->header.seq;
3278 serial = np->header.serial;
3279 flags = np->header.flags;
3281 /* If the call is in an error state, send an abort message */
3283 return rxi_SendCallAbort(call, np, istack, 0);
3285 /* The RX_JUMBO_PACKET is set in all but the last packet in each
3286 * AFS 3.5 jumbogram. */
3287 if (flags & RX_JUMBO_PACKET) {
3288 tnp = rxi_SplitJumboPacket(np, host, port, isFirst);
3293 if (np->header.spare != 0) {
3294 MUTEX_ENTER(&call->conn->conn_data_lock);
3295 call->conn->flags |= RX_CONN_USING_PACKET_CKSUM;
3296 MUTEX_EXIT(&call->conn->conn_data_lock);
3299 /* The usual case is that this is the expected next packet */
3300 if (seq == call->rnext) {
3302 /* Check to make sure it is not a duplicate of one already queued */
3303 if (queue_IsNotEmpty(&call->rq)
3304 && queue_First(&call->rq, rx_packet)->header.seq == seq) {
3305 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3306 dpf(("packet %x dropped on receipt - duplicate", np));
3307 rxevent_Cancel(call->delayedAckEvent, call,
3308 RX_CALL_REFCOUNT_DELAY);
3309 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3315 /* It's the next packet. Stick it on the receive queue
3316 * for this call. Set newPackets to make sure we wake
3317 * the reader once all packets have been processed */
3318 np->flags |= RX_PKTFLAG_RQ;
3319 queue_Prepend(&call->rq, np);
3320 #ifdef RXDEBUG_PACKET
3322 #endif /* RXDEBUG_PACKET */
3324 np = NULL; /* We can't use this anymore */
3327 /* If an ack is requested then set a flag to make sure we
3328 * send an acknowledgement for this packet */
3329 if (flags & RX_REQUEST_ACK) {
3330 ackNeeded = RX_ACK_REQUESTED;
3333 /* Keep track of whether we have received the last packet */
3334 if (flags & RX_LAST_PACKET) {
3335 call->flags |= RX_CALL_HAVE_LAST;
3339 /* Check whether we have all of the packets for this call */
3340 if (call->flags & RX_CALL_HAVE_LAST) {
3341 afs_uint32 tseq; /* temporary sequence number */
3342 struct rx_packet *tp; /* Temporary packet pointer */
3343 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3345 for (tseq = seq, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3346 if (tseq != tp->header.seq)
3348 if (tp->header.flags & RX_LAST_PACKET) {
3349 call->flags |= RX_CALL_RECEIVE_DONE;
3356 /* Provide asynchronous notification for those who want it
3357 * (e.g. multi rx) */
3358 if (call->arrivalProc) {
3359 (*call->arrivalProc) (call, call->arrivalProcHandle,
3360 call->arrivalProcArg);
3361 call->arrivalProc = (void (*)())0;
3364 /* Update last packet received */
3367 /* If there is no server process serving this call, grab
3368 * one, if available. We only need to do this once. If a
3369 * server thread is available, this thread becomes a server
3370 * thread and the server thread becomes a listener thread. */
3372 TryAttach(call, socket, tnop, newcallp, 0);
3375 /* This is not the expected next packet. */
3377 /* Determine whether this is a new or old packet, and if it's
3378 * a new one, whether it fits into the current receive window.
3379 * Also figure out whether the packet was delivered in sequence.
3380 * We use the prev variable to determine whether the new packet
3381 * is the successor of its immediate predecessor in the
3382 * receive queue, and the missing flag to determine whether
3383 * any of this packets predecessors are missing. */
3385 afs_uint32 prev; /* "Previous packet" sequence number */
3386 struct rx_packet *tp; /* Temporary packet pointer */
3387 struct rx_packet *nxp; /* Next pointer, for queue_Scan */
3388 int missing; /* Are any predecessors missing? */
3390 /* If the new packet's sequence number has been sent to the
3391 * application already, then this is a duplicate */
3392 if (seq < call->rnext) {
3393 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3394 rxevent_Cancel(call->delayedAckEvent, call,
3395 RX_CALL_REFCOUNT_DELAY);
3396 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE, istack);
3402 /* If the sequence number is greater than what can be
3403 * accomodated by the current window, then send a negative
3404 * acknowledge and drop the packet */
3405 if ((call->rnext + call->rwind) <= seq) {
3406 rxevent_Cancel(call->delayedAckEvent, call,
3407 RX_CALL_REFCOUNT_DELAY);
3408 np = rxi_SendAck(call, np, serial, RX_ACK_EXCEEDS_WINDOW,
3415 /* Look for the packet in the queue of old received packets */
3416 for (prev = call->rnext - 1, missing =
3417 0, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3418 /*Check for duplicate packet */
3419 if (seq == tp->header.seq) {
3420 rx_MutexIncrement(rx_stats.dupPacketsRead, rx_stats_mutex);
3421 rxevent_Cancel(call->delayedAckEvent, call,
3422 RX_CALL_REFCOUNT_DELAY);
3423 np = rxi_SendAck(call, np, serial, RX_ACK_DUPLICATE,
3429 /* If we find a higher sequence packet, break out and
3430 * insert the new packet here. */
3431 if (seq < tp->header.seq)
3433 /* Check for missing packet */
3434 if (tp->header.seq != prev + 1) {
3438 prev = tp->header.seq;
3441 /* Keep track of whether we have received the last packet. */
3442 if (flags & RX_LAST_PACKET) {
3443 call->flags |= RX_CALL_HAVE_LAST;
3446 /* It's within the window: add it to the the receive queue.
3447 * tp is left by the previous loop either pointing at the
3448 * packet before which to insert the new packet, or at the
3449 * queue head if the queue is empty or the packet should be
3451 np->flags |= RX_PKTFLAG_RQ;
3452 #ifdef RXDEBUG_PACKET
3454 #endif /* RXDEBUG_PACKET */
3455 queue_InsertBefore(tp, np);
3459 /* Check whether we have all of the packets for this call */
3460 if ((call->flags & RX_CALL_HAVE_LAST)
3461 && !(call->flags & RX_CALL_RECEIVE_DONE)) {
3462 afs_uint32 tseq; /* temporary sequence number */
3465 call->rnext, queue_Scan(&call->rq, tp, nxp, rx_packet)) {
3466 if (tseq != tp->header.seq)
3468 if (tp->header.flags & RX_LAST_PACKET) {
3469 call->flags |= RX_CALL_RECEIVE_DONE;
3476 /* We need to send an ack of the packet is out of sequence,
3477 * or if an ack was requested by the peer. */
3478 if (seq != prev + 1 || missing) {
3479 ackNeeded = RX_ACK_OUT_OF_SEQUENCE;
3480 } else if (flags & RX_REQUEST_ACK) {
3481 ackNeeded = RX_ACK_REQUESTED;
3484 /* Acknowledge the last packet for each call */
3485 if (flags & RX_LAST_PACKET) {
3496 * If the receiver is waiting for an iovec, fill the iovec
3497 * using the data from the receive queue */
3498 if (call->flags & RX_CALL_IOVEC_WAIT) {
3499 didHardAck = rxi_FillReadVec(call, serial);
3500 /* the call may have been aborted */
3509 /* Wakeup the reader if any */
3510 if ((call->flags & RX_CALL_READER_WAIT)
3511 && (!(call->flags & RX_CALL_IOVEC_WAIT) || !(call->iovNBytes)
3512 || (call->iovNext >= call->iovMax)
3513 || (call->flags & RX_CALL_RECEIVE_DONE))) {
3514 call->flags &= ~RX_CALL_READER_WAIT;
3515 #ifdef RX_ENABLE_LOCKS
3516 CV_BROADCAST(&call->cv_rq);
3518 osi_rxWakeup(&call->rq);
3524 * Send an ack when requested by the peer, or once every
3525 * rxi_SoftAckRate packets until the last packet has been
3526 * received. Always send a soft ack for the last packet in
3527 * the server's reply. */
3529 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3530 np = rxi_SendAck(call, np, serial, ackNeeded, istack);
3531 } else if (call->nSoftAcks > (u_short) rxi_SoftAckRate) {
3532 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3533 np = rxi_SendAck(call, np, serial, RX_ACK_IDLE, istack);
3534 } else if (call->nSoftAcks) {
3535 clock_GetTime(&now);
3537 if (haveLast && !(flags & RX_CLIENT_INITIATED)) {
3538 clock_Add(&when, &rx_lastAckDelay);
3540 clock_Add(&when, &rx_softAckDelay);
3542 if (!call->delayedAckEvent
3543 || clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
3544 rxevent_Cancel(call->delayedAckEvent, call,
3545 RX_CALL_REFCOUNT_DELAY);
3546 CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
3547 call->delayedAckEvent =
3548 rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
3550 } else if (call->flags & RX_CALL_RECEIVE_DONE) {
3551 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
3558 static void rxi_ComputeRate();
3562 rxi_UpdatePeerReach(struct rx_connection *conn, struct rx_call *acall)
3564 struct rx_peer *peer = conn->peer;
3566 MUTEX_ENTER(&peer->peer_lock);
3567 peer->lastReachTime = clock_Sec();
3568 MUTEX_EXIT(&peer->peer_lock);
3570 MUTEX_ENTER(&conn->conn_data_lock);
3571 if (conn->flags & RX_CONN_ATTACHWAIT) {
3574 conn->flags &= ~RX_CONN_ATTACHWAIT;
3575 MUTEX_EXIT(&conn->conn_data_lock);
3577 for (i = 0; i < RX_MAXCALLS; i++) {
3578 struct rx_call *call = conn->call[i];
3581 MUTEX_ENTER(&call->lock);
3582 /* tnop can be null if newcallp is null */
3583 TryAttach(call, (osi_socket) - 1, NULL, NULL, 1);
3585 MUTEX_EXIT(&call->lock);
3589 MUTEX_EXIT(&conn->conn_data_lock);
3593 rx_ack_reason(int reason)
3596 case RX_ACK_REQUESTED:
3598 case RX_ACK_DUPLICATE:
3600 case RX_ACK_OUT_OF_SEQUENCE:
3602 case RX_ACK_EXCEEDS_WINDOW:
3604 case RX_ACK_NOSPACE:
3608 case RX_ACK_PING_RESPONSE:
3620 /* rxi_ComputePeerNetStats
3622 * Called exclusively by rxi_ReceiveAckPacket to compute network link
3623 * estimates (like RTT and throughput) based on ack packets. Caller
3624 * must ensure that the packet in question is the right one (i.e.
3625 * serial number matches).
3628 rxi_ComputePeerNetStats(struct rx_call *call, struct rx_packet *p,
3629 struct rx_ackPacket *ap, struct rx_packet *np)
3631 struct rx_peer *peer = call->conn->peer;
3633 /* Use RTT if not delayed by client. */
3634 if (ap->reason != RX_ACK_DELAY)
3635 rxi_ComputeRoundTripTime(p, &p->timeSent, peer);
3637 rxi_ComputeRate(peer, call, p, np, ap->reason);
3641 /* The real smarts of the whole thing. */
3643 rxi_ReceiveAckPacket(register struct rx_call *call, struct rx_packet *np,
3646 struct rx_ackPacket *ap;
3648 register struct rx_packet *tp;
3649 register struct rx_packet *nxp; /* Next packet pointer for queue_Scan */
3650 register struct rx_connection *conn = call->conn;
3651 struct rx_peer *peer = conn->peer;
3654 /* because there are CM's that are bogus, sending weird values for this. */
3655 afs_uint32 skew = 0;
3660 int newAckCount = 0;
3661 u_short maxMTU = 0; /* Set if peer supports AFS 3.4a jumbo datagrams */
3662 int maxDgramPackets = 0; /* Set if peer supports AFS 3.5 jumbo datagrams */
3664 rx_MutexIncrement(rx_stats.ackPacketsRead, rx_stats_mutex);
3665 ap = (struct rx_ackPacket *)rx_DataOf(np);
3666 nbytes = rx_Contiguous(np) - (int)((ap->acks) - (u_char *) ap);
3668 return np; /* truncated ack packet */
3670 /* depends on ack packet struct */
3671 nAcks = MIN((unsigned)nbytes, (unsigned)ap->nAcks);
3672 first = ntohl(ap->firstPacket);
3673 serial = ntohl(ap->serial);
3674 /* temporarily disabled -- needs to degrade over time
3675 * skew = ntohs(ap->maxSkew); */
3677 /* Ignore ack packets received out of order */
3678 if (first < call->tfirst) {
3682 if (np->header.flags & RX_SLOW_START_OK) {
3683 call->flags |= RX_CALL_SLOW_START_OK;
3686 if (ap->reason == RX_ACK_PING_RESPONSE)
3687 rxi_UpdatePeerReach(conn, call);
3691 if (rxdebug_active) {
3695 len = _snprintf(msg, sizeof(msg),
3696 "tid[%d] RACK: reason %s serial %u previous %u seq %u skew %d first %u acks %u space %u ",
3697 GetCurrentThreadId(), rx_ack_reason(ap->reason),
3698 ntohl(ap->serial), ntohl(ap->previousPacket),
3699 (unsigned int)np->header.seq, (unsigned int)skew,
3700 ntohl(ap->firstPacket), ap->nAcks, ntohs(ap->bufferSpace) );
3704 for (offset = 0; offset < nAcks && len < sizeof(msg); offset++)
3705 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
3709 OutputDebugString(msg);
3711 #else /* AFS_NT40_ENV */
3714 "RACK: reason %x previous %u seq %u serial %u skew %d first %u",
3715 ap->reason, ntohl(ap->previousPacket),
3716 (unsigned int)np->header.seq, (unsigned int)serial,
3717 (unsigned int)skew, ntohl(ap->firstPacket));
3720 for (offset = 0; offset < nAcks; offset++)
3721 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
3726 #endif /* AFS_NT40_ENV */
3729 /* Update the outgoing packet skew value to the latest value of
3730 * the peer's incoming packet skew value. The ack packet, of
3731 * course, could arrive out of order, but that won't affect things
3733 MUTEX_ENTER(&peer->peer_lock);
3734 peer->outPacketSkew = skew;
3736 /* Check for packets that no longer need to be transmitted, and
3737 * discard them. This only applies to packets positively
3738 * acknowledged as having been sent to the peer's upper level.
3739 * All other packets must be retained. So only packets with
3740 * sequence numbers < ap->firstPacket are candidates. */
3741 for (queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3742 if (tp->header.seq >= first)
3744 call->tfirst = tp->header.seq + 1;
3746 && (tp->header.serial == serial || tp->firstSerial == serial))
3747 rxi_ComputePeerNetStats(call, tp, ap, np);
3748 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3751 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3752 /* XXX Hack. Because we have to release the global rx lock when sending
3753 * packets (osi_NetSend) we drop all acks while we're traversing the tq
3754 * in rxi_Start sending packets out because packets may move to the
3755 * freePacketQueue as result of being here! So we drop these packets until
3756 * we're safely out of the traversing. Really ugly!
3757 * To make it even uglier, if we're using fine grain locking, we can
3758 * set the ack bits in the packets and have rxi_Start remove the packets
3759 * when it's done transmitting.
3761 if (call->flags & RX_CALL_TQ_BUSY) {
3762 #ifdef RX_ENABLE_LOCKS
3763 tp->flags |= RX_PKTFLAG_ACKED;
3764 call->flags |= RX_CALL_TQ_SOME_ACKED;
3765 #else /* RX_ENABLE_LOCKS */
3767 #endif /* RX_ENABLE_LOCKS */
3769 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3772 tp->flags &= ~RX_PKTFLAG_TQ;
3773 #ifdef RXDEBUG_PACKET
3775 #endif /* RXDEBUG_PACKET */
3776 rxi_FreePacket(tp); /* rxi_FreePacket mustn't wake up anyone, preemptively. */
3781 /* Give rate detector a chance to respond to ping requests */
3782 if (ap->reason == RX_ACK_PING_RESPONSE) {
3783 rxi_ComputeRate(peer, call, 0, np, ap->reason);
3787 /* N.B. we don't turn off any timers here. They'll go away by themselves, anyway */
3789 /* Now go through explicit acks/nacks and record the results in
3790 * the waiting packets. These are packets that can't be released
3791 * yet, even with a positive acknowledge. This positive
3792 * acknowledge only means the packet has been received by the
3793 * peer, not that it will be retained long enough to be sent to
3794 * the peer's upper level. In addition, reset the transmit timers
3795 * of any missing packets (those packets that must be missing
3796 * because this packet was out of sequence) */
3798 call->nSoftAcked = 0;
3799 for (missing = 0, queue_Scan(&call->tq, tp, nxp, rx_packet)) {
3800 /* Update round trip time if the ack was stimulated on receipt
3802 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
3803 #ifdef RX_ENABLE_LOCKS
3804 if (tp->header.seq >= first)
3805 #endif /* RX_ENABLE_LOCKS */
3806 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
3808 && (tp->header.serial == serial || tp->firstSerial == serial))
3809 rxi_ComputePeerNetStats(call, tp, ap, np);
3811 /* Set the acknowledge flag per packet based on the
3812 * information in the ack packet. An acknowlegded packet can
3813 * be downgraded when the server has discarded a packet it
3814 * soacked previously, or when an ack packet is received
3815 * out of sequence. */
3816 if (tp->header.seq < first) {
3817 /* Implicit ack information */
3818 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3821 tp->flags |= RX_PKTFLAG_ACKED;
3822 } else if (tp->header.seq < first + nAcks) {
3823 /* Explicit ack information: set it in the packet appropriately */
3824 if (ap->acks[tp->header.seq - first] == RX_ACK_TYPE_ACK) {
3825 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
3827 tp->flags |= RX_PKTFLAG_ACKED;
3834 } else /* RX_ACK_TYPE_NACK */ {
3835 tp->flags &= ~RX_PKTFLAG_ACKED;
3839 tp->flags &= ~RX_PKTFLAG_ACKED;
3843 /* If packet isn't yet acked, and it has been transmitted at least
3844 * once, reset retransmit time using latest timeout
3845 * ie, this should readjust the retransmit timer for all outstanding
3846 * packets... So we don't just retransmit when we should know better*/
3848 if (!(tp->flags & RX_PKTFLAG_ACKED) && !clock_IsZero(&tp->retryTime)) {
3849 tp->retryTime = tp->timeSent;
3850 clock_Add(&tp->retryTime, &peer->timeout);
3851 /* shift by eight because one quarter-sec ~ 256 milliseconds */
3852 clock_Addmsec(&(tp->retryTime), ((afs_uint32) tp->backoff) << 8);
3856 /* If the window has been extended by this acknowledge packet,
3857 * then wakeup a sender waiting in alloc for window space, or try
3858 * sending packets now, if he's been sitting on packets due to
3859 * lack of window space */
3860 if (call->tnext < (call->tfirst + call->twind)) {
3861 #ifdef RX_ENABLE_LOCKS
3862 CV_SIGNAL(&call->cv_twind);
3864 if (call->flags & RX_CALL_WAIT_WINDOW_ALLOC) {
3865 call->flags &= ~RX_CALL_WAIT_WINDOW_ALLOC;
3866 osi_rxWakeup(&call->twind);
3869 if (call->flags & RX_CALL_WAIT_WINDOW_SEND) {
3870 call->flags &= ~RX_CALL_WAIT_WINDOW_SEND;
3874 /* if the ack packet has a receivelen field hanging off it,
3875 * update our state */
3876 if (np->length >= rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32)) {
3879 /* If the ack packet has a "recommended" size that is less than
3880 * what I am using now, reduce my size to match */
3881 rx_packetread(np, rx_AckDataSize(ap->nAcks) + sizeof(afs_int32),
3882 (int)sizeof(afs_int32), &tSize);
3883 tSize = (afs_uint32) ntohl(tSize);
3884 peer->natMTU = rxi_AdjustIfMTU(MIN(tSize, peer->ifMTU));
3886 /* Get the maximum packet size to send to this peer */
3887 rx_packetread(np, rx_AckDataSize(ap->nAcks), (int)sizeof(afs_int32),
3889 tSize = (afs_uint32) ntohl(tSize);
3890 tSize = (afs_uint32) MIN(tSize, rx_MyMaxSendSize);
3891 tSize = rxi_AdjustMaxMTU(peer->natMTU, tSize);
3893 /* sanity check - peer might have restarted with different params.
3894 * If peer says "send less", dammit, send less... Peer should never
3895 * be unable to accept packets of the size that prior AFS versions would
3896 * send without asking. */
3897 if (peer->maxMTU != tSize) {
3898 if (peer->maxMTU > tSize) /* possible cong., maxMTU decreased */
3900 peer->maxMTU = tSize;
3901 peer->MTU = MIN(tSize, peer->MTU);
3902 call->MTU = MIN(call->MTU, tSize);
3905 if (np->length == rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32)) {
3908 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3909 (int)sizeof(afs_int32), &tSize);
3910 tSize = (afs_uint32) ntohl(tSize); /* peer's receive window, if it's */
3911 if (tSize < call->twind) { /* smaller than our send */
3912 call->twind = tSize; /* window, we must send less... */
3913 call->ssthresh = MIN(call->twind, call->ssthresh);
3914 call->conn->twind[call->channel] = call->twind;
3917 /* Only send jumbograms to 3.4a fileservers. 3.3a RX gets the
3918 * network MTU confused with the loopback MTU. Calculate the
3919 * maximum MTU here for use in the slow start code below.
3921 maxMTU = peer->maxMTU;
3922 /* Did peer restart with older RX version? */
3923 if (peer->maxDgramPackets > 1) {
3924 peer->maxDgramPackets = 1;
3926 } else if (np->length >=
3927 rx_AckDataSize(ap->nAcks) + 4 * sizeof(afs_int32)) {
3930 rx_AckDataSize(ap->nAcks) + 2 * sizeof(afs_int32),
3931 sizeof(afs_int32), &tSize);
3932 tSize = (afs_uint32) ntohl(tSize);
3934 * As of AFS 3.5 we set the send window to match the receive window.
3936 if (tSize < call->twind) {
3937 call->twind = tSize;
3938 call->conn->twind[call->channel] = call->twind;
3939 call->ssthresh = MIN(call->twind, call->ssthresh);
3940 } else if (tSize > call->twind) {
3941 call->twind = tSize;
3942 call->conn->twind[call->channel] = call->twind;
3946 * As of AFS 3.5, a jumbogram is more than one fixed size
3947 * packet transmitted in a single UDP datagram. If the remote
3948 * MTU is smaller than our local MTU then never send a datagram
3949 * larger than the natural MTU.
3952 rx_AckDataSize(ap->nAcks) + 3 * sizeof(afs_int32),
3953 sizeof(afs_int32), &tSize);
3954 maxDgramPackets = (afs_uint32) ntohl(tSize);
3955 maxDgramPackets = MIN(maxDgramPackets, rxi_nDgramPackets);
3956 maxDgramPackets = MIN(maxDgramPackets, peer->ifDgramPackets);
3957 if (peer->natMTU < peer->ifMTU)
3958 maxDgramPackets = MIN(maxDgramPackets, rxi_AdjustDgramPackets(1, peer->natMTU));
3959 if (maxDgramPackets > 1) {
3960 peer->maxDgramPackets = maxDgramPackets;
3961 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
3963 peer->maxDgramPackets = 1;
3964 call->MTU = peer->natMTU;
3966 } else if (peer->maxDgramPackets > 1) {
3967 /* Restarted with lower version of RX */
3968 peer->maxDgramPackets = 1;
3970 } else if (peer->maxDgramPackets > 1
3971 || peer->maxMTU != OLD_MAX_PACKET_SIZE) {
3972 /* Restarted with lower version of RX */
3973 peer->maxMTU = OLD_MAX_PACKET_SIZE;
3974 peer->natMTU = OLD_MAX_PACKET_SIZE;
3975 peer->MTU = OLD_MAX_PACKET_SIZE;
3976 peer->maxDgramPackets = 1;
3977 peer->nDgramPackets = 1;
3979 call->MTU = OLD_MAX_PACKET_SIZE;
3984 * Calculate how many datagrams were successfully received after
3985 * the first missing packet and adjust the negative ack counter
3990 nNacked = (nNacked + call->nDgramPackets - 1) / call->nDgramPackets;
3991 if (call->nNacks < nNacked) {
3992 call->nNacks = nNacked;
3995 call->nAcks += newAckCount;
3999 if (call->flags & RX_CALL_FAST_RECOVER) {
4001 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4003 call->flags &= ~RX_CALL_FAST_RECOVER;
4004 call->cwind = call->nextCwind;
4005 call->nextCwind = 0;
4008 call->nCwindAcks = 0;
4009 } else if (nNacked && call->nNacks >= (u_short) rx_nackThreshold) {
4010 /* Three negative acks in a row trigger congestion recovery */
4011 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4012 MUTEX_EXIT(&peer->peer_lock);
4013 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
4014 /* someone else is waiting to start recovery */
4017 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
4018 rxi_WaitforTQBusy(call);
4019 MUTEX_ENTER(&peer->peer_lock);
4020 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4021 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
4022 call->flags |= RX_CALL_FAST_RECOVER;
4023 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
4025 MIN((int)(call->ssthresh + rx_nackThreshold), rx_maxSendWindow);
4026 call->nDgramPackets = MAX(2, (int)call->nDgramPackets) >> 1;
4027 call->nextCwind = call->ssthresh;
4030 peer->MTU = call->MTU;
4031 peer->cwind = call->nextCwind;
4032 peer->nDgramPackets = call->nDgramPackets;
4034 call->congestSeq = peer->congestSeq;
4035 /* Reset the resend times on the packets that were nacked
4036 * so we will retransmit as soon as the window permits*/
4037 for (acked = 0, queue_ScanBackwards(&call->tq, tp, nxp, rx_packet)) {
4039 if (!(tp->flags & RX_PKTFLAG_ACKED)) {
4040 clock_Zero(&tp->retryTime);
4042 } else if (tp->flags & RX_PKTFLAG_ACKED) {
4047 /* If cwind is smaller than ssthresh, then increase
4048 * the window one packet for each ack we receive (exponential
4050 * If cwind is greater than or equal to ssthresh then increase
4051 * the congestion window by one packet for each cwind acks we
4052 * receive (linear growth). */
4053 if (call->cwind < call->ssthresh) {
4055 MIN((int)call->ssthresh, (int)(call->cwind + newAckCount));
4056 call->nCwindAcks = 0;
4058 call->nCwindAcks += newAckCount;
4059 if (call->nCwindAcks >= call->cwind) {
4060 call->nCwindAcks = 0;
4061 call->cwind = MIN((int)(call->cwind + 1), rx_maxSendWindow);
4065 * If we have received several acknowledgements in a row then
4066 * it is time to increase the size of our datagrams
4068 if ((int)call->nAcks > rx_nDgramThreshold) {
4069 if (peer->maxDgramPackets > 1) {
4070 if (call->nDgramPackets < peer->maxDgramPackets) {
4071 call->nDgramPackets++;
4073 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4074 } else if (call->MTU < peer->maxMTU) {
4075 call->MTU += peer->natMTU;
4076 call->MTU = MIN(call->MTU, peer->maxMTU);
4082 MUTEX_EXIT(&peer->peer_lock); /* rxi_Start will lock peer. */
4084 /* Servers need to hold the call until all response packets have
4085 * been acknowledged. Soft acks are good enough since clients
4086 * are not allowed to clear their receive queues. */
4087 if (call->state == RX_STATE_HOLD
4088 && call->tfirst + call->nSoftAcked >= call->tnext) {
4089 call->state = RX_STATE_DALLY;
4090 rxi_ClearTransmitQueue(call, 0);
4091 rxevent_Cancel(call->keepAliveEvent, call, RX_CALL_REFCOUNT_ALIVE);
4092 } else if (!queue_IsEmpty(&call->tq)) {
4093 rxi_Start(0, call, 0, istack);
4098 /* Received a response to a challenge packet */
4100 rxi_ReceiveResponsePacket(register struct rx_connection *conn,
4101 register struct rx_packet *np, int istack)
4105 /* Ignore the packet if we're the client */
4106 if (conn->type == RX_CLIENT_CONNECTION)
4109 /* If already authenticated, ignore the packet (it's probably a retry) */
4110 if (RXS_CheckAuthentication(conn->securityObject, conn) == 0)
4113 /* Otherwise, have the security object evaluate the response packet */
4114 error = RXS_CheckResponse(conn->securityObject, conn, np);
4116 /* If the response is invalid, reset the connection, sending
4117 * an abort to the peer */
4121 rxi_ConnectionError(conn, error);
4122 MUTEX_ENTER(&conn->conn_data_lock);
4123 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4124 MUTEX_EXIT(&conn->conn_data_lock);
4127 /* If the response is valid, any calls waiting to attach
4128 * servers can now do so */
4131 for (i = 0; i < RX_MAXCALLS; i++) {
4132 struct rx_call *call = conn->call[i];
4134 MUTEX_ENTER(&call->lock);
4135 if (call->state == RX_STATE_PRECALL)
4136 rxi_AttachServerProc(call, (osi_socket) - 1, NULL, NULL);
4137 /* tnop can be null if newcallp is null */
4138 MUTEX_EXIT(&call->lock);
4142 /* Update the peer reachability information, just in case
4143 * some calls went into attach-wait while we were waiting
4144 * for authentication..
4146 rxi_UpdatePeerReach(conn, NULL);
4151 /* A client has received an authentication challenge: the security
4152 * object is asked to cough up a respectable response packet to send
4153 * back to the server. The server is responsible for retrying the
4154 * challenge if it fails to get a response. */
4157 rxi_ReceiveChallengePacket(register struct rx_connection *conn,
4158 register struct rx_packet *np, int istack)
4162 /* Ignore the challenge if we're the server */
4163 if (conn->type == RX_SERVER_CONNECTION)
4166 /* Ignore the challenge if the connection is otherwise idle; someone's
4167 * trying to use us as an oracle. */
4168 if (!rxi_HasActiveCalls(conn))
4171 /* Send the security object the challenge packet. It is expected to fill
4172 * in the response. */
4173 error = RXS_GetResponse(conn->securityObject, conn, np);
4175 /* If the security object is unable to return a valid response, reset the
4176 * connection and send an abort to the peer. Otherwise send the response
4177 * packet to the peer connection. */
4179 rxi_ConnectionError(conn, error);
4180 MUTEX_ENTER(&conn->conn_data_lock);
4181 np = rxi_SendConnectionAbort(conn, np, istack, 0);
4182 MUTEX_EXIT(&conn->conn_data_lock);
4184 np = rxi_SendSpecial((struct rx_call *)0, conn, np,
4185 RX_PACKET_TYPE_RESPONSE, NULL, -1, istack);
4191 /* Find an available server process to service the current request in
4192 * the given call structure. If one isn't available, queue up this
4193 * call so it eventually gets one */
4195 rxi_AttachServerProc(register struct rx_call *call,
4196 register osi_socket socket, register int *tnop,
4197 register struct rx_call **newcallp)
4199 register struct rx_serverQueueEntry *sq;
4200 register struct rx_service *service = call->conn->service;
4201 register int haveQuota = 0;
4203 /* May already be attached */
4204 if (call->state == RX_STATE_ACTIVE)
4207 MUTEX_ENTER(&rx_serverPool_lock);
4209 haveQuota = QuotaOK(service);
4210 if ((!haveQuota) || queue_IsEmpty(&rx_idleServerQueue)) {
4211 /* If there are no processes available to service this call,
4212 * put the call on the incoming call queue (unless it's
4213 * already on the queue).
4215 #ifdef RX_ENABLE_LOCKS
4217 ReturnToServerPool(service);
4218 #endif /* RX_ENABLE_LOCKS */
4220 if (!(call->flags & RX_CALL_WAIT_PROC)) {
4221 call->flags |= RX_CALL_WAIT_PROC;
4222 MUTEX_ENTER(&rx_stats_mutex);
4225 MUTEX_EXIT(&rx_stats_mutex);
4226 rxi_calltrace(RX_CALL_ARRIVAL, call);
4227 SET_CALL_QUEUE_LOCK(call, &rx_serverPool_lock);
4228 queue_Append(&rx_incomingCallQueue, call);
4231 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
4233 /* If hot threads are enabled, and both newcallp and sq->socketp
4234 * are non-null, then this thread will process the call, and the
4235 * idle server thread will start listening on this threads socket.
4238 if (rx_enable_hot_thread && newcallp && sq->socketp) {
4241 *sq->socketp = socket;
4242 clock_GetTime(&call->startTime);
4243 CALL_HOLD(call, RX_CALL_REFCOUNT_BEGIN);
4247 if (call->flags & RX_CALL_WAIT_PROC) {
4248 /* Conservative: I don't think this should happen */
4249 call->flags &= ~RX_CALL_WAIT_PROC;
4250 if (queue_IsOnQueue(call)) {
4252 rx_MutexDecrement(rx_nWaiting, rx_stats_mutex);
4255 call->state = RX_STATE_ACTIVE;
4256 call->mode = RX_MODE_RECEIVING;
4257 #ifdef RX_KERNEL_TRACE
4259 int glockOwner = ISAFS_GLOCK();
4262 afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
4263 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
4269 if (call->flags & RX_CALL_CLEARED) {
4270 /* send an ack now to start the packet flow up again */
4271 call->flags &= ~RX_CALL_CLEARED;
4272 rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4274 #ifdef RX_ENABLE_LOCKS
4277 service->nRequestsRunning++;
4278 if (service->nRequestsRunning <= service->minProcs)
4284 MUTEX_EXIT(&rx_serverPool_lock);
4287 /* Delay the sending of an acknowledge event for a short while, while
4288 * a new call is being prepared (in the case of a client) or a reply
4289 * is being prepared (in the case of a server). Rather than sending
4290 * an ack packet, an ACKALL packet is sent. */
4292 rxi_AckAll(struct rxevent *event, register struct rx_call *call, char *dummy)
4294 #ifdef RX_ENABLE_LOCKS
4296 MUTEX_ENTER(&call->lock);
4297 call->delayedAckEvent = NULL;
4298 CALL_RELE(call, RX_CALL_REFCOUNT_ACKALL);
4300 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4301 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4303 MUTEX_EXIT(&call->lock);
4304 #else /* RX_ENABLE_LOCKS */
4306 call->delayedAckEvent = NULL;
4307 rxi_SendSpecial(call, call->conn, (struct rx_packet *)0,
4308 RX_PACKET_TYPE_ACKALL, NULL, 0, 0);
4309 #endif /* RX_ENABLE_LOCKS */
4313 rxi_SendDelayedAck(struct rxevent *event, void *arg1, void *unused)
4315 struct rx_call *call = arg1;
4316 #ifdef RX_ENABLE_LOCKS
4318 MUTEX_ENTER(&call->lock);
4319 if (event == call->delayedAckEvent)
4320 call->delayedAckEvent = NULL;
4321 CALL_RELE(call, RX_CALL_REFCOUNT_DELAY);
4323 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4325 MUTEX_EXIT(&call->lock);
4326 #else /* RX_ENABLE_LOCKS */
4328 call->delayedAckEvent = NULL;
4329 (void)rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
4330 #endif /* RX_ENABLE_LOCKS */
4334 #ifdef RX_ENABLE_LOCKS
4335 /* Set ack in all packets in transmit queue. rxi_Start will deal with
4336 * clearing them out.
4339 rxi_SetAcksInTransmitQueue(register struct rx_call *call)
4341 register struct rx_packet *p, *tp;
4344 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4345 p->flags |= RX_PKTFLAG_ACKED;
4349 call->flags |= RX_CALL_TQ_CLEARME;
4350 call->flags |= RX_CALL_TQ_SOME_ACKED;
4353 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4354 call->tfirst = call->tnext;
4355 call->nSoftAcked = 0;
4357 if (call->flags & RX_CALL_FAST_RECOVER) {
4358 call->flags &= ~RX_CALL_FAST_RECOVER;
4359 call->cwind = call->nextCwind;
4360 call->nextCwind = 0;
4363 CV_SIGNAL(&call->cv_twind);
4365 #endif /* RX_ENABLE_LOCKS */
4367 /* Clear out the transmit queue for the current call (all packets have
4368 * been received by peer) */
4370 rxi_ClearTransmitQueue(register struct rx_call *call, register int force)
4372 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4373 register struct rx_packet *p, *tp;
4375 if (!force && (call->flags & RX_CALL_TQ_BUSY)) {
4377 for (queue_Scan(&call->tq, p, tp, rx_packet)) {
4378 p->flags |= RX_PKTFLAG_ACKED;
4382 call->flags |= RX_CALL_TQ_CLEARME;
4383 call->flags |= RX_CALL_TQ_SOME_ACKED;
4386 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4387 #ifdef RXDEBUG_PACKET
4389 #endif /* RXDEBUG_PACKET */
4390 rxi_FreePackets(0, &call->tq);
4391 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4392 call->flags &= ~RX_CALL_TQ_CLEARME;
4394 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4396 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
4397 call->tfirst = call->tnext; /* implicitly acknowledge all data already sent */
4398 call->nSoftAcked = 0;
4400 if (call->flags & RX_CALL_FAST_RECOVER) {
4401 call->flags &= ~RX_CALL_FAST_RECOVER;
4402 call->cwind = call->nextCwind;
4404 #ifdef RX_ENABLE_LOCKS
4405 CV_SIGNAL(&call->cv_twind);
4407 osi_rxWakeup(&call->twind);
4412 rxi_ClearReceiveQueue(register struct rx_call *call)
4414 if (queue_IsNotEmpty(&call->rq)) {
4417 count = rxi_FreePackets(0, &call->rq);
4418 rx_packetReclaims += count;
4419 #ifdef RXDEBUG_PACKET
4421 if ( call->rqc != 0 )
4422 dpf(("rxi_ClearReceiveQueue call %x rqc %u != 0", call, call->rqc));
4424 call->flags &= ~(RX_CALL_RECEIVE_DONE | RX_CALL_HAVE_LAST);
4426 if (call->state == RX_STATE_PRECALL) {
4427 call->flags |= RX_CALL_CLEARED;
4431 /* Send an abort packet for the specified call */
4433 rxi_SendCallAbort(register struct rx_call *call, struct rx_packet *packet,
4434 int istack, int force)
4437 struct clock when, now;
4442 /* Clients should never delay abort messages */
4443 if (rx_IsClientConn(call->conn))
4446 if (call->abortCode != call->error) {
4447 call->abortCode = call->error;
4448 call->abortCount = 0;
4451 if (force || rxi_callAbortThreshhold == 0
4452 || call->abortCount < rxi_callAbortThreshhold) {
4453 if (call->delayedAbortEvent) {
4454 rxevent_Cancel(call->delayedAbortEvent, call,
4455 RX_CALL_REFCOUNT_ABORT);
4457 error = htonl(call->error);
4460 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
4461 (char *)&error, sizeof(error), istack);
4462 } else if (!call->delayedAbortEvent) {
4463 clock_GetTime(&now);
4465 clock_Addmsec(&when, rxi_callAbortDelay);
4466 CALL_HOLD(call, RX_CALL_REFCOUNT_ABORT);
4467 call->delayedAbortEvent =
4468 rxevent_PostNow(&when, &now, rxi_SendDelayedCallAbort, call, 0);
4473 /* Send an abort packet for the specified connection. Packet is an
4474 * optional pointer to a packet that can be used to send the abort.
4475 * Once the number of abort messages reaches the threshhold, an
4476 * event is scheduled to send the abort. Setting the force flag
4477 * overrides sending delayed abort messages.
4479 * NOTE: Called with conn_data_lock held. conn_data_lock is dropped
4480 * to send the abort packet.
4483 rxi_SendConnectionAbort(register struct rx_connection *conn,
4484 struct rx_packet *packet, int istack, int force)
4487 struct clock when, now;
4489 if (!rx_ConnError(conn))
4492 /* Clients should never delay abort messages */
4493 if (rx_IsClientConn(conn))
4496 if (force || rxi_connAbortThreshhold == 0
4497 || conn->abortCount < rxi_connAbortThreshhold) {
4498 if (conn->delayedAbortEvent) {
4499 rxevent_Cancel(conn->delayedAbortEvent, (struct rx_call *)0, 0);
4501 error = htonl(rx_ConnError(conn));
4503 MUTEX_EXIT(&conn->conn_data_lock);
4505 rxi_SendSpecial((struct rx_call *)0, conn, packet,
4506 RX_PACKET_TYPE_ABORT, (char *)&error,
4507 sizeof(error), istack);
4508 MUTEX_ENTER(&conn->conn_data_lock);
4509 } else if (!conn->delayedAbortEvent) {
4510 clock_GetTime(&now);
4512 clock_Addmsec(&when, rxi_connAbortDelay);
4513 conn->delayedAbortEvent =
4514 rxevent_PostNow(&when, &now, rxi_SendDelayedConnAbort, conn, 0);
4520 * Associate an error all of the calls owned by a connection. Called
4521 * with error non-zero. This is only for really fatal things, like
4522 * bad authentication responses. The connection itself is set in
4523 * error at this point, so that future packets received will be
4527 rxi_ConnectionError(register struct rx_connection *conn,
4528 register afs_int32 error)
4532 struct rx_connection *tconn;
4534 dpf(("rxi_ConnectionError conn %x error %d", conn, error));
4536 MUTEX_ENTER(&conn->conn_data_lock);
4537 if (conn->challengeEvent)
4538 rxevent_Cancel(conn->challengeEvent, (struct rx_call *)0, 0);
4539 if (conn->checkReachEvent) {
4540 rxevent_Cancel(conn->checkReachEvent, (struct rx_call *)0, 0);
4541 conn->checkReachEvent = 0;
4542 conn->flags &= ~RX_CONN_ATTACHWAIT;
4545 MUTEX_EXIT(&conn->conn_data_lock);
4547 for ( tconn = rx_IsClonedConn(conn) ? conn->parent : conn;
4549 tconn = tconn->next_clone) {
4550 for (i = 0; i < RX_MAXCALLS; i++) {
4551 struct rx_call *call = tconn->call[i];
4553 MUTEX_ENTER(&call->lock);
4554 rxi_CallError(call, error);
4555 MUTEX_EXIT(&call->lock);
4559 rx_SetConnError(conn, error);
4560 rx_MutexIncrement(rx_stats.fatalErrors, rx_stats_mutex);
4565 rxi_CallError(register struct rx_call *call, afs_int32 error)
4568 osirx_AssertMine(&call->lock, "rxi_CallError");
4570 dpf(("rxi_CallError call %x error %d call->error %d", call, error, call->error));
4572 error = call->error;
4574 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4575 if (!((call->flags & RX_CALL_TQ_BUSY) || (call->tqWaiters > 0))) {
4576 rxi_ResetCall(call, 0);
4579 rxi_ResetCall(call, 0);
4581 call->error = error;
4582 call->mode = RX_MODE_ERROR;
4585 /* Reset various fields in a call structure, and wakeup waiting
4586 * processes. Some fields aren't changed: state & mode are not
4587 * touched (these must be set by the caller), and bufptr, nLeft, and
4588 * nFree are not reset, since these fields are manipulated by
4589 * unprotected macros, and may only be reset by non-interrupting code.
4592 /* this code requires that call->conn be set properly as a pre-condition. */
4593 #endif /* ADAPT_WINDOW */
4596 rxi_ResetCall(register struct rx_call *call, register int newcall)
4599 register struct rx_peer *peer;
4600 struct rx_packet *packet;
4602 osirx_AssertMine(&call->lock, "rxi_ResetCall");
4604 dpf(("rxi_ResetCall(call %x, newcall %d)\n", call, newcall));
4606 /* Notify anyone who is waiting for asynchronous packet arrival */
4607 if (call->arrivalProc) {
4608 (*call->arrivalProc) (call, call->arrivalProcHandle,
4609 call->arrivalProcArg);
4610 call->arrivalProc = (void (*)())0;
4613 if (call->delayedAbortEvent) {
4614 rxevent_Cancel(call->delayedAbortEvent, call, RX_CALL_REFCOUNT_ABORT);
4615 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
4617 rxi_SendCallAbort(call, packet, 0, 1);
4618 rxi_FreePacket(packet);
4623 * Update the peer with the congestion information in this call
4624 * so other calls on this connection can pick up where this call
4625 * left off. If the congestion sequence numbers don't match then
4626 * another call experienced a retransmission.
4628 peer = call->conn->peer;
4629 MUTEX_ENTER(&peer->peer_lock);
4631 if (call->congestSeq == peer->congestSeq) {
4632 peer->cwind = MAX(peer->cwind, call->cwind);
4633 peer->MTU = MAX(peer->MTU, call->MTU);
4634 peer->nDgramPackets =
4635 MAX(peer->nDgramPackets, call->nDgramPackets);
4638 call->abortCode = 0;
4639 call->abortCount = 0;
4641 if (peer->maxDgramPackets > 1) {
4642 call->MTU = RX_HEADER_SIZE + RX_JUMBOBUFFERSIZE;
4644 call->MTU = peer->MTU;
4646 call->cwind = MIN((int)peer->cwind, (int)peer->nDgramPackets);
4647 call->ssthresh = rx_maxSendWindow;
4648 call->nDgramPackets = peer->nDgramPackets;
4649 call->congestSeq = peer->congestSeq;
4650 MUTEX_EXIT(&peer->peer_lock);
4652 flags = call->flags;
4653 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
4654 if (flags & RX_CALL_TQ_BUSY) {
4655 call->flags = RX_CALL_TQ_CLEARME | RX_CALL_TQ_BUSY;
4656 call->flags |= (flags & RX_CALL_TQ_WAIT);
4658 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
4660 rxi_ClearTransmitQueue(call, 1);
4661 /* why init the queue if you just emptied it? queue_Init(&call->tq); */
4662 if (call->tqWaiters || (flags & RX_CALL_TQ_WAIT)) {
4663 dpf(("rcall %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
4666 while (call->tqWaiters) {
4667 #ifdef RX_ENABLE_LOCKS
4668 CV_BROADCAST(&call->cv_tq);
4669 #else /* RX_ENABLE_LOCKS */
4670 osi_rxWakeup(&call->tq);
4671 #endif /* RX_ENABLE_LOCKS */
4676 rxi_ClearReceiveQueue(call);
4677 /* why init the queue if you just emptied it? queue_Init(&call->rq); */
4679 if (call->currentPacket) {
4680 call->currentPacket->flags &= ~RX_PKTFLAG_CP;
4681 call->currentPacket->flags |= RX_PKTFLAG_IOVQ;
4682 queue_Prepend(&call->iovq, call->currentPacket);
4683 #ifdef RXDEBUG_PACKET
4685 #endif /* RXDEBUG_PACKET */
4686 call->currentPacket = (struct rx_packet *)0;
4688 call->curlen = call->nLeft = call->nFree = 0;
4690 #ifdef RXDEBUG_PACKET
4693 rxi_FreePackets(0, &call->iovq);
4696 call->twind = call->conn->twind[call->channel];
4697 call->rwind = call->conn->rwind[call->channel];
4698 call->nSoftAcked = 0;
4699 call->nextCwind = 0;
4702 call->nCwindAcks = 0;
4703 call->nSoftAcks = 0;
4704 call->nHardAcks = 0;
4706 call->tfirst = call->rnext = call->tnext = 1;
4708 call->lastAcked = 0;
4709 call->localStatus = call->remoteStatus = 0;
4711 if (flags & RX_CALL_READER_WAIT) {
4712 #ifdef RX_ENABLE_LOCKS
4713 CV_BROADCAST(&call->cv_rq);
4715 osi_rxWakeup(&call->rq);
4718 if (flags & RX_CALL_WAIT_PACKETS) {
4719 MUTEX_ENTER(&rx_freePktQ_lock);
4720 rxi_PacketsUnWait(); /* XXX */
4721 MUTEX_EXIT(&rx_freePktQ_lock);
4723 #ifdef RX_ENABLE_LOCKS
4724 CV_SIGNAL(&call->cv_twind);
4726 if (flags & RX_CALL_WAIT_WINDOW_ALLOC)
4727 osi_rxWakeup(&call->twind);
4730 #ifdef RX_ENABLE_LOCKS
4731 /* The following ensures that we don't mess with any queue while some
4732 * other thread might also be doing so. The call_queue_lock field is
4733 * is only modified under the call lock. If the call is in the process
4734 * of being removed from a queue, the call is not locked until the
4735 * the queue lock is dropped and only then is the call_queue_lock field
4736 * zero'd out. So it's safe to lock the queue if call_queue_lock is set.
4737 * Note that any other routine which removes a call from a queue has to
4738 * obtain the queue lock before examing the queue and removing the call.
4740 if (call->call_queue_lock) {
4741 MUTEX_ENTER(call->call_queue_lock);
4742 if (queue_IsOnQueue(call)) {
4744 if (flags & RX_CALL_WAIT_PROC) {
4745 rx_MutexDecrement(rx_nWaiting, rx_stats_mutex);
4748 MUTEX_EXIT(call->call_queue_lock);
4749 CLEAR_CALL_QUEUE_LOCK(call);
4751 #else /* RX_ENABLE_LOCKS */
4752 if (queue_IsOnQueue(call)) {
4754 if (flags & RX_CALL_WAIT_PROC)
4757 #endif /* RX_ENABLE_LOCKS */
4759 rxi_KeepAliveOff(call);
4760 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
4763 /* Send an acknowledge for the indicated packet (seq,serial) of the
4764 * indicated call, for the indicated reason (reason). This
4765 * acknowledge will specifically acknowledge receiving the packet, and
4766 * will also specify which other packets for this call have been
4767 * received. This routine returns the packet that was used to the
4768 * caller. The caller is responsible for freeing it or re-using it.
4769 * This acknowledgement also returns the highest sequence number
4770 * actually read out by the higher level to the sender; the sender
4771 * promises to keep around packets that have not been read by the
4772 * higher level yet (unless, of course, the sender decides to abort
4773 * the call altogether). Any of p, seq, serial, pflags, or reason may
4774 * be set to zero without ill effect. That is, if they are zero, they
4775 * will not convey any information.
4776 * NOW there is a trailer field, after the ack where it will safely be
4777 * ignored by mundanes, which indicates the maximum size packet this
4778 * host can swallow. */
4780 register struct rx_packet *optionalPacket; use to send ack (or null)
4781 int seq; Sequence number of the packet we are acking
4782 int serial; Serial number of the packet
4783 int pflags; Flags field from packet header
4784 int reason; Reason an acknowledge was prompted
4788 rxi_SendAck(register struct rx_call *call,
4789 register struct rx_packet *optionalPacket, int serial, int reason,
4792 struct rx_ackPacket *ap;
4793 register struct rx_packet *rqp;
4794 register struct rx_packet *nxp; /* For queue_Scan */
4795 register struct rx_packet *p;
4798 #ifdef RX_ENABLE_TSFPQ
4799 struct rx_ts_info_t * rx_ts_info;
4803 * Open the receive window once a thread starts reading packets
4805 if (call->rnext > 1) {
4806 call->conn->rwind[call->channel] = call->rwind = rx_maxReceiveWindow;
4809 call->nHardAcks = 0;
4810 call->nSoftAcks = 0;
4811 if (call->rnext > call->lastAcked)
4812 call->lastAcked = call->rnext;
4816 rx_computelen(p, p->length); /* reset length, you never know */
4817 } /* where that's been... */
4818 #ifdef RX_ENABLE_TSFPQ
4820 RX_TS_INFO_GET(rx_ts_info);
4821 if ((p = rx_ts_info->local_special_packet)) {
4822 rx_computelen(p, p->length);
4823 } else if ((p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4824 rx_ts_info->local_special_packet = p;
4825 } else { /* We won't send the ack, but don't panic. */
4826 return optionalPacket;
4830 else if (!(p = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL))) {
4831 /* We won't send the ack, but don't panic. */
4832 return optionalPacket;
4837 rx_AckDataSize(call->rwind) + 4 * sizeof(afs_int32) -
4840 if (rxi_AllocDataBuf(p, templ, RX_PACKET_CLASS_SPECIAL) > 0) {
4841 #ifndef RX_ENABLE_TSFPQ
4842 if (!optionalPacket)
4845 return optionalPacket;
4847 templ = rx_AckDataSize(call->rwind) + 2 * sizeof(afs_int32);
4848 if (rx_Contiguous(p) < templ) {
4849 #ifndef RX_ENABLE_TSFPQ
4850 if (!optionalPacket)
4853 return optionalPacket;
4858 /* MTUXXX failing to send an ack is very serious. We should */
4859 /* try as hard as possible to send even a partial ack; it's */
4860 /* better than nothing. */
4861 ap = (struct rx_ackPacket *)rx_DataOf(p);
4862 ap->bufferSpace = htonl(0); /* Something should go here, sometime */
4863 ap->reason = reason;
4865 /* The skew computation used to be bogus, I think it's better now. */
4866 /* We should start paying attention to skew. XXX */
4867 ap->serial = htonl(serial);
4868 ap->maxSkew = 0; /* used to be peer->inPacketSkew */
4870 ap->firstPacket = htonl(call->rnext); /* First packet not yet forwarded to reader */
4871 ap->previousPacket = htonl(call->rprev); /* Previous packet received */
4873 /* No fear of running out of ack packet here because there can only be at most
4874 * one window full of unacknowledged packets. The window size must be constrained
4875 * to be less than the maximum ack size, of course. Also, an ack should always
4876 * fit into a single packet -- it should not ever be fragmented. */
4877 for (offset = 0, queue_Scan(&call->rq, rqp, nxp, rx_packet)) {
4878 if (!rqp || !call->rq.next
4879 || (rqp->header.seq > (call->rnext + call->rwind))) {
4880 #ifndef RX_ENABLE_TSFPQ
4881 if (!optionalPacket)
4884 rxi_CallError(call, RX_CALL_DEAD);
4885 return optionalPacket;
4888 while (rqp->header.seq > call->rnext + offset)
4889 ap->acks[offset++] = RX_ACK_TYPE_NACK;
4890 ap->acks[offset++] = RX_ACK_TYPE_ACK;
4892 if ((offset > (u_char) rx_maxReceiveWindow) || (offset > call->rwind)) {
4893 #ifndef RX_ENABLE_TSFPQ
4894 if (!optionalPacket)
4897 rxi_CallError(call, RX_CALL_DEAD);
4898 return optionalPacket;
4903 p->length = rx_AckDataSize(offset) + 4 * sizeof(afs_int32);
4905 /* these are new for AFS 3.3 */
4906 templ = rxi_AdjustMaxMTU(call->conn->peer->ifMTU, rx_maxReceiveSize);
4907 templ = htonl(templ);
4908 rx_packetwrite(p, rx_AckDataSize(offset), sizeof(afs_int32), &templ);
4909 templ = htonl(call->conn->peer->ifMTU);
4910 rx_packetwrite(p, rx_AckDataSize(offset) + sizeof(afs_int32),
4911 sizeof(afs_int32), &templ);
4913 /* new for AFS 3.4 */
4914 templ = htonl(call->rwind);
4915 rx_packetwrite(p, rx_AckDataSize(offset) + 2 * sizeof(afs_int32),
4916 sizeof(afs_int32), &templ);
4918 /* new for AFS 3.5 */
4919 templ = htonl(call->conn->peer->ifDgramPackets);
4920 rx_packetwrite(p, rx_AckDataSize(offset) + 3 * sizeof(afs_int32),
4921 sizeof(afs_int32), &templ);
4923 p->header.serviceId = call->conn->serviceId;
4924 p->header.cid = (call->conn->cid | call->channel);
4925 p->header.callNumber = *call->callNumber;
4927 p->header.securityIndex = call->conn->securityIndex;
4928 p->header.epoch = call->conn->epoch;
4929 p->header.type = RX_PACKET_TYPE_ACK;
4930 p->header.flags = RX_SLOW_START_OK;
4931 if (reason == RX_ACK_PING) {
4932 p->header.flags |= RX_REQUEST_ACK;
4934 clock_GetTime(&call->pingRequestTime);
4937 if (call->conn->type == RX_CLIENT_CONNECTION)
4938 p->header.flags |= RX_CLIENT_INITIATED;
4942 if (rxdebug_active) {
4946 len = _snprintf(msg, sizeof(msg),
4947 "tid[%d] SACK: reason %s serial %u previous %u seq %u first %u acks %u space %u ",
4948 GetCurrentThreadId(), rx_ack_reason(ap->reason),
4949 ntohl(ap->serial), ntohl(ap->previousPacket),
4950 (unsigned int)p->header.seq, ntohl(ap->firstPacket),
4951 ap->nAcks, ntohs(ap->bufferSpace) );
4955 for (offset = 0; offset < ap->nAcks && len < sizeof(msg); offset++)
4956 msg[len++] = (ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*');
4960 OutputDebugString(msg);
4962 #else /* AFS_NT40_ENV */
4964 fprintf(rx_Log, "SACK: reason %x previous %u seq %u first %u ",
4965 ap->reason, ntohl(ap->previousPacket),
4966 (unsigned int)p->header.seq, ntohl(ap->firstPacket));
4968 for (offset = 0; offset < ap->nAcks; offset++)
4969 putc(ap->acks[offset] == RX_ACK_TYPE_NACK ? '-' : '*',
4974 #endif /* AFS_NT40_ENV */
4977 register int i, nbytes = p->length;
4979 for (i = 1; i < p->niovecs; i++) { /* vec 0 is ALWAYS header */
4980 if (nbytes <= p->wirevec[i].iov_len) {
4981 register int savelen, saven;
4983 savelen = p->wirevec[i].iov_len;
4985 p->wirevec[i].iov_len = nbytes;
4987 rxi_Send(call, p, istack);
4988 p->wirevec[i].iov_len = savelen;
4992 nbytes -= p->wirevec[i].iov_len;
4995 rx_MutexIncrement(rx_stats.ackPacketsSent, rx_stats_mutex);
4996 #ifndef RX_ENABLE_TSFPQ
4997 if (!optionalPacket)
5000 return optionalPacket; /* Return packet for re-use by caller */
5003 /* Send all of the packets in the list in single datagram */
5005 rxi_SendList(struct rx_call *call, struct rx_packet **list, int len,
5006 int istack, int moreFlag, struct clock *now,
5007 struct clock *retryTime, int resending)
5012 struct rx_connection *conn = call->conn;
5013 struct rx_peer *peer = conn->peer;
5015 MUTEX_ENTER(&peer->peer_lock);
5018 peer->reSends += len;
5019 rx_MutexIncrement(rx_stats.dataPacketsSent, rx_stats_mutex);
5020 MUTEX_EXIT(&peer->peer_lock);
5022 if (list[len - 1]->header.flags & RX_LAST_PACKET) {
5026 /* Set the packet flags and schedule the resend events */
5027 /* Only request an ack for the last packet in the list */
5028 for (i = 0; i < len; i++) {
5029 list[i]->retryTime = *retryTime;
5030 if (list[i]->header.serial) {
5031 /* Exponentially backoff retry times */
5032 if (list[i]->backoff < MAXBACKOFF) {
5033 /* so it can't stay == 0 */
5034 list[i]->backoff = (list[i]->backoff << 1) + 1;
5037 clock_Addmsec(&(list[i]->retryTime),
5038 ((afs_uint32) list[i]->backoff) << 8);
5041 /* Wait a little extra for the ack on the last packet */
5042 if (lastPacket && !(list[i]->header.flags & RX_CLIENT_INITIATED)) {
5043 clock_Addmsec(&(list[i]->retryTime), 400);
5046 /* Record the time sent */
5047 list[i]->timeSent = *now;
5049 /* Ask for an ack on retransmitted packets, on every other packet
5050 * if the peer doesn't support slow start. Ask for an ack on every
5051 * packet until the congestion window reaches the ack rate. */
5052 if (list[i]->header.serial) {
5054 rx_MutexIncrement(rx_stats.dataPacketsReSent, rx_stats_mutex);
5056 /* improved RTO calculation- not Karn */
5057 list[i]->firstSent = *now;
5058 if (!lastPacket && (call->cwind <= (u_short) (conn->ackRate + 1)
5059 || (!(call->flags & RX_CALL_SLOW_START_OK)
5060 && (list[i]->header.seq & 1)))) {
5065 MUTEX_ENTER(&peer->peer_lock);
5069 rx_MutexIncrement(rx_stats.dataPacketsSent, rx_stats_mutex);
5070 MUTEX_EXIT(&peer->peer_lock);
5072 /* Tag this packet as not being the last in this group,
5073 * for the receiver's benefit */
5074 if (i < len - 1 || moreFlag) {
5075 list[i]->header.flags |= RX_MORE_PACKETS;
5078 /* Install the new retransmit time for the packet, and
5079 * record the time sent */
5080 list[i]->timeSent = *now;
5084 list[len - 1]->header.flags |= RX_REQUEST_ACK;
5087 /* Since we're about to send a data packet to the peer, it's
5088 * safe to nuke any scheduled end-of-packets ack */
5089 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5091 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5092 MUTEX_EXIT(&call->lock);
5094 rxi_SendPacketList(call, conn, list, len, istack);
5096 rxi_SendPacket(call, conn, list[0], istack);
5098 MUTEX_ENTER(&call->lock);
5099 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5101 /* Update last send time for this call (for keep-alive
5102 * processing), and for the connection (so that we can discover
5103 * idle connections) */
5104 call->lastSendData = conn->lastSendTime = call->lastSendTime = clock_Sec();
5107 /* When sending packets we need to follow these rules:
5108 * 1. Never send more than maxDgramPackets in a jumbogram.
5109 * 2. Never send a packet with more than two iovecs in a jumbogram.
5110 * 3. Never send a retransmitted packet in a jumbogram.
5111 * 4. Never send more than cwind/4 packets in a jumbogram
5112 * We always keep the last list we should have sent so we
5113 * can set the RX_MORE_PACKETS flags correctly.
5116 rxi_SendXmitList(struct rx_call *call, struct rx_packet **list, int len,
5117 int istack, struct clock *now, struct clock *retryTime,
5120 int i, cnt, lastCnt = 0;
5121 struct rx_packet **listP, **lastP = 0;
5122 struct rx_peer *peer = call->conn->peer;
5123 int morePackets = 0;
5125 for (cnt = 0, listP = &list[0], i = 0; i < len; i++) {
5126 /* Does the current packet force us to flush the current list? */
5128 && (list[i]->header.serial || (list[i]->flags & RX_PKTFLAG_ACKED)
5129 || list[i]->length > RX_JUMBOBUFFERSIZE)) {
5131 rxi_SendList(call, lastP, lastCnt, istack, 1, now, retryTime,
5133 /* If the call enters an error state stop sending, or if
5134 * we entered congestion recovery mode, stop sending */
5135 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5143 /* Add the current packet to the list if it hasn't been acked.
5144 * Otherwise adjust the list pointer to skip the current packet. */
5145 if (!(list[i]->flags & RX_PKTFLAG_ACKED)) {
5147 /* Do we need to flush the list? */
5148 if (cnt >= (int)peer->maxDgramPackets
5149 || cnt >= (int)call->nDgramPackets || cnt >= (int)call->cwind
5150 || list[i]->header.serial
5151 || list[i]->length != RX_JUMBOBUFFERSIZE) {
5153 rxi_SendList(call, lastP, lastCnt, istack, 1, now,
5154 retryTime, resending);
5155 /* If the call enters an error state stop sending, or if
5156 * we entered congestion recovery mode, stop sending */
5158 || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5163 listP = &list[i + 1];
5168 osi_Panic("rxi_SendList error");
5170 listP = &list[i + 1];
5174 /* Send the whole list when the call is in receive mode, when
5175 * the call is in eof mode, when we are in fast recovery mode,
5176 * and when we have the last packet */
5177 if ((list[len - 1]->header.flags & RX_LAST_PACKET)
5178 || call->mode == RX_MODE_RECEIVING || call->mode == RX_MODE_EOF
5179 || (call->flags & RX_CALL_FAST_RECOVER)) {
5180 /* Check for the case where the current list contains
5181 * an acked packet. Since we always send retransmissions
5182 * in a separate packet, we only need to check the first
5183 * packet in the list */
5184 if (cnt > 0 && !(listP[0]->flags & RX_PKTFLAG_ACKED)) {
5188 rxi_SendList(call, lastP, lastCnt, istack, morePackets, now,
5189 retryTime, resending);
5190 /* If the call enters an error state stop sending, or if
5191 * we entered congestion recovery mode, stop sending */
5192 if (call->error || (call->flags & RX_CALL_FAST_RECOVER_WAIT))
5196 rxi_SendList(call, listP, cnt, istack, 0, now, retryTime,
5199 } else if (lastCnt > 0) {
5200 rxi_SendList(call, lastP, lastCnt, istack, 0, now, retryTime,
5205 #ifdef RX_ENABLE_LOCKS
5206 /* Call rxi_Start, below, but with the call lock held. */
5208 rxi_StartUnlocked(struct rxevent *event,
5209 void *arg0, void *arg1, int istack)
5211 struct rx_call *call = arg0;
5213 MUTEX_ENTER(&call->lock);
5214 rxi_Start(event, call, arg1, istack);
5215 MUTEX_EXIT(&call->lock);
5217 #endif /* RX_ENABLE_LOCKS */
5219 /* This routine is called when new packets are readied for
5220 * transmission and when retransmission may be necessary, or when the
5221 * transmission window or burst count are favourable. This should be
5222 * better optimized for new packets, the usual case, now that we've
5223 * got rid of queues of send packets. XXXXXXXXXXX */
5225 rxi_Start(struct rxevent *event,
5226 void *arg0, void *arg1, int istack)
5228 struct rx_call *call = arg0;
5230 struct rx_packet *p;
5231 register struct rx_packet *nxp; /* Next pointer for queue_Scan */
5232 struct rx_peer *peer = call->conn->peer;
5233 struct clock now, usenow, retryTime;
5237 struct rx_packet **xmitList;
5240 /* If rxi_Start is being called as a result of a resend event,
5241 * then make sure that the event pointer is removed from the call
5242 * structure, since there is no longer a per-call retransmission
5244 if (event && event == call->resendEvent) {
5245 CALL_RELE(call, RX_CALL_REFCOUNT_RESEND);
5246 call->resendEvent = NULL;
5248 if (queue_IsEmpty(&call->tq)) {
5252 /* Timeouts trigger congestion recovery */
5253 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5254 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5255 /* someone else is waiting to start recovery */
5258 call->flags |= RX_CALL_FAST_RECOVER_WAIT;
5259 rxi_WaitforTQBusy(call);
5260 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5261 call->flags &= ~RX_CALL_FAST_RECOVER_WAIT;
5262 call->flags |= RX_CALL_FAST_RECOVER;
5263 if (peer->maxDgramPackets > 1) {
5264 call->MTU = RX_JUMBOBUFFERSIZE + RX_HEADER_SIZE;
5266 call->MTU = MIN(peer->natMTU, peer->maxMTU);
5268 call->ssthresh = MAX(4, MIN((int)call->cwind, (int)call->twind)) >> 1;
5269 call->nDgramPackets = 1;
5271 call->nextCwind = 1;
5274 MUTEX_ENTER(&peer->peer_lock);
5275 peer->MTU = call->MTU;
5276 peer->cwind = call->cwind;
5277 peer->nDgramPackets = 1;
5279 call->congestSeq = peer->congestSeq;
5280 MUTEX_EXIT(&peer->peer_lock);
5281 /* Clear retry times on packets. Otherwise, it's possible for
5282 * some packets in the queue to force resends at rates faster
5283 * than recovery rates.
5285 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5286 if (!(p->flags & RX_PKTFLAG_ACKED)) {
5287 clock_Zero(&p->retryTime);
5292 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5293 rx_MutexIncrement(rx_tq_debug.rxi_start_in_error, rx_stats_mutex);
5298 if (queue_IsNotEmpty(&call->tq)) { /* If we have anything to send */
5299 /* Get clock to compute the re-transmit time for any packets
5300 * in this burst. Note, if we back off, it's reasonable to
5301 * back off all of the packets in the same manner, even if
5302 * some of them have been retransmitted more times than more
5304 * Do a dance to avoid blocking after setting now. */
5305 clock_Zero(&retryTime);
5306 MUTEX_ENTER(&peer->peer_lock);
5307 clock_Add(&retryTime, &peer->timeout);
5308 MUTEX_EXIT(&peer->peer_lock);
5309 clock_GetTime(&now);
5310 clock_Add(&retryTime, &now);
5312 /* Send (or resend) any packets that need it, subject to
5313 * window restrictions and congestion burst control
5314 * restrictions. Ask for an ack on the last packet sent in
5315 * this burst. For now, we're relying upon the window being
5316 * considerably bigger than the largest number of packets that
5317 * are typically sent at once by one initial call to
5318 * rxi_Start. This is probably bogus (perhaps we should ask
5319 * for an ack when we're half way through the current
5320 * window?). Also, for non file transfer applications, this
5321 * may end up asking for an ack for every packet. Bogus. XXXX
5324 * But check whether we're here recursively, and let the other guy
5327 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5328 if (!(call->flags & RX_CALL_TQ_BUSY)) {
5329 call->flags |= RX_CALL_TQ_BUSY;
5331 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5333 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5334 call->flags &= ~RX_CALL_NEED_START;
5335 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5337 maxXmitPackets = MIN(call->twind, call->cwind);
5338 xmitList = (struct rx_packet **)
5339 #if defined(KERNEL) && !defined(UKERNEL) && defined(AFS_FBSD80_ENV)
5340 /* XXXX else we must drop any mtx we hold */
5341 afs_osi_Alloc_NoSleep(maxXmitPackets * sizeof(struct rx_packet *));
5343 osi_Alloc(maxXmitPackets * sizeof(struct rx_packet *));
5345 if (xmitList == NULL)
5346 osi_Panic("rxi_Start, failed to allocate xmit list");
5347 for (queue_Scan(&call->tq, p, nxp, rx_packet)) {
5348 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5349 /* We shouldn't be sending packets if a thread is waiting
5350 * to initiate congestion recovery */
5354 && (call->flags & RX_CALL_FAST_RECOVER)) {
5355 /* Only send one packet during fast recovery */
5358 if ((p->flags & RX_PKTFLAG_FREE)
5359 || (!queue_IsEnd(&call->tq, nxp)
5360 && (nxp->flags & RX_PKTFLAG_FREE))
5361 || (p == (struct rx_packet *)&rx_freePacketQueue)
5362 || (nxp == (struct rx_packet *)&rx_freePacketQueue)) {
5363 osi_Panic("rxi_Start: xmit queue clobbered");
5365 if (p->flags & RX_PKTFLAG_ACKED) {
5366 /* Since we may block, don't trust this */
5367 usenow.sec = usenow.usec = 0;
5368 rx_MutexIncrement(rx_stats.ignoreAckedPacket, rx_stats_mutex);
5369 continue; /* Ignore this packet if it has been acknowledged */
5372 /* Turn off all flags except these ones, which are the same
5373 * on each transmission */
5374 p->header.flags &= RX_PRESET_FLAGS;
5376 if (p->header.seq >=
5377 call->tfirst + MIN((int)call->twind,
5378 (int)(call->nSoftAcked +
5380 call->flags |= RX_CALL_WAIT_WINDOW_SEND; /* Wait for transmit window */
5381 /* Note: if we're waiting for more window space, we can
5382 * still send retransmits; hence we don't return here, but
5383 * break out to schedule a retransmit event */
5384 dpf(("call %d waiting for window",
5385 *(call->callNumber)));
5389 /* Transmit the packet if it needs to be sent. */
5390 if (!clock_Lt(&now, &p->retryTime)) {
5391 if (nXmitPackets == maxXmitPackets) {
5392 rxi_SendXmitList(call, xmitList, nXmitPackets,
5393 istack, &now, &retryTime,
5395 osi_Free(xmitList, maxXmitPackets *
5396 sizeof(struct rx_packet *));
5399 xmitList[nXmitPackets++] = p;
5403 /* xmitList now hold pointers to all of the packets that are
5404 * ready to send. Now we loop to send the packets */
5405 if (nXmitPackets > 0) {
5406 rxi_SendXmitList(call, xmitList, nXmitPackets, istack,
5407 &now, &retryTime, resending);
5410 maxXmitPackets * sizeof(struct rx_packet *));
5412 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5414 * TQ references no longer protected by this flag; they must remain
5415 * protected by the global lock.
5417 if (call->flags & RX_CALL_FAST_RECOVER_WAIT) {
5418 call->flags &= ~RX_CALL_TQ_BUSY;
5419 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5420 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5421 #ifdef RX_ENABLE_LOCKS
5422 osirx_AssertMine(&call->lock, "rxi_Start start");
5423 CV_BROADCAST(&call->cv_tq);
5424 #else /* RX_ENABLE_LOCKS */
5425 osi_rxWakeup(&call->tq);
5426 #endif /* RX_ENABLE_LOCKS */
5431 /* We went into the error state while sending packets. Now is
5432 * the time to reset the call. This will also inform the using
5433 * process that the call is in an error state.
5435 rx_MutexIncrement(rx_tq_debug.rxi_start_aborted, rx_stats_mutex);
5436 call->flags &= ~RX_CALL_TQ_BUSY;
5437 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5438 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5439 #ifdef RX_ENABLE_LOCKS
5440 osirx_AssertMine(&call->lock, "rxi_Start middle");
5441 CV_BROADCAST(&call->cv_tq);
5442 #else /* RX_ENABLE_LOCKS */
5443 osi_rxWakeup(&call->tq);
5444 #endif /* RX_ENABLE_LOCKS */
5446 rxi_CallError(call, call->error);
5449 #ifdef RX_ENABLE_LOCKS
5450 if (call->flags & RX_CALL_TQ_SOME_ACKED) {
5451 register int missing;
5452 call->flags &= ~RX_CALL_TQ_SOME_ACKED;
5453 /* Some packets have received acks. If they all have, we can clear
5454 * the transmit queue.
5457 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5458 if (p->header.seq < call->tfirst
5459 && (p->flags & RX_PKTFLAG_ACKED)) {
5461 p->flags &= ~RX_PKTFLAG_TQ;
5462 #ifdef RXDEBUG_PACKET
5470 call->flags |= RX_CALL_TQ_CLEARME;
5472 #endif /* RX_ENABLE_LOCKS */
5473 /* Don't bother doing retransmits if the TQ is cleared. */
5474 if (call->flags & RX_CALL_TQ_CLEARME) {
5475 rxi_ClearTransmitQueue(call, 1);
5477 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5480 /* Always post a resend event, if there is anything in the
5481 * queue, and resend is possible. There should be at least
5482 * one unacknowledged packet in the queue ... otherwise none
5483 * of these packets should be on the queue in the first place.
5485 if (call->resendEvent) {
5486 /* Cancel the existing event and post a new one */
5487 rxevent_Cancel(call->resendEvent, call,
5488 RX_CALL_REFCOUNT_RESEND);
5491 /* The retry time is the retry time on the first unacknowledged
5492 * packet inside the current window */
5494 0, queue_Scan(&call->tq, p, nxp, rx_packet)) {
5495 /* Don't set timers for packets outside the window */
5496 if (p->header.seq >= call->tfirst + call->twind) {
5500 if (!(p->flags & RX_PKTFLAG_ACKED)
5501 && !clock_IsZero(&p->retryTime)) {
5503 retryTime = p->retryTime;
5508 /* Post a new event to re-run rxi_Start when retries may be needed */
5509 if (haveEvent && !(call->flags & RX_CALL_NEED_START)) {
5510 #ifdef RX_ENABLE_LOCKS
5511 CALL_HOLD(call, RX_CALL_REFCOUNT_RESEND);
5513 rxevent_PostNow2(&retryTime, &usenow,
5515 (void *)call, 0, istack);
5516 #else /* RX_ENABLE_LOCKS */
5518 rxevent_PostNow2(&retryTime, &usenow, rxi_Start,
5519 (void *)call, 0, istack);
5520 #endif /* RX_ENABLE_LOCKS */
5523 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5524 } while (call->flags & RX_CALL_NEED_START);
5526 * TQ references no longer protected by this flag; they must remain
5527 * protected by the global lock.
5529 call->flags &= ~RX_CALL_TQ_BUSY;
5530 if (call->tqWaiters || (call->flags & RX_CALL_TQ_WAIT)) {
5531 dpf(("call %x has %d waiters and flags %d\n", call, call->tqWaiters, call->flags));
5532 #ifdef RX_ENABLE_LOCKS
5533 osirx_AssertMine(&call->lock, "rxi_Start end");
5534 CV_BROADCAST(&call->cv_tq);
5535 #else /* RX_ENABLE_LOCKS */
5536 osi_rxWakeup(&call->tq);
5537 #endif /* RX_ENABLE_LOCKS */
5540 call->flags |= RX_CALL_NEED_START;
5542 #endif /* AFS_GLOBAL_RXLOCK_KERNEL */
5544 if (call->resendEvent) {
5545 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5550 /* Also adjusts the keep alive parameters for the call, to reflect
5551 * that we have just sent a packet (so keep alives aren't sent
5554 rxi_Send(register struct rx_call *call, register struct rx_packet *p,
5557 register struct rx_connection *conn = call->conn;
5559 /* Stamp each packet with the user supplied status */
5560 p->header.userStatus = call->localStatus;
5562 /* Allow the security object controlling this call's security to
5563 * make any last-minute changes to the packet */
5564 RXS_SendPacket(conn->securityObject, call, p);
5566 /* Since we're about to send SOME sort of packet to the peer, it's
5567 * safe to nuke any scheduled end-of-packets ack */
5568 rxevent_Cancel(call->delayedAckEvent, call, RX_CALL_REFCOUNT_DELAY);
5570 /* Actually send the packet, filling in more connection-specific fields */
5571 CALL_HOLD(call, RX_CALL_REFCOUNT_SEND);
5572 MUTEX_EXIT(&call->lock);
5573 rxi_SendPacket(call, conn, p, istack);
5574 MUTEX_ENTER(&call->lock);
5575 CALL_RELE(call, RX_CALL_REFCOUNT_SEND);
5577 /* Update last send time for this call (for keep-alive
5578 * processing), and for the connection (so that we can discover
5579 * idle connections) */
5580 conn->lastSendTime = call->lastSendTime = clock_Sec();
5581 /* Don't count keepalives here, so idleness can be tracked. */
5582 if ((p->header.type != RX_PACKET_TYPE_ACK) || (((struct rx_ackPacket *)rx_DataOf(p))->reason != RX_ACK_PING))
5583 call->lastSendData = call->lastSendTime;
5587 /* Check if a call needs to be destroyed. Called by keep-alive code to ensure
5588 * that things are fine. Also called periodically to guarantee that nothing
5589 * falls through the cracks (e.g. (error + dally) connections have keepalive
5590 * turned off. Returns 0 if conn is well, -1 otherwise. If otherwise, call
5592 * haveCTLock Set if calling from rxi_ReapConnections
5594 #ifdef RX_ENABLE_LOCKS
5596 rxi_CheckCall(register struct rx_call *call, int haveCTLock)
5597 #else /* RX_ENABLE_LOCKS */
5599 rxi_CheckCall(register struct rx_call *call)
5600 #endif /* RX_ENABLE_LOCKS */
5602 register struct rx_connection *conn = call->conn;
5604 afs_uint32 deadTime;
5606 #ifdef AFS_GLOBAL_RXLOCK_KERNEL
5607 if (call->flags & RX_CALL_TQ_BUSY) {
5608 /* Call is active and will be reset by rxi_Start if it's
5609 * in an error state.
5614 /* dead time + RTT + 8*MDEV, rounded up to next second. */
5616 (((afs_uint32) rx_ConnSecondsUntilDead(conn) << 10) +
5617 ((afs_uint32) conn->peer->rtt >> 3) +
5618 ((afs_uint32) conn->peer->rtt_dev << 1) + 1023) >> 10;
5620 /* These are computed to the second (+- 1 second). But that's
5621 * good enough for these values, which should be a significant
5622 * number of seconds. */
5623 if (now > (call->lastReceiveTime + deadTime)) {
5624 if (call->state == RX_STATE_ACTIVE) {
5626 #if defined(KERNEL) && defined(AFS_SUN57_ENV)
5628 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5629 netstack_t *ns = netstack_find_by_stackid(GLOBAL_NETSTACKID);
5630 ip_stack_t *ipst = ns->netstack_ip;
5632 ire = ire_cache_lookup(call->conn->peer->host
5633 #if defined(AFS_SUN510_ENV) && defined(ALL_ZONES)
5635 #if defined(AFS_SUN510_ENV) && (defined(ICL_3_ARG) || defined(GLOBAL_NETSTACKID))
5637 #if defined(AFS_SUN510_ENV) && defined(GLOBAL_NETSTACKID)
5644 if (ire && ire->ire_max_frag > 0)
5645 rxi_SetPeerMtu(call->conn->peer->host, 0, ire->ire_max_frag);
5646 #if defined(GLOBAL_NETSTACKID)
5650 #endif /* ADAPT_PMTU */
5651 rxi_CallError(call, RX_CALL_DEAD);
5654 #ifdef RX_ENABLE_LOCKS
5655 /* Cancel pending events */
5656 rxevent_Cancel(call->delayedAckEvent, call,
5657 RX_CALL_REFCOUNT_DELAY);
5658 rxevent_Cancel(call->resendEvent, call, RX_CALL_REFCOUNT_RESEND);
5659 rxevent_Cancel(call->keepAliveEvent, call,
5660 RX_CALL_REFCOUNT_ALIVE);
5661 if (call->refCount == 0) {
5662 rxi_FreeCall(call, haveCTLock);
5666 #else /* RX_ENABLE_LOCKS */
5669 #endif /* RX_ENABLE_LOCKS */
5671 /* Non-active calls are destroyed if they are not responding
5672 * to pings; active calls are simply flagged in error, so the
5673 * attached process can die reasonably gracefully. */
5675 /* see if we have a non-activity timeout */
5676 if (call->startWait && rx_ConnIdleDeadTime(conn)
5677 && ((call->startWait + rx_ConnIdleDeadTime(conn)) < now)) {
5678 if (call->state == RX_STATE_ACTIVE) {
5679 rxi_CallError(call, RX_CALL_TIMEOUT);
5683 if (call->lastSendData && rx_ConnIdleDeadTime(conn)
5684 && (rx_ConnIdleDeadErr(conn) != 0)
5685 && ((call->lastSendData + rx_ConnIdleDeadTime(conn)) < now)) {
5686 if (call->state == RX_STATE_ACTIVE) {
5687 rxi_CallError(call, conn->idleDeadErr);
5691 /* see if we have a hard timeout */
5692 if (rx_ConnHardDeadTime(conn)
5693 && (now > (rx_ConnHardDeadTime(conn) + call->startTime.sec))) {
5694 if (call->state == RX_STATE_ACTIVE)
5695 rxi_CallError(call, RX_CALL_TIMEOUT);
5702 /* When a call is in progress, this routine is called occasionally to
5703 * make sure that some traffic has arrived (or been sent to) the peer.
5704 * If nothing has arrived in a reasonable amount of time, the call is
5705 * declared dead; if nothing has been sent for a while, we send a
5706 * keep-alive packet (if we're actually trying to keep the call alive)
5709 rxi_KeepAliveEvent(struct rxevent *event, void *arg1, void *dummy)
5711 struct rx_call *call = arg1;
5712 struct rx_connection *conn;
5715 MUTEX_ENTER(&call->lock);
5716 CALL_RELE(call, RX_CALL_REFCOUNT_ALIVE);
5717 if (event == call->keepAliveEvent)
5718 call->keepAliveEvent = NULL;
5721 #ifdef RX_ENABLE_LOCKS
5722 if (rxi_CheckCall(call, 0)) {
5723 MUTEX_EXIT(&call->lock);
5726 #else /* RX_ENABLE_LOCKS */
5727 if (rxi_CheckCall(call))
5729 #endif /* RX_ENABLE_LOCKS */
5731 /* Don't try to keep alive dallying calls */
5732 if (call->state == RX_STATE_DALLY) {
5733 MUTEX_EXIT(&call->lock);
5738 if ((now - call->lastSendTime) > rx_ConnSecondsUntilPing(conn)) {
5739 /* Don't try to send keepalives if there is unacknowledged data */
5740 /* the rexmit code should be good enough, this little hack
5741 * doesn't quite work XXX */
5742 (void)rxi_SendAck(call, NULL, 0, RX_ACK_PING, 0);
5744 rxi_ScheduleKeepAliveEvent(call);
5745 MUTEX_EXIT(&call->lock);
5750 rxi_ScheduleKeepAliveEvent(register struct rx_call *call)
5752 if (!call->keepAliveEvent) {
5753 struct clock when, now;
5754 clock_GetTime(&now);
5756 when.sec += rx_ConnSecondsUntilPing(call->conn);
5757 CALL_HOLD(call, RX_CALL_REFCOUNT_ALIVE);
5758 call->keepAliveEvent =
5759 rxevent_PostNow(&when, &now, rxi_KeepAliveEvent, call, 0);
5763 /* N.B. rxi_KeepAliveOff: is defined earlier as a macro */
5765 rxi_KeepAliveOn(register struct rx_call *call)
5767 /* Pretend last packet received was received now--i.e. if another
5768 * packet isn't received within the keep alive time, then the call
5769 * will die; Initialize last send time to the current time--even
5770 * if a packet hasn't been sent yet. This will guarantee that a
5771 * keep-alive is sent within the ping time */
5772 call->lastReceiveTime = call->lastSendTime = clock_Sec();
5773 rxi_ScheduleKeepAliveEvent(call);
5776 /* This routine is called to send connection abort messages
5777 * that have been delayed to throttle looping clients. */
5779 rxi_SendDelayedConnAbort(struct rxevent *event,
5780 void *arg1, void *unused)
5782 struct rx_connection *conn = arg1;
5785 struct rx_packet *packet;
5787 MUTEX_ENTER(&conn->conn_data_lock);
5788 conn->delayedAbortEvent = NULL;
5789 error = htonl(rx_ConnError(conn));
5791 MUTEX_EXIT(&conn->conn_data_lock);
5792 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5795 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5796 RX_PACKET_TYPE_ABORT, (char *)&error,
5798 rxi_FreePacket(packet);
5802 /* This routine is called to send call abort messages
5803 * that have been delayed to throttle looping clients. */
5805 rxi_SendDelayedCallAbort(struct rxevent *event,
5806 void *arg1, void *dummy)
5808 struct rx_call *call = arg1;
5811 struct rx_packet *packet;
5813 MUTEX_ENTER(&call->lock);
5814 call->delayedAbortEvent = NULL;
5815 error = htonl(call->error);
5817 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5820 rxi_SendSpecial(call, call->conn, packet, RX_PACKET_TYPE_ABORT,
5821 (char *)&error, sizeof(error), 0);
5822 rxi_FreePacket(packet);
5824 CALL_RELE(call, RX_CALL_REFCOUNT_ABORT);
5825 MUTEX_EXIT(&call->lock);
5828 /* This routine is called periodically (every RX_AUTH_REQUEST_TIMEOUT
5829 * seconds) to ask the client to authenticate itself. The routine
5830 * issues a challenge to the client, which is obtained from the
5831 * security object associated with the connection */
5833 rxi_ChallengeEvent(struct rxevent *event,
5834 void *arg0, void *arg1, int tries)
5836 struct rx_connection *conn = arg0;
5838 conn->challengeEvent = NULL;
5839 if (RXS_CheckAuthentication(conn->securityObject, conn) != 0) {
5840 register struct rx_packet *packet;
5841 struct clock when, now;
5844 /* We've failed to authenticate for too long.
5845 * Reset any calls waiting for authentication;
5846 * they are all in RX_STATE_PRECALL.
5850 MUTEX_ENTER(&conn->conn_call_lock);
5851 for (i = 0; i < RX_MAXCALLS; i++) {
5852 struct rx_call *call = conn->call[i];
5854 MUTEX_ENTER(&call->lock);
5855 if (call->state == RX_STATE_PRECALL) {
5856 rxi_CallError(call, RX_CALL_DEAD);
5857 rxi_SendCallAbort(call, NULL, 0, 0);
5859 MUTEX_EXIT(&call->lock);
5862 MUTEX_EXIT(&conn->conn_call_lock);
5866 packet = rxi_AllocPacket(RX_PACKET_CLASS_SPECIAL);
5868 /* If there's no packet available, do this later. */
5869 RXS_GetChallenge(conn->securityObject, conn, packet);
5870 rxi_SendSpecial((struct rx_call *)0, conn, packet,
5871 RX_PACKET_TYPE_CHALLENGE, NULL, -1, 0);
5872 rxi_FreePacket(packet);
5874 clock_GetTime(&now);
5876 when.sec += RX_CHALLENGE_TIMEOUT;
5877 conn->challengeEvent =
5878 rxevent_PostNow2(&when, &now, rxi_ChallengeEvent, conn, 0,
5883 /* Call this routine to start requesting the client to authenticate
5884 * itself. This will continue until authentication is established,
5885 * the call times out, or an invalid response is returned. The
5886 * security object associated with the connection is asked to create
5887 * the challenge at this time. N.B. rxi_ChallengeOff is a macro,
5888 * defined earlier. */
5890 rxi_ChallengeOn(register struct rx_connection *conn)
5892 if (!conn->challengeEvent) {
5893 RXS_CreateChallenge(conn->securityObject, conn);
5894 rxi_ChallengeEvent(NULL, conn, 0, RX_CHALLENGE_MAXTRIES);
5899 /* Compute round trip time of the packet provided, in *rttp.
5902 /* rxi_ComputeRoundTripTime is called with peer locked. */
5903 /* sentp and/or peer may be null */
5905 rxi_ComputeRoundTripTime(register struct rx_packet *p,
5906 register struct clock *sentp,
5907 register struct rx_peer *peer)
5909 struct clock thisRtt, *rttp = &thisRtt;
5911 register int rtt_timeout;
5913 clock_GetTime(rttp);
5915 if (clock_Lt(rttp, sentp)) {
5917 return; /* somebody set the clock back, don't count this time. */
5919 clock_Sub(rttp, sentp);
5920 MUTEX_ENTER(&rx_stats_mutex);
5921 if (clock_Lt(rttp, &rx_stats.minRtt))
5922 rx_stats.minRtt = *rttp;
5923 if (clock_Gt(rttp, &rx_stats.maxRtt)) {
5924 if (rttp->sec > 60) {
5925 MUTEX_EXIT(&rx_stats_mutex);
5926 return; /* somebody set the clock ahead */
5928 rx_stats.maxRtt = *rttp;
5930 clock_Add(&rx_stats.totalRtt, rttp);
5931 rx_stats.nRttSamples++;
5932 MUTEX_EXIT(&rx_stats_mutex);
5934 /* better rtt calculation courtesy of UMich crew (dave,larry,peter,?) */
5936 /* Apply VanJacobson round-trip estimations */
5941 * srtt (peer->rtt) is in units of one-eighth-milliseconds.
5942 * srtt is stored as fixed point with 3 bits after the binary
5943 * point (i.e., scaled by 8). The following magic is
5944 * equivalent to the smoothing algorithm in rfc793 with an
5945 * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point).
5946 * srtt*8 = srtt*8 + rtt - srtt
5947 * srtt = srtt + rtt/8 - srtt/8
5950 delta = MSEC(rttp) - (peer->rtt >> 3);
5954 * We accumulate a smoothed rtt variance (actually, a smoothed
5955 * mean difference), then set the retransmit timer to smoothed
5956 * rtt + 4 times the smoothed variance (was 2x in van's original
5957 * paper, but 4x works better for me, and apparently for him as
5959 * rttvar is stored as
5960 * fixed point with 2 bits after the binary point (scaled by
5961 * 4). The following is equivalent to rfc793 smoothing with
5962 * an alpha of .75 (rttvar = rttvar*3/4 + |delta| / 4). This
5963 * replaces rfc793's wired-in beta.
5964 * dev*4 = dev*4 + (|actual - expected| - dev)
5970 delta -= (peer->rtt_dev >> 2);
5971 peer->rtt_dev += delta;
5973 /* I don't have a stored RTT so I start with this value. Since I'm
5974 * probably just starting a call, and will be pushing more data down
5975 * this, I expect congestion to increase rapidly. So I fudge a
5976 * little, and I set deviance to half the rtt. In practice,
5977 * deviance tends to approach something a little less than
5978 * half the smoothed rtt. */
5979 peer->rtt = (MSEC(rttp) << 3) + 8;
5980 peer->rtt_dev = peer->rtt >> 2; /* rtt/2: they're scaled differently */
5982 /* the timeout is RTT + 4*MDEV + 0.35 sec This is because one end or
5983 * the other of these connections is usually in a user process, and can
5984 * be switched and/or swapped out. So on fast, reliable networks, the
5985 * timeout would otherwise be too short.
5987 rtt_timeout = (peer->rtt >> 3) + peer->rtt_dev + 350;
5988 clock_Zero(&(peer->timeout));
5989 clock_Addmsec(&(peer->timeout), rtt_timeout);
5991 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)));
5995 /* Find all server connections that have not been active for a long time, and
5998 rxi_ReapConnections(struct rxevent *unused, void *unused1, void *unused2)
6000 struct clock now, when;
6001 clock_GetTime(&now);
6003 /* Find server connection structures that haven't been used for
6004 * greater than rx_idleConnectionTime */
6006 struct rx_connection **conn_ptr, **conn_end;
6007 int i, havecalls = 0;
6008 RWLOCK_WRLOCK(&rx_connHashTable_lock);
6009 for (conn_ptr = &rx_connHashTable[0], conn_end =
6010 &rx_connHashTable[rx_hashTableSize]; conn_ptr < conn_end;
6012 struct rx_connection *conn, *next;
6013 struct rx_call *call;
6017 for (conn = *conn_ptr; conn; conn = next) {
6018 /* XXX -- Shouldn't the connection be locked? */
6021 for (i = 0; i < RX_MAXCALLS; i++) {
6022 call = conn->call[i];
6025 MUTEX_ENTER(&call->lock);
6026 #ifdef RX_ENABLE_LOCKS
6027 result = rxi_CheckCall(call, 1);
6028 #else /* RX_ENABLE_LOCKS */
6029 result = rxi_CheckCall(call);
6030 #endif /* RX_ENABLE_LOCKS */
6031 MUTEX_EXIT(&call->lock);
6033 /* If CheckCall freed the call, it might
6034 * have destroyed the connection as well,
6035 * which screws up the linked lists.
6041 if (conn->type == RX_SERVER_CONNECTION) {
6042 /* This only actually destroys the connection if
6043 * there are no outstanding calls */
6044 MUTEX_ENTER(&conn->conn_data_lock);
6045 if (!havecalls && !conn->refCount
6046 && ((conn->lastSendTime + rx_idleConnectionTime) <
6048 conn->refCount++; /* it will be decr in rx_DestroyConn */
6049 MUTEX_EXIT(&conn->conn_data_lock);
6050 #ifdef RX_ENABLE_LOCKS
6051 rxi_DestroyConnectionNoLock(conn);
6052 #else /* RX_ENABLE_LOCKS */
6053 rxi_DestroyConnection(conn);
6054 #endif /* RX_ENABLE_LOCKS */
6056 #ifdef RX_ENABLE_LOCKS
6058 MUTEX_EXIT(&conn->conn_data_lock);
6060 #endif /* RX_ENABLE_LOCKS */
6064 #ifdef RX_ENABLE_LOCKS
6065 while (rx_connCleanup_list) {
6066 struct rx_connection *conn;
6067 conn = rx_connCleanup_list;
6068 rx_connCleanup_list = rx_connCleanup_list->next;
6069 RWLOCK_UNLOCK(&rx_connHashTable_lock);
6070 rxi_CleanupConnection(conn);
6071 RWLOCK_WRLOCK(&rx_connHashTable_lock);
6073 RWLOCK_UNLOCK(&rx_connHashTable_lock);
6074 #endif /* RX_ENABLE_LOCKS */
6077 /* Find any peer structures that haven't been used (haven't had an
6078 * associated connection) for greater than rx_idlePeerTime */
6080 struct rx_peer **peer_ptr, **peer_end;
6082 MUTEX_ENTER(&rx_rpc_stats);
6083 RWLOCK_RDLOCK(&rx_peerHashTable_lock);
6084 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6085 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6087 struct rx_peer *peer, *next, *prev;
6088 for (prev = peer = *peer_ptr; peer; peer = next) {
6090 code = MUTEX_TRYENTER(&peer->peer_lock);
6091 if ((code) && (peer->refCount == 0)
6092 && ((peer->idleWhen + rx_idlePeerTime) < now.sec)) {
6093 rx_interface_stat_p rpc_stat, nrpc_stat;
6095 MUTEX_EXIT(&peer->peer_lock);
6096 MUTEX_DESTROY(&peer->peer_lock);
6098 (&peer->rpcStats, rpc_stat, nrpc_stat,
6099 rx_interface_stat)) {
6100 unsigned int num_funcs;
6103 queue_Remove(&rpc_stat->queue_header);
6104 queue_Remove(&rpc_stat->all_peers);
6105 num_funcs = rpc_stat->stats[0].func_total;
6107 sizeof(rx_interface_stat_t) +
6108 rpc_stat->stats[0].func_total *
6109 sizeof(rx_function_entry_v1_t);
6111 rxi_Free(rpc_stat, space);
6112 rxi_rpc_peer_stat_cnt -= num_funcs;
6114 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6115 RWLOCK_UPLOCK(&rx_peerHashTable_lock);
6116 if (peer == *peer_ptr) {
6124 MUTEX_EXIT(&peer->peer_lock);
6130 RWLOCK_UNLOCK(&rx_peerHashTable_lock);
6131 MUTEX_EXIT(&rx_rpc_stats);
6134 /* THIS HACK IS A TEMPORARY HACK. The idea is that the race condition in
6135 * rxi_AllocSendPacket, if it hits, will be handled at the next conn
6136 * GC, just below. Really, we shouldn't have to keep moving packets from
6137 * one place to another, but instead ought to always know if we can
6138 * afford to hold onto a packet in its particular use. */
6139 MUTEX_ENTER(&rx_freePktQ_lock);
6140 if (rx_waitingForPackets) {
6141 rx_waitingForPackets = 0;
6142 #ifdef RX_ENABLE_LOCKS
6143 CV_BROADCAST(&rx_waitingForPackets_cv);
6145 osi_rxWakeup(&rx_waitingForPackets);
6148 MUTEX_EXIT(&rx_freePktQ_lock);
6151 when.sec += RX_REAP_TIME; /* Check every RX_REAP_TIME seconds */
6152 rxevent_Post(&when, rxi_ReapConnections, 0, 0);
6156 /* rxs_Release - This isn't strictly necessary but, since the macro name from
6157 * rx.h is sort of strange this is better. This is called with a security
6158 * object before it is discarded. Each connection using a security object has
6159 * its own refcount to the object so it won't actually be freed until the last
6160 * connection is destroyed.
6162 * This is the only rxs module call. A hold could also be written but no one
6166 rxs_Release(struct rx_securityClass *aobj)
6168 return RXS_Close(aobj);
6172 #define RXRATE_PKT_OH (RX_HEADER_SIZE + RX_IPUDP_SIZE)
6173 #define RXRATE_SMALL_PKT (RXRATE_PKT_OH + sizeof(struct rx_ackPacket))
6174 #define RXRATE_AVG_SMALL_PKT (RXRATE_PKT_OH + (sizeof(struct rx_ackPacket)/2))
6175 #define RXRATE_LARGE_PKT (RXRATE_SMALL_PKT + 256)
6177 /* Adjust our estimate of the transmission rate to this peer, given
6178 * that the packet p was just acked. We can adjust peer->timeout and
6179 * call->twind. Pragmatically, this is called
6180 * only with packets of maximal length.
6181 * Called with peer and call locked.
6185 rxi_ComputeRate(register struct rx_peer *peer, register struct rx_call *call,
6186 struct rx_packet *p, struct rx_packet *ackp, u_char ackReason)
6188 afs_int32 xferSize, xferMs;
6189 register afs_int32 minTime;
6192 /* Count down packets */
6193 if (peer->rateFlag > 0)
6195 /* Do nothing until we're enabled */
6196 if (peer->rateFlag != 0)
6201 /* Count only when the ack seems legitimate */
6202 switch (ackReason) {
6203 case RX_ACK_REQUESTED:
6205 p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize;
6209 case RX_ACK_PING_RESPONSE:
6210 if (p) /* want the response to ping-request, not data send */
6212 clock_GetTime(&newTO);
6213 if (clock_Gt(&newTO, &call->pingRequestTime)) {
6214 clock_Sub(&newTO, &call->pingRequestTime);
6215 xferMs = (newTO.sec * 1000) + (newTO.usec / 1000);
6219 xferSize = rx_AckDataSize(rx_Window) + RX_HEADER_SIZE;
6226 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));
6228 /* Track only packets that are big enough. */
6229 if ((p->length + RX_HEADER_SIZE + call->conn->securityMaxTrailerSize) <
6233 /* absorb RTT data (in milliseconds) for these big packets */
6234 if (peer->smRtt == 0) {
6235 peer->smRtt = xferMs;
6237 peer->smRtt = ((peer->smRtt * 15) + xferMs + 4) >> 4;
6242 if (peer->countDown) {
6246 peer->countDown = 10; /* recalculate only every so often */
6248 /* In practice, we can measure only the RTT for full packets,
6249 * because of the way Rx acks the data that it receives. (If it's
6250 * smaller than a full packet, it often gets implicitly acked
6251 * either by the call response (from a server) or by the next call
6252 * (from a client), and either case confuses transmission times
6253 * with processing times.) Therefore, replace the above
6254 * more-sophisticated processing with a simpler version, where the
6255 * smoothed RTT is kept for full-size packets, and the time to
6256 * transmit a windowful of full-size packets is simply RTT *
6257 * windowSize. Again, we take two steps:
6258 - ensure the timeout is large enough for a single packet's RTT;
6259 - ensure that the window is small enough to fit in the desired timeout.*/
6261 /* First, the timeout check. */
6262 minTime = peer->smRtt;
6263 /* Get a reasonable estimate for a timeout period */
6265 newTO.sec = minTime / 1000;
6266 newTO.usec = (minTime - (newTO.sec * 1000)) * 1000;
6268 /* Increase the timeout period so that we can always do at least
6269 * one packet exchange */
6270 if (clock_Gt(&newTO, &peer->timeout)) {
6272 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));
6274 peer->timeout = newTO;
6277 /* Now, get an estimate for the transmit window size. */
6278 minTime = peer->timeout.sec * 1000 + (peer->timeout.usec / 1000);
6279 /* Now, convert to the number of full packets that could fit in a
6280 * reasonable fraction of that interval */
6281 minTime /= (peer->smRtt << 1);
6282 xferSize = minTime; /* (make a copy) */
6284 /* Now clamp the size to reasonable bounds. */
6287 else if (minTime > rx_Window)
6288 minTime = rx_Window;
6289 /* if (minTime != peer->maxWindow) {
6290 dpf(("CONG peer %lx/%u: windowsize %lu ==> %lu (to %lu.%06lu, rtt %u, ps %u)",
6291 ntohl(peer->host), ntohs(peer->port), peer->maxWindow, minTime,
6292 peer->timeout.sec, peer->timeout.usec, peer->smRtt,
6294 peer->maxWindow = minTime;
6295 elide... call->twind = minTime;
6299 /* Cut back on the peer timeout if it had earlier grown unreasonably.
6300 * Discern this by calculating the timeout necessary for rx_Window
6302 if ((xferSize > rx_Window) && (peer->timeout.sec >= 3)) {
6303 /* calculate estimate for transmission interval in milliseconds */
6304 minTime = rx_Window * peer->smRtt;
6305 if (minTime < 1000) {
6306 dpf(("CONG peer %lx/%u: cut TO %lu.%06lu by 0.5 (rtt %u, ps %u)",
6307 ntohl(peer->host), ntohs(peer->port), peer->timeout.sec,
6308 peer->timeout.usec, peer->smRtt, peer->packetSize));
6310 newTO.sec = 0; /* cut back on timeout by half a second */
6311 newTO.usec = 500000;
6312 clock_Sub(&peer->timeout, &newTO);
6317 } /* end of rxi_ComputeRate */
6318 #endif /* ADAPT_WINDOW */
6326 #define TRACE_OPTION_DEBUGLOG 4
6334 code = RegOpenKeyEx(HKEY_LOCAL_MACHINE, AFSREG_CLT_SVC_PARAM_SUBKEY,
6335 0, KEY_QUERY_VALUE, &parmKey);
6336 if (code != ERROR_SUCCESS)
6339 dummyLen = sizeof(TraceOption);
6340 code = RegQueryValueEx(parmKey, "TraceOption", NULL, NULL,
6341 (BYTE *) &TraceOption, &dummyLen);
6342 if (code == ERROR_SUCCESS) {
6343 rxdebug_active = (TraceOption & TRACE_OPTION_DEBUGLOG) ? 1 : 0;
6345 RegCloseKey (parmKey);
6346 #endif /* AFS_NT40_ENV */
6351 rx_DebugOnOff(int on)
6353 rxdebug_active = on;
6355 #endif /* AFS_NT40_ENV */
6358 /* Don't call this debugging routine directly; use dpf */
6360 rxi_DebugPrint(char *format, ...)
6368 va_start(ap, format);
6370 len = _snprintf(tformat, sizeof(tformat), "tid[%d] %s", GetCurrentThreadId(), format);
6373 len = _vsnprintf(msg, sizeof(msg)-2, tformat, ap);
6375 if (msg[len-1] != '\n') {
6379 OutputDebugString(msg);
6386 va_start(ap, format);
6388 clock_GetTime(&now);
6389 fprintf(rx_Log, " %u.%.3u:", (unsigned int)now.sec,
6390 (unsigned int)now.usec / 1000);
6391 vfprintf(rx_Log, format, ap);
6398 * This function is used to process the rx_stats structure that is local
6399 * to a process as well as an rx_stats structure received from a remote
6400 * process (via rxdebug). Therefore, it needs to do minimal version
6404 rx_PrintTheseStats(FILE * file, struct rx_statistics *s, int size,
6405 afs_int32 freePackets, char version)
6409 if (size != sizeof(struct rx_statistics)) {
6411 "Unexpected size of stats structure: was %d, expected %lud\n",
6412 size, sizeof(struct rx_statistics));
6415 fprintf(file, "rx stats: free packets %d, allocs %d, ", (int)freePackets,
6418 if (version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6419 fprintf(file, "alloc-failures(rcv %d/%d,send %d/%d,ack %d)\n",
6420 s->receivePktAllocFailures, s->receiveCbufPktAllocFailures,
6421 s->sendPktAllocFailures, s->sendCbufPktAllocFailures,
6422 s->specialPktAllocFailures);
6424 fprintf(file, "alloc-failures(rcv %d,send %d,ack %d)\n",
6425 s->receivePktAllocFailures, s->sendPktAllocFailures,
6426 s->specialPktAllocFailures);
6430 " greedy %d, " "bogusReads %d (last from host %x), "
6431 "noPackets %d, " "noBuffers %d, " "selects %d, "
6432 "sendSelects %d\n", s->socketGreedy, s->bogusPacketOnRead,
6433 s->bogusHost, s->noPacketOnRead, s->noPacketBuffersOnRead,
6434 s->selects, s->sendSelects);
6436 fprintf(file, " packets read: ");
6437 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6438 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsRead[i]);
6440 fprintf(file, "\n");
6443 " other read counters: data %d, " "ack %d, " "dup %d "
6444 "spurious %d " "dally %d\n", s->dataPacketsRead,
6445 s->ackPacketsRead, s->dupPacketsRead, s->spuriousPacketsRead,
6446 s->ignorePacketDally);
6448 fprintf(file, " packets sent: ");
6449 for (i = 0; i < RX_N_PACKET_TYPES; i++) {
6450 fprintf(file, "%s %d ", rx_packetTypes[i], s->packetsSent[i]);
6452 fprintf(file, "\n");
6455 " other send counters: ack %d, " "data %d (not resends), "
6456 "resends %d, " "pushed %d, " "acked&ignored %d\n",
6457 s->ackPacketsSent, s->dataPacketsSent, s->dataPacketsReSent,
6458 s->dataPacketsPushed, s->ignoreAckedPacket);
6461 " \t(these should be small) sendFailed %d, " "fatalErrors %d\n",
6462 s->netSendFailures, (int)s->fatalErrors);
6464 if (s->nRttSamples) {
6465 fprintf(file, " Average rtt is %0.3f, with %d samples\n",
6466 clock_Float(&s->totalRtt) / s->nRttSamples, s->nRttSamples);
6468 fprintf(file, " Minimum rtt is %0.3f, maximum is %0.3f\n",
6469 clock_Float(&s->minRtt), clock_Float(&s->maxRtt));
6473 " %d server connections, " "%d client connections, "
6474 "%d peer structs, " "%d call structs, " "%d free call structs\n",
6475 s->nServerConns, s->nClientConns, s->nPeerStructs,
6476 s->nCallStructs, s->nFreeCallStructs);
6478 #if !defined(AFS_PTHREAD_ENV) && !defined(AFS_USE_GETTIMEOFDAY)
6479 fprintf(file, " %d clock updates\n", clock_nUpdates);
6484 /* for backward compatibility */
6486 rx_PrintStats(FILE * file)
6488 MUTEX_ENTER(&rx_stats_mutex);
6489 rx_PrintTheseStats(file, &rx_stats, sizeof(rx_stats), rx_nFreePackets,
6491 MUTEX_EXIT(&rx_stats_mutex);
6495 rx_PrintPeerStats(FILE * file, struct rx_peer *peer)
6497 fprintf(file, "Peer %x.%d. " "Burst size %d, " "burst wait %u.%d.\n",
6498 ntohl(peer->host), (int)peer->port, (int)peer->burstSize,
6499 (int)peer->burstWait.sec, (int)peer->burstWait.usec);
6502 " Rtt %d, " "retry time %u.%06d, " "total sent %d, "
6503 "resent %d\n", peer->rtt, (int)peer->timeout.sec,
6504 (int)peer->timeout.usec, peer->nSent, peer->reSends);
6507 " Packet size %d, " "max in packet skew %d, "
6508 "max out packet skew %d\n", peer->ifMTU, (int)peer->inPacketSkew,
6509 (int)peer->outPacketSkew);
6512 #ifdef AFS_PTHREAD_ENV
6514 * This mutex protects the following static variables:
6518 #define LOCK_RX_DEBUG assert(pthread_mutex_lock(&rx_debug_mutex)==0)
6519 #define UNLOCK_RX_DEBUG assert(pthread_mutex_unlock(&rx_debug_mutex)==0)
6521 #define LOCK_RX_DEBUG
6522 #define UNLOCK_RX_DEBUG
6523 #endif /* AFS_PTHREAD_ENV */
6526 MakeDebugCall(osi_socket socket, afs_uint32 remoteAddr, afs_uint16 remotePort,
6527 u_char type, void *inputData, size_t inputLength,
6528 void *outputData, size_t outputLength)
6530 static afs_int32 counter = 100;
6531 time_t waitTime, waitCount, startTime;
6532 struct rx_header theader;
6534 register afs_int32 code;
6535 struct timeval tv_now, tv_wake, tv_delta;
6536 struct sockaddr_in taddr, faddr;
6545 startTime = time(0);
6551 tp = &tbuffer[sizeof(struct rx_header)];
6552 taddr.sin_family = AF_INET;
6553 taddr.sin_port = remotePort;
6554 taddr.sin_addr.s_addr = remoteAddr;
6555 #ifdef STRUCT_SOCKADDR_HAS_SA_LEN
6556 taddr.sin_len = sizeof(struct sockaddr_in);
6559 memset(&theader, 0, sizeof(theader));
6560 theader.epoch = htonl(999);
6562 theader.callNumber = htonl(counter);
6565 theader.type = type;
6566 theader.flags = RX_CLIENT_INITIATED | RX_LAST_PACKET;
6567 theader.serviceId = 0;
6569 memcpy(tbuffer, &theader, sizeof(theader));
6570 memcpy(tp, inputData, inputLength);
6572 sendto(socket, tbuffer, inputLength + sizeof(struct rx_header), 0,
6573 (struct sockaddr *)&taddr, sizeof(struct sockaddr_in));
6575 /* see if there's a packet available */
6576 gettimeofday(&tv_wake,0);
6577 tv_wake.tv_sec += waitTime;
6580 FD_SET(socket, &imask);
6581 tv_delta.tv_sec = tv_wake.tv_sec;
6582 tv_delta.tv_usec = tv_wake.tv_usec;
6583 gettimeofday(&tv_now, 0);
6585 if (tv_delta.tv_usec < tv_now.tv_usec) {
6587 tv_delta.tv_usec += 1000000;
6590 tv_delta.tv_usec -= tv_now.tv_usec;
6592 if (tv_delta.tv_sec < tv_now.tv_sec) {
6596 tv_delta.tv_sec -= tv_now.tv_sec;
6598 code = select(socket + 1, &imask, 0, 0, &tv_delta);
6599 if (code == 1 && FD_ISSET(socket, &imask)) {
6600 /* now receive a packet */
6601 faddrLen = sizeof(struct sockaddr_in);
6603 recvfrom(socket, tbuffer, sizeof(tbuffer), 0,
6604 (struct sockaddr *)&faddr, &faddrLen);
6607 memcpy(&theader, tbuffer, sizeof(struct rx_header));
6608 if (counter == ntohl(theader.callNumber))
6616 /* see if we've timed out */
6624 code -= sizeof(struct rx_header);
6625 if (code > outputLength)
6626 code = outputLength;
6627 memcpy(outputData, tp, code);
6632 rx_GetServerDebug(osi_socket socket, afs_uint32 remoteAddr,
6633 afs_uint16 remotePort, struct rx_debugStats * stat,
6634 afs_uint32 * supportedValues)
6636 struct rx_debugIn in;
6639 *supportedValues = 0;
6640 in.type = htonl(RX_DEBUGI_GETSTATS);
6643 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6644 &in, sizeof(in), stat, sizeof(*stat));
6647 * If the call was successful, fixup the version and indicate
6648 * what contents of the stat structure are valid.
6649 * Also do net to host conversion of fields here.
6653 if (stat->version >= RX_DEBUGI_VERSION_W_SECSTATS) {
6654 *supportedValues |= RX_SERVER_DEBUG_SEC_STATS;
6656 if (stat->version >= RX_DEBUGI_VERSION_W_GETALLCONN) {
6657 *supportedValues |= RX_SERVER_DEBUG_ALL_CONN;
6659 if (stat->version >= RX_DEBUGI_VERSION_W_RXSTATS) {
6660 *supportedValues |= RX_SERVER_DEBUG_RX_STATS;
6662 if (stat->version >= RX_DEBUGI_VERSION_W_WAITERS) {
6663 *supportedValues |= RX_SERVER_DEBUG_WAITER_CNT;
6665 if (stat->version >= RX_DEBUGI_VERSION_W_IDLETHREADS) {
6666 *supportedValues |= RX_SERVER_DEBUG_IDLE_THREADS;
6668 if (stat->version >= RX_DEBUGI_VERSION_W_NEWPACKETTYPES) {
6669 *supportedValues |= RX_SERVER_DEBUG_NEW_PACKETS;
6671 if (stat->version >= RX_DEBUGI_VERSION_W_GETPEER) {
6672 *supportedValues |= RX_SERVER_DEBUG_ALL_PEER;
6674 if (stat->version >= RX_DEBUGI_VERSION_W_WAITED) {
6675 *supportedValues |= RX_SERVER_DEBUG_WAITED_CNT;
6677 if (stat->version >= RX_DEBUGI_VERSION_W_PACKETS) {
6678 *supportedValues |= RX_SERVER_DEBUG_PACKETS_CNT;
6680 stat->nFreePackets = ntohl(stat->nFreePackets);
6681 stat->packetReclaims = ntohl(stat->packetReclaims);
6682 stat->callsExecuted = ntohl(stat->callsExecuted);
6683 stat->nWaiting = ntohl(stat->nWaiting);
6684 stat->idleThreads = ntohl(stat->idleThreads);
6685 stat->nWaited = ntohl(stat->nWaited);
6686 stat->nPackets = ntohl(stat->nPackets);
6693 rx_GetServerStats(osi_socket socket, afs_uint32 remoteAddr,
6694 afs_uint16 remotePort, struct rx_statistics * stat,
6695 afs_uint32 * supportedValues)
6697 struct rx_debugIn in;
6699 afs_int32 *lp = (afs_int32 *) stat;
6703 * supportedValues is currently unused, but added to allow future
6704 * versioning of this function.
6707 *supportedValues = 0;
6708 in.type = htonl(RX_DEBUGI_RXSTATS);
6710 memset(stat, 0, sizeof(*stat));
6712 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6713 &in, sizeof(in), stat, sizeof(*stat));
6718 * Do net to host conversion here
6721 for (i = 0; i < sizeof(*stat) / sizeof(afs_int32); i++, lp++) {
6730 rx_GetServerVersion(osi_socket socket, afs_uint32 remoteAddr,
6731 afs_uint16 remotePort, size_t version_length,
6735 return MakeDebugCall(socket, remoteAddr, remotePort,
6736 RX_PACKET_TYPE_VERSION, a, 1, version,
6741 rx_GetServerConnections(osi_socket socket, afs_uint32 remoteAddr,
6742 afs_uint16 remotePort, afs_int32 * nextConnection,
6743 int allConnections, afs_uint32 debugSupportedValues,
6744 struct rx_debugConn * conn,
6745 afs_uint32 * supportedValues)
6747 struct rx_debugIn in;
6752 * supportedValues is currently unused, but added to allow future
6753 * versioning of this function.
6756 *supportedValues = 0;
6757 if (allConnections) {
6758 in.type = htonl(RX_DEBUGI_GETALLCONN);
6760 in.type = htonl(RX_DEBUGI_GETCONN);
6762 in.index = htonl(*nextConnection);
6763 memset(conn, 0, sizeof(*conn));
6765 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6766 &in, sizeof(in), conn, sizeof(*conn));
6769 *nextConnection += 1;
6772 * Convert old connection format to new structure.
6775 if (debugSupportedValues & RX_SERVER_DEBUG_OLD_CONN) {
6776 struct rx_debugConn_vL *vL = (struct rx_debugConn_vL *)conn;
6777 #define MOVEvL(a) (conn->a = vL->a)
6779 /* any old or unrecognized version... */
6780 for (i = 0; i < RX_MAXCALLS; i++) {
6781 MOVEvL(callState[i]);
6782 MOVEvL(callMode[i]);
6783 MOVEvL(callFlags[i]);
6784 MOVEvL(callOther[i]);
6786 if (debugSupportedValues & RX_SERVER_DEBUG_SEC_STATS) {
6787 MOVEvL(secStats.type);
6788 MOVEvL(secStats.level);
6789 MOVEvL(secStats.flags);
6790 MOVEvL(secStats.expires);
6791 MOVEvL(secStats.packetsReceived);
6792 MOVEvL(secStats.packetsSent);
6793 MOVEvL(secStats.bytesReceived);
6794 MOVEvL(secStats.bytesSent);
6799 * Do net to host conversion here
6801 * I don't convert host or port since we are most likely
6802 * going to want these in NBO.
6804 conn->cid = ntohl(conn->cid);
6805 conn->serial = ntohl(conn->serial);
6806 for (i = 0; i < RX_MAXCALLS; i++) {
6807 conn->callNumber[i] = ntohl(conn->callNumber[i]);
6809 rx_SetConnError(conn, ntohl(rx_ConnError(conn)));
6810 conn->secStats.flags = ntohl(conn->secStats.flags);
6811 conn->secStats.expires = ntohl(conn->secStats.expires);
6812 conn->secStats.packetsReceived =
6813 ntohl(conn->secStats.packetsReceived);
6814 conn->secStats.packetsSent = ntohl(conn->secStats.packetsSent);
6815 conn->secStats.bytesReceived = ntohl(conn->secStats.bytesReceived);
6816 conn->secStats.bytesSent = ntohl(conn->secStats.bytesSent);
6817 conn->epoch = ntohl(conn->epoch);
6818 conn->natMTU = ntohl(conn->natMTU);
6825 rx_GetServerPeers(osi_socket socket, afs_uint32 remoteAddr,
6826 afs_uint16 remotePort, afs_int32 * nextPeer,
6827 afs_uint32 debugSupportedValues, struct rx_debugPeer * peer,
6828 afs_uint32 * supportedValues)
6830 struct rx_debugIn in;
6834 * supportedValues is currently unused, but added to allow future
6835 * versioning of this function.
6838 *supportedValues = 0;
6839 in.type = htonl(RX_DEBUGI_GETPEER);
6840 in.index = htonl(*nextPeer);
6841 memset(peer, 0, sizeof(*peer));
6843 rc = MakeDebugCall(socket, remoteAddr, remotePort, RX_PACKET_TYPE_DEBUG,
6844 &in, sizeof(in), peer, sizeof(*peer));
6850 * Do net to host conversion here
6852 * I don't convert host or port since we are most likely
6853 * going to want these in NBO.
6855 peer->ifMTU = ntohs(peer->ifMTU);
6856 peer->idleWhen = ntohl(peer->idleWhen);
6857 peer->refCount = ntohs(peer->refCount);
6858 peer->burstWait.sec = ntohl(peer->burstWait.sec);
6859 peer->burstWait.usec = ntohl(peer->burstWait.usec);
6860 peer->rtt = ntohl(peer->rtt);
6861 peer->rtt_dev = ntohl(peer->rtt_dev);
6862 peer->timeout.sec = ntohl(peer->timeout.sec);
6863 peer->timeout.usec = ntohl(peer->timeout.usec);
6864 peer->nSent = ntohl(peer->nSent);
6865 peer->reSends = ntohl(peer->reSends);
6866 peer->inPacketSkew = ntohl(peer->inPacketSkew);
6867 peer->outPacketSkew = ntohl(peer->outPacketSkew);
6868 peer->rateFlag = ntohl(peer->rateFlag);
6869 peer->natMTU = ntohs(peer->natMTU);
6870 peer->maxMTU = ntohs(peer->maxMTU);
6871 peer->maxDgramPackets = ntohs(peer->maxDgramPackets);
6872 peer->ifDgramPackets = ntohs(peer->ifDgramPackets);
6873 peer->MTU = ntohs(peer->MTU);
6874 peer->cwind = ntohs(peer->cwind);
6875 peer->nDgramPackets = ntohs(peer->nDgramPackets);
6876 peer->congestSeq = ntohs(peer->congestSeq);
6877 peer->bytesSent.high = ntohl(peer->bytesSent.high);
6878 peer->bytesSent.low = ntohl(peer->bytesSent.low);
6879 peer->bytesReceived.high = ntohl(peer->bytesReceived.high);
6880 peer->bytesReceived.low = ntohl(peer->bytesReceived.low);
6885 #endif /* RXDEBUG */
6890 struct rx_serverQueueEntry *np;
6893 register struct rx_call *call;
6894 register struct rx_serverQueueEntry *sq;
6898 if (rxinit_status == 1) {
6900 return; /* Already shutdown. */
6904 #ifndef AFS_PTHREAD_ENV
6905 FD_ZERO(&rx_selectMask);
6906 #endif /* AFS_PTHREAD_ENV */
6907 rxi_dataQuota = RX_MAX_QUOTA;
6908 #ifndef AFS_PTHREAD_ENV
6910 #endif /* AFS_PTHREAD_ENV */
6913 #ifndef AFS_PTHREAD_ENV
6914 #ifndef AFS_USE_GETTIMEOFDAY
6916 #endif /* AFS_USE_GETTIMEOFDAY */
6917 #endif /* AFS_PTHREAD_ENV */
6919 while (!queue_IsEmpty(&rx_freeCallQueue)) {
6920 call = queue_First(&rx_freeCallQueue, rx_call);
6922 rxi_Free(call, sizeof(struct rx_call));
6925 while (!queue_IsEmpty(&rx_idleServerQueue)) {
6926 sq = queue_First(&rx_idleServerQueue, rx_serverQueueEntry);
6932 struct rx_peer **peer_ptr, **peer_end;
6933 for (peer_ptr = &rx_peerHashTable[0], peer_end =
6934 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
6936 struct rx_peer *peer, *next;
6937 for (peer = *peer_ptr; peer; peer = next) {
6938 rx_interface_stat_p rpc_stat, nrpc_stat;
6941 (&peer->rpcStats, rpc_stat, nrpc_stat,
6942 rx_interface_stat)) {
6946 queue_Remove(&rpc_stat->queue_header);
6947 queue_Remove(&rpc_stat->all_peers);
6948 num_funcs = rpc_stat->stats[0].func_total;
6950 sizeof(rx_interface_stat_t) +
6951 rpc_stat->stats[0].func_total *
6952 sizeof(rx_function_entry_v1_t);
6954 rxi_Free(rpc_stat, space);
6955 rx_MutexAdd(rxi_rpc_peer_stat_cnt, -num_funcs, rx_rpc_stats);
6959 rx_MutexDecrement(rx_stats.nPeerStructs, rx_stats_mutex);
6963 for (i = 0; i < RX_MAX_SERVICES; i++) {
6965 rxi_Free(rx_services[i], sizeof(*rx_services[i]));
6967 for (i = 0; i < rx_hashTableSize; i++) {
6968 register struct rx_connection *tc, *ntc;
6969 RWLOCK_RDLOCK(&rx_connHashTable_lock);
6970 for (tc = rx_connHashTable[i]; tc; tc = ntc) {
6972 for (j = 0; j < RX_MAXCALLS; j++) {
6974 rxi_Free(tc->call[j], sizeof(*tc->call[j]));
6977 rxi_Free(tc, sizeof(*tc));
6979 RWLOCK_UNLOCK(&rx_connHashTable_lock);
6982 MUTEX_ENTER(&freeSQEList_lock);
6984 while ((np = rx_FreeSQEList)) {
6985 rx_FreeSQEList = *(struct rx_serverQueueEntry **)np;
6986 MUTEX_DESTROY(&np->lock);
6987 rxi_Free(np, sizeof(*np));
6990 MUTEX_EXIT(&freeSQEList_lock);
6991 MUTEX_DESTROY(&freeSQEList_lock);
6992 MUTEX_DESTROY(&rx_freeCallQueue_lock);
6993 RWLOCK_DESTROY(&rx_connHashTable_lock);
6994 RWLOCK_DESTROY(&rx_peerHashTable_lock);
6995 MUTEX_DESTROY(&rx_serverPool_lock);
6997 osi_Free(rx_connHashTable,
6998 rx_hashTableSize * sizeof(struct rx_connection *));
6999 osi_Free(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7001 UNPIN(rx_connHashTable,
7002 rx_hashTableSize * sizeof(struct rx_connection *));
7003 UNPIN(rx_peerHashTable, rx_hashTableSize * sizeof(struct rx_peer *));
7005 rxi_FreeAllPackets();
7007 MUTEX_ENTER(&rx_stats_mutex);
7008 rxi_dataQuota = RX_MAX_QUOTA;
7009 rxi_availProcs = rxi_totalMin = rxi_minDeficit = 0;
7010 MUTEX_EXIT(&rx_stats_mutex);
7016 #ifdef RX_ENABLE_LOCKS
7018 osirx_AssertMine(afs_kmutex_t * lockaddr, char *msg)
7020 if (!MUTEX_ISMINE(lockaddr))
7021 osi_Panic("Lock not held: %s", msg);
7023 #endif /* RX_ENABLE_LOCKS */
7028 * Routines to implement connection specific data.
7032 rx_KeyCreate(rx_destructor_t rtn)
7035 MUTEX_ENTER(&rxi_keyCreate_lock);
7036 key = rxi_keyCreate_counter++;
7037 rxi_keyCreate_destructor = (rx_destructor_t *)
7038 realloc((void *)rxi_keyCreate_destructor,
7039 (key + 1) * sizeof(rx_destructor_t));
7040 rxi_keyCreate_destructor[key] = rtn;
7041 MUTEX_EXIT(&rxi_keyCreate_lock);
7046 rx_SetSpecific(struct rx_connection *conn, int key, void *ptr)
7049 struct rx_connection *tconn =
7050 (rx_IsClonedConn(conn)) ? conn->parent : conn;
7052 MUTEX_ENTER(&tconn->conn_data_lock);
7053 if (!tconn->specific) {
7054 tconn->specific = (void **)malloc((key + 1) * sizeof(void *));
7055 for (i = 0; i < key; i++)
7056 tconn->specific[i] = NULL;
7057 tconn->nSpecific = key + 1;
7058 tconn->specific[key] = ptr;
7059 } else if (key >= tconn->nSpecific) {
7060 tconn->specific = (void **)
7061 realloc(tconn->specific, (key + 1) * sizeof(void *));
7062 for (i = tconn->nSpecific; i < key; i++)
7063 tconn->specific[i] = NULL;
7064 tconn->nSpecific = key + 1;
7065 tconn->specific[key] = ptr;
7067 if (tconn->specific[key] && rxi_keyCreate_destructor[key])
7068 (*rxi_keyCreate_destructor[key]) (conn->specific[key]);
7069 tconn->specific[key] = ptr;
7071 MUTEX_EXIT(&tconn->conn_data_lock);
7075 rx_GetSpecific(struct rx_connection *conn, int key)
7078 struct rx_connection *tconn =
7079 (rx_IsClonedConn(conn)) ? conn->parent : conn;
7081 MUTEX_ENTER(&tconn->conn_data_lock);
7082 if (key >= tconn->nSpecific)
7085 ptr = tconn->specific[key];
7086 MUTEX_EXIT(&tconn->conn_data_lock);
7090 #endif /* !KERNEL */
7093 * processStats is a queue used to store the statistics for the local
7094 * process. Its contents are similar to the contents of the rpcStats
7095 * queue on a rx_peer structure, but the actual data stored within
7096 * this queue contains totals across the lifetime of the process (assuming
7097 * the stats have not been reset) - unlike the per peer structures
7098 * which can come and go based upon the peer lifetime.
7101 static struct rx_queue processStats = { &processStats, &processStats };
7104 * peerStats is a queue used to store the statistics for all peer structs.
7105 * Its contents are the union of all the peer rpcStats queues.
7108 static struct rx_queue peerStats = { &peerStats, &peerStats };
7111 * rxi_monitor_processStats is used to turn process wide stat collection
7115 static int rxi_monitor_processStats = 0;
7118 * rxi_monitor_peerStats is used to turn per peer stat collection on and off
7121 static int rxi_monitor_peerStats = 0;
7124 * rxi_AddRpcStat - given all of the information for a particular rpc
7125 * call, create (if needed) and update the stat totals for the rpc.
7129 * IN stats - the queue of stats that will be updated with the new value
7131 * IN rxInterface - a unique number that identifies the rpc interface
7133 * IN currentFunc - the index of the function being invoked
7135 * IN totalFunc - the total number of functions in this interface
7137 * IN queueTime - the amount of time this function waited for a thread
7139 * IN execTime - the amount of time this function invocation took to execute
7141 * IN bytesSent - the number bytes sent by this invocation
7143 * IN bytesRcvd - the number bytes received by this invocation
7145 * IN isServer - if true, this invocation was made to a server
7147 * IN remoteHost - the ip address of the remote host
7149 * IN remotePort - the port of the remote host
7151 * IN addToPeerList - if != 0, add newly created stat to the global peer list
7153 * INOUT counter - if a new stats structure is allocated, the counter will
7154 * be updated with the new number of allocated stat structures
7162 rxi_AddRpcStat(struct rx_queue *stats, afs_uint32 rxInterface,
7163 afs_uint32 currentFunc, afs_uint32 totalFunc,
7164 struct clock *queueTime, struct clock *execTime,
7165 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd, int isServer,
7166 afs_uint32 remoteHost, afs_uint32 remotePort,
7167 int addToPeerList, unsigned int *counter)
7170 rx_interface_stat_p rpc_stat, nrpc_stat;
7173 * See if there's already a structure for this interface
7176 for (queue_Scan(stats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7177 if ((rpc_stat->stats[0].interfaceId == rxInterface)
7178 && (rpc_stat->stats[0].remote_is_server == isServer))
7183 * Didn't find a match so allocate a new structure and add it to the
7187 if (queue_IsEnd(stats, rpc_stat) || (rpc_stat == NULL)
7188 || (rpc_stat->stats[0].interfaceId != rxInterface)
7189 || (rpc_stat->stats[0].remote_is_server != isServer)) {
7194 sizeof(rx_interface_stat_t) +
7195 totalFunc * sizeof(rx_function_entry_v1_t);
7197 rpc_stat = (rx_interface_stat_p) rxi_Alloc(space);
7198 if (rpc_stat == NULL) {
7202 *counter += totalFunc;
7203 for (i = 0; i < totalFunc; i++) {
7204 rpc_stat->stats[i].remote_peer = remoteHost;
7205 rpc_stat->stats[i].remote_port = remotePort;
7206 rpc_stat->stats[i].remote_is_server = isServer;
7207 rpc_stat->stats[i].interfaceId = rxInterface;
7208 rpc_stat->stats[i].func_total = totalFunc;
7209 rpc_stat->stats[i].func_index = i;
7210 hzero(rpc_stat->stats[i].invocations);
7211 hzero(rpc_stat->stats[i].bytes_sent);
7212 hzero(rpc_stat->stats[i].bytes_rcvd);
7213 rpc_stat->stats[i].queue_time_sum.sec = 0;
7214 rpc_stat->stats[i].queue_time_sum.usec = 0;
7215 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7216 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7217 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7218 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7219 rpc_stat->stats[i].queue_time_max.sec = 0;
7220 rpc_stat->stats[i].queue_time_max.usec = 0;
7221 rpc_stat->stats[i].execution_time_sum.sec = 0;
7222 rpc_stat->stats[i].execution_time_sum.usec = 0;
7223 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7224 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7225 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7226 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7227 rpc_stat->stats[i].execution_time_max.sec = 0;
7228 rpc_stat->stats[i].execution_time_max.usec = 0;
7230 queue_Prepend(stats, rpc_stat);
7231 if (addToPeerList) {
7232 queue_Prepend(&peerStats, &rpc_stat->all_peers);
7237 * Increment the stats for this function
7240 hadd32(rpc_stat->stats[currentFunc].invocations, 1);
7241 hadd(rpc_stat->stats[currentFunc].bytes_sent, *bytesSent);
7242 hadd(rpc_stat->stats[currentFunc].bytes_rcvd, *bytesRcvd);
7243 clock_Add(&rpc_stat->stats[currentFunc].queue_time_sum, queueTime);
7244 clock_AddSq(&rpc_stat->stats[currentFunc].queue_time_sum_sqr, queueTime);
7245 if (clock_Lt(queueTime, &rpc_stat->stats[currentFunc].queue_time_min)) {
7246 rpc_stat->stats[currentFunc].queue_time_min = *queueTime;
7248 if (clock_Gt(queueTime, &rpc_stat->stats[currentFunc].queue_time_max)) {
7249 rpc_stat->stats[currentFunc].queue_time_max = *queueTime;
7251 clock_Add(&rpc_stat->stats[currentFunc].execution_time_sum, execTime);
7252 clock_AddSq(&rpc_stat->stats[currentFunc].execution_time_sum_sqr,
7254 if (clock_Lt(execTime, &rpc_stat->stats[currentFunc].execution_time_min)) {
7255 rpc_stat->stats[currentFunc].execution_time_min = *execTime;
7257 if (clock_Gt(execTime, &rpc_stat->stats[currentFunc].execution_time_max)) {
7258 rpc_stat->stats[currentFunc].execution_time_max = *execTime;
7266 * rx_IncrementTimeAndCount - increment the times and count for a particular
7271 * IN peer - the peer who invoked the rpc
7273 * IN rxInterface - a unique number that identifies the rpc interface
7275 * IN currentFunc - the index of the function being invoked
7277 * IN totalFunc - the total number of functions in this interface
7279 * IN queueTime - the amount of time this function waited for a thread
7281 * IN execTime - the amount of time this function invocation took to execute
7283 * IN bytesSent - the number bytes sent by this invocation
7285 * IN bytesRcvd - the number bytes received by this invocation
7287 * IN isServer - if true, this invocation was made to a server
7295 rx_IncrementTimeAndCount(struct rx_peer *peer, afs_uint32 rxInterface,
7296 afs_uint32 currentFunc, afs_uint32 totalFunc,
7297 struct clock *queueTime, struct clock *execTime,
7298 afs_hyper_t * bytesSent, afs_hyper_t * bytesRcvd,
7302 if (!(rxi_monitor_peerStats || rxi_monitor_processStats))
7305 MUTEX_ENTER(&rx_rpc_stats);
7306 MUTEX_ENTER(&peer->peer_lock);
7308 if (rxi_monitor_peerStats) {
7309 rxi_AddRpcStat(&peer->rpcStats, rxInterface, currentFunc, totalFunc,
7310 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7311 peer->host, peer->port, 1, &rxi_rpc_peer_stat_cnt);
7314 if (rxi_monitor_processStats) {
7315 rxi_AddRpcStat(&processStats, rxInterface, currentFunc, totalFunc,
7316 queueTime, execTime, bytesSent, bytesRcvd, isServer,
7317 0xffffffff, 0xffffffff, 0, &rxi_rpc_process_stat_cnt);
7320 MUTEX_EXIT(&peer->peer_lock);
7321 MUTEX_EXIT(&rx_rpc_stats);
7326 * rx_MarshallProcessRPCStats - marshall an array of rpc statistics
7330 * IN callerVersion - the rpc stat version of the caller.
7332 * IN count - the number of entries to marshall.
7334 * IN stats - pointer to stats to be marshalled.
7336 * OUT ptr - Where to store the marshalled data.
7343 rx_MarshallProcessRPCStats(afs_uint32 callerVersion, int count,
7344 rx_function_entry_v1_t * stats, afs_uint32 ** ptrP)
7350 * We only support the first version
7352 for (ptr = *ptrP, i = 0; i < count; i++, stats++) {
7353 *(ptr++) = stats->remote_peer;
7354 *(ptr++) = stats->remote_port;
7355 *(ptr++) = stats->remote_is_server;
7356 *(ptr++) = stats->interfaceId;
7357 *(ptr++) = stats->func_total;
7358 *(ptr++) = stats->func_index;
7359 *(ptr++) = hgethi(stats->invocations);
7360 *(ptr++) = hgetlo(stats->invocations);
7361 *(ptr++) = hgethi(stats->bytes_sent);
7362 *(ptr++) = hgetlo(stats->bytes_sent);
7363 *(ptr++) = hgethi(stats->bytes_rcvd);
7364 *(ptr++) = hgetlo(stats->bytes_rcvd);
7365 *(ptr++) = stats->queue_time_sum.sec;
7366 *(ptr++) = stats->queue_time_sum.usec;
7367 *(ptr++) = stats->queue_time_sum_sqr.sec;
7368 *(ptr++) = stats->queue_time_sum_sqr.usec;
7369 *(ptr++) = stats->queue_time_min.sec;
7370 *(ptr++) = stats->queue_time_min.usec;
7371 *(ptr++) = stats->queue_time_max.sec;
7372 *(ptr++) = stats->queue_time_max.usec;
7373 *(ptr++) = stats->execution_time_sum.sec;
7374 *(ptr++) = stats->execution_time_sum.usec;
7375 *(ptr++) = stats->execution_time_sum_sqr.sec;
7376 *(ptr++) = stats->execution_time_sum_sqr.usec;
7377 *(ptr++) = stats->execution_time_min.sec;
7378 *(ptr++) = stats->execution_time_min.usec;
7379 *(ptr++) = stats->execution_time_max.sec;
7380 *(ptr++) = stats->execution_time_max.usec;
7386 * rx_RetrieveProcessRPCStats - retrieve all of the rpc statistics for
7391 * IN callerVersion - the rpc stat version of the caller
7393 * OUT myVersion - the rpc stat version of this function
7395 * OUT clock_sec - local time seconds
7397 * OUT clock_usec - local time microseconds
7399 * OUT allocSize - the number of bytes allocated to contain stats
7401 * OUT statCount - the number stats retrieved from this process.
7403 * OUT stats - the actual stats retrieved from this process.
7407 * Returns void. If successful, stats will != NULL.
7411 rx_RetrieveProcessRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7412 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7413 size_t * allocSize, afs_uint32 * statCount,
7414 afs_uint32 ** stats)
7424 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7427 * Check to see if stats are enabled
7430 MUTEX_ENTER(&rx_rpc_stats);
7431 if (!rxi_monitor_processStats) {
7432 MUTEX_EXIT(&rx_rpc_stats);
7436 clock_GetTime(&now);
7437 *clock_sec = now.sec;
7438 *clock_usec = now.usec;
7441 * Allocate the space based upon the caller version
7443 * If the client is at an older version than we are,
7444 * we return the statistic data in the older data format, but
7445 * we still return our version number so the client knows we
7446 * are maintaining more data than it can retrieve.
7449 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7450 space = rxi_rpc_process_stat_cnt * sizeof(rx_function_entry_v1_t);
7451 *statCount = rxi_rpc_process_stat_cnt;
7454 * This can't happen yet, but in the future version changes
7455 * can be handled by adding additional code here
7459 if (space > (size_t) 0) {
7461 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7464 rx_interface_stat_p rpc_stat, nrpc_stat;
7468 (&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7470 * Copy the data based upon the caller version
7472 rx_MarshallProcessRPCStats(callerVersion,
7473 rpc_stat->stats[0].func_total,
7474 rpc_stat->stats, &ptr);
7480 MUTEX_EXIT(&rx_rpc_stats);
7485 * rx_RetrievePeerRPCStats - retrieve all of the rpc statistics for the peers
7489 * IN callerVersion - the rpc stat version of the caller
7491 * OUT myVersion - the rpc stat version of this function
7493 * OUT clock_sec - local time seconds
7495 * OUT clock_usec - local time microseconds
7497 * OUT allocSize - the number of bytes allocated to contain stats
7499 * OUT statCount - the number of stats retrieved from the individual
7502 * OUT stats - the actual stats retrieved from the individual peer structures.
7506 * Returns void. If successful, stats will != NULL.
7510 rx_RetrievePeerRPCStats(afs_uint32 callerVersion, afs_uint32 * myVersion,
7511 afs_uint32 * clock_sec, afs_uint32 * clock_usec,
7512 size_t * allocSize, afs_uint32 * statCount,
7513 afs_uint32 ** stats)
7523 *myVersion = RX_STATS_RETRIEVAL_VERSION;
7526 * Check to see if stats are enabled
7529 MUTEX_ENTER(&rx_rpc_stats);
7530 if (!rxi_monitor_peerStats) {
7531 MUTEX_EXIT(&rx_rpc_stats);
7535 clock_GetTime(&now);
7536 *clock_sec = now.sec;
7537 *clock_usec = now.usec;
7540 * Allocate the space based upon the caller version
7542 * If the client is at an older version than we are,
7543 * we return the statistic data in the older data format, but
7544 * we still return our version number so the client knows we
7545 * are maintaining more data than it can retrieve.
7548 if (callerVersion >= RX_STATS_RETRIEVAL_FIRST_EDITION) {
7549 space = rxi_rpc_peer_stat_cnt * sizeof(rx_function_entry_v1_t);
7550 *statCount = rxi_rpc_peer_stat_cnt;
7553 * This can't happen yet, but in the future version changes
7554 * can be handled by adding additional code here
7558 if (space > (size_t) 0) {
7560 ptr = *stats = (afs_uint32 *) rxi_Alloc(space);
7563 rx_interface_stat_p rpc_stat, nrpc_stat;
7567 (&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7569 * We have to fix the offset of rpc_stat since we are
7570 * keeping this structure on two rx_queues. The rx_queue
7571 * package assumes that the rx_queue member is the first
7572 * member of the structure. That is, rx_queue assumes that
7573 * any one item is only on one queue at a time. We are
7574 * breaking that assumption and so we have to do a little
7575 * math to fix our pointers.
7578 fix_offset = (char *)rpc_stat;
7579 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7580 rpc_stat = (rx_interface_stat_p) fix_offset;
7583 * Copy the data based upon the caller version
7585 rx_MarshallProcessRPCStats(callerVersion,
7586 rpc_stat->stats[0].func_total,
7587 rpc_stat->stats, &ptr);
7593 MUTEX_EXIT(&rx_rpc_stats);
7598 * rx_FreeRPCStats - free memory allocated by
7599 * rx_RetrieveProcessRPCStats and rx_RetrievePeerRPCStats
7603 * IN stats - stats previously returned by rx_RetrieveProcessRPCStats or
7604 * rx_RetrievePeerRPCStats
7606 * IN allocSize - the number of bytes in stats.
7614 rx_FreeRPCStats(afs_uint32 * stats, size_t allocSize)
7616 rxi_Free(stats, allocSize);
7620 * rx_queryProcessRPCStats - see if process rpc stat collection is
7621 * currently enabled.
7627 * Returns 0 if stats are not enabled != 0 otherwise
7631 rx_queryProcessRPCStats(void)
7634 MUTEX_ENTER(&rx_rpc_stats);
7635 rc = rxi_monitor_processStats;
7636 MUTEX_EXIT(&rx_rpc_stats);
7641 * rx_queryPeerRPCStats - see if peer stat collection is currently enabled.
7647 * Returns 0 if stats are not enabled != 0 otherwise
7651 rx_queryPeerRPCStats(void)
7654 MUTEX_ENTER(&rx_rpc_stats);
7655 rc = rxi_monitor_peerStats;
7656 MUTEX_EXIT(&rx_rpc_stats);
7661 * rx_enableProcessRPCStats - begin rpc stat collection for entire process
7671 rx_enableProcessRPCStats(void)
7673 MUTEX_ENTER(&rx_rpc_stats);
7674 rx_enable_stats = 1;
7675 rxi_monitor_processStats = 1;
7676 MUTEX_EXIT(&rx_rpc_stats);
7680 * rx_enablePeerRPCStats - begin rpc stat collection per peer structure
7690 rx_enablePeerRPCStats(void)
7692 MUTEX_ENTER(&rx_rpc_stats);
7693 rx_enable_stats = 1;
7694 rxi_monitor_peerStats = 1;
7695 MUTEX_EXIT(&rx_rpc_stats);
7699 * rx_disableProcessRPCStats - stop rpc stat collection for entire process
7709 rx_disableProcessRPCStats(void)
7711 rx_interface_stat_p rpc_stat, nrpc_stat;
7714 MUTEX_ENTER(&rx_rpc_stats);
7717 * Turn off process statistics and if peer stats is also off, turn
7721 rxi_monitor_processStats = 0;
7722 if (rxi_monitor_peerStats == 0) {
7723 rx_enable_stats = 0;
7726 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7727 unsigned int num_funcs = 0;
7730 queue_Remove(rpc_stat);
7731 num_funcs = rpc_stat->stats[0].func_total;
7733 sizeof(rx_interface_stat_t) +
7734 rpc_stat->stats[0].func_total * sizeof(rx_function_entry_v1_t);
7736 rxi_Free(rpc_stat, space);
7737 rxi_rpc_process_stat_cnt -= num_funcs;
7739 MUTEX_EXIT(&rx_rpc_stats);
7743 * rx_disablePeerRPCStats - stop rpc stat collection for peers
7753 rx_disablePeerRPCStats(void)
7755 struct rx_peer **peer_ptr, **peer_end;
7758 MUTEX_ENTER(&rx_rpc_stats);
7761 * Turn off peer statistics and if process stats is also off, turn
7765 rxi_monitor_peerStats = 0;
7766 if (rxi_monitor_processStats == 0) {
7767 rx_enable_stats = 0;
7770 RWLOCK_RDLOCK(&rx_peerHashTable_lock);
7771 for (peer_ptr = &rx_peerHashTable[0], peer_end =
7772 &rx_peerHashTable[rx_hashTableSize]; peer_ptr < peer_end;
7774 struct rx_peer *peer, *next, *prev;
7775 for (prev = peer = *peer_ptr; peer; peer = next) {
7777 code = MUTEX_TRYENTER(&peer->peer_lock);
7779 rx_interface_stat_p rpc_stat, nrpc_stat;
7782 (&peer->rpcStats, rpc_stat, nrpc_stat,
7783 rx_interface_stat)) {
7784 unsigned int num_funcs = 0;
7787 queue_Remove(&rpc_stat->queue_header);
7788 queue_Remove(&rpc_stat->all_peers);
7789 num_funcs = rpc_stat->stats[0].func_total;
7791 sizeof(rx_interface_stat_t) +
7792 rpc_stat->stats[0].func_total *
7793 sizeof(rx_function_entry_v1_t);
7795 rxi_Free(rpc_stat, space);
7796 rxi_rpc_peer_stat_cnt -= num_funcs;
7798 MUTEX_EXIT(&peer->peer_lock);
7799 if (prev == *peer_ptr) {
7809 RWLOCK_UNLOCK(&rx_peerHashTable_lock);
7810 MUTEX_EXIT(&rx_rpc_stats);
7814 * rx_clearProcessRPCStats - clear the contents of the rpc stats according
7819 * IN clearFlag - flag indicating which stats to clear
7827 rx_clearProcessRPCStats(afs_uint32 clearFlag)
7829 rx_interface_stat_p rpc_stat, nrpc_stat;
7831 MUTEX_ENTER(&rx_rpc_stats);
7833 for (queue_Scan(&processStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7834 unsigned int num_funcs = 0, i;
7835 num_funcs = rpc_stat->stats[0].func_total;
7836 for (i = 0; i < num_funcs; i++) {
7837 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7838 hzero(rpc_stat->stats[i].invocations);
7840 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7841 hzero(rpc_stat->stats[i].bytes_sent);
7843 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7844 hzero(rpc_stat->stats[i].bytes_rcvd);
7846 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7847 rpc_stat->stats[i].queue_time_sum.sec = 0;
7848 rpc_stat->stats[i].queue_time_sum.usec = 0;
7850 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7851 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7852 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7854 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7855 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7856 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7858 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7859 rpc_stat->stats[i].queue_time_max.sec = 0;
7860 rpc_stat->stats[i].queue_time_max.usec = 0;
7862 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7863 rpc_stat->stats[i].execution_time_sum.sec = 0;
7864 rpc_stat->stats[i].execution_time_sum.usec = 0;
7866 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7867 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7868 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7870 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7871 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7872 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7874 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7875 rpc_stat->stats[i].execution_time_max.sec = 0;
7876 rpc_stat->stats[i].execution_time_max.usec = 0;
7881 MUTEX_EXIT(&rx_rpc_stats);
7885 * rx_clearPeerRPCStats - clear the contents of the rpc stats according
7890 * IN clearFlag - flag indicating which stats to clear
7898 rx_clearPeerRPCStats(afs_uint32 clearFlag)
7900 rx_interface_stat_p rpc_stat, nrpc_stat;
7902 MUTEX_ENTER(&rx_rpc_stats);
7904 for (queue_Scan(&peerStats, rpc_stat, nrpc_stat, rx_interface_stat)) {
7905 unsigned int num_funcs = 0, i;
7908 * We have to fix the offset of rpc_stat since we are
7909 * keeping this structure on two rx_queues. The rx_queue
7910 * package assumes that the rx_queue member is the first
7911 * member of the structure. That is, rx_queue assumes that
7912 * any one item is only on one queue at a time. We are
7913 * breaking that assumption and so we have to do a little
7914 * math to fix our pointers.
7917 fix_offset = (char *)rpc_stat;
7918 fix_offset -= offsetof(rx_interface_stat_t, all_peers);
7919 rpc_stat = (rx_interface_stat_p) fix_offset;
7921 num_funcs = rpc_stat->stats[0].func_total;
7922 for (i = 0; i < num_funcs; i++) {
7923 if (clearFlag & AFS_RX_STATS_CLEAR_INVOCATIONS) {
7924 hzero(rpc_stat->stats[i].invocations);
7926 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_SENT) {
7927 hzero(rpc_stat->stats[i].bytes_sent);
7929 if (clearFlag & AFS_RX_STATS_CLEAR_BYTES_RCVD) {
7930 hzero(rpc_stat->stats[i].bytes_rcvd);
7932 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SUM) {
7933 rpc_stat->stats[i].queue_time_sum.sec = 0;
7934 rpc_stat->stats[i].queue_time_sum.usec = 0;
7936 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_SQUARE) {
7937 rpc_stat->stats[i].queue_time_sum_sqr.sec = 0;
7938 rpc_stat->stats[i].queue_time_sum_sqr.usec = 0;
7940 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MIN) {
7941 rpc_stat->stats[i].queue_time_min.sec = 9999999;
7942 rpc_stat->stats[i].queue_time_min.usec = 9999999;
7944 if (clearFlag & AFS_RX_STATS_CLEAR_QUEUE_TIME_MAX) {
7945 rpc_stat->stats[i].queue_time_max.sec = 0;
7946 rpc_stat->stats[i].queue_time_max.usec = 0;
7948 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SUM) {
7949 rpc_stat->stats[i].execution_time_sum.sec = 0;
7950 rpc_stat->stats[i].execution_time_sum.usec = 0;
7952 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_SQUARE) {
7953 rpc_stat->stats[i].execution_time_sum_sqr.sec = 0;
7954 rpc_stat->stats[i].execution_time_sum_sqr.usec = 0;
7956 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MIN) {
7957 rpc_stat->stats[i].execution_time_min.sec = 9999999;
7958 rpc_stat->stats[i].execution_time_min.usec = 9999999;
7960 if (clearFlag & AFS_RX_STATS_CLEAR_EXEC_TIME_MAX) {
7961 rpc_stat->stats[i].execution_time_max.sec = 0;
7962 rpc_stat->stats[i].execution_time_max.usec = 0;
7967 MUTEX_EXIT(&rx_rpc_stats);
7971 * rxi_rxstat_userok points to a routine that returns 1 if the caller
7972 * is authorized to enable/disable/clear RX statistics.
7974 static int (*rxi_rxstat_userok) (struct rx_call * call) = NULL;
7977 rx_SetRxStatUserOk(int (*proc) (struct rx_call * call))
7979 rxi_rxstat_userok = proc;
7983 rx_RxStatUserOk(struct rx_call *call)
7985 if (!rxi_rxstat_userok)
7987 return rxi_rxstat_userok(call);
7992 * DllMain() -- Entry-point function called by the DllMainCRTStartup()
7993 * function in the MSVC runtime DLL (msvcrt.dll).
7995 * Note: the system serializes calls to this function.
7998 DllMain(HINSTANCE dllInstHandle, /* instance handle for this DLL module */
7999 DWORD reason, /* reason function is being called */
8000 LPVOID reserved) /* reserved for future use */
8003 case DLL_PROCESS_ATTACH:
8004 /* library is being attached to a process */
8008 case DLL_PROCESS_DETACH:
8017 int rx_DumpCalls(FILE *outputFile, char *cookie)
8019 #ifdef RXDEBUG_PACKET
8021 #ifdef KDUMP_RX_LOCK
8022 struct rx_call_rx_lock *c;
8028 sprintf(output, "%s - Start dumping all Rx Calls - count=%u\r\n", cookie, rx_stats.nCallStructs);
8029 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8031 for (c = rx_allCallsp; c; c = c->allNextp) {
8032 u_short rqc, tqc, iovqc;
8033 struct rx_packet *p, *np;
8035 MUTEX_ENTER(&c->lock);
8036 queue_Count(&c->rq, p, np, rx_packet, rqc);
8037 queue_Count(&c->tq, p, np, rx_packet, tqc);
8038 queue_Count(&c->iovq, p, np, rx_packet, iovqc);
8040 sprintf(output, "%s - call=0x%p, id=%u, state=%u, mode=%u, conn=%p, epoch=%u, cid=%u, callNum=%u, connFlags=0x%x, flags=0x%x, "
8041 "rqc=%u,%u, tqc=%u,%u, iovqc=%u,%u, "
8042 "lstatus=%u, rstatus=%u, error=%d, timeout=%u, "
8043 "resendEvent=%d, timeoutEvt=%d, keepAliveEvt=%d, delayedAckEvt=%d, delayedAbortEvt=%d, abortCode=%d, abortCount=%d, "
8044 "lastSendTime=%u, lastRecvTime=%u, lastSendData=%u"
8045 #ifdef RX_ENABLE_LOCKS
8048 #ifdef RX_REFCOUNT_CHECK
8049 ", refCountBegin=%u, refCountResend=%u, refCountDelay=%u, "
8050 "refCountAlive=%u, refCountPacket=%u, refCountSend=%u, refCountAckAll=%u, refCountAbort=%u"
8053 cookie, c, c->call_id, (afs_uint32)c->state, (afs_uint32)c->mode, c->conn, c->conn?c->conn->epoch:0, c->conn?c->conn->cid:0,
8054 c->callNumber?*c->callNumber:0, c->conn?c->conn->flags:0, c->flags,
8055 (afs_uint32)c->rqc, (afs_uint32)rqc, (afs_uint32)c->tqc, (afs_uint32)tqc, (afs_uint32)c->iovqc, (afs_uint32)iovqc,
8056 (afs_uint32)c->localStatus, (afs_uint32)c->remoteStatus, c->error, c->timeout,
8057 c->resendEvent?1:0, c->timeoutEvent?1:0, c->keepAliveEvent?1:0, c->delayedAckEvent?1:0, c->delayedAbortEvent?1:0,
8058 c->abortCode, c->abortCount, c->lastSendTime, c->lastReceiveTime, c->lastSendData
8059 #ifdef RX_ENABLE_LOCKS
8060 , (afs_uint32)c->refCount
8062 #ifdef RX_REFCOUNT_CHECK
8063 , c->refCDebug[0],c->refCDebug[1],c->refCDebug[2],c->refCDebug[3],c->refCDebug[4],c->refCDebug[5],c->refCDebug[6],c->refCDebug[7]
8066 MUTEX_EXIT(&c->lock);
8068 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8070 sprintf(output, "%s - End dumping all Rx Calls\r\n", cookie);
8071 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
8072 #endif /* RXDEBUG_PACKET */
8075 #endif /* AFS_NT40_ENV */