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),